aboutsummaryrefslogtreecommitdiff
path: root/src/Ryujinx.Graphics.Gpu/Engine/Threed
diff options
context:
space:
mode:
authorTSR Berry <20988865+TSRBerry@users.noreply.github.com>2023-04-08 01:22:00 +0200
committerMary <thog@protonmail.com>2023-04-27 23:51:14 +0200
commitcee712105850ac3385cd0091a923438167433f9f (patch)
tree4a5274b21d8b7f938c0d0ce18736d3f2993b11b1 /src/Ryujinx.Graphics.Gpu/Engine/Threed
parentcd124bda587ef09668a971fa1cac1c3f0cfc9f21 (diff)
Move solution and projects to src
Diffstat (limited to 'src/Ryujinx.Graphics.Gpu/Engine/Threed')
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendFunctions.cs4226
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendManager.cs115
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendPreGenTable.cs273
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendUcode.cs126
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/UcodeAssembler.cs305
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/ConditionalRendering.cs130
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs183
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs856
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs65
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs194
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/IndirectDrawType.cs38
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/RenderTargetUpdateFlags.cs41
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs190
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/SpecializationStateUpdater.cs346
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs177
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs1448
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs620
-rw-r--r--src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs1048
18 files changed, 10381 insertions, 0 deletions
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendFunctions.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendFunctions.cs
new file mode 100644
index 00000000..a40b9cc4
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendFunctions.cs
@@ -0,0 +1,4226 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System.Globalization;
+using System.Runtime.InteropServices;
+using System.Text;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+ static class AdvancedBlendFunctions
+ {
+ public static readonly AdvancedBlendUcode[] Table = new AdvancedBlendUcode[]
+ {
+ new AdvancedBlendUcode(AdvancedBlendOp.PlusClamped, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedPlusClampedPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedPlusClampedAlphaPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.PlusDarker, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedPlusDarkerPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedMultiplyPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedScreenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedOverlayPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedDarkenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedLightenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedColorDodgePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedColorBurnPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedHardLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedSoftLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedDifferencePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Minus, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedMinusPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.MinusClamped, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedMinusClampedPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedExclusionPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Contrast, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedContrastPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedInvertPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedInvertRGBPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.InvertOvg, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedInvertOvgPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedLinearDodgePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedLinearBurnPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedVividLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedLinearLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedPinLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedHardMixPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Red, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedRedPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Green, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedGreenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Blue, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedBluePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedHslHuePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedHslSaturationPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedHslColorPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Uncorrelated, true, GenUncorrelatedHslLuminosityPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Src, AdvancedBlendOverlap.Disjoint, true, GenDisjointSrcPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Dst, AdvancedBlendOverlap.Disjoint, true, GenDisjointDstPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Disjoint, true, GenDisjointSrcOverPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Disjoint, true, GenDisjointDstOverPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Disjoint, true, GenDisjointSrcInPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstIn, AdvancedBlendOverlap.Disjoint, true, GenDisjointDstInPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Disjoint, true, GenDisjointSrcOutPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstOut, AdvancedBlendOverlap.Disjoint, true, GenDisjointDstOutPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Disjoint, true, GenDisjointSrcAtopPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Disjoint, true, GenDisjointDstAtopPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Disjoint, true, GenDisjointXorPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Plus, AdvancedBlendOverlap.Disjoint, true, GenDisjointPlusPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Disjoint, true, GenDisjointMultiplyPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Disjoint, true, GenDisjointScreenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Disjoint, true, GenDisjointOverlayPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Disjoint, true, GenDisjointDarkenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Disjoint, true, GenDisjointLightenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Disjoint, true, GenDisjointColorDodgePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Disjoint, true, GenDisjointColorBurnPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Disjoint, true, GenDisjointHardLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Disjoint, true, GenDisjointSoftLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Disjoint, true, GenDisjointDifferencePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Disjoint, true, GenDisjointExclusionPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Disjoint, true, GenDisjointInvertPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Disjoint, true, GenDisjointInvertRGBPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Disjoint, true, GenDisjointLinearDodgePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Disjoint, true, GenDisjointLinearBurnPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Disjoint, true, GenDisjointVividLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Disjoint, true, GenDisjointLinearLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Disjoint, true, GenDisjointPinLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Disjoint, true, GenDisjointHardMixPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Disjoint, true, GenDisjointHslHuePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Disjoint, true, GenDisjointHslSaturationPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Disjoint, true, GenDisjointHslColorPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Disjoint, true, GenDisjointHslLuminosityPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Src, AdvancedBlendOverlap.Conjoint, true, GenConjointSrcPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Dst, AdvancedBlendOverlap.Conjoint, true, GenConjointDstPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Conjoint, true, GenConjointSrcOverPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Conjoint, true, GenConjointDstOverPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Conjoint, true, GenConjointSrcInPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstIn, AdvancedBlendOverlap.Conjoint, true, GenConjointDstInPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Conjoint, true, GenConjointSrcOutPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstOut, AdvancedBlendOverlap.Conjoint, true, GenConjointDstOutPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Conjoint, true, GenConjointSrcAtopPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Conjoint, true, GenConjointDstAtopPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Conjoint, true, GenConjointXorPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Conjoint, true, GenConjointMultiplyPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Conjoint, true, GenConjointScreenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Conjoint, true, GenConjointOverlayPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Conjoint, true, GenConjointDarkenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Conjoint, true, GenConjointLightenPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Conjoint, true, GenConjointColorDodgePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Conjoint, true, GenConjointColorBurnPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Conjoint, true, GenConjointHardLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Conjoint, true, GenConjointSoftLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Conjoint, true, GenConjointDifferencePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Conjoint, true, GenConjointExclusionPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Conjoint, true, GenConjointInvertPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Conjoint, true, GenConjointInvertRGBPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Conjoint, true, GenConjointLinearDodgePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Conjoint, true, GenConjointLinearBurnPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Conjoint, true, GenConjointVividLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Conjoint, true, GenConjointLinearLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Conjoint, true, GenConjointPinLightPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Conjoint, true, GenConjointHardMixPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Conjoint, true, GenConjointHslHuePremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Conjoint, true, GenConjointHslSaturationPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Conjoint, true, GenConjointHslColorPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Conjoint, true, GenConjointHslLuminosityPremul),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDstOver),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSrcIn),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSrcOut),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSrcAtop),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDstAtop),
+ new AdvancedBlendUcode(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedXor),
+ new AdvancedBlendUcode(AdvancedBlendOp.PlusClamped, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPlusClamped),
+ new AdvancedBlendUcode(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPlusClampedAlpha),
+ new AdvancedBlendUcode(AdvancedBlendOp.PlusDarker, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPlusDarker),
+ new AdvancedBlendUcode(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedMultiply),
+ new AdvancedBlendUcode(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedScreen),
+ new AdvancedBlendUcode(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedOverlay),
+ new AdvancedBlendUcode(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDarken),
+ new AdvancedBlendUcode(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLighten),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedColorDodge),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedColorBurn),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHardLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSoftLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDifference),
+ new AdvancedBlendUcode(AdvancedBlendOp.Minus, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedMinus),
+ new AdvancedBlendUcode(AdvancedBlendOp.MinusClamped, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedMinusClamped),
+ new AdvancedBlendUcode(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedExclusion),
+ new AdvancedBlendUcode(AdvancedBlendOp.Contrast, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedContrast),
+ new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedInvertRGB),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLinearDodge),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLinearBurn),
+ new AdvancedBlendUcode(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedVividLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLinearLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPinLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHardMix),
+ new AdvancedBlendUcode(AdvancedBlendOp.Red, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedRed),
+ new AdvancedBlendUcode(AdvancedBlendOp.Green, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedGreen),
+ new AdvancedBlendUcode(AdvancedBlendOp.Blue, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedBlue),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslHue),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslSaturation),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslColor),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslLuminosity),
+ new AdvancedBlendUcode(AdvancedBlendOp.Src, AdvancedBlendOverlap.Disjoint, false, GenDisjointSrc),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Disjoint, false, GenDisjointSrcOver),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Disjoint, false, GenDisjointDstOver),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Disjoint, false, GenDisjointSrcIn),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Disjoint, false, GenDisjointSrcOut),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Disjoint, false, GenDisjointSrcAtop),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Disjoint, false, GenDisjointDstAtop),
+ new AdvancedBlendUcode(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Disjoint, false, GenDisjointXor),
+ new AdvancedBlendUcode(AdvancedBlendOp.Plus, AdvancedBlendOverlap.Disjoint, false, GenDisjointPlus),
+ new AdvancedBlendUcode(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Disjoint, false, GenDisjointMultiply),
+ new AdvancedBlendUcode(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Disjoint, false, GenDisjointScreen),
+ new AdvancedBlendUcode(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Disjoint, false, GenDisjointOverlay),
+ new AdvancedBlendUcode(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Disjoint, false, GenDisjointDarken),
+ new AdvancedBlendUcode(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Disjoint, false, GenDisjointLighten),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Disjoint, false, GenDisjointColorDodge),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Disjoint, false, GenDisjointColorBurn),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Disjoint, false, GenDisjointHardLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Disjoint, false, GenDisjointSoftLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Disjoint, false, GenDisjointDifference),
+ new AdvancedBlendUcode(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Disjoint, false, GenDisjointExclusion),
+ new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Disjoint, false, GenDisjointInvertRGB),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Disjoint, false, GenDisjointLinearDodge),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Disjoint, false, GenDisjointLinearBurn),
+ new AdvancedBlendUcode(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Disjoint, false, GenDisjointVividLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Disjoint, false, GenDisjointLinearLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Disjoint, false, GenDisjointPinLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Disjoint, false, GenDisjointHardMix),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Disjoint, false, GenDisjointHslHue),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Disjoint, false, GenDisjointHslSaturation),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Disjoint, false, GenDisjointHslColor),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Disjoint, false, GenDisjointHslLuminosity),
+ new AdvancedBlendUcode(AdvancedBlendOp.Src, AdvancedBlendOverlap.Conjoint, false, GenConjointSrc),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Conjoint, false, GenConjointSrcOver),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Conjoint, false, GenConjointDstOver),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Conjoint, false, GenConjointSrcIn),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Conjoint, false, GenConjointSrcOut),
+ new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Conjoint, false, GenConjointSrcAtop),
+ new AdvancedBlendUcode(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Conjoint, false, GenConjointDstAtop),
+ new AdvancedBlendUcode(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Conjoint, false, GenConjointXor),
+ new AdvancedBlendUcode(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Conjoint, false, GenConjointMultiply),
+ new AdvancedBlendUcode(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Conjoint, false, GenConjointScreen),
+ new AdvancedBlendUcode(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Conjoint, false, GenConjointOverlay),
+ new AdvancedBlendUcode(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Conjoint, false, GenConjointDarken),
+ new AdvancedBlendUcode(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Conjoint, false, GenConjointLighten),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Conjoint, false, GenConjointColorDodge),
+ new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Conjoint, false, GenConjointColorBurn),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Conjoint, false, GenConjointHardLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Conjoint, false, GenConjointSoftLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Conjoint, false, GenConjointDifference),
+ new AdvancedBlendUcode(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Conjoint, false, GenConjointExclusion),
+ new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Conjoint, false, GenConjointInvertRGB),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Conjoint, false, GenConjointLinearDodge),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Conjoint, false, GenConjointLinearBurn),
+ new AdvancedBlendUcode(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Conjoint, false, GenConjointVividLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Conjoint, false, GenConjointLinearLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Conjoint, false, GenConjointPinLight),
+ new AdvancedBlendUcode(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Conjoint, false, GenConjointHardMix),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Conjoint, false, GenConjointHslHue),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Conjoint, false, GenConjointHslSaturation),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Conjoint, false, GenConjointHslColor),
+ new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Conjoint, false, GenConjointHslLuminosity)
+ };
+
+ public static string GenTable()
+ {
+ // This can be used to generate the table on AdvancedBlendPreGenTable.
+
+ StringBuilder sb = new StringBuilder();
+
+ sb.AppendLine($"private static Dictionary<Hash128, AdvancedBlendEntry> _entries = new()");
+ sb.AppendLine("{");
+
+ foreach (var entry in Table)
+ {
+ Hash128 hash = XXHash128.ComputeHash(MemoryMarshal.Cast<uint, byte>(entry.Code));
+
+ string[] constants = new string[entry.Constants != null ? entry.Constants.Length : 0];
+
+ for (int i = 0; i < constants.Length; i++)
+ {
+ RgbFloat rgb = entry.Constants[i];
+
+ constants[i] = string.Format(CultureInfo.InvariantCulture, "new " + nameof(RgbFloat) + "({0}f, {1}f, {2}f)", rgb.R, rgb.G, rgb.B);
+ }
+
+ string constantList = constants.Length > 0 ? $"new[] {{ {string.Join(", ", constants)} }}" : $"Array.Empty<{nameof(RgbFloat)}>()";
+
+ static string EnumValue(string name, object value)
+ {
+ if (value.ToString() == "0")
+ {
+ return "0";
+ }
+
+ return $"{name}.{value}";
+ }
+
+ string alpha = $"new {nameof(FixedFunctionAlpha)}({EnumValue(nameof(BlendUcodeEnable), entry.Alpha.Enable)}, {EnumValue(nameof(BlendOp), entry.Alpha.AlphaOp)}, {EnumValue(nameof(BlendFactor), entry.Alpha.AlphaSrcFactor)}, {EnumValue(nameof(BlendFactor), entry.Alpha.AlphaDstFactor)})";
+
+ sb.AppendLine($" {{ new Hash128(0x{hash.Low:X16}, 0x{hash.High:X16}), new AdvancedBlendEntry({nameof(AdvancedBlendOp)}.{entry.Op}, {nameof(AdvancedBlendOverlap)}.{entry.Overlap}, {(entry.SrcPreMultiplied ? "true" : "false")}, {constantList}, {alpha}) }},");
+ }
+
+ sb.AppendLine("};");
+
+ return sb.ToString();
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPlusClampedPremul(ref UcodeAssembler asm)
+ {
+ asm.Add(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.SrcRGB);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPlusClampedAlphaPremul(ref UcodeAssembler asm)
+ {
+ asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.SrcRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPlusDarkerPremul(ref UcodeAssembler asm)
+ {
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.SrcRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstRGB);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.SrcAAA);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedMultiplyPremul(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstRGB);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedScreenPremul(ref UcodeAssembler asm)
+ {
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.DstRGB);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedOverlayPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedDarkenPremul(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLightenPremul(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedColorDodgePremul(ref UcodeAssembler asm)
+ {
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.SrcRGB);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.SrcAAA);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.DstRGB);
+ asm.Min(CC.GT, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+ asm.Mul(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.SrcAAA);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstRGB, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedColorBurnPremul(ref UcodeAssembler asm)
+ {
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstAAA, OpBD.SrcAAA, OpAC.SrcAAA, OpBD.DstRGB);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcRGB);
+ asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA, OpAC.SrcAAA, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstAAA, OpBD.DstRGB);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHardLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedSoftLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+ asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(5, 16f, 16f, 16f);
+ asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(6, 12f, 12f, 12f);
+ asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(7, 3f, 3f, 3f);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedDifferencePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.Temp2, OpBD.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedMinusPremul(ref UcodeAssembler asm)
+ {
+ asm.Sub(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.SrcRGB);
+ return new FixedFunctionAlpha(BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedMinusClampedPremul(ref UcodeAssembler asm)
+ {
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.SrcRGB);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedExclusionPremul(ref UcodeAssembler asm)
+ {
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.DstRGB);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.DstRGB);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedContrastPremul(ref UcodeAssembler asm)
+ {
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.ConstantRGB, OpAC.DstAAA, OpBD.ConstantOne);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.SrcAAA, OpBD.ConstantOne);
+ asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+ asm.SetConstant(1, 0.5f, 0.5f, 0.5f);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantRGB);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedInvertPremul(ref UcodeAssembler asm)
+ {
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA, OpAC.SrcAAA, OpBD.DstRGB);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedInvertRGBPremul(ref UcodeAssembler asm)
+ {
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.DstRGB);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedInvertOvgPremul(ref UcodeAssembler asm)
+ {
+ asm.Sub(CC.T, Dest.PBR, OpBD.ConstantOne, OpBD.DstRGB);
+ asm.Mmadd(CC.T, Dest.Temp0, OpAC.SrcAAA, OpBD.PBR, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLinearDodgePremul(ref UcodeAssembler asm)
+ {
+ asm.Mmadd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLinearBurnPremul(ref UcodeAssembler asm)
+ {
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedVividLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantRGB);
+ asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+ asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantOne);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLinearLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Madd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPinLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Add(CC.LE, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHardMixPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedRedPremul(ref UcodeAssembler asm)
+ {
+ asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+ asm.Mov(CC.T, Dest.Temp0.R, OpBD.SrcRGB);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedGreenPremul(ref UcodeAssembler asm)
+ {
+ asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+ asm.Mov(CC.T, Dest.Temp0.G, OpBD.SrcRGB);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedBluePremul(ref UcodeAssembler asm)
+ {
+ asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+ asm.Mov(CC.T, Dest.Temp0.B, OpBD.SrcRGB);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslHuePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp2, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslSaturationPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslColorPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp2, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslLuminosityPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointDstPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcOverPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp2);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDstOverPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp1);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcInPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Sub(CC.T, Dest.Temp1.RToA, OpBD.DstAAA, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDstInPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Sub(CC.T, Dest.Temp1.RToA, OpBD.DstAAA, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcOutPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDstOutPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcAtopPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointDstAtopPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointXorPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ asm.Min(CC.T, Dest.Temp1, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Add(CC.T, Dest.Temp1.RToA, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointPlusPremul(ref UcodeAssembler asm)
+ {
+ asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.SrcRGB);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointMultiplyPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointScreenPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointOverlayPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDarkenPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointLightenPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointColorDodgePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointColorBurnPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantZero);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp2);
+ asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHardLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSoftLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+ asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(5, 16f, 16f, 16f);
+ asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(6, 12f, 12f, 12f);
+ asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(7, 3f, 3f, 3f);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDifferencePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.Temp2, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointExclusionPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointInvertPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp0, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointInvertRGBPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.ConstantOne, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp0, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointLinearDodgePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointLinearBurnPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointVividLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantRGB);
+ asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+ asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantOne);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointLinearLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Madd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointPinLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Add(CC.LE, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHardMixPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslHuePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp2, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslSaturationPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslColorPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp2, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslLuminosityPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDstPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcOverPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDstOverPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcInPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDstInPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcOutPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenConjointDstOutPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcAtopPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDstAtopPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointXorPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ asm.Sub(CC.T, Dest.Temp1.CC, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Sub(CC.LT, Dest.Temp1, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mov(CC.T, Dest.Temp1.RToA, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenConjointMultiplyPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointScreenPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointOverlayPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDarkenPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLightenPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointColorDodgePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointColorBurnPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantZero);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp2);
+ asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHardLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSoftLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+ asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(5, 16f, 16f, 16f);
+ asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(6, 12f, 12f, 12f);
+ asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(7, 3f, 3f, 3f);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDifferencePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.Temp2, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointExclusionPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointInvertPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointInvertRGBPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.ConstantOne, OpAC.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLinearDodgePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLinearBurnPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointVividLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantRGB);
+ asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+ asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+ asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantOne);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLinearLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Madd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointPinLightPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Add(CC.LE, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+ asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHardMixPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslHuePremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp2, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslSaturationPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslColorPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp2, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslLuminosityPremul(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedDstOver(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedSrcIn(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.DstAlphaGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedSrcOut(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.OneMinusDstAAA);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedSrcAtop(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedDstAtop(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedXor(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.PBR, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPlusClamped(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Add(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPlusClampedAlpha(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPlusDarker(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+ asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstRGB);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.SrcAAA);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedMultiply(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstRGB);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedScreen(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.DstRGB);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedOverlay(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedDarken(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLighten(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedColorDodge(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.PBR);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.SrcAAA);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.DstRGB);
+ asm.Min(CC.GT, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+ asm.Mul(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.SrcAAA);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstRGB, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedColorBurn(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstAAA, OpBD.SrcAAA, OpAC.SrcAAA, OpBD.DstRGB);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.Temp2);
+ asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA, OpAC.SrcAAA, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstAAA, OpBD.DstRGB);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHardLight(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedSoftLight(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+ asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(5, 16f, 16f, 16f);
+ asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(6, 12f, 12f, 12f);
+ asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(7, 3f, 3f, 3f);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedDifference(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.SrcRGB);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.SrcRGB, OpBD.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedMinus(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedMinusClamped(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedExclusion(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.DstRGB);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.DstRGB);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedContrast(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.ConstantRGB, OpAC.DstAAA, OpBD.ConstantOne);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.SrcAAA, OpBD.ConstantOne);
+ asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+ asm.SetConstant(1, 0.5f, 0.5f, 0.5f);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantRGB);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedInvertRGB(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.PBR, OpBD.DstRGB);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLinearDodge(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLinearBurn(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedVividLight(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+ asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+ asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.LT, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantOne);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedLinearLight(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Madd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedPinLight(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Add(CC.LE, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHardMix(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedRed(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+ asm.Mov(CC.T, Dest.Temp0.R, OpBD.Temp2);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedGreen(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+ asm.Mov(CC.T, Dest.Temp0.G, OpBD.Temp2);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedBlue(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+ asm.Mov(CC.T, Dest.Temp0.B, OpBD.Temp2);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslHue(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.SrcRGB, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslSaturation(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslColor(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.SrcRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenUncorrelatedHslLuminosity(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrc(ref UcodeAssembler asm)
+ {
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcOver(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.SrcRGB);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDstOver(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp1);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcIn(ref UcodeAssembler asm)
+ {
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Sub(CC.T, Dest.Temp1.RToA, OpBD.DstAAA, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcOut(ref UcodeAssembler asm)
+ {
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSrcAtop(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointDstAtop(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointXor(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ asm.Min(CC.T, Dest.Temp1, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Add(CC.T, Dest.Temp1.RToA, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointPlus(ref UcodeAssembler asm)
+ {
+ asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+ asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointMultiply(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointScreen(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointOverlay(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDarken(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointLighten(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointColorDodge(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointColorBurn(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHardLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointSoftLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+ asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(5, 16f, 16f, 16f);
+ asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(6, 12f, 12f, 12f);
+ asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(7, 3f, 3f, 3f);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointDifference(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.SrcRGB);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.SrcRGB, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointExclusion(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointInvertRGB(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp0, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenDisjointLinearDodge(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointLinearBurn(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointVividLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+ asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+ asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.LT, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantOne);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointLinearLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Madd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointPinLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Add(CC.LE, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHardMix(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslHue(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.SrcRGB, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslSaturation(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslColor(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.SrcRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenDisjointHslLuminosity(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenConjointSrc(ref UcodeAssembler asm)
+ {
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcOver(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDstOver(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcIn(ref UcodeAssembler asm)
+ {
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcOut(ref UcodeAssembler asm)
+ {
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenConjointSrcAtop(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDstAtop(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointXor(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ asm.Sub(CC.T, Dest.Temp1.CC, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Sub(CC.LT, Dest.Temp1, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mov(CC.T, Dest.Temp1.RToA, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+ return FixedFunctionAlpha.Disabled;
+ }
+
+ private static FixedFunctionAlpha GenConjointMultiply(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointScreen(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointOverlay(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDarken(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLighten(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointColorDodge(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointColorBurn(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHardLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointSoftLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+ asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+ asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+ asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+ asm.SetConstant(5, 16f, 16f, 16f);
+ asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(6, 12f, 12f, 12f);
+ asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.SetConstant(7, 3f, 3f, 3f);
+ asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointDifference(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.SrcRGB);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointExclusion(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointInvertRGB(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+ asm.Mul(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+ return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLinearDodge(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLinearBurn(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointVividLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+ asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+ asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+ asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+ asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Add(CC.LT, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantOne);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointLinearLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 2f, 2f, 2f);
+ asm.Madd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+ asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointPinLight(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+ asm.Add(CC.LE, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+ asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHardMix(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslHue(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.SrcRGB, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslSaturation(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+ asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+ asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslColor(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp2, OpBD.SrcRGB, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+
+ private static FixedFunctionAlpha GenConjointHslLuminosity(ref UcodeAssembler asm)
+ {
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+ asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+ asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+ asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+ asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+ asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+ asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+ asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+ asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+ asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+ asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+ asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+ asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+ asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+ asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+ asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+ asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+ asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+ asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+ asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+ return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+ }
+ }
+} \ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendManager.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendManager.cs
new file mode 100644
index 00000000..8072c6af
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendManager.cs
@@ -0,0 +1,115 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+ /// <summary>
+ /// Advanced blend manager.
+ /// </summary>
+ class AdvancedBlendManager
+ {
+ private const int InstructionRamSize = 128;
+ private const int InstructionRamSizeMask = InstructionRamSize - 1;
+
+ private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+ private readonly uint[] _code;
+ private int _ip;
+
+ /// <summary>
+ /// Creates a new instance of the advanced blend manager.
+ /// </summary>
+ /// <param name="state">GPU state of the channel owning this manager</param>
+ public AdvancedBlendManager(DeviceStateWithShadow<ThreedClassState> state)
+ {
+ _state = state;
+ _code = new uint[InstructionRamSize];
+ }
+
+ /// <summary>
+ /// Sets the start offset of the blend microcode in memory.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void LoadBlendUcodeStart(int argument)
+ {
+ _ip = argument;
+ }
+
+ /// <summary>
+ /// Pushes one word of blend microcode.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void LoadBlendUcodeInstruction(int argument)
+ {
+ _code[_ip++ & InstructionRamSizeMask] = (uint)argument;
+ }
+
+ /// <summary>
+ /// Tries to identify the current advanced blend function being used,
+ /// given the current state and microcode that was uploaded.
+ /// </summary>
+ /// <param name="descriptor">Advanced blend descriptor</param>
+ /// <returns>True if the function was found, false otherwise</returns>
+ public bool TryGetAdvancedBlend(out AdvancedBlendDescriptor descriptor)
+ {
+ Span<uint> currentCode = new Span<uint>(_code);
+ byte codeLength = (byte)_state.State.BlendUcodeSize;
+
+ if (currentCode.Length > codeLength)
+ {
+ currentCode = currentCode.Slice(0, codeLength);
+ }
+
+ Hash128 hash = XXHash128.ComputeHash(MemoryMarshal.Cast<uint, byte>(currentCode));
+
+ descriptor = default;
+
+ if (!AdvancedBlendPreGenTable.Entries.TryGetValue(hash, out var entry))
+ {
+ return false;
+ }
+
+ if (entry.Constants != null)
+ {
+ bool constantsMatch = true;
+
+ for (int i = 0; i < entry.Constants.Length; i++)
+ {
+ RgbFloat constant = entry.Constants[i];
+ RgbHalf constant2 = _state.State.BlendUcodeConstants[i];
+
+ if ((Half)constant.R != constant2.UnpackR() ||
+ (Half)constant.G != constant2.UnpackG() ||
+ (Half)constant.B != constant2.UnpackB())
+ {
+ constantsMatch = false;
+ break;
+ }
+ }
+
+ if (!constantsMatch)
+ {
+ return false;
+ }
+ }
+
+ if (entry.Alpha.Enable != _state.State.BlendUcodeEnable)
+ {
+ return false;
+ }
+
+ if (entry.Alpha.Enable == BlendUcodeEnable.EnableRGBA &&
+ (entry.Alpha.AlphaOp != _state.State.BlendStateCommon.AlphaOp ||
+ entry.Alpha.AlphaSrcFactor != _state.State.BlendStateCommon.AlphaSrcFactor ||
+ entry.Alpha.AlphaDstFactor != _state.State.BlendStateCommon.AlphaDstFactor))
+ {
+ return false;
+ }
+
+ descriptor = new AdvancedBlendDescriptor(entry.Op, entry.Overlap, entry.SrcPreMultiplied);
+ return true;
+ }
+ }
+} \ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendPreGenTable.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendPreGenTable.cs
new file mode 100644
index 00000000..d35d8abf
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendPreGenTable.cs
@@ -0,0 +1,273 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+ /// <summary>
+ /// Advanced blend function entry.
+ /// </summary>
+ struct AdvancedBlendEntry
+ {
+ /// <summary>
+ /// Advanced blend operation.
+ /// </summary>
+ public AdvancedBlendOp Op { get; }
+
+ /// <summary>
+ /// Advanced blend overlap mode.
+ /// </summary>
+ public AdvancedBlendOverlap Overlap { get; }
+
+ /// <summary>
+ /// Whenever the source input is pre-multiplied.
+ /// </summary>
+ public bool SrcPreMultiplied { get; }
+
+ /// <summary>
+ /// Constants used by the microcode.
+ /// </summary>
+ public RgbFloat[] Constants { get; }
+
+ /// <summary>
+ /// Fixed function alpha state.
+ /// </summary>
+ public FixedFunctionAlpha Alpha { get; }
+
+ /// <summary>
+ /// Creates a new advanced blend function entry.
+ /// </summary>
+ /// <param name="op">Advanced blend operation</param>
+ /// <param name="overlap">Advanced blend overlap mode</param>
+ /// <param name="srcPreMultiplied">Whenever the source input is pre-multiplied</param>
+ /// <param name="constants">Constants used by the microcode</param>
+ /// <param name="alpha">Fixed function alpha state</param>
+ public AdvancedBlendEntry(
+ AdvancedBlendOp op,
+ AdvancedBlendOverlap overlap,
+ bool srcPreMultiplied,
+ RgbFloat[] constants,
+ FixedFunctionAlpha alpha)
+ {
+ Op = op;
+ Overlap = overlap;
+ SrcPreMultiplied = srcPreMultiplied;
+ Constants = constants;
+ Alpha = alpha;
+ }
+ }
+
+ /// <summary>
+ /// Pre-generated hash table with advanced blend functions used by the driver.
+ /// </summary>
+ static class AdvancedBlendPreGenTable
+ {
+ /// <summary>
+ /// Advanced blend functions dictionary.
+ /// </summary>
+ public static readonly IReadOnlyDictionary<Hash128, AdvancedBlendEntry> Entries = new Dictionary<Hash128, AdvancedBlendEntry>()
+ {
+ { new Hash128(0x19ECF57B83DE31F7, 0x5BAE759246F264C0), new AdvancedBlendEntry(AdvancedBlendOp.PlusClamped, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xDE1B14A356A1A9ED, 0x59D803593C607C1D), new AdvancedBlendEntry(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x1A3C3A6D32DEC368, 0xBCAE519EC6AAA045), new AdvancedBlendEntry(AdvancedBlendOp.PlusDarker, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x6FD380261A63B240, 0x17C3B335DBB9E3DB), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x1D39164823D3A2D1, 0xC45350959CE1C8FB), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x18DF09FF53B129FE, 0xC02EDA33C36019F6), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x5973E583271EBF06, 0x711497D75D1272E0), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x4759E0E5DA54D5E8, 0x1FDD57C0C38AFA1F), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x337684D43CCE97FA, 0x0139E30CC529E1C9), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xDA59E85D8428992D, 0x1D3D7C64C9EF0132), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x9455B949298CE805, 0xE73D3301518BE98A), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xBDD3B4DEDBE336AA, 0xBFA4DCD50D535DEE), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x22D4E970A028649A, 0x4F3FCB055FCED965), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xA346A91311D72114, 0x151A27A3FB0A1904), new AdvancedBlendEntry(AdvancedBlendOp.Minus, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x8A307241061FACD6, 0xA39D1826440B8EE7), new AdvancedBlendEntry(AdvancedBlendOp.MinusClamped, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xB3BE569485EFFFE0, 0x0BA4E269B3CFB165), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x36FCA3277DC11822, 0x2BC0F6CAC2029672), new AdvancedBlendEntry(AdvancedBlendOp.Contrast, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(2f, 2f, 2f), new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x4A6226AF2DE9BD7F, 0xEB890D7DA716F73A), new AdvancedBlendEntry(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0xF364CAA94E160FEB, 0xBF364512C72A3797), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x6BF791AB4AC19C87, 0x6FA17A994EA0FCDE), new AdvancedBlendEntry(AdvancedBlendOp.InvertOvg, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x053C75A0AE0BB222, 0x03C791FEEB59754C), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x25762AB40B6CBDE9, 0x595E9A968AC4F01C), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xC2D05E2DBE16955D, 0xB8659C7A3FCFA7CE), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x223F220B8F74CBFB, 0xD3DD19D7C39209A5), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xD0DAE57A9F1FE78A, 0x353796BCFB8CE30B), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x601C8CBEC07FF8FF, 0xB8E22882360E8695), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x3A55B7B78C76A7A8, 0x206F503B2D9FFEAA), new AdvancedBlendEntry(AdvancedBlendOp.Red, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x80BC65C7831388E5, 0xC652457B2C766AEC), new AdvancedBlendEntry(AdvancedBlendOp.Green, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x3D3A912E5833EE13, 0x307895951349EE33), new AdvancedBlendEntry(AdvancedBlendOp.Blue, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x289105BE92E81803, 0xFD8F1F03D15C53B4), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x007AE3BD140764EB, 0x0EE05A0D2E80BBAE), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x77F7EE0DB3FDDB96, 0xDEA47C881306DB3E), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x66F4E9A7D73CA157, 0x1486058A177DB11C), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x593E9F331612D618, 0x9D217BEFA4EB919A), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x0A5194C5E6891106, 0xDD8EC6586106557C), new AdvancedBlendEntry(AdvancedBlendOp.Dst, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x8D77173D5E06E916, 0x06AB190E7D10F4D4), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x655B4EBC148981DA, 0x455999EF2B9BD28A), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x98F5437D5F518929, 0xBFF4A6E83183DB63), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x6ADDEFE3B9CEF2FD, 0xB6F6272AFECB1AAB), new AdvancedBlendEntry(AdvancedBlendOp.DstIn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x80953F0953BF05B1, 0xD59ABFAA34F8196F), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xA401D9AA2A39C121, 0xFC0C8005C22AD7E3), new AdvancedBlendEntry(AdvancedBlendOp.DstOut, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x06274FB7CA9CDD22, 0x6CE8188B1A9AB6EF), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x0B079BE7F7F70817, 0xB72E7736CA51E321), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x66215C99403CEDDE, 0x900B733D62204C48), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x12DEF2AD900CAD6C, 0x58CF5CC3004910DF), new AdvancedBlendEntry(AdvancedBlendOp.Plus, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x272BA3A49F64DAE4, 0xAC70B96C00A99EAF), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x206C34AAA7D3F545, 0xDA4B30CACAA483A0), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x3D93494920D257BE, 0xDCC573BE1F5F4449), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x0D7417D80191107B, 0xEAF40547827E005F), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xEC1B03E8C883F9C9, 0x2D3CA044C58C01B4), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x58A19A0135D68B31, 0x82F35B97AED068E5), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x20489F9AB36CC0E3, 0x20499874219E35EE), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xBB176935E5EE05BF, 0x95B26D4D30EA7A14), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x5FF9393C908ACFED, 0x068B0BD875773ABF), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x03181F8711C9802C, 0x6B02C7C6B224FE7B), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x2EE2209021F6B977, 0xF3AFA1491B8B89FC), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xD8BA4DD2EDE4DC9E, 0x01006114977CF715), new AdvancedBlendEntry(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0xD156B99835A2D8ED, 0x2D0BEE9E135EA7A7), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x20CE8C898ED4BE27, 0x1514900B6F5E8F66), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xCDE5F743820BA2D9, 0x917845FE2ECB083D), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xEB03DF4A0C1D14CD, 0xBAE2E831C6E8FFE4), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x1DC9E49AABC779AC, 0x4053A1441EB713D3), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xFBDEF776248F7B3E, 0xE05EEFD65AC47CB7), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x415A1A48E03AA6E7, 0x046D7EE33CA46B9A), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x59A6901EC9BB2041, 0x2F3E19CE5EEC3EBE), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x044B2B6E105221DA, 0x3089BBC033F994AF), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x374A5A24AA8E6CC5, 0x29930FAA6215FA2B), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x30CD0F7AF0CF26F9, 0x06CCA6744DE7DCF5), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x1A6C9A1F6FE494A5, 0xA0CFAF77617E54DD), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x081AF6DAAB1C8717, 0xBFEDCE59AE3DC9AC), new AdvancedBlendEntry(AdvancedBlendOp.Dst, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x3518E44573AB68BA, 0xC96EE71AF9F8F546), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xF89E81FE8D73C96F, 0x4583A04577A0F21C), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xDF4026421CB61119, 0x14115A1F5139AFC7), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x91A20262C3E3A695, 0x0B3A102BFCDC6B1C), new AdvancedBlendEntry(AdvancedBlendOp.DstIn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x44F4C7CCFEB9EBFA, 0xF68394E6D56E5C2F), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xB89F17C7021E9760, 0x430357EE0F7188EF), new AdvancedBlendEntry(AdvancedBlendOp.DstOut, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xDA2D20EA4242B8A0, 0x0D1EC05B72E3838F), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x855DFEE1208D11B9, 0x77C6E3DDCFE30B85), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x9B3808439683FD58, 0x123DCBE4705AB25E), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xA42CF045C248A00A, 0x0C6C63C24EA0B0C1), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x320A83B6D00C8059, 0x796EDAB3EB7314BC), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x45253AC9ABFFC613, 0x8F92EA70195FB573), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x1A5D263B588274B6, 0x167D305F6C794179), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x709C1A837FE966AC, 0x75D8CE49E8A78EDB), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x8265C26F85E4145F, 0x932E6CCBF37CB600), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x3F252B3FEF983F27, 0x9370D7EEFEFA1A9E), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x66A334A4AEA41078, 0xCB52254E1E395231), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xFDD05C53B25F0035, 0xB7E3ECEE166C222F), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x25D932A77FFED81A, 0xA50D797B0FCA94E8), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x4A953B6F5F7D341C, 0xDC05CFB50DDB5DC1), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x838CB660C4F41F6D, 0x9E7D958697543495), new AdvancedBlendEntry(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x4DF6EC1348A8F797, 0xA128E0CD69DB5A64), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x178CDFAB9A015295, 0x2BF40EA72E596D57), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x338FC99050E56AFD, 0x2AF41CF82BE602BF), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x62E02ED60D1E978E, 0xBF726B3E68C11E4D), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xFBAF92DD4C101502, 0x7AF2EDA6596B819D), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x0EF1241F65D4B50A, 0xE8D85DFA6AEDDB84), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x77FE024B5C9D4A18, 0xF19D48A932F6860F), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x9C88CBFA2E09D857, 0x0A0361704CBEEE1D), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x5B94127FA190E640, 0x8D1FEFF837A91268), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xB9C9105B7E063DDB, 0xF6A70E1D511B96FD), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xF0751AAE332B3ED1, 0xC40146F5C83C2533), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x579EB12F595F75AD, 0x151BF0504703B81B), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xF9CA152C03AC8C62, 0x1581336205E5CF47), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.DstAlphaGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x98ACD8BB5E195D0F, 0x91F937672BE899F0), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0xBF97F10FC301F44C, 0x75721789F0D48548), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x1B982263B8B08A10, 0x3350C76E2E1B27DF), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0xFF20AC79F64EDED8, 0xAF9025B2D97B9273), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x9FFD986600FB112F, 0x384FDDF4E060139A), new AdvancedBlendEntry(AdvancedBlendOp.PlusClamped, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x0425E40B5B8B3B52, 0x5880CBED7CAB631C), new AdvancedBlendEntry(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x16DAC8593F28623A, 0x233DBC82325B8AED), new AdvancedBlendEntry(AdvancedBlendOp.PlusDarker, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xB37E5F234B9F0948, 0xD5F957A2ECD98FD6), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xCA0FDADD1D20DBE3, 0x1A5C15CCBF1AC538), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x1C48304D73A9DF3A, 0x891DB93FA36E3450), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x53200F2279B7FA39, 0x051C2462EBF6789C), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xB88BFB80714DCD5C, 0xEBD6938D744E6A41), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xE33DC2A25FC1A976, 0x08B3DBB1F3027D45), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xCE97E71615370316, 0xE131AE49D3A4D62B), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xE059FD265149B256, 0x94AF817AC348F61F), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x16D31333D477E231, 0x9A98AAC84F72CC62), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x47FC3B0776366D3C, 0xE96D9BD83B277874), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x7230401E3FEA1F3B, 0xF0D15F05D3D1E309), new AdvancedBlendEntry(AdvancedBlendOp.Minus, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x188212F9303742F5, 0x100C51CB96E03591), new AdvancedBlendEntry(AdvancedBlendOp.MinusClamped, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x52B755D296B44DC5, 0x4003B87275625973), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xD873ED973ADF7EAD, 0x73E68B57D92034E7), new AdvancedBlendEntry(AdvancedBlendOp.Contrast, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(2f, 2f, 2f), new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x471F9FA34B945ACB, 0x10524D1410B3C402), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x99F569454EA0EF32, 0x6FC70A8B3A07DC8B), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x5AD55F950067AC7E, 0x4BA60A4FBABDD0AC), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x03FF2C858C9C4C5B, 0xE95AE7F561FB60E9), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x6DC0E510C7BCF9D2, 0xAE805D7CECDCB5C1), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x44832332CED5C054, 0x2F8D5536C085B30A), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x4AB4D387618AC51F, 0x495B46E0555F4B32), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x99282B49405A01A8, 0xD6FA93F864F24A8E), new AdvancedBlendEntry(AdvancedBlendOp.Red, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x37B30C1064FBD23E, 0x5D068366F42317C2), new AdvancedBlendEntry(AdvancedBlendOp.Green, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x760FAE9D59E04BC2, 0xA40AD483EA01435E), new AdvancedBlendEntry(AdvancedBlendOp.Blue, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0xE786950FD9D1C6EF, 0xF9FDD5AF6451D239), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x052458BB4788B0CA, 0x8AC58FDCA1F45EF5), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x6AFC3837D1D31920, 0xB9D49C2FE49642C6), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0xAFC2911949317E01, 0xD5B63636F5CB3422), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+ { new Hash128(0x13B46DF507CC2C53, 0x86DE26517E6BF0A7), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x5C372442474BE410, 0x79ECD3C0C496EF2E), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x74AAB45DBF5336E9, 0x01BFC4E181DAD442), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x43239E282A36C85C, 0x36FB65560E46AD0F), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x1A3BA8A7583B8F7A, 0xE64E41D548033180), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x32BBB9859E9B565D, 0x3D5CE94FE55F18B5), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0xD947A0766AE3C0FC, 0x391E5D53E86F4ED6), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0xBD9A7C08BDFD8CE6, 0x905407634901355E), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x8395475BCB0D7A8C, 0x48AF5DD501D44A70), new AdvancedBlendEntry(AdvancedBlendOp.Plus, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x80AAC23FEBD4A3E5, 0xEA8C70F0B4DE52DE), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x2F3AD1B0F1B3FD09, 0xC0EBC784BFAB8EA3), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x52B54032F2F70BFF, 0xC941D6FDED674765), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xCA7B86F72EC6A99B, 0x55868A131AFE359E), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x377919B60BD133CA, 0x0FD611627664EF40), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x9D4A0C5EE1153887, 0x7B869EBA218C589B), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x311F2A858545D123, 0xB4D09C802480AD62), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xCF78AA6A83AFA689, 0x9DC48B0C2182A3E1), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xC3018CD6F1CF62D1, 0x016E32DD9087B1BB), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x9CB62CE0E956EE29, 0x0FB67F503E60B3AD), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x3589A13C16EF3BFA, 0x15B29BFC91F3BDFB), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x3502CA5FB7529917, 0xFA51BFD0D1688071), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x62ADC25AD6D0A923, 0x76CB6D238276D3A3), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x09FDEB1116A9D52C, 0x85BB8627CD5C2733), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x0709FED1B65E18EB, 0x5BC3AA4D99EC19CF), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xB18D28AE5DE4C723, 0xE820AA2B75C9C02E), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x6743C51621497480, 0x4B164E40858834AE), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x63D1E181E34A2944, 0x1AE292C9D9F12819), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x079523298250BFF6, 0xC0C793510603CDB5), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x4C9D0A973C805EA6, 0xD1FF59AD5156B93C), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x1E914678F3057BCD, 0xD503AE389C12D229), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0x9FDBADE5556C5311, 0x03F0CBC798FC5C94), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xE39451534635403C, 0x606CC1CA1F452388), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x1D39F0F0A1008AA6, 0xBFDF2B97E6C3F125), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xDB81BED30D5BDBEA, 0xAF0B2856EB93AD2C), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x83F69CCF1D0A79B6, 0x70D31332797430AC), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x7B87F807AB7A8F5C, 0x1241A2A01FB31771), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xF557172E20D5272D, 0xC1961F8C7A5D2820), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0xA8476B3944DBBC9B, 0x84A2F6AF97B15FDF), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+ { new Hash128(0x3259602B55414DA3, 0x72AACCC00B5A9D10), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+ { new Hash128(0xC0CB8C10F36EDCD6, 0x8C2D088AD8191E1C), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x81806C451C6255EF, 0x5AA8AC9A08941A15), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xE55A6537F4568198, 0xCA8735390B799B19), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x5C044BA14536DDA3, 0xBCE0123ED7D510EC), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x6788346C405BE130, 0x372A4BB199C01F9F), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x510EDC2A34E2856B, 0xE1727A407E294254), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x4B7BE01BD398C7A8, 0x5BFF79BC00672C18), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x213B43845540CFEC, 0xDA857411CF1CCFCE), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x765AFA6732E783F1, 0x8F1CABF1BC78A014), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xA4A5DE1CC06F6CB1, 0xA0634A0011001709), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x81F32BD8816EA796, 0x697EE86683165170), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xB870C209EAA5F092, 0xAF5FD923909CAA1F), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+ { new Hash128(0x3649A9F5C936FB83, 0xDD7C834897AA182A), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xD72A2B1097A5995C, 0x3D41B2763A913654), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x551E212B9F6C454A, 0xB0DFA05BEB3C37FA), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x681B5A313B7416BF, 0xCB1CBAEEB4D81500), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x9343A18BD4B16777, 0xEDB4AC1C8972C3A4), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xC960BF6D8519DE28, 0x78D8557FD405D119), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x65A7B01FDC73A46C, 0x297E096ED5CC4D8A), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0xD9C99BA4A6CDC13B, 0x3CFF0ACEDC2EE150), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x6BC00DA6EB922BD1, 0x5FD4C11F2A685234), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ { new Hash128(0x8652300E32D93050, 0x9460E7B449132371), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+ };
+ }
+} \ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendUcode.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendUcode.cs
new file mode 100644
index 00000000..f06b4bf7
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendUcode.cs
@@ -0,0 +1,126 @@
+using Ryujinx.Graphics.GAL;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+ /// <summary>
+ /// Fixed function alpha state used for a advanced blend function.
+ /// </summary>
+ struct FixedFunctionAlpha
+ {
+ /// <summary>
+ /// Fixed function alpha state with alpha blending disabled.
+ /// </summary>
+ public static FixedFunctionAlpha Disabled => new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, default, default, default);
+
+ /// <summary>
+ /// Individual enable bits for the RGB and alpha components.
+ /// </summary>
+ public BlendUcodeEnable Enable { get; }
+
+ /// <summary>
+ /// Alpha blend operation.
+ /// </summary>
+ public BlendOp AlphaOp { get; }
+
+ /// <summary>
+ /// Value multiplied with the blend source operand.
+ /// </summary>
+ public BlendFactor AlphaSrcFactor { get; }
+
+ /// <summary>
+ /// Value multiplied with the blend destination operand.
+ /// </summary>
+ public BlendFactor AlphaDstFactor { get; }
+
+ /// <summary>
+ /// Creates a new blend fixed function alpha state.
+ /// </summary>
+ /// <param name="enable">Individual enable bits for the RGB and alpha components</param>
+ /// <param name="alphaOp">Alpha blend operation</param>
+ /// <param name="alphaSrc">Value multiplied with the blend source operand</param>
+ /// <param name="alphaDst">Value multiplied with the blend destination operand</param>
+ public FixedFunctionAlpha(BlendUcodeEnable enable, BlendOp alphaOp, BlendFactor alphaSrc, BlendFactor alphaDst)
+ {
+ Enable = enable;
+ AlphaOp = alphaOp;
+ AlphaSrcFactor = alphaSrc;
+ AlphaDstFactor = alphaDst;
+ }
+
+ /// <summary>
+ /// Creates a new blend fixed function alpha state.
+ /// </summary>
+ /// <param name="alphaOp">Alpha blend operation</param>
+ /// <param name="alphaSrc">Value multiplied with the blend source operand</param>
+ /// <param name="alphaDst">Value multiplied with the blend destination operand</param>
+ public FixedFunctionAlpha(BlendOp alphaOp, BlendFactor alphaSrc, BlendFactor alphaDst) : this(BlendUcodeEnable.EnableRGB, alphaOp, alphaSrc, alphaDst)
+ {
+ }
+ }
+
+ /// <summary>
+ /// Blend microcode assembly function delegate.
+ /// </summary>
+ /// <param name="asm">Assembler</param>
+ /// <returns>Fixed function alpha state for the microcode</returns>
+ delegate FixedFunctionAlpha GenUcodeFunc(ref UcodeAssembler asm);
+
+ /// <summary>
+ /// Advanced blend microcode state.
+ /// </summary>
+ struct AdvancedBlendUcode
+ {
+ /// <summary>
+ /// Advanced blend operation.
+ /// </summary>
+ public AdvancedBlendOp Op { get; }
+
+ /// <summary>
+ /// Advanced blend overlap mode.
+ /// </summary>
+ public AdvancedBlendOverlap Overlap { get; }
+
+ /// <summary>
+ /// Whenever the source input is pre-multiplied.
+ /// </summary>
+ public bool SrcPreMultiplied { get; }
+
+ /// <summary>
+ /// Fixed function alpha state.
+ /// </summary>
+ public FixedFunctionAlpha Alpha { get; }
+
+ /// <summary>
+ /// Microcode.
+ /// </summary>
+ public uint[] Code { get; }
+
+ /// <summary>
+ /// Constants used by the microcode.
+ /// </summary>
+ public RgbFloat[] Constants { get; }
+
+ /// <summary>
+ /// Creates a new advanced blend state.
+ /// </summary>
+ /// <param name="op">Advanced blend operation</param>
+ /// <param name="overlap">Advanced blend overlap mode</param>
+ /// <param name="srcPreMultiplied">Whenever the source input is pre-multiplied</param>
+ /// <param name="genFunc">Function that will generate the advanced blend microcode</param>
+ public AdvancedBlendUcode(
+ AdvancedBlendOp op,
+ AdvancedBlendOverlap overlap,
+ bool srcPreMultiplied,
+ GenUcodeFunc genFunc)
+ {
+ Op = op;
+ Overlap = overlap;
+ SrcPreMultiplied = srcPreMultiplied;
+
+ UcodeAssembler asm = new UcodeAssembler();
+ Alpha = genFunc(ref asm);
+ Code = asm.GetCode();
+ Constants = asm.GetConstants();
+ }
+ }
+} \ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/UcodeAssembler.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/UcodeAssembler.cs
new file mode 100644
index 00000000..f854787e
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/UcodeAssembler.cs
@@ -0,0 +1,305 @@
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+ /// <summary>
+ /// Blend microcode instruction.
+ /// </summary>
+ enum Instruction
+ {
+ Mmadd = 0,
+ Mmsub = 1,
+ Min = 2,
+ Max = 3,
+ Rcp = 4,
+ Add = 5,
+ Sub = 6
+ }
+
+ /// <summary>
+ /// Blend microcode condition code.
+ /// </summary>
+ enum CC
+ {
+ F = 0,
+ T = 1,
+ EQ = 2,
+ NE = 3,
+ LT = 4,
+ LE = 5,
+ GT = 6,
+ GE = 7
+ }
+
+ /// <summary>
+ /// Blend microcode opend B or D value.
+ /// </summary>
+ enum OpBD
+ {
+ ConstantZero = 0x0,
+ ConstantOne = 0x1,
+ SrcRGB = 0x2,
+ SrcAAA = 0x3,
+ OneMinusSrcAAA = 0x4,
+ DstRGB = 0x5,
+ DstAAA = 0x6,
+ OneMinusDstAAA = 0x7,
+ Temp0 = 0x9,
+ Temp1 = 0xa,
+ Temp2 = 0xb,
+ PBR = 0xc,
+ ConstantRGB = 0xd
+ }
+
+ /// <summary>
+ /// Blend microcode operand A or C value.
+ /// </summary>
+ enum OpAC
+ {
+ SrcRGB = 0,
+ DstRGB = 1,
+ SrcAAA = 2,
+ DstAAA = 3,
+ Temp0 = 4,
+ Temp1 = 5,
+ Temp2 = 6,
+ PBR = 7
+ }
+
+ /// <summary>
+ /// Blend microcode destination operand.
+ /// </summary>
+ enum OpDst
+ {
+ Temp0 = 0,
+ Temp1 = 1,
+ Temp2 = 2,
+ PBR = 3
+ }
+
+ /// <summary>
+ /// Blend microcode input swizzle.
+ /// </summary>
+ enum Swizzle
+ {
+ RGB = 0,
+ GBR = 1,
+ RRR = 2,
+ GGG = 3,
+ BBB = 4,
+ RToA = 5
+ }
+
+ /// <summary>
+ /// Blend microcode output components.
+ /// </summary>
+ enum WriteMask
+ {
+ RGB = 0,
+ R = 1,
+ G = 2,
+ B = 3
+ }
+
+ /// <summary>
+ /// Floating-point RGB color values.
+ /// </summary>
+ struct RgbFloat
+ {
+ /// <summary>
+ /// Red component value.
+ /// </summary>
+ public float R { get; }
+
+ /// <summary>
+ /// Green component value.
+ /// </summary>
+ public float G { get; }
+
+ /// <summary>
+ /// Blue component value.
+ /// </summary>
+ public float B { get; }
+
+ /// <summary>
+ /// Creates a new floating-point RGB value.
+ /// </summary>
+ /// <param name="r">Red component value</param>
+ /// <param name="g">Green component value</param>
+ /// <param name="b">Blue component value</param>
+ public RgbFloat(float r, float g, float b)
+ {
+ R = r;
+ G = g;
+ B = b;
+ }
+ }
+
+ /// <summary>
+ /// Blend microcode destination operand, including swizzle, write mask and condition code update flag.
+ /// </summary>
+ struct Dest
+ {
+ public static Dest Temp0 => new Dest(OpDst.Temp0, Swizzle.RGB, WriteMask.RGB, false);
+ public static Dest Temp1 => new Dest(OpDst.Temp1, Swizzle.RGB, WriteMask.RGB, false);
+ public static Dest Temp2 => new Dest(OpDst.Temp2, Swizzle.RGB, WriteMask.RGB, false);
+ public static Dest PBR => new Dest(OpDst.PBR, Swizzle.RGB, WriteMask.RGB, false);
+
+ public Dest GBR => new Dest(Dst, Swizzle.GBR, WriteMask, WriteCC);
+ public Dest RRR => new Dest(Dst, Swizzle.RRR, WriteMask, WriteCC);
+ public Dest GGG => new Dest(Dst, Swizzle.GGG, WriteMask, WriteCC);
+ public Dest BBB => new Dest(Dst, Swizzle.BBB, WriteMask, WriteCC);
+ public Dest RToA => new Dest(Dst, Swizzle.RToA, WriteMask, WriteCC);
+
+ public Dest R => new Dest(Dst, Swizzle, WriteMask.R, WriteCC);
+ public Dest G => new Dest(Dst, Swizzle, WriteMask.G, WriteCC);
+ public Dest B => new Dest(Dst, Swizzle, WriteMask.B, WriteCC);
+
+ public Dest CC => new Dest(Dst, Swizzle, WriteMask, true);
+
+ public OpDst Dst { get; }
+ public Swizzle Swizzle { get; }
+ public WriteMask WriteMask { get; }
+ public bool WriteCC { get; }
+
+ /// <summary>
+ /// Creates a new blend microcode destination operand.
+ /// </summary>
+ /// <param name="dst">Operand</param>
+ /// <param name="swizzle">Swizzle</param>
+ /// <param name="writeMask">Write maks</param>
+ /// <param name="writeCC">Indicates if condition codes should be updated</param>
+ public Dest(OpDst dst, Swizzle swizzle, WriteMask writeMask, bool writeCC)
+ {
+ Dst = dst;
+ Swizzle = swizzle;
+ WriteMask = writeMask;
+ WriteCC = writeCC;
+ }
+ }
+
+ /// <summary>
+ /// Blend microcode operaiton.
+ /// </summary>
+ struct UcodeOp
+ {
+ public readonly uint Word;
+
+ /// <summary>
+ /// Creates a new blend microcode operation.
+ /// </summary>
+ /// <param name="cc">Condition code that controls whenever the operation is executed or not</param>
+ /// <param name="inst">Instruction</param>
+ /// <param name="constIndex">Index on the constant table of the constant used by any constant operand</param>
+ /// <param name="dest">Destination operand</param>
+ /// <param name="srcA">First input operand</param>
+ /// <param name="srcB">Second input operand</param>
+ /// <param name="srcC">Third input operand</param>
+ /// <param name="srcD">Fourth input operand</param>
+ public UcodeOp(CC cc, Instruction inst, int constIndex, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+ {
+ Word = (uint)cc |
+ ((uint)inst << 3) |
+ ((uint)constIndex << 6) |
+ ((uint)srcA << 9) |
+ ((uint)srcB << 12) |
+ ((uint)srcC << 16) |
+ ((uint)srcD << 19) |
+ ((uint)dest.Swizzle << 23) |
+ ((uint)dest.WriteMask << 26) |
+ ((uint)dest.Dst << 28) |
+ (dest.WriteCC ? (1u << 31) : 0);
+ }
+ }
+
+ /// <summary>
+ /// Blend microcode assembler.
+ /// </summary>
+ struct UcodeAssembler
+ {
+ private List<uint> _code;
+ private RgbFloat[] _constants;
+ private int _constantIndex;
+
+ public void Mul(CC cc, Dest dest, OpAC srcA, OpBD srcB)
+ {
+ Assemble(cc, Instruction.Mmadd, dest, srcA, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+ }
+
+ public void Madd(CC cc, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC)
+ {
+ Assemble(cc, Instruction.Mmadd, dest, srcA, srcB, srcC, OpBD.ConstantOne);
+ }
+
+ public void Mmadd(CC cc, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+ {
+ Assemble(cc, Instruction.Mmadd, dest, srcA, srcB, srcC, srcD);
+ }
+
+ public void Mmsub(CC cc, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+ {
+ Assemble(cc, Instruction.Mmsub, dest, srcA, srcB, srcC, srcD);
+ }
+
+ public void Min(CC cc, Dest dest, OpAC srcA, OpBD srcB)
+ {
+ Assemble(cc, Instruction.Min, dest, srcA, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+ }
+
+ public void Max(CC cc, Dest dest, OpAC srcA, OpBD srcB)
+ {
+ Assemble(cc, Instruction.Max, dest, srcA, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+ }
+
+ public void Rcp(CC cc, Dest dest, OpAC srcA)
+ {
+ Assemble(cc, Instruction.Rcp, dest, srcA, OpBD.ConstantZero, OpAC.SrcRGB, OpBD.ConstantZero);
+ }
+
+ public void Mov(CC cc, Dest dest, OpBD srcB)
+ {
+ Assemble(cc, Instruction.Add, dest, OpAC.SrcRGB, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+ }
+
+ public void Add(CC cc, Dest dest, OpBD srcB, OpBD srcD)
+ {
+ Assemble(cc, Instruction.Add, dest, OpAC.SrcRGB, srcB, OpAC.SrcRGB, srcD);
+ }
+
+ public void Sub(CC cc, Dest dest, OpBD srcB, OpBD srcD)
+ {
+ Assemble(cc, Instruction.Sub, dest, OpAC.SrcRGB, srcB, OpAC.SrcRGB, srcD);
+ }
+
+ private void Assemble(CC cc, Instruction inst, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+ {
+ (_code ??= new List<uint>()).Add(new UcodeOp(cc, inst, _constantIndex, dest, srcA, srcB, srcC, srcD).Word);
+ }
+
+ public void SetConstant(int index, float r, float g, float b)
+ {
+ if (_constants == null)
+ {
+ _constants = new RgbFloat[index + 1];
+ }
+ else if (_constants.Length <= index)
+ {
+ Array.Resize(ref _constants, index + 1);
+ }
+
+ _constants[index] = new RgbFloat(r, g, b);
+ _constantIndex = index;
+ }
+
+ public uint[] GetCode()
+ {
+ return _code?.ToArray();
+ }
+
+ public RgbFloat[] GetConstants()
+ {
+ return _constants;
+ }
+ }
+} \ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConditionalRendering.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConditionalRendering.cs
new file mode 100644
index 00000000..a6b62a4a
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConditionalRendering.cs
@@ -0,0 +1,130 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Memory;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Helper methods used for conditional rendering.
+ /// </summary>
+ static class ConditionalRendering
+ {
+ /// <summary>
+ /// Checks if draws and clears should be performed, according
+ /// to currently set conditional rendering conditions.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ /// <param name="memoryManager">Memory manager bound to the channel currently executing</param>
+ /// <param name="address">Conditional rendering buffer address</param>
+ /// <param name="condition">Conditional rendering condition</param>
+ /// <returns>True if rendering is enabled, false otherwise</returns>
+ public static ConditionalRenderEnabled GetRenderEnable(GpuContext context, MemoryManager memoryManager, GpuVa address, Condition condition)
+ {
+ switch (condition)
+ {
+ case Condition.Always:
+ return ConditionalRenderEnabled.True;
+ case Condition.Never:
+ return ConditionalRenderEnabled.False;
+ case Condition.ResultNonZero:
+ return CounterNonZero(context, memoryManager, address.Pack());
+ case Condition.Equal:
+ return CounterCompare(context, memoryManager, address.Pack(), true);
+ case Condition.NotEqual:
+ return CounterCompare(context, memoryManager, address.Pack(), false);
+ }
+
+ Logger.Warning?.Print(LogClass.Gpu, $"Invalid conditional render condition \"{condition}\".");
+
+ return ConditionalRenderEnabled.True;
+ }
+
+ /// <summary>
+ /// Checks if the counter value at a given GPU memory address is non-zero.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ /// <param name="memoryManager">Memory manager bound to the channel currently executing</param>
+ /// <param name="gpuVa">GPU virtual address of the counter value</param>
+ /// <returns>True if the value is not zero, false otherwise. Returns host if handling with host conditional rendering</returns>
+ private static ConditionalRenderEnabled CounterNonZero(GpuContext context, MemoryManager memoryManager, ulong gpuVa)
+ {
+ ICounterEvent evt = memoryManager.CounterCache.FindEvent(gpuVa);
+
+ if (evt == null)
+ {
+ return ConditionalRenderEnabled.False;
+ }
+
+ if (context.Renderer.Pipeline.TryHostConditionalRendering(evt, 0L, false))
+ {
+ return ConditionalRenderEnabled.Host;
+ }
+ else
+ {
+ evt.Flush();
+ return (memoryManager.Read<ulong>(gpuVa, true) != 0) ? ConditionalRenderEnabled.True : ConditionalRenderEnabled.False;
+ }
+ }
+
+ /// <summary>
+ /// Checks if the counter at a given GPU memory address passes a specified equality comparison.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ /// <param name="memoryManager">Memory manager bound to the channel currently executing</param>
+ /// <param name="gpuVa">GPU virtual address</param>
+ /// <param name="isEqual">True to check if the values are equal, false to check if they are not equal</param>
+ /// <returns>True if the condition is met, false otherwise. Returns host if handling with host conditional rendering</returns>
+ private static ConditionalRenderEnabled CounterCompare(GpuContext context, MemoryManager memoryManager, ulong gpuVa, bool isEqual)
+ {
+ ICounterEvent evt = FindEvent(memoryManager.CounterCache, gpuVa);
+ ICounterEvent evt2 = FindEvent(memoryManager.CounterCache, gpuVa + 16);
+
+ bool useHost;
+
+ if (evt != null && evt2 == null)
+ {
+ useHost = context.Renderer.Pipeline.TryHostConditionalRendering(evt, memoryManager.Read<ulong>(gpuVa + 16), isEqual);
+ }
+ else if (evt == null && evt2 != null)
+ {
+ useHost = context.Renderer.Pipeline.TryHostConditionalRendering(evt2, memoryManager.Read<ulong>(gpuVa), isEqual);
+ }
+ else if (evt != null && evt2 != null)
+ {
+ useHost = context.Renderer.Pipeline.TryHostConditionalRendering(evt, evt2, isEqual);
+ }
+ else
+ {
+ useHost = false;
+ }
+
+ if (useHost)
+ {
+ return ConditionalRenderEnabled.Host;
+ }
+ else
+ {
+ evt?.Flush();
+ evt2?.Flush();
+
+ ulong x = memoryManager.Read<ulong>(gpuVa, true);
+ ulong y = memoryManager.Read<ulong>(gpuVa + 16, true);
+
+ return (isEqual ? x == y : x != y) ? ConditionalRenderEnabled.True : ConditionalRenderEnabled.False;
+ }
+ }
+
+ /// <summary>
+ /// Tries to find a counter that is supposed to be written at the specified address,
+ /// returning the related event.
+ /// </summary>
+ /// <param name="counterCache">GPU counter cache to search on</param>
+ /// <param name="gpuVa">GPU virtual address where the counter is supposed to be written</param>
+ /// <returns>The counter event, or null if not present</returns>
+ private static ICounterEvent FindEvent(CounterCache counterCache, ulong gpuVa)
+ {
+ return counterCache.FindEvent(gpuVa);
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs
new file mode 100644
index 00000000..5c936616
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs
@@ -0,0 +1,183 @@
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Constant buffer updater.
+ /// </summary>
+ class ConstantBufferUpdater
+ {
+ private const int UniformDataCacheSize = 512;
+
+ private readonly GpuChannel _channel;
+ private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+ // State associated with direct uniform buffer updates.
+ // This state is used to attempt to batch together consecutive updates.
+ private ulong _ubBeginCpuAddress = 0;
+ private ulong _ubFollowUpAddress = 0;
+ private ulong _ubByteCount = 0;
+ private int _ubIndex = 0;
+ private int[] _ubData = new int[UniformDataCacheSize];
+
+ /// <summary>
+ /// Creates a new instance of the constant buffer updater.
+ /// </summary>
+ /// <param name="channel">GPU channel</param>
+ /// <param name="state">Channel state</param>
+ public ConstantBufferUpdater(GpuChannel channel, DeviceStateWithShadow<ThreedClassState> state)
+ {
+ _channel = channel;
+ _state = state;
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the vertex shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void BindVertex(int argument)
+ {
+ Bind(argument, ShaderType.Vertex);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the tessellation control shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void BindTessControl(int argument)
+ {
+ Bind(argument, ShaderType.TessellationControl);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the tessellation evaluation shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void BindTessEvaluation(int argument)
+ {
+ Bind(argument, ShaderType.TessellationEvaluation);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the geometry shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void BindGeometry(int argument)
+ {
+ Bind(argument, ShaderType.Geometry);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the fragment shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void BindFragment(int argument)
+ {
+ Bind(argument, ShaderType.Fragment);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the specified shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ /// <param name="type">Shader stage that will access the uniform buffer</param>
+ private void Bind(int argument, ShaderType type)
+ {
+ bool enable = (argument & 1) != 0;
+
+ int index = (argument >> 4) & 0x1f;
+
+ FlushUboDirty();
+
+ if (enable)
+ {
+ var uniformBuffer = _state.State.UniformBufferState;
+
+ ulong address = uniformBuffer.Address.Pack();
+
+ _channel.BufferManager.SetGraphicsUniformBuffer((int)type, index, address, (uint)uniformBuffer.Size);
+ }
+ else
+ {
+ _channel.BufferManager.SetGraphicsUniformBuffer((int)type, index, 0, 0);
+ }
+ }
+
+ /// <summary>
+ /// Flushes any queued UBO updates.
+ /// </summary>
+ public void FlushUboDirty()
+ {
+ if (_ubFollowUpAddress != 0)
+ {
+ var memoryManager = _channel.MemoryManager;
+
+ Span<byte> data = MemoryMarshal.Cast<int, byte>(_ubData.AsSpan(0, (int)(_ubByteCount / 4)));
+
+ if (memoryManager.Physical.WriteWithRedundancyCheck(_ubBeginCpuAddress, data))
+ {
+ memoryManager.Physical.BufferCache.ForceDirty(memoryManager, _ubFollowUpAddress - _ubByteCount, _ubByteCount);
+ }
+
+ _ubFollowUpAddress = 0;
+ _ubIndex = 0;
+ }
+ }
+
+ /// <summary>
+ /// Updates the uniform buffer data with inline data.
+ /// </summary>
+ /// <param name="argument">New uniform buffer data word</param>
+ public void Update(int argument)
+ {
+ var uniformBuffer = _state.State.UniformBufferState;
+
+ ulong address = uniformBuffer.Address.Pack() + (uint)uniformBuffer.Offset;
+
+ if (_ubFollowUpAddress != address || _ubIndex == _ubData.Length)
+ {
+ FlushUboDirty();
+
+ _ubByteCount = 0;
+ _ubBeginCpuAddress = _channel.MemoryManager.Translate(address);
+ }
+
+ _ubData[_ubIndex++] = argument;
+
+ _ubFollowUpAddress = address + 4;
+ _ubByteCount += 4;
+
+ _state.State.UniformBufferState.Offset += 4;
+ }
+
+ /// <summary>
+ /// Updates the uniform buffer data with inline data.
+ /// </summary>
+ /// <param name="data">Data to be written to the uniform buffer</param>
+ public void Update(ReadOnlySpan<int> data)
+ {
+ var uniformBuffer = _state.State.UniformBufferState;
+
+ ulong address = uniformBuffer.Address.Pack() + (uint)uniformBuffer.Offset;
+
+ ulong size = (ulong)data.Length * 4;
+
+ if (_ubFollowUpAddress != address || _ubIndex + data.Length > _ubData.Length)
+ {
+ FlushUboDirty();
+
+ _ubByteCount = 0;
+ _ubBeginCpuAddress = _channel.MemoryManager.Translate(address);
+ }
+
+ data.CopyTo(_ubData.AsSpan(_ubIndex));
+ _ubIndex += data.Length;
+
+ _ubFollowUpAddress = address + size;
+ _ubByteCount += size;
+
+ _state.State.UniformBufferState.Offset += data.Length * 4;
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs
new file mode 100644
index 00000000..7438ba03
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs
@@ -0,0 +1,856 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Memory;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Draw manager.
+ /// </summary>
+ class DrawManager
+ {
+ // Since we don't know the index buffer size for indirect draws,
+ // we must assume a minimum and maximum size and use that for buffer data update purposes.
+ private const int MinIndirectIndexCount = 0x10000;
+ private const int MaxIndirectIndexCount = 0x4000000;
+
+ private readonly GpuContext _context;
+ private readonly GpuChannel _channel;
+ private readonly DeviceStateWithShadow<ThreedClassState> _state;
+ private readonly DrawState _drawState;
+ private readonly SpecializationStateUpdater _currentSpecState;
+ private bool _topologySet;
+
+ private bool _instancedDrawPending;
+ private bool _instancedIndexed;
+ private bool _instancedIndexedInline;
+
+ private int _instancedFirstIndex;
+ private int _instancedFirstVertex;
+ private int _instancedFirstInstance;
+ private int _instancedIndexCount;
+ private int _instancedDrawStateFirst;
+ private int _instancedDrawStateCount;
+
+ private int _instanceIndex;
+
+ private const int VertexBufferFirstMethodOffset = 0x35d;
+ private const int IndexBufferCountMethodOffset = 0x5f8;
+
+ /// <summary>
+ /// Creates a new instance of the draw manager.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ /// <param name="channel">GPU channel</param>
+ /// <param name="state">Channel state</param>
+ /// <param name="drawState">Draw state</param>
+ /// <param name="spec">Specialization state updater</param>
+ public DrawManager(GpuContext context, GpuChannel channel, DeviceStateWithShadow<ThreedClassState> state, DrawState drawState, SpecializationStateUpdater spec)
+ {
+ _context = context;
+ _channel = channel;
+ _state = state;
+ _drawState = drawState;
+ _currentSpecState = spec;
+ }
+
+ /// <summary>
+ /// Marks the entire state as dirty, forcing a full host state update before the next draw.
+ /// </summary>
+ public void ForceStateDirty()
+ {
+ _topologySet = false;
+ }
+
+ /// <summary>
+ /// Pushes four 8-bit index buffer elements.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void VbElementU8(int argument)
+ {
+ _drawState.IbStreamer.VbElementU8(_context.Renderer, argument);
+ }
+
+ /// <summary>
+ /// Pushes two 16-bit index buffer elements.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void VbElementU16(int argument)
+ {
+ _drawState.IbStreamer.VbElementU16(_context.Renderer, argument);
+ }
+
+ /// <summary>
+ /// Pushes one 32-bit index buffer element.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void VbElementU32(int argument)
+ {
+ _drawState.IbStreamer.VbElementU32(_context.Renderer, argument);
+ }
+
+ /// <summary>
+ /// Finishes the draw call.
+ /// This draws geometry on the bound buffers based on the current GPU state.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawEnd(ThreedClass engine, int argument)
+ {
+ DrawEnd(
+ engine,
+ _state.State.IndexBufferState.First,
+ (int)_state.State.IndexBufferCount,
+ _state.State.VertexBufferDrawState.First,
+ _state.State.VertexBufferDrawState.Count);
+ }
+
+ /// <summary>
+ /// Finishes the draw call.
+ /// This draws geometry on the bound buffers based on the current GPU state.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="firstIndex">Index of the first index buffer element used on the draw</param>
+ /// <param name="indexCount">Number of index buffer elements used on the draw</param>
+ /// <param name="drawFirstVertex">Index of the first vertex used on the draw</param>
+ /// <param name="drawVertexCount">Number of vertices used on the draw</param>
+ private void DrawEnd(ThreedClass engine, int firstIndex, int indexCount, int drawFirstVertex, int drawVertexCount)
+ {
+ ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+ _context,
+ _channel.MemoryManager,
+ _state.State.RenderEnableAddress,
+ _state.State.RenderEnableCondition);
+
+ if (renderEnable == ConditionalRenderEnabled.False || _instancedDrawPending)
+ {
+ if (renderEnable == ConditionalRenderEnabled.False)
+ {
+ PerformDeferredDraws();
+ }
+
+ _drawState.DrawIndexed = false;
+
+ if (renderEnable == ConditionalRenderEnabled.Host)
+ {
+ _context.Renderer.Pipeline.EndHostConditionalRendering();
+ }
+
+ return;
+ }
+
+ _drawState.FirstIndex = firstIndex;
+ _drawState.IndexCount = indexCount;
+ _drawState.DrawFirstVertex = drawFirstVertex;
+ _drawState.DrawVertexCount = drawVertexCount;
+ _currentSpecState.SetHasConstantBufferDrawParameters(false);
+
+ engine.UpdateState();
+
+ bool instanced = _drawState.VsUsesInstanceId || _drawState.IsAnyVbInstanced;
+
+ if (instanced)
+ {
+ _instancedDrawPending = true;
+
+ int ibCount = _drawState.IbStreamer.InlineIndexCount;
+
+ _instancedIndexed = _drawState.DrawIndexed;
+ _instancedIndexedInline = ibCount != 0;
+
+ _instancedFirstIndex = firstIndex;
+ _instancedFirstVertex = (int)_state.State.FirstVertex;
+ _instancedFirstInstance = (int)_state.State.FirstInstance;
+
+ _instancedIndexCount = ibCount != 0 ? ibCount : indexCount;
+
+ _instancedDrawStateFirst = drawFirstVertex;
+ _instancedDrawStateCount = drawVertexCount;
+
+ _drawState.DrawIndexed = false;
+
+ if (renderEnable == ConditionalRenderEnabled.Host)
+ {
+ _context.Renderer.Pipeline.EndHostConditionalRendering();
+ }
+
+ return;
+ }
+
+ int firstInstance = (int)_state.State.FirstInstance;
+
+ int inlineIndexCount = _drawState.IbStreamer.GetAndResetInlineIndexCount(_context.Renderer);
+
+ if (inlineIndexCount != 0)
+ {
+ int firstVertex = (int)_state.State.FirstVertex;
+
+ BufferRange br = new BufferRange(_drawState.IbStreamer.GetInlineIndexBuffer(), 0, inlineIndexCount * 4);
+
+ _channel.BufferManager.SetIndexBuffer(br, IndexType.UInt);
+
+ _context.Renderer.Pipeline.DrawIndexed(inlineIndexCount, 1, firstIndex, firstVertex, firstInstance);
+ }
+ else if (_drawState.DrawIndexed)
+ {
+ int firstVertex = (int)_state.State.FirstVertex;
+
+ _context.Renderer.Pipeline.DrawIndexed(indexCount, 1, firstIndex, firstVertex, firstInstance);
+ }
+ else
+ {
+ var drawState = _state.State.VertexBufferDrawState;
+
+ _context.Renderer.Pipeline.Draw(drawVertexCount, 1, drawFirstVertex, firstInstance);
+ }
+
+ _drawState.DrawIndexed = false;
+
+ if (renderEnable == ConditionalRenderEnabled.Host)
+ {
+ _context.Renderer.Pipeline.EndHostConditionalRendering();
+ }
+ }
+
+ /// <summary>
+ /// Starts draw.
+ /// This sets primitive type and instanced draw parameters.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void DrawBegin(int argument)
+ {
+ bool incrementInstance = (argument & (1 << 26)) != 0;
+ bool resetInstance = (argument & (1 << 27)) == 0;
+
+ PrimitiveType type = (PrimitiveType)(argument & 0xffff);
+ DrawBegin(incrementInstance, resetInstance, type);
+ }
+
+ /// <summary>
+ /// Starts draw.
+ /// This sets primitive type and instanced draw parameters.
+ /// </summary>
+ /// <param name="incrementInstance">Indicates if the current instance should be incremented</param>
+ /// <param name="resetInstance">Indicates if the current instance should be set to zero</param>
+ /// <param name="primitiveType">Primitive type</param>
+ private void DrawBegin(bool incrementInstance, bool resetInstance, PrimitiveType primitiveType)
+ {
+ if (incrementInstance)
+ {
+ _instanceIndex++;
+ }
+ else if (resetInstance)
+ {
+ PerformDeferredDraws();
+
+ _instanceIndex = 0;
+ }
+
+ PrimitiveTopology topology;
+
+ if (_state.State.PrimitiveTypeOverrideEnable)
+ {
+ PrimitiveTypeOverride typeOverride = _state.State.PrimitiveTypeOverride;
+ topology = typeOverride.Convert();
+ }
+ else
+ {
+ topology = primitiveType.Convert();
+ }
+
+ UpdateTopology(topology);
+ }
+
+ /// <summary>
+ /// Updates the current primitive topology if needed.
+ /// </summary>
+ /// <param name="topology">New primitive topology</param>
+ private void UpdateTopology(PrimitiveTopology topology)
+ {
+ if (_drawState.Topology != topology || !_topologySet)
+ {
+ _context.Renderer.Pipeline.SetPrimitiveTopology(topology);
+ _currentSpecState.SetTopology(topology);
+ _drawState.Topology = topology;
+ _topologySet = true;
+ }
+ }
+
+ /// <summary>
+ /// Sets the index buffer count.
+ /// This also sets internal state that indicates that the next draw is an indexed draw.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void SetIndexBufferCount(int argument)
+ {
+ _drawState.DrawIndexed = true;
+ }
+
+ // TODO: Verify if the index type is implied from the method that is called,
+ // or if it uses the state index type on hardware.
+
+ /// <summary>
+ /// Performs a indexed draw with 8-bit index buffer elements.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawIndexBuffer8BeginEndInstanceFirst(ThreedClass engine, int argument)
+ {
+ DrawIndexBufferBeginEndInstance(engine, argument, false);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 16-bit index buffer elements.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawIndexBuffer16BeginEndInstanceFirst(ThreedClass engine, int argument)
+ {
+ DrawIndexBufferBeginEndInstance(engine, argument, false);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 32-bit index buffer elements.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawIndexBuffer32BeginEndInstanceFirst(ThreedClass engine, int argument)
+ {
+ DrawIndexBufferBeginEndInstance(engine, argument, false);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 8-bit index buffer elements,
+ /// while also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawIndexBuffer8BeginEndInstanceSubsequent(ThreedClass engine, int argument)
+ {
+ DrawIndexBufferBeginEndInstance(engine, argument, true);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 16-bit index buffer elements,
+ /// while also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawIndexBuffer16BeginEndInstanceSubsequent(ThreedClass engine, int argument)
+ {
+ DrawIndexBufferBeginEndInstance(engine, argument, true);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 32-bit index buffer elements,
+ /// while also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawIndexBuffer32BeginEndInstanceSubsequent(ThreedClass engine, int argument)
+ {
+ DrawIndexBufferBeginEndInstance(engine, argument, true);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with a low number of index buffer elements,
+ /// while optionally also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ /// <param name="instanced">True to increment the current instance value, false otherwise</param>
+ private void DrawIndexBufferBeginEndInstance(ThreedClass engine, int argument, bool instanced)
+ {
+ DrawBegin(instanced, !instanced, (PrimitiveType)((argument >> 28) & 0xf));
+
+ int firstIndex = argument & 0xffff;
+ int indexCount = (argument >> 16) & 0xfff;
+
+ bool oldDrawIndexed = _drawState.DrawIndexed;
+
+ _drawState.DrawIndexed = true;
+ engine.ForceStateDirty(IndexBufferCountMethodOffset * 4);
+
+ DrawEnd(engine, firstIndex, indexCount, 0, 0);
+
+ _drawState.DrawIndexed = oldDrawIndexed;
+ }
+
+ /// <summary>
+ /// Performs a non-indexed draw with the specified topology, index and count.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawVertexArrayBeginEndInstanceFirst(ThreedClass engine, int argument)
+ {
+ DrawVertexArrayBeginEndInstance(engine, argument, false);
+ }
+
+ /// <summary>
+ /// Performs a non-indexed draw with the specified topology, index and count,
+ /// while incrementing the current instance.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawVertexArrayBeginEndInstanceSubsequent(ThreedClass engine, int argument)
+ {
+ DrawVertexArrayBeginEndInstance(engine, argument, true);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with a low number of index buffer elements,
+ /// while optionally also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ /// <param name="instanced">True to increment the current instance value, false otherwise</param>
+ private void DrawVertexArrayBeginEndInstance(ThreedClass engine, int argument, bool instanced)
+ {
+ DrawBegin(instanced, !instanced, (PrimitiveType)((argument >> 28) & 0xf));
+
+ int firstVertex = argument & 0xffff;
+ int vertexCount = (argument >> 16) & 0xfff;
+
+ bool oldDrawIndexed = _drawState.DrawIndexed;
+
+ _drawState.DrawIndexed = false;
+ engine.ForceStateDirty(VertexBufferFirstMethodOffset * 4);
+
+ DrawEnd(engine, 0, 0, firstVertex, vertexCount);
+
+ _drawState.DrawIndexed = oldDrawIndexed;
+ }
+
+ /// <summary>
+ /// Performs a texture draw with a source texture and sampler ID, along with source
+ /// and destination coordinates and sizes.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void DrawTexture(ThreedClass engine, int argument)
+ {
+ static float FixedToFloat(int fixedValue)
+ {
+ return fixedValue * (1f / 4096);
+ }
+
+ float dstX0 = FixedToFloat(_state.State.DrawTextureDstX);
+ float dstY0 = FixedToFloat(_state.State.DrawTextureDstY);
+ float dstWidth = FixedToFloat(_state.State.DrawTextureDstWidth);
+ float dstHeight = FixedToFloat(_state.State.DrawTextureDstHeight);
+
+ // TODO: Confirm behaviour on hardware.
+ // When this is active, the origin appears to be on the bottom.
+ if (_state.State.YControl.HasFlag(YControl.NegateY))
+ {
+ dstY0 -= dstHeight;
+ }
+
+ float dstX1 = dstX0 + dstWidth;
+ float dstY1 = dstY0 + dstHeight;
+
+ float srcX0 = FixedToFloat(_state.State.DrawTextureSrcX);
+ float srcY0 = FixedToFloat(_state.State.DrawTextureSrcY);
+ float srcX1 = ((float)_state.State.DrawTextureDuDx / (1UL << 32)) * dstWidth + srcX0;
+ float srcY1 = ((float)_state.State.DrawTextureDvDy / (1UL << 32)) * dstHeight + srcY0;
+
+ engine.UpdateState(ulong.MaxValue & ~(1UL << StateUpdater.ShaderStateIndex));
+
+ _channel.TextureManager.UpdateRenderTargets();
+
+ int textureId = _state.State.DrawTextureTextureId;
+ int samplerId = _state.State.DrawTextureSamplerId;
+
+ (var texture, var sampler) = _channel.TextureManager.GetGraphicsTextureAndSampler(textureId, samplerId);
+
+ srcX0 *= texture.ScaleFactor;
+ srcY0 *= texture.ScaleFactor;
+ srcX1 *= texture.ScaleFactor;
+ srcY1 *= texture.ScaleFactor;
+
+ float dstScale = _channel.TextureManager.RenderTargetScale;
+
+ dstX0 *= dstScale;
+ dstY0 *= dstScale;
+ dstX1 *= dstScale;
+ dstY1 *= dstScale;
+
+ _context.Renderer.Pipeline.DrawTexture(
+ texture?.HostTexture,
+ sampler?.GetHostSampler(texture),
+ new Extents2DF(srcX0, srcY0, srcX1, srcY1),
+ new Extents2DF(dstX0, dstY0, dstX1, dstY1));
+ }
+
+ /// <summary>
+ /// Performs a indexed or non-indexed draw.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="topology">Primitive topology</param>
+ /// <param name="count">Index count for indexed draws, vertex count for non-indexed draws</param>
+ /// <param name="instanceCount">Instance count</param>
+ /// <param name="firstIndex">First index on the index buffer for indexed draws, ignored for non-indexed draws</param>
+ /// <param name="firstVertex">First vertex on the vertex buffer</param>
+ /// <param name="firstInstance">First instance</param>
+ /// <param name="indexed">True if the draw is indexed, false otherwise</param>
+ public void Draw(
+ ThreedClass engine,
+ PrimitiveTopology topology,
+ int count,
+ int instanceCount,
+ int firstIndex,
+ int firstVertex,
+ int firstInstance,
+ bool indexed)
+ {
+ UpdateTopology(topology);
+
+ ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+ _context,
+ _channel.MemoryManager,
+ _state.State.RenderEnableAddress,
+ _state.State.RenderEnableCondition);
+
+ if (renderEnable == ConditionalRenderEnabled.False)
+ {
+ _drawState.DrawIndexed = false;
+ return;
+ }
+
+ if (indexed)
+ {
+ _drawState.FirstIndex = firstIndex;
+ _drawState.IndexCount = count;
+ _state.State.FirstVertex = (uint)firstVertex;
+ engine.ForceStateDirty(IndexBufferCountMethodOffset * 4);
+ }
+ else
+ {
+ _drawState.DrawFirstVertex = firstVertex;
+ _drawState.DrawVertexCount = count;
+ engine.ForceStateDirty(VertexBufferFirstMethodOffset * 4);
+ }
+
+ _state.State.FirstInstance = (uint)firstInstance;
+
+ _drawState.DrawIndexed = indexed;
+ _currentSpecState.SetHasConstantBufferDrawParameters(true);
+
+ engine.UpdateState();
+
+ if (indexed)
+ {
+ _context.Renderer.Pipeline.DrawIndexed(count, instanceCount, firstIndex, firstVertex, firstInstance);
+ _state.State.FirstVertex = 0;
+ }
+ else
+ {
+ _context.Renderer.Pipeline.Draw(count, instanceCount, firstVertex, firstInstance);
+ }
+
+ _state.State.FirstInstance = 0;
+
+ _drawState.DrawIndexed = false;
+
+ if (renderEnable == ConditionalRenderEnabled.Host)
+ {
+ _context.Renderer.Pipeline.EndHostConditionalRendering();
+ }
+ }
+
+ /// <summary>
+ /// Performs a indirect draw, with parameters from a GPU buffer.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="topology">Primitive topology</param>
+ /// <param name="indirectBufferAddress">Address of the buffer with the draw parameters, such as count, first index, etc</param>
+ /// <param name="parameterBufferAddress">Address of the buffer with the draw count</param>
+ /// <param name="maxDrawCount">Maximum number of draws that can be made</param>
+ /// <param name="stride">Distance in bytes between each entry on the data pointed to by <paramref name="indirectBufferAddress"/></param>
+ /// <param name="indexCount">Maximum number of indices that the draw can consume</param>
+ /// <param name="drawType">Type of the indirect draw, which can be indexed or non-indexed, with or without a draw count</param>
+ public void DrawIndirect(
+ ThreedClass engine,
+ PrimitiveTopology topology,
+ ulong indirectBufferAddress,
+ ulong parameterBufferAddress,
+ int maxDrawCount,
+ int stride,
+ int indexCount,
+ IndirectDrawType drawType)
+ {
+ UpdateTopology(topology);
+
+ ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+ _context,
+ _channel.MemoryManager,
+ _state.State.RenderEnableAddress,
+ _state.State.RenderEnableCondition);
+
+ if (renderEnable == ConditionalRenderEnabled.False)
+ {
+ _drawState.DrawIndexed = false;
+ return;
+ }
+
+ PhysicalMemory memory = _channel.MemoryManager.Physical;
+
+ bool hasCount = (drawType & IndirectDrawType.Count) != 0;
+ bool indexed = (drawType & IndirectDrawType.Indexed) != 0;
+
+ if (indexed)
+ {
+ indexCount = Math.Clamp(indexCount, MinIndirectIndexCount, MaxIndirectIndexCount);
+ _drawState.FirstIndex = 0;
+ _drawState.IndexCount = indexCount;
+ engine.ForceStateDirty(IndexBufferCountMethodOffset * 4);
+ }
+
+ _drawState.DrawIndexed = indexed;
+ _drawState.DrawIndirect = true;
+ _currentSpecState.SetHasConstantBufferDrawParameters(true);
+
+ engine.UpdateState();
+
+ if (hasCount)
+ {
+ var indirectBuffer = memory.BufferCache.GetBufferRange(indirectBufferAddress, (ulong)maxDrawCount * (ulong)stride);
+ var parameterBuffer = memory.BufferCache.GetBufferRange(parameterBufferAddress, 4);
+
+ if (indexed)
+ {
+ _context.Renderer.Pipeline.DrawIndexedIndirectCount(indirectBuffer, parameterBuffer, maxDrawCount, stride);
+ }
+ else
+ {
+ _context.Renderer.Pipeline.DrawIndirectCount(indirectBuffer, parameterBuffer, maxDrawCount, stride);
+ }
+ }
+ else
+ {
+ var indirectBuffer = memory.BufferCache.GetBufferRange(indirectBufferAddress, (ulong)stride);
+
+ if (indexed)
+ {
+ _context.Renderer.Pipeline.DrawIndexedIndirect(indirectBuffer);
+ }
+ else
+ {
+ _context.Renderer.Pipeline.DrawIndirect(indirectBuffer);
+ }
+ }
+
+ _drawState.DrawIndexed = false;
+ _drawState.DrawIndirect = false;
+
+ if (renderEnable == ConditionalRenderEnabled.Host)
+ {
+ _context.Renderer.Pipeline.EndHostConditionalRendering();
+ }
+ }
+
+ /// <summary>
+ /// Perform any deferred draws.
+ /// This is used for instanced draws.
+ /// Since each instance is a separate draw, we defer the draw and accumulate the instance count.
+ /// Once we detect the last instanced draw, then we perform the host instanced draw,
+ /// with the accumulated instance count.
+ /// </summary>
+ public void PerformDeferredDraws()
+ {
+ // Perform any pending instanced draw.
+ if (_instancedDrawPending)
+ {
+ _instancedDrawPending = false;
+
+ bool indexedInline = _instancedIndexedInline;
+
+ if (_instancedIndexed || indexedInline)
+ {
+ if (indexedInline)
+ {
+ int inlineIndexCount = _drawState.IbStreamer.GetAndResetInlineIndexCount(_context.Renderer);
+ BufferRange br = new BufferRange(_drawState.IbStreamer.GetInlineIndexBuffer(), 0, inlineIndexCount * 4);
+
+ _channel.BufferManager.SetIndexBuffer(br, IndexType.UInt);
+ }
+
+ _context.Renderer.Pipeline.DrawIndexed(
+ _instancedIndexCount,
+ _instanceIndex + 1,
+ _instancedFirstIndex,
+ _instancedFirstVertex,
+ _instancedFirstInstance);
+ }
+ else
+ {
+ _context.Renderer.Pipeline.Draw(
+ _instancedDrawStateCount,
+ _instanceIndex + 1,
+ _instancedDrawStateFirst,
+ _instancedFirstInstance);
+ }
+ }
+ }
+
+ /// <summary>
+ /// Clears the current color and depth-stencil buffers.
+ /// Which buffers should be cleared can also be specified with the argument.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ public void Clear(ThreedClass engine, int argument)
+ {
+ Clear(engine, argument, 1);
+ }
+
+ /// <summary>
+ /// Clears the current color and depth-stencil buffers.
+ /// Which buffers should be cleared can also specified with the arguments.
+ /// </summary>
+ /// <param name="engine">3D engine where this method is being called</param>
+ /// <param name="argument">Method call argument</param>
+ /// <param name="layerCount">For array and 3D textures, indicates how many layers should be cleared</param>
+ public void Clear(ThreedClass engine, int argument, int layerCount)
+ {
+ ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+ _context,
+ _channel.MemoryManager,
+ _state.State.RenderEnableAddress,
+ _state.State.RenderEnableCondition);
+
+ if (renderEnable == ConditionalRenderEnabled.False)
+ {
+ return;
+ }
+
+ bool clearDepth = (argument & 1) != 0;
+ bool clearStencil = (argument & 2) != 0;
+ uint componentMask = (uint)((argument >> 2) & 0xf);
+ int index = (argument >> 6) & 0xf;
+ int layer = (argument >> 10) & 0x3ff;
+
+ RenderTargetUpdateFlags updateFlags = RenderTargetUpdateFlags.SingleColor;
+
+ if (layer != 0 || layerCount > 1)
+ {
+ updateFlags |= RenderTargetUpdateFlags.Layered;
+ }
+
+ if (clearDepth || clearStencil)
+ {
+ updateFlags |= RenderTargetUpdateFlags.UpdateDepthStencil;
+ }
+
+ engine.UpdateRenderTargetState(updateFlags, singleUse: componentMask != 0 ? index : -1);
+
+ // If there is a mismatch on the host clip region and the one explicitly defined by the guest
+ // on the screen scissor state, then we need to force only one texture to be bound to avoid
+ // host clipping.
+ var screenScissorState = _state.State.ScreenScissorState;
+
+ // Must happen after UpdateRenderTargetState to have up-to-date clip region values.
+ bool clipMismatch = (screenScissorState.X | screenScissorState.Y) != 0 ||
+ screenScissorState.Width != _channel.TextureManager.ClipRegionWidth ||
+ screenScissorState.Height != _channel.TextureManager.ClipRegionHeight;
+
+ bool clearAffectedByStencilMask = (_state.State.ClearFlags & 1) != 0;
+ bool clearAffectedByScissor = (_state.State.ClearFlags & 0x100) != 0;
+ bool needsCustomScissor = !clearAffectedByScissor || clipMismatch;
+
+ // Scissor and rasterizer discard also affect clears.
+ ulong updateMask = 1UL << StateUpdater.RasterizerStateIndex;
+
+ if (!needsCustomScissor)
+ {
+ updateMask |= 1UL << StateUpdater.ScissorStateIndex;
+ }
+
+ engine.UpdateState(updateMask);
+
+ if (needsCustomScissor)
+ {
+ int scissorX = screenScissorState.X;
+ int scissorY = screenScissorState.Y;
+ int scissorW = screenScissorState.Width;
+ int scissorH = screenScissorState.Height;
+
+ if (clearAffectedByScissor && _state.State.ScissorState[0].Enable)
+ {
+ ref var scissorState = ref _state.State.ScissorState[0];
+
+ scissorX = Math.Max(scissorX, scissorState.X1);
+ scissorY = Math.Max(scissorY, scissorState.Y1);
+ scissorW = Math.Min(scissorW, scissorState.X2 - scissorState.X1);
+ scissorH = Math.Min(scissorH, scissorState.Y2 - scissorState.Y1);
+ }
+
+ float scale = _channel.TextureManager.RenderTargetScale;
+ if (scale != 1f)
+ {
+ scissorX = (int)(scissorX * scale);
+ scissorY = (int)(scissorY * scale);
+ scissorW = (int)MathF.Ceiling(scissorW * scale);
+ scissorH = (int)MathF.Ceiling(scissorH * scale);
+ }
+
+ Span<Rectangle<int>> scissors = stackalloc Rectangle<int>[]
+ {
+ new Rectangle<int>(scissorX, scissorY, scissorW, scissorH)
+ };
+
+ _context.Renderer.Pipeline.SetScissors(scissors);
+ }
+
+ _channel.TextureManager.UpdateRenderTargets();
+
+ if (componentMask != 0)
+ {
+ var clearColor = _state.State.ClearColors;
+
+ ColorF color = new ColorF(clearColor.Red, clearColor.Green, clearColor.Blue, clearColor.Alpha);
+
+ _context.Renderer.Pipeline.ClearRenderTargetColor(index, layer, layerCount, componentMask, color);
+ }
+
+ if (clearDepth || clearStencil)
+ {
+ float depthValue = _state.State.ClearDepthValue;
+ int stencilValue = (int)_state.State.ClearStencilValue;
+
+ int stencilMask = 0;
+
+ if (clearStencil)
+ {
+ stencilMask = clearAffectedByStencilMask ? _state.State.StencilTestState.FrontMask : 0xff;
+ }
+
+ if (clipMismatch)
+ {
+ _channel.TextureManager.UpdateRenderTargetDepthStencil();
+ }
+
+ _context.Renderer.Pipeline.ClearRenderTargetDepthStencil(
+ layer,
+ layerCount,
+ depthValue,
+ clearDepth,
+ stencilValue,
+ stencilMask);
+ }
+
+ if (needsCustomScissor)
+ {
+ engine.UpdateScissorState();
+ }
+
+ engine.UpdateRenderTargetState(RenderTargetUpdateFlags.UpdateAll);
+
+ if (renderEnable == ConditionalRenderEnabled.Host)
+ {
+ _context.Renderer.Pipeline.EndHostConditionalRendering();
+ }
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs
new file mode 100644
index 00000000..42ec2442
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs
@@ -0,0 +1,65 @@
+using Ryujinx.Graphics.GAL;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Draw state.
+ /// </summary>
+ class DrawState
+ {
+ /// <summary>
+ /// First index to be used for the draw on the index buffer.
+ /// </summary>
+ public int FirstIndex;
+
+ /// <summary>
+ /// Number of indices to be used for the draw on the index buffer.
+ /// </summary>
+ public int IndexCount;
+
+ /// <summary>
+ /// First vertex used on non-indexed draws. This value is stored somewhere else on indexed draws.
+ /// </summary>
+ public int DrawFirstVertex;
+
+ /// <summary>
+ /// Vertex count used on non-indexed draws. Indexed draws have a index count instead.
+ /// </summary>
+ public int DrawVertexCount;
+
+ /// <summary>
+ /// Indicates if the next draw will be a indexed draw.
+ /// </summary>
+ public bool DrawIndexed;
+
+ /// <summary>
+ /// Indicates if the next draw will be a indirect draw.
+ /// </summary>
+ public bool DrawIndirect;
+
+ /// <summary>
+ /// Indicates if any of the currently used vertex shaders reads the instance ID.
+ /// </summary>
+ public bool VsUsesInstanceId;
+
+ /// <summary>
+ /// Indicates if any of the currently used vertex buffers is instanced.
+ /// </summary>
+ public bool IsAnyVbInstanced;
+
+ /// <summary>
+ /// Indicates that the draw is writing the base vertex, base instance and draw index to Constant Buffer 0.
+ /// </summary>
+ public bool HasConstantBufferDrawParameters;
+
+ /// <summary>
+ /// Primitive topology for the next draw.
+ /// </summary>
+ public PrimitiveTopology Topology;
+
+ /// <summary>
+ /// Index buffer data streamer for inline index buffer updates, such as those used in legacy OpenGL.
+ /// </summary>
+ public IbStreamer IbStreamer = new IbStreamer();
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs
new file mode 100644
index 00000000..80d8c00b
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs
@@ -0,0 +1,194 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Holds inline index buffer state.
+ /// The inline index buffer data is sent to the GPU through the command buffer.
+ /// </summary>
+ struct IbStreamer
+ {
+ private const int BufferCapacity = 256; // Must be a power of 2.
+
+ private BufferHandle _inlineIndexBuffer;
+ private int _inlineIndexBufferSize;
+ private int _inlineIndexCount;
+ private uint[] _buffer;
+ private int _bufferOffset;
+
+ /// <summary>
+ /// Indicates if any index buffer data has been pushed.
+ /// </summary>
+ public bool HasInlineIndexData => _inlineIndexCount != 0;
+
+ /// <summary>
+ /// Total numbers of indices that have been pushed.
+ /// </summary>
+ public int InlineIndexCount => _inlineIndexCount;
+
+ /// <summary>
+ /// Gets the handle for the host buffer currently holding the inline index buffer data.
+ /// </summary>
+ /// <returns>Host buffer handle</returns>
+ public BufferHandle GetInlineIndexBuffer()
+ {
+ return _inlineIndexBuffer;
+ }
+
+ /// <summary>
+ /// Gets the number of elements on the current inline index buffer,
+ /// while also reseting it to zero for the next draw.
+ /// </summary>
+ /// <param name="renderer">Host renderer</param>
+ /// <returns>Inline index bufffer count</returns>
+ public int GetAndResetInlineIndexCount(IRenderer renderer)
+ {
+ UpdateRemaining(renderer);
+ int temp = _inlineIndexCount;
+ _inlineIndexCount = 0;
+ return temp;
+ }
+
+ /// <summary>
+ /// Pushes four 8-bit index buffer elements.
+ /// </summary>
+ /// <param name="renderer">Host renderer</param>
+ /// <param name="argument">Method call argument</param>
+ public void VbElementU8(IRenderer renderer, int argument)
+ {
+ byte i0 = (byte)argument;
+ byte i1 = (byte)(argument >> 8);
+ byte i2 = (byte)(argument >> 16);
+ byte i3 = (byte)(argument >> 24);
+
+ int offset = _inlineIndexCount;
+
+ PushData(renderer, offset, i0);
+ PushData(renderer, offset + 1, i1);
+ PushData(renderer, offset + 2, i2);
+ PushData(renderer, offset + 3, i3);
+
+ _inlineIndexCount += 4;
+ }
+
+ /// <summary>
+ /// Pushes two 16-bit index buffer elements.
+ /// </summary>
+ /// <param name="renderer">Host renderer</param>
+ /// <param name="argument">Method call argument</param>
+ public void VbElementU16(IRenderer renderer, int argument)
+ {
+ ushort i0 = (ushort)argument;
+ ushort i1 = (ushort)(argument >> 16);
+
+ int offset = _inlineIndexCount;
+
+ PushData(renderer, offset, i0);
+ PushData(renderer, offset + 1, i1);
+
+ _inlineIndexCount += 2;
+ }
+
+ /// <summary>
+ /// Pushes one 32-bit index buffer element.
+ /// </summary>
+ /// <param name="renderer">Host renderer</param>
+ /// <param name="argument">Method call argument</param>
+ public void VbElementU32(IRenderer renderer, int argument)
+ {
+ uint i0 = (uint)argument;
+
+ int offset = _inlineIndexCount++;
+
+ PushData(renderer, offset, i0);
+ }
+
+ /// <summary>
+ /// Pushes a 32-bit value to the index buffer.
+ /// </summary>
+ /// <param name="renderer">Host renderer</param>
+ /// <param name="offset">Offset where the data should be written, in 32-bit words</param>
+ /// <param name="value">Index value to be written</param>
+ private void PushData(IRenderer renderer, int offset, uint value)
+ {
+ if (_buffer == null)
+ {
+ _buffer = new uint[BufferCapacity];
+ }
+
+ // We upload data in chunks.
+ // If we are at the start of a chunk, then the buffer might be full,
+ // in that case we need to submit any existing data before overwriting the buffer.
+ int subOffset = offset & (BufferCapacity - 1);
+
+ if (subOffset == 0 && offset != 0)
+ {
+ int baseOffset = (offset - BufferCapacity) * sizeof(uint);
+ BufferHandle buffer = GetInlineIndexBuffer(renderer, baseOffset, BufferCapacity * sizeof(uint));
+ renderer.SetBufferData(buffer, baseOffset, MemoryMarshal.Cast<uint, byte>(_buffer));
+ }
+
+ _buffer[subOffset] = value;
+ }
+
+ /// <summary>
+ /// Makes sure that any pending data is submitted to the GPU before the index buffer is used.
+ /// </summary>
+ /// <param name="renderer">Host renderer</param>
+ private void UpdateRemaining(IRenderer renderer)
+ {
+ int offset = _inlineIndexCount;
+ if (offset == 0)
+ {
+ return;
+ }
+
+ int count = offset & (BufferCapacity - 1);
+ if (count == 0)
+ {
+ count = BufferCapacity;
+ }
+
+ int baseOffset = (offset - count) * sizeof(uint);
+ int length = count * sizeof(uint);
+ BufferHandle buffer = GetInlineIndexBuffer(renderer, baseOffset, length);
+ renderer.SetBufferData(buffer, baseOffset, MemoryMarshal.Cast<uint, byte>(_buffer).Slice(0, length));
+ }
+
+ /// <summary>
+ /// Gets the handle of a buffer large enough to hold the data that will be written to <paramref name="offset"/>.
+ /// </summary>
+ /// <param name="renderer">Host renderer</param>
+ /// <param name="offset">Offset where the data will be written</param>
+ /// <param name="length">Number of bytes that will be written</param>
+ /// <returns>Buffer handle</returns>
+ private BufferHandle GetInlineIndexBuffer(IRenderer renderer, int offset, int length)
+ {
+ // Calculate a reasonable size for the buffer that can fit all the data,
+ // and that also won't require frequent resizes if we need to push more data.
+ int size = BitUtils.AlignUp(offset + length + 0x10, 0x200);
+
+ if (_inlineIndexBuffer == BufferHandle.Null)
+ {
+ _inlineIndexBuffer = renderer.CreateBuffer(size);
+ _inlineIndexBufferSize = size;
+ }
+ else if (_inlineIndexBufferSize < size)
+ {
+ BufferHandle oldBuffer = _inlineIndexBuffer;
+ int oldSize = _inlineIndexBufferSize;
+
+ _inlineIndexBuffer = renderer.CreateBuffer(size);
+ _inlineIndexBufferSize = size;
+
+ renderer.Pipeline.CopyBuffer(oldBuffer, _inlineIndexBuffer, 0, 0, oldSize);
+ renderer.DeleteBuffer(oldBuffer);
+ }
+
+ return _inlineIndexBuffer;
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/IndirectDrawType.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IndirectDrawType.cs
new file mode 100644
index 00000000..d78aa498
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IndirectDrawType.cs
@@ -0,0 +1,38 @@
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Indirect draw type, which can be indexed or non-indexed, with or without a draw count.
+ /// </summary>
+ enum IndirectDrawType
+ {
+ /// <summary>
+ /// Non-indexed draw without draw count.
+ /// </summary>
+ DrawIndirect = 0,
+
+ /// <summary>
+ /// Indexed draw without draw count.
+ /// </summary>
+ DrawIndexedIndirect = Indexed,
+
+ /// <summary>
+ /// Non-indexed draw with draw count.
+ /// </summary>
+ DrawIndirectCount = Count,
+
+ /// <summary>
+ /// Indexed draw with draw count.
+ /// </summary>
+ DrawIndexedIndirectCount = Indexed | Count,
+
+ /// <summary>
+ /// Indexed flag.
+ /// </summary>
+ Indexed = 1 << 0,
+
+ /// <summary>
+ /// Draw count flag.
+ /// </summary>
+ Count = 1 << 1
+ }
+} \ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/RenderTargetUpdateFlags.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/RenderTargetUpdateFlags.cs
new file mode 100644
index 00000000..cf2e818c
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/RenderTargetUpdateFlags.cs
@@ -0,0 +1,41 @@
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Flags indicating how the render targets should be updated.
+ /// </summary>
+ [Flags]
+ enum RenderTargetUpdateFlags
+ {
+ /// <summary>
+ /// No flags.
+ /// </summary>
+ None = 0,
+
+ /// <summary>
+ /// Get render target index from the control register.
+ /// </summary>
+ UseControl = 1 << 0,
+
+ /// <summary>
+ /// Indicates that all render targets are 2D array textures.
+ /// </summary>
+ Layered = 1 << 1,
+
+ /// <summary>
+ /// Indicates that only a single color target will be used.
+ /// </summary>
+ SingleColor = 1 << 2,
+
+ /// <summary>
+ /// Indicates that the depth-stencil target will be used.
+ /// </summary>
+ UpdateDepthStencil = 1 << 3,
+
+ /// <summary>
+ /// Default update flags for draw.
+ /// </summary>
+ UpdateAll = UseControl | UpdateDepthStencil
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs
new file mode 100644
index 00000000..63a2c841
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs
@@ -0,0 +1,190 @@
+using Ryujinx.Graphics.GAL;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Semaphore updater.
+ /// </summary>
+ class SemaphoreUpdater
+ {
+ /// <summary>
+ /// GPU semaphore operation.
+ /// </summary>
+ private enum SemaphoreOperation
+ {
+ Release = 0,
+ Acquire = 1,
+ Counter = 2
+ }
+
+ /// <summary>
+ /// Counter type for GPU counter reset.
+ /// </summary>
+ private enum ResetCounterType
+ {
+ SamplesPassed = 1,
+ ZcullStats = 2,
+ TransformFeedbackPrimitivesWritten = 0x10,
+ InputVertices = 0x12,
+ InputPrimitives = 0x13,
+ VertexShaderInvocations = 0x15,
+ TessControlShaderInvocations = 0x16,
+ TessEvaluationShaderInvocations = 0x17,
+ TessEvaluationShaderPrimitives = 0x18,
+ GeometryShaderInvocations = 0x1a,
+ GeometryShaderPrimitives = 0x1b,
+ ClipperInputPrimitives = 0x1c,
+ ClipperOutputPrimitives = 0x1d,
+ FragmentShaderInvocations = 0x1e,
+ PrimitivesGenerated = 0x1f
+ }
+
+ /// <summary>
+ /// Counter type for GPU counter reporting.
+ /// </summary>
+ private enum ReportCounterType
+ {
+ Payload = 0,
+ InputVertices = 1,
+ InputPrimitives = 3,
+ VertexShaderInvocations = 5,
+ GeometryShaderInvocations = 7,
+ GeometryShaderPrimitives = 9,
+ ZcullStats0 = 0xa,
+ TransformFeedbackPrimitivesWritten = 0xb,
+ ZcullStats1 = 0xc,
+ ZcullStats2 = 0xe,
+ ClipperInputPrimitives = 0xf,
+ ZcullStats3 = 0x10,
+ ClipperOutputPrimitives = 0x11,
+ PrimitivesGenerated = 0x12,
+ FragmentShaderInvocations = 0x13,
+ SamplesPassed = 0x15,
+ TransformFeedbackOffset = 0x1a,
+ TessControlShaderInvocations = 0x1b,
+ TessEvaluationShaderInvocations = 0x1d,
+ TessEvaluationShaderPrimitives = 0x1f
+ }
+
+ private readonly GpuContext _context;
+ private readonly GpuChannel _channel;
+ private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+ /// <summary>
+ /// Creates a new instance of the semaphore updater.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ /// <param name="channel">GPU channel</param>
+ /// <param name="state">Channel state</param>
+ public SemaphoreUpdater(GpuContext context, GpuChannel channel, DeviceStateWithShadow<ThreedClassState> state)
+ {
+ _context = context;
+ _channel = channel;
+ _state = state;
+ }
+
+ /// <summary>
+ /// Resets the value of an internal GPU counter back to zero.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void ResetCounter(int argument)
+ {
+ ResetCounterType type = (ResetCounterType)argument;
+
+ switch (type)
+ {
+ case ResetCounterType.SamplesPassed:
+ _context.Renderer.ResetCounter(CounterType.SamplesPassed);
+ break;
+ case ResetCounterType.PrimitivesGenerated:
+ _context.Renderer.ResetCounter(CounterType.PrimitivesGenerated);
+ break;
+ case ResetCounterType.TransformFeedbackPrimitivesWritten:
+ _context.Renderer.ResetCounter(CounterType.TransformFeedbackPrimitivesWritten);
+ break;
+ }
+ }
+
+ /// <summary>
+ /// Writes a GPU counter to guest memory.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void Report(int argument)
+ {
+ SemaphoreOperation op = (SemaphoreOperation)(argument & 3);
+ ReportCounterType type = (ReportCounterType)((argument >> 23) & 0x1f);
+
+ switch (op)
+ {
+ case SemaphoreOperation.Release: ReleaseSemaphore(); break;
+ case SemaphoreOperation.Counter: ReportCounter(type); break;
+ }
+ }
+
+ /// <summary>
+ /// Writes (or Releases) a GPU semaphore value to guest memory.
+ /// </summary>
+ private void ReleaseSemaphore()
+ {
+ _channel.MemoryManager.Write(_state.State.SemaphoreAddress.Pack(), _state.State.SemaphorePayload);
+
+ _context.AdvanceSequence();
+ }
+
+ /// <summary>
+ /// Packed GPU counter data (including GPU timestamp) in memory.
+ /// </summary>
+ private struct CounterData
+ {
+ public ulong Counter;
+ public ulong Timestamp;
+ }
+
+ /// <summary>
+ /// Writes a GPU counter to guest memory.
+ /// This also writes the current timestamp value.
+ /// </summary>
+ /// <param name="type">Counter to be written to memory</param>
+ private void ReportCounter(ReportCounterType type)
+ {
+ ulong gpuVa = _state.State.SemaphoreAddress.Pack();
+
+ ulong ticks = _context.GetTimestamp();
+
+ ICounterEvent counter = null;
+
+ void resultHandler(object evt, ulong result)
+ {
+ CounterData counterData = new CounterData
+ {
+ Counter = result,
+ Timestamp = ticks
+ };
+
+ if (counter?.Invalid != true)
+ {
+ _channel.MemoryManager.Write(gpuVa, counterData);
+ }
+ }
+
+ switch (type)
+ {
+ case ReportCounterType.Payload:
+ resultHandler(null, (ulong)_state.State.SemaphorePayload);
+ break;
+ case ReportCounterType.SamplesPassed:
+ counter = _context.Renderer.ReportCounter(CounterType.SamplesPassed, resultHandler, false);
+ break;
+ case ReportCounterType.PrimitivesGenerated:
+ counter = _context.Renderer.ReportCounter(CounterType.PrimitivesGenerated, resultHandler, false);
+ break;
+ case ReportCounterType.TransformFeedbackPrimitivesWritten:
+ counter = _context.Renderer.ReportCounter(CounterType.TransformFeedbackPrimitivesWritten, resultHandler, false);
+ break;
+ }
+
+ _channel.MemoryManager.CounterCache.AddOrUpdate(gpuVa, counter);
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/SpecializationStateUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SpecializationStateUpdater.cs
new file mode 100644
index 00000000..a8af5497
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SpecializationStateUpdater.cs
@@ -0,0 +1,346 @@
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Shader;
+using Ryujinx.Graphics.Shader;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Maintains a "current" specialiation state, and provides a flag to check if it has changed meaningfully.
+ /// </summary>
+ internal class SpecializationStateUpdater
+ {
+ private readonly GpuContext _context;
+ private GpuChannelGraphicsState _graphics;
+ private GpuChannelPoolState _pool;
+
+ private bool _usesDrawParameters;
+ private bool _usesTopology;
+
+ private bool _changed;
+
+ /// <summary>
+ /// Creates a new instance of the specialization state updater class.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ public SpecializationStateUpdater(GpuContext context)
+ {
+ _context = context;
+ }
+
+ /// <summary>
+ /// Signal that the specialization state has changed.
+ /// </summary>
+ private void Signal()
+ {
+ _changed = true;
+ }
+
+ /// <summary>
+ /// Checks if the specialization state has changed since the last check.
+ /// </summary>
+ /// <returns>True if it has changed, false otherwise</returns>
+ public bool HasChanged()
+ {
+ if (_changed)
+ {
+ _changed = false;
+ return true;
+ }
+ else
+ {
+ return false;
+ }
+ }
+
+ /// <summary>
+ /// Sets the active shader, clearing the dirty state and recording if certain specializations are noteworthy.
+ /// </summary>
+ /// <param name="gs">The active shader</param>
+ public void SetShader(CachedShaderProgram gs)
+ {
+ _usesDrawParameters = gs.Shaders[1]?.Info.UsesDrawParameters ?? false;
+ _usesTopology = gs.SpecializationState.IsPrimitiveTopologyQueried();
+
+ _changed = false;
+ }
+
+ /// <summary>
+ /// Get the current graphics state.
+ /// </summary>
+ /// <returns>GPU graphics state</returns>
+ public ref GpuChannelGraphicsState GetGraphicsState()
+ {
+ return ref _graphics;
+ }
+
+ /// <summary>
+ /// Get the current pool state.
+ /// </summary>
+ /// <returns>GPU pool state</returns>
+ public ref GpuChannelPoolState GetPoolState()
+ {
+ return ref _pool;
+ }
+
+ /// <summary>
+ /// Early Z force enable.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetEarlyZForce(bool value)
+ {
+ _graphics.EarlyZForce = value;
+
+ Signal();
+ }
+
+ /// <summary>
+ /// Primitive topology of current draw.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetTopology(PrimitiveTopology value)
+ {
+ if (value != _graphics.Topology)
+ {
+ _graphics.Topology = value;
+
+ if (_usesTopology)
+ {
+ Signal();
+ }
+ }
+ }
+
+ /// <summary>
+ /// Tessellation mode.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetTessellationMode(TessMode value)
+ {
+ if (value.Packed != _graphics.TessellationMode.Packed)
+ {
+ _graphics.TessellationMode = value;
+
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Updates alpha-to-coverage state, and sets it as changed.
+ /// </summary>
+ /// <param name="enable">Whether alpha-to-coverage is enabled</param>
+ /// <param name="ditherEnable">Whether alpha-to-coverage dithering is enabled</param>
+ public void SetAlphaToCoverageEnable(bool enable, bool ditherEnable)
+ {
+ _graphics.AlphaToCoverageEnable = enable;
+ _graphics.AlphaToCoverageDitherEnable = ditherEnable;
+
+ Signal();
+ }
+
+ /// <summary>
+ /// Indicates whether the viewport transform is disabled.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetViewportTransformDisable(bool value)
+ {
+ if (value != _graphics.ViewportTransformDisable)
+ {
+ _graphics.ViewportTransformDisable = value;
+
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Depth mode zero to one or minus one to one.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetDepthMode(bool value)
+ {
+ if (value != _graphics.DepthMode)
+ {
+ _graphics.DepthMode = value;
+
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Indicates if the point size is set on the shader or is fixed.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetProgramPointSizeEnable(bool value)
+ {
+ if (value != _graphics.ProgramPointSizeEnable)
+ {
+ _graphics.ProgramPointSizeEnable = value;
+
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Point size used if <see cref="SetProgramPointSizeEnable" /> is provided false.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetPointSize(float value)
+ {
+ if (value != _graphics.PointSize)
+ {
+ _graphics.PointSize = value;
+
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Updates alpha test specialization state, and sets it as changed.
+ /// </summary>
+ /// <param name="enable">Whether alpha test is enabled</param>
+ /// <param name="reference">The value to compare with the fragment output alpha</param>
+ /// <param name="op">The comparison that decides if the fragment should be discarded</param>
+ public void SetAlphaTest(bool enable, float reference, CompareOp op)
+ {
+ _graphics.AlphaTestEnable = enable;
+ _graphics.AlphaTestReference = reference;
+ _graphics.AlphaTestCompare = op;
+
+ Signal();
+ }
+
+ /// <summary>
+ /// Updates the type of the vertex attributes consumed by the shader.
+ /// </summary>
+ /// <param name="state">The new state</param>
+ public void SetAttributeTypes(ref Array32<VertexAttribState> state)
+ {
+ bool changed = false;
+ ref Array32<AttributeType> attributeTypes = ref _graphics.AttributeTypes;
+
+ for (int location = 0; location < state.Length; location++)
+ {
+ VertexAttribType type = state[location].UnpackType();
+
+ AttributeType value = type switch
+ {
+ VertexAttribType.Sint => AttributeType.Sint,
+ VertexAttribType.Uint => AttributeType.Uint,
+ _ => AttributeType.Float
+ };
+
+ if (attributeTypes[location] != value)
+ {
+ attributeTypes[location] = value;
+ changed = true;
+ }
+ }
+
+ if (changed)
+ {
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Updates the type of the outputs produced by the fragment shader based on the current render target state.
+ /// </summary>
+ /// <param name="rtControl">The render target control register</param>
+ /// <param name="state">The color attachment state</param>
+ public void SetFragmentOutputTypes(RtControl rtControl, ref Array8<RtColorState> state)
+ {
+ bool changed = false;
+ int count = rtControl.UnpackCount();
+
+ for (int index = 0; index < Constants.TotalRenderTargets; index++)
+ {
+ int rtIndex = rtControl.UnpackPermutationIndex(index);
+
+ var colorState = state[rtIndex];
+
+ if (index < count && StateUpdater.IsRtEnabled(colorState))
+ {
+ Format format = colorState.Format.Convert().Format;
+
+ AttributeType type = format.IsInteger() ? (format.IsSint() ? AttributeType.Sint : AttributeType.Uint) : AttributeType.Float;
+
+ if (type != _graphics.FragmentOutputTypes[index])
+ {
+ _graphics.FragmentOutputTypes[index] = type;
+ changed = true;
+ }
+ }
+ }
+
+ if (changed && _context.Capabilities.NeedsFragmentOutputSpecialization)
+ {
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Indicates that the draw is writing the base vertex, base instance and draw index to Constant Buffer 0.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ public void SetHasConstantBufferDrawParameters(bool value)
+ {
+ if (value != _graphics.HasConstantBufferDrawParameters)
+ {
+ _graphics.HasConstantBufferDrawParameters = value;
+
+ if (_usesDrawParameters)
+ {
+ Signal();
+ }
+ }
+ }
+
+ /// <summary>
+ /// Indicates that any storage buffer use is unaligned.
+ /// </summary>
+ /// <param name="value">The new value</param>
+ /// <returns>True if the unaligned state changed, false otherwise</returns>
+ public bool SetHasUnalignedStorageBuffer(bool value)
+ {
+ if (value != _graphics.HasUnalignedStorageBuffer)
+ {
+ _graphics.HasUnalignedStorageBuffer = value;
+
+ Signal();
+
+ return true;
+ }
+
+ return false;
+ }
+
+ /// <summary>
+ /// Sets the GPU pool state.
+ /// </summary>
+ /// <param name="state">The new state</param>
+ public void SetPoolState(GpuChannelPoolState state)
+ {
+ if (!state.Equals(_pool))
+ {
+ _pool = state;
+
+ Signal();
+ }
+ }
+
+ /// <summary>
+ /// Sets the dual-source blend enabled state.
+ /// </summary>
+ /// <param name="enabled">True if blending is enabled and using dual-source blend</param>
+ public void SetDualSourceBlendEnabled(bool enabled)
+ {
+ if (enabled != _graphics.DualSourceBlendEnable)
+ {
+ _graphics.DualSourceBlendEnable = enabled;
+
+ Signal();
+ }
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs
new file mode 100644
index 00000000..7c730967
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs
@@ -0,0 +1,177 @@
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+using System.Diagnostics;
+using System.Diagnostics.CodeAnalysis;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// State update callback entry, with the callback function and associated field names.
+ /// </summary>
+ readonly struct StateUpdateCallbackEntry
+ {
+ /// <summary>
+ /// Callback function, to be called if the register was written as the state needs to be updated.
+ /// </summary>
+ public Action Callback { get; }
+
+ /// <summary>
+ /// Name of the state fields (registers) associated with the callback function.
+ /// </summary>
+ public string[] FieldNames { get; }
+
+ /// <summary>
+ /// Creates a new state update callback entry.
+ /// </summary>
+ /// <param name="callback">Callback function, to be called if the register was written as the state needs to be updated</param>
+ /// <param name="fieldNames">Name of the state fields (registers) associated with the callback function</param>
+ public StateUpdateCallbackEntry(Action callback, params string[] fieldNames)
+ {
+ Callback = callback;
+ FieldNames = fieldNames;
+ }
+ }
+
+ /// <summary>
+ /// GPU state update tracker.
+ /// </summary>
+ /// <typeparam name="TState">State type</typeparam>
+ class StateUpdateTracker<[DynamicallyAccessedMembers(DynamicallyAccessedMemberTypes.PublicFields)] TState>
+ {
+ private const int BlockSize = 0xe00;
+ private const int RegisterSize = sizeof(uint);
+
+ private readonly byte[] _registerToGroupMapping;
+ private readonly Action[] _callbacks;
+ private ulong _dirtyMask;
+
+ /// <summary>
+ /// Creates a new instance of the state update tracker.
+ /// </summary>
+ /// <param name="entries">Update tracker callback entries</param>
+ public StateUpdateTracker(StateUpdateCallbackEntry[] entries)
+ {
+ _registerToGroupMapping = new byte[BlockSize];
+ _callbacks = new Action[entries.Length];
+
+ var fieldToDelegate = new Dictionary<string, int>();
+
+ for (int entryIndex = 0; entryIndex < entries.Length; entryIndex++)
+ {
+ var entry = entries[entryIndex];
+
+ foreach (var fieldName in entry.FieldNames)
+ {
+ fieldToDelegate.Add(fieldName, entryIndex);
+ }
+
+ _callbacks[entryIndex] = entry.Callback;
+ }
+
+ var fields = typeof(TState).GetFields();
+ int offset = 0;
+
+ for (int fieldIndex = 0; fieldIndex < fields.Length; fieldIndex++)
+ {
+ var field = fields[fieldIndex];
+
+ int sizeOfField = SizeCalculator.SizeOf(field.FieldType);
+
+ if (fieldToDelegate.TryGetValue(field.Name, out int entryIndex))
+ {
+ for (int i = 0; i < ((sizeOfField + 3) & ~3); i += 4)
+ {
+ _registerToGroupMapping[(offset + i) / RegisterSize] = (byte)(entryIndex + 1);
+ }
+ }
+
+ offset += sizeOfField;
+ }
+
+ Debug.Assert(offset == Unsafe.SizeOf<TState>());
+ }
+
+ /// <summary>
+ /// Sets a register as modified.
+ /// </summary>
+ /// <param name="offset">Register offset in bytes</param>
+ [MethodImpl(MethodImplOptions.AggressiveInlining)]
+ public void SetDirty(int offset)
+ {
+ uint index = (uint)offset / RegisterSize;
+
+ if (index < BlockSize)
+ {
+ int groupIndex = Unsafe.Add(ref MemoryMarshal.GetArrayDataReference(_registerToGroupMapping), (IntPtr)index);
+ if (groupIndex != 0)
+ {
+ groupIndex--;
+ _dirtyMask |= 1UL << groupIndex;
+ }
+ }
+ }
+
+ /// <summary>
+ /// Forces a register group as dirty, by index.
+ /// </summary>
+ /// <param name="groupIndex">Index of the group to be dirtied</param>
+ public void ForceDirty(int groupIndex)
+ {
+ if ((uint)groupIndex >= _callbacks.Length)
+ {
+ throw new ArgumentOutOfRangeException(nameof(groupIndex));
+ }
+
+ _dirtyMask |= 1UL << groupIndex;
+ }
+
+ /// <summary>
+ /// Forces all register groups as dirty, triggering a full update on the next call to <see cref="Update"/>.
+ /// </summary>
+ public void SetAllDirty()
+ {
+ Debug.Assert(_callbacks.Length <= sizeof(ulong) * 8);
+ _dirtyMask = ulong.MaxValue >> ((sizeof(ulong) * 8) - _callbacks.Length);
+ }
+
+ /// <summary>
+ /// Check if the given register group is dirty without clearing it.
+ /// </summary>
+ /// <param name="groupIndex">Index of the group to check</param>
+ /// <returns>True if dirty, false otherwise</returns>
+ public bool IsDirty(int groupIndex)
+ {
+ return (_dirtyMask & (1UL << groupIndex)) != 0;
+ }
+
+ /// <summary>
+ /// Check all the groups specified by <paramref name="checkMask"/> for modification, and update if modified.
+ /// </summary>
+ /// <param name="checkMask">Mask, where each bit set corresponds to a group index that should be checked</param>
+ [MethodImpl(MethodImplOptions.AggressiveInlining)]
+ public void Update(ulong checkMask)
+ {
+ ulong mask = _dirtyMask & checkMask;
+ if (mask == 0)
+ {
+ return;
+ }
+
+ do
+ {
+ int groupIndex = BitOperations.TrailingZeroCount(mask);
+
+ _callbacks[groupIndex]();
+
+ mask &= ~(1UL << groupIndex);
+ }
+ while (mask != 0);
+
+ _dirtyMask &= ~checkMask;
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs
new file mode 100644
index 00000000..00e09a31
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs
@@ -0,0 +1,1448 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Threed.Blender;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Image;
+using Ryujinx.Graphics.Gpu.Shader;
+using Ryujinx.Graphics.Shader;
+using Ryujinx.Graphics.Texture;
+using System;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// GPU state updater.
+ /// </summary>
+ class StateUpdater
+ {
+ public const int ShaderStateIndex = 26;
+ public const int RasterizerStateIndex = 15;
+ public const int ScissorStateIndex = 16;
+ public const int VertexBufferStateIndex = 0;
+ public const int PrimitiveRestartStateIndex = 12;
+ public const int RenderTargetStateIndex = 27;
+
+ private readonly GpuContext _context;
+ private readonly GpuChannel _channel;
+ private readonly DeviceStateWithShadow<ThreedClassState> _state;
+ private readonly DrawState _drawState;
+ private readonly AdvancedBlendManager _blendManager;
+
+ private readonly StateUpdateTracker<ThreedClassState> _updateTracker;
+
+ private readonly ShaderProgramInfo[] _currentProgramInfo;
+ private ShaderSpecializationState _shaderSpecState;
+ private SpecializationStateUpdater _currentSpecState;
+
+ private ProgramPipelineState _pipeline;
+
+ private bool _vsUsesDrawParameters;
+ private bool _vtgWritesRtLayer;
+ private byte _vsClipDistancesWritten;
+ private uint _vbEnableMask;
+
+ private bool _prevDrawIndexed;
+ private bool _prevDrawIndirect;
+ private IndexType _prevIndexType;
+ private uint _prevFirstVertex;
+ private bool _prevTfEnable;
+
+ private uint _prevRtNoAlphaMask;
+
+ /// <summary>
+ /// Creates a new instance of the state updater.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ /// <param name="channel">GPU channel</param>
+ /// <param name="state">3D engine state</param>
+ /// <param name="drawState">Draw state</param>
+ /// <param name="blendManager">Advanced blend manager</param>
+ /// <param name="spec">Specialization state updater</param>
+ public StateUpdater(
+ GpuContext context,
+ GpuChannel channel,
+ DeviceStateWithShadow<ThreedClassState> state,
+ DrawState drawState,
+ AdvancedBlendManager blendManager,
+ SpecializationStateUpdater spec)
+ {
+ _context = context;
+ _channel = channel;
+ _state = state;
+ _drawState = drawState;
+ _blendManager = blendManager;
+ _currentProgramInfo = new ShaderProgramInfo[Constants.ShaderStages];
+ _currentSpecState = spec;
+
+ // ShaderState must be updated after other state updates, as specialization/pipeline state is used when fetching shaders.
+ // Render target state must appear after shader state as it depends on information from the currently bound shader.
+ // Rasterizer and scissor states are checked by render target clear, their indexes
+ // must be updated on the constants "RasterizerStateIndex" and "ScissorStateIndex" if modified.
+ // The vertex buffer state may be forced dirty when a indexed draw starts, the "VertexBufferStateIndex"
+ // constant must be updated if modified.
+ // The order of the other state updates doesn't matter.
+ _updateTracker = new StateUpdateTracker<ThreedClassState>(new[]
+ {
+ new StateUpdateCallbackEntry(UpdateVertexBufferState,
+ nameof(ThreedClassState.VertexBufferDrawState),
+ nameof(ThreedClassState.VertexBufferInstanced),
+ nameof(ThreedClassState.VertexBufferState),
+ nameof(ThreedClassState.VertexBufferEndAddress)),
+
+ // Must be done after vertex buffer updates.
+ new StateUpdateCallbackEntry(UpdateVertexAttribState, nameof(ThreedClassState.VertexAttribState)),
+
+ new StateUpdateCallbackEntry(UpdateBlendState,
+ nameof(ThreedClassState.BlendUcodeEnable),
+ nameof(ThreedClassState.BlendUcodeSize),
+ nameof(ThreedClassState.BlendIndependent),
+ nameof(ThreedClassState.BlendConstant),
+ nameof(ThreedClassState.BlendStateCommon),
+ nameof(ThreedClassState.BlendEnableCommon),
+ nameof(ThreedClassState.BlendEnable),
+ nameof(ThreedClassState.BlendState)),
+
+ new StateUpdateCallbackEntry(UpdateFaceState, nameof(ThreedClassState.FaceState)),
+
+ new StateUpdateCallbackEntry(UpdateStencilTestState,
+ nameof(ThreedClassState.StencilBackMasks),
+ nameof(ThreedClassState.StencilTestState),
+ nameof(ThreedClassState.StencilBackTestState)),
+
+ new StateUpdateCallbackEntry(UpdateDepthTestState,
+ nameof(ThreedClassState.DepthTestEnable),
+ nameof(ThreedClassState.DepthWriteEnable),
+ nameof(ThreedClassState.DepthTestFunc)),
+
+ new StateUpdateCallbackEntry(UpdateTessellationState,
+ nameof(ThreedClassState.TessMode),
+ nameof(ThreedClassState.TessOuterLevel),
+ nameof(ThreedClassState.TessInnerLevel),
+ nameof(ThreedClassState.PatchVertices)),
+
+ new StateUpdateCallbackEntry(UpdateViewportTransform,
+ nameof(ThreedClassState.DepthMode),
+ nameof(ThreedClassState.ViewportTransform),
+ nameof(ThreedClassState.ViewportExtents),
+ nameof(ThreedClassState.YControl),
+ nameof(ThreedClassState.ViewportTransformEnable)),
+
+ new StateUpdateCallbackEntry(UpdateLogicOpState, nameof(ThreedClassState.LogicOpState)),
+
+ new StateUpdateCallbackEntry(UpdateDepthClampState, nameof(ThreedClassState.ViewVolumeClipControl)),
+
+ new StateUpdateCallbackEntry(UpdatePolygonMode,
+ nameof(ThreedClassState.PolygonModeFront),
+ nameof(ThreedClassState.PolygonModeBack)),
+
+ new StateUpdateCallbackEntry(UpdateDepthBiasState,
+ nameof(ThreedClassState.DepthBiasState),
+ nameof(ThreedClassState.DepthBiasFactor),
+ nameof(ThreedClassState.DepthBiasUnits),
+ nameof(ThreedClassState.DepthBiasClamp)),
+
+ new StateUpdateCallbackEntry(UpdatePrimitiveRestartState, nameof(ThreedClassState.PrimitiveRestartState)),
+
+ new StateUpdateCallbackEntry(UpdateLineState,
+ nameof(ThreedClassState.LineWidthSmooth),
+ nameof(ThreedClassState.LineSmoothEnable)),
+
+ new StateUpdateCallbackEntry(UpdateRtColorMask,
+ nameof(ThreedClassState.RtColorMaskShared),
+ nameof(ThreedClassState.RtColorMask)),
+
+ new StateUpdateCallbackEntry(UpdateRasterizerState, nameof(ThreedClassState.RasterizeEnable)),
+
+ new StateUpdateCallbackEntry(UpdateScissorState,
+ nameof(ThreedClassState.ScissorState),
+ nameof(ThreedClassState.ScreenScissorState)),
+
+ new StateUpdateCallbackEntry(UpdateTfBufferState, nameof(ThreedClassState.TfBufferState)),
+ new StateUpdateCallbackEntry(UpdateUserClipState, nameof(ThreedClassState.ClipDistanceEnable)),
+
+ new StateUpdateCallbackEntry(UpdateAlphaTestState,
+ nameof(ThreedClassState.AlphaTestEnable),
+ nameof(ThreedClassState.AlphaTestRef),
+ nameof(ThreedClassState.AlphaTestFunc)),
+
+ new StateUpdateCallbackEntry(UpdateSamplerPoolState,
+ nameof(ThreedClassState.SamplerPoolState),
+ nameof(ThreedClassState.SamplerIndex)),
+
+ new StateUpdateCallbackEntry(UpdateTexturePoolState, nameof(ThreedClassState.TexturePoolState)),
+
+ new StateUpdateCallbackEntry(UpdatePointState,
+ nameof(ThreedClassState.PointSize),
+ nameof(ThreedClassState.VertexProgramPointSize),
+ nameof(ThreedClassState.PointSpriteEnable),
+ nameof(ThreedClassState.PointCoordReplace)),
+
+ new StateUpdateCallbackEntry(UpdateIndexBufferState,
+ nameof(ThreedClassState.IndexBufferState),
+ nameof(ThreedClassState.IndexBufferCount)),
+
+ new StateUpdateCallbackEntry(UpdateMultisampleState,
+ nameof(ThreedClassState.AlphaToCoverageDitherEnable),
+ nameof(ThreedClassState.MultisampleControl)),
+
+ new StateUpdateCallbackEntry(UpdateEarlyZState,
+ nameof(ThreedClassState.EarlyZForce)),
+
+ new StateUpdateCallbackEntry(UpdateShaderState,
+ nameof(ThreedClassState.ShaderBaseAddress),
+ nameof(ThreedClassState.ShaderState)),
+
+ new StateUpdateCallbackEntry(UpdateRenderTargetState,
+ nameof(ThreedClassState.RtColorState),
+ nameof(ThreedClassState.RtDepthStencilState),
+ nameof(ThreedClassState.RtControl),
+ nameof(ThreedClassState.RtDepthStencilSize),
+ nameof(ThreedClassState.RtDepthStencilEnable)),
+ });
+ }
+
+ /// <summary>
+ /// Sets a register at a specific offset as dirty.
+ /// This must be called if the register value was modified.
+ /// </summary>
+ /// <param name="offset">Register offset</param>
+ [MethodImpl(MethodImplOptions.AggressiveInlining)]
+ public void SetDirty(int offset)
+ {
+ _updateTracker.SetDirty(offset);
+ }
+
+ /// <summary>
+ /// Force all the guest state to be marked as dirty.
+ /// The next call to <see cref="Update"/> will update all the host state.
+ /// </summary>
+ public void SetAllDirty()
+ {
+ _updateTracker.SetAllDirty();
+ }
+
+ /// <summary>
+ /// Updates host state for any modified guest state, since the last time this function was called.
+ /// </summary>
+ [MethodImpl(MethodImplOptions.AggressiveInlining)]
+ public void Update()
+ {
+ // The vertex buffer size is calculated using a different
+ // method when doing indexed draws, so we need to make sure
+ // to update the vertex buffers if we are doing a regular
+ // draw after a indexed one and vice-versa.
+ if (_drawState.DrawIndexed != _prevDrawIndexed)
+ {
+ _updateTracker.ForceDirty(VertexBufferStateIndex);
+
+ // If PrimitiveRestartDrawArrays is false and this is a non-indexed draw, we need to ensure primitive restart is disabled.
+ // If PrimitiveRestartDrawArrays is false and this is a indexed draw, we need to ensure primitive restart enable matches GPU state.
+ // If PrimitiveRestartDrawArrays is true, then primitive restart enable should always match GPU state.
+ // That is because "PrimitiveRestartDrawArrays" is not configurable on the backend, it is always
+ // true on OpenGL and always false on Vulkan.
+ if (!_state.State.PrimitiveRestartDrawArrays && _state.State.PrimitiveRestartState.Enable)
+ {
+ _updateTracker.ForceDirty(PrimitiveRestartStateIndex);
+ }
+
+ _prevDrawIndexed = _drawState.DrawIndexed;
+ }
+
+ // Some draw parameters are used to restrict the vertex buffer size,
+ // but they can't be used on indirect draws because their values are unknown in this case.
+ // When switching between indirect and non-indirect draw, we need to
+ // make sure the vertex buffer sizes are still correct.
+ if (_drawState.DrawIndirect != _prevDrawIndirect)
+ {
+ _updateTracker.ForceDirty(VertexBufferStateIndex);
+ }
+
+ // In some cases, the index type is also used to guess the
+ // vertex buffer size, so we must update it if the type changed too.
+ if (_drawState.DrawIndexed &&
+ (_prevIndexType != _state.State.IndexBufferState.Type ||
+ _prevFirstVertex != _state.State.FirstVertex))
+ {
+ _updateTracker.ForceDirty(VertexBufferStateIndex);
+ _prevIndexType = _state.State.IndexBufferState.Type;
+ _prevFirstVertex = _state.State.FirstVertex;
+ }
+
+ bool tfEnable = _state.State.TfEnable;
+
+ if (!tfEnable && _prevTfEnable)
+ {
+ _context.Renderer.Pipeline.EndTransformFeedback();
+ _prevTfEnable = false;
+ }
+
+ if (_updateTracker.IsDirty(RenderTargetStateIndex))
+ {
+ UpdateRenderTargetSpecialization();
+ }
+
+ _updateTracker.Update(ulong.MaxValue);
+
+ // If any state that the shader depends on changed,
+ // then we may need to compile/bind a different version
+ // of the shader for the new state.
+ if (_shaderSpecState != null && _currentSpecState.HasChanged())
+ {
+ if (!_shaderSpecState.MatchesGraphics(_channel, ref _currentSpecState.GetPoolState(), ref _currentSpecState.GetGraphicsState(), _vsUsesDrawParameters, false))
+ {
+ // Shader must be reloaded. _vtgWritesRtLayer should not change.
+ UpdateShaderState();
+ }
+ }
+
+ CommitBindings();
+
+ if (tfEnable && !_prevTfEnable)
+ {
+ _context.Renderer.Pipeline.BeginTransformFeedback(_drawState.Topology);
+ _prevTfEnable = true;
+ }
+ }
+
+ /// <summary>
+ /// Updates the host state for any modified guest state group with the respective bit set on <paramref name="mask"/>.
+ /// </summary>
+ /// <param name="mask">Mask, where each bit set corresponds to a group index that should be checked and updated</param>
+ public void Update(ulong mask)
+ {
+ _updateTracker.Update(mask);
+ }
+
+ /// <summary>
+ /// Ensures that the bindings are visible to the host GPU.
+ /// Note: this actually performs the binding using the host graphics API.
+ /// </summary>
+ private void CommitBindings()
+ {
+ UpdateStorageBuffers();
+
+ bool unalignedChanged = _currentSpecState.SetHasUnalignedStorageBuffer(_channel.BufferManager.HasUnalignedStorageBuffers);
+
+ if (!_channel.TextureManager.CommitGraphicsBindings(_shaderSpecState) || unalignedChanged)
+ {
+ // Shader must be reloaded. _vtgWritesRtLayer should not change.
+ UpdateShaderState();
+ }
+
+ _channel.BufferManager.CommitGraphicsBindings();
+ }
+
+ /// <summary>
+ /// Updates storage buffer bindings.
+ /// </summary>
+ private void UpdateStorageBuffers()
+ {
+ for (int stage = 0; stage < Constants.ShaderStages; stage++)
+ {
+ ShaderProgramInfo info = _currentProgramInfo[stage];
+
+ if (info == null)
+ {
+ continue;
+ }
+
+ for (int index = 0; index < info.SBuffers.Count; index++)
+ {
+ BufferDescriptor sb = info.SBuffers[index];
+
+ ulong sbDescAddress = _channel.BufferManager.GetGraphicsUniformBufferAddress(stage, 0);
+
+ int sbDescOffset = 0x110 + stage * 0x100 + sb.Slot * 0x10;
+
+ sbDescAddress += (ulong)sbDescOffset;
+
+ SbDescriptor sbDescriptor = _channel.MemoryManager.Physical.Read<SbDescriptor>(sbDescAddress);
+
+ _channel.BufferManager.SetGraphicsStorageBuffer(stage, sb.Slot, sbDescriptor.PackAddress(), (uint)sbDescriptor.Size, sb.Flags);
+ }
+ }
+ }
+
+ /// <summary>
+ /// Updates tessellation state based on the guest GPU state.
+ /// </summary>
+ private void UpdateTessellationState()
+ {
+ _pipeline.PatchControlPoints = (uint)_state.State.PatchVertices;
+
+ _context.Renderer.Pipeline.SetPatchParameters(
+ _state.State.PatchVertices,
+ _state.State.TessOuterLevel.AsSpan(),
+ _state.State.TessInnerLevel.AsSpan());
+
+ _currentSpecState.SetTessellationMode(_state.State.TessMode);
+ }
+
+ /// <summary>
+ /// Updates transform feedback buffer state based on the guest GPU state.
+ /// </summary>
+ private void UpdateTfBufferState()
+ {
+ for (int index = 0; index < Constants.TotalTransformFeedbackBuffers; index++)
+ {
+ TfBufferState tfb = _state.State.TfBufferState[index];
+
+ if (!tfb.Enable)
+ {
+ _channel.BufferManager.SetTransformFeedbackBuffer(index, 0, 0);
+
+ continue;
+ }
+
+ _channel.BufferManager.SetTransformFeedbackBuffer(index, tfb.Address.Pack(), (uint)tfb.Size);
+ }
+ }
+
+ /// <summary>
+ /// Updates Rasterizer primitive discard state based on guest gpu state.
+ /// </summary>
+ private void UpdateRasterizerState()
+ {
+ bool enable = _state.State.RasterizeEnable;
+ _pipeline.RasterizerDiscard = !enable;
+ _context.Renderer.Pipeline.SetRasterizerDiscard(!enable);
+ }
+
+ /// <summary>
+ /// Updates render targets (color and depth-stencil buffers) based on current render target state.
+ /// </summary>
+ private void UpdateRenderTargetState()
+ {
+ UpdateRenderTargetState(RenderTargetUpdateFlags.UpdateAll);
+ }
+
+ /// <summary>
+ /// Updates render targets (color and depth-stencil buffers) based on current render target state.
+ /// </summary>
+ /// <param name="updateFlags">Flags indicating which render targets should be updated and how</param>
+ /// <param name="singleUse">If this is not -1, it indicates that only the given indexed target will be used.</param>
+ public void UpdateRenderTargetState(RenderTargetUpdateFlags updateFlags, int singleUse = -1)
+ {
+ var memoryManager = _channel.MemoryManager;
+ var rtControl = _state.State.RtControl;
+
+ bool useControl = updateFlags.HasFlag(RenderTargetUpdateFlags.UseControl);
+ bool layered = updateFlags.HasFlag(RenderTargetUpdateFlags.Layered);
+ bool singleColor = updateFlags.HasFlag(RenderTargetUpdateFlags.SingleColor);
+
+ int count = useControl ? rtControl.UnpackCount() : Constants.TotalRenderTargets;
+
+ var msaaMode = _state.State.RtMsaaMode;
+
+ int samplesInX = msaaMode.SamplesInX();
+ int samplesInY = msaaMode.SamplesInY();
+
+ var scissor = _state.State.ScreenScissorState;
+ Size sizeHint = new Size((scissor.X + scissor.Width) * samplesInX, (scissor.Y + scissor.Height) * samplesInY, 1);
+
+ int clipRegionWidth = int.MaxValue;
+ int clipRegionHeight = int.MaxValue;
+
+ bool changedScale = false;
+ uint rtNoAlphaMask = 0;
+
+ for (int index = 0; index < Constants.TotalRenderTargets; index++)
+ {
+ int rtIndex = useControl ? rtControl.UnpackPermutationIndex(index) : index;
+
+ var colorState = _state.State.RtColorState[rtIndex];
+
+ if (index >= count || !IsRtEnabled(colorState) || (singleColor && index != singleUse))
+ {
+ changedScale |= _channel.TextureManager.SetRenderTargetColor(index, null);
+
+ continue;
+ }
+
+ if (colorState.Format.NoAlpha())
+ {
+ rtNoAlphaMask |= 1u << index;
+ }
+
+ Image.Texture color = memoryManager.Physical.TextureCache.FindOrCreateTexture(
+ memoryManager,
+ colorState,
+ _vtgWritesRtLayer || layered,
+ samplesInX,
+ samplesInY,
+ sizeHint);
+
+ changedScale |= _channel.TextureManager.SetRenderTargetColor(index, color);
+
+ if (color != null)
+ {
+ if (clipRegionWidth > color.Width / samplesInX)
+ {
+ clipRegionWidth = color.Width / samplesInX;
+ }
+
+ if (clipRegionHeight > color.Height / samplesInY)
+ {
+ clipRegionHeight = color.Height / samplesInY;
+ }
+ }
+ }
+
+ bool dsEnable = _state.State.RtDepthStencilEnable;
+
+ Image.Texture depthStencil = null;
+
+ if (dsEnable && updateFlags.HasFlag(RenderTargetUpdateFlags.UpdateDepthStencil))
+ {
+ var dsState = _state.State.RtDepthStencilState;
+ var dsSize = _state.State.RtDepthStencilSize;
+
+ depthStencil = memoryManager.Physical.TextureCache.FindOrCreateTexture(
+ memoryManager,
+ dsState,
+ dsSize,
+ _vtgWritesRtLayer || layered,
+ samplesInX,
+ samplesInY,
+ sizeHint);
+
+ if (depthStencil != null)
+ {
+ if (clipRegionWidth > depthStencil.Width / samplesInX)
+ {
+ clipRegionWidth = depthStencil.Width / samplesInX;
+ }
+
+ if (clipRegionHeight > depthStencil.Height / samplesInY)
+ {
+ clipRegionHeight = depthStencil.Height / samplesInY;
+ }
+ }
+ }
+
+ changedScale |= _channel.TextureManager.SetRenderTargetDepthStencil(depthStencil);
+
+ if (changedScale)
+ {
+ float oldScale = _channel.TextureManager.RenderTargetScale;
+ _channel.TextureManager.UpdateRenderTargetScale(singleUse);
+
+ if (oldScale != _channel.TextureManager.RenderTargetScale)
+ {
+ _context.Renderer.Pipeline.SetRenderTargetScale(_channel.TextureManager.RenderTargetScale);
+
+ UpdateViewportTransform();
+ UpdateScissorState();
+ }
+ }
+
+ _channel.TextureManager.SetClipRegion(clipRegionWidth, clipRegionHeight);
+
+ if (useControl && _prevRtNoAlphaMask != rtNoAlphaMask)
+ {
+ _prevRtNoAlphaMask = rtNoAlphaMask;
+
+ UpdateBlendState();
+ }
+ }
+
+ /// <summary>
+ /// Updates specialization state based on render target state.
+ /// </summary>
+ public void UpdateRenderTargetSpecialization()
+ {
+ _currentSpecState.SetFragmentOutputTypes(_state.State.RtControl, ref _state.State.RtColorState);
+ }
+
+ /// <summary>
+ /// Checks if a render target color buffer is used.
+ /// </summary>
+ /// <param name="colorState">Color buffer information</param>
+ /// <returns>True if the specified buffer is enabled/used, false otherwise</returns>
+ internal static bool IsRtEnabled(RtColorState colorState)
+ {
+ // Colors are disabled by writing 0 to the format.
+ return colorState.Format != 0 && colorState.WidthOrStride != 0;
+ }
+
+ /// <summary>
+ /// Updates host scissor test state based on current GPU state.
+ /// </summary>
+ public void UpdateScissorState()
+ {
+ const int MinX = 0;
+ const int MinY = 0;
+ const int MaxW = 0xffff;
+ const int MaxH = 0xffff;
+
+ Span<Rectangle<int>> regions = stackalloc Rectangle<int>[Constants.TotalViewports];
+
+ for (int index = 0; index < Constants.TotalViewports; index++)
+ {
+ ScissorState scissor = _state.State.ScissorState[index];
+
+ bool enable = scissor.Enable && (scissor.X1 != MinX ||
+ scissor.Y1 != MinY ||
+ scissor.X2 != MaxW ||
+ scissor.Y2 != MaxH);
+
+ if (enable)
+ {
+ int x = scissor.X1;
+ int y = scissor.Y1;
+ int width = scissor.X2 - x;
+ int height = scissor.Y2 - y;
+
+ if (_state.State.YControl.HasFlag(YControl.NegateY))
+ {
+ ref var screenScissor = ref _state.State.ScreenScissorState;
+ y = screenScissor.Height - height - y;
+
+ if (y < 0)
+ {
+ height += y;
+ y = 0;
+ }
+ }
+
+ float scale = _channel.TextureManager.RenderTargetScale;
+ if (scale != 1f)
+ {
+ x = (int)(x * scale);
+ y = (int)(y * scale);
+ width = (int)MathF.Ceiling(width * scale);
+ height = (int)MathF.Ceiling(height * scale);
+ }
+
+ regions[index] = new Rectangle<int>(x, y, width, height);
+ }
+ else
+ {
+ regions[index] = new Rectangle<int>(MinX, MinY, MaxW, MaxH);
+ }
+ }
+
+ _context.Renderer.Pipeline.SetScissors(regions);
+ }
+
+ /// <summary>
+ /// Updates host depth clamp state based on current GPU state.
+ /// </summary>
+ /// <param name="state">Current GPU state</param>
+ private void UpdateDepthClampState()
+ {
+ ViewVolumeClipControl clip = _state.State.ViewVolumeClipControl;
+ bool clamp = (clip & ViewVolumeClipControl.DepthClampDisabled) == 0;
+
+ _pipeline.DepthClampEnable = clamp;
+ _context.Renderer.Pipeline.SetDepthClamp(clamp);
+ }
+
+ /// <summary>
+ /// Updates host alpha test state based on current GPU state.
+ /// </summary>
+ private void UpdateAlphaTestState()
+ {
+ _context.Renderer.Pipeline.SetAlphaTest(
+ _state.State.AlphaTestEnable,
+ _state.State.AlphaTestRef,
+ _state.State.AlphaTestFunc);
+
+ _currentSpecState.SetAlphaTest(
+ _state.State.AlphaTestEnable,
+ _state.State.AlphaTestRef,
+ _state.State.AlphaTestFunc);
+ }
+
+ /// <summary>
+ /// Updates host depth test state based on current GPU state.
+ /// </summary>
+ private void UpdateDepthTestState()
+ {
+ DepthTestDescriptor descriptor = new DepthTestDescriptor(
+ _state.State.DepthTestEnable,
+ _state.State.DepthWriteEnable,
+ _state.State.DepthTestFunc);
+
+ _pipeline.DepthTest = descriptor;
+ _context.Renderer.Pipeline.SetDepthTest(descriptor);
+ }
+
+ /// <summary>
+ /// Updates host viewport transform and clipping state based on current GPU state.
+ /// </summary>
+ private void UpdateViewportTransform()
+ {
+ var yControl = _state.State.YControl;
+ var face = _state.State.FaceState;
+
+ bool disableTransform = _state.State.ViewportTransformEnable == 0;
+
+ UpdateFrontFace(yControl, face.FrontFace);
+ UpdateDepthMode();
+
+ bool flipY = yControl.HasFlag(YControl.NegateY);
+
+ Span<Viewport> viewports = stackalloc Viewport[Constants.TotalViewports];
+
+ for (int index = 0; index < Constants.TotalViewports; index++)
+ {
+ if (disableTransform)
+ {
+ ref var scissor = ref _state.State.ScreenScissorState;
+
+ float rScale = _channel.TextureManager.RenderTargetScale;
+ var scissorRect = new Rectangle<float>(0, 0, (scissor.X + scissor.Width) * rScale, (scissor.Y + scissor.Height) * rScale);
+
+ viewports[index] = new Viewport(scissorRect, ViewportSwizzle.PositiveX, ViewportSwizzle.PositiveY, ViewportSwizzle.PositiveZ, ViewportSwizzle.PositiveW, 0, 1);
+ continue;
+ }
+
+ ref var transform = ref _state.State.ViewportTransform[index];
+ ref var extents = ref _state.State.ViewportExtents[index];
+
+ float scaleX = MathF.Abs(transform.ScaleX);
+ float scaleY = transform.ScaleY;
+
+ if (flipY)
+ {
+ scaleY = -scaleY;
+ }
+
+ if (!_context.Capabilities.SupportsViewportSwizzle && transform.UnpackSwizzleY() == ViewportSwizzle.NegativeY)
+ {
+ scaleY = -scaleY;
+ }
+
+ float x = transform.TranslateX - scaleX;
+ float y = transform.TranslateY - scaleY;
+
+ float width = scaleX * 2;
+ float height = scaleY * 2;
+
+ float scale = _channel.TextureManager.RenderTargetScale;
+ if (scale != 1f)
+ {
+ x *= scale;
+ y *= scale;
+ width *= scale;
+ height *= scale;
+ }
+
+ Rectangle<float> region = new Rectangle<float>(x, y, width, height);
+
+ ViewportSwizzle swizzleX = transform.UnpackSwizzleX();
+ ViewportSwizzle swizzleY = transform.UnpackSwizzleY();
+ ViewportSwizzle swizzleZ = transform.UnpackSwizzleZ();
+ ViewportSwizzle swizzleW = transform.UnpackSwizzleW();
+
+ float depthNear = extents.DepthNear;
+ float depthFar = extents.DepthFar;
+
+ if (transform.ScaleZ < 0)
+ {
+ float temp = depthNear;
+ depthNear = depthFar;
+ depthFar = temp;
+ }
+
+ viewports[index] = new Viewport(region, swizzleX, swizzleY, swizzleZ, swizzleW, depthNear, depthFar);
+ }
+
+ _context.Renderer.Pipeline.SetDepthMode(GetDepthMode());
+ _context.Renderer.Pipeline.SetViewports(viewports, disableTransform);
+
+ _currentSpecState.SetViewportTransformDisable(_state.State.ViewportTransformEnable == 0);
+ _currentSpecState.SetDepthMode(GetDepthMode() == DepthMode.MinusOneToOne);
+ }
+
+ /// <summary>
+ /// Updates the depth mode (0 to 1 or -1 to 1) based on the current viewport and depth mode register state.
+ /// </summary>
+ private void UpdateDepthMode()
+ {
+ _context.Renderer.Pipeline.SetDepthMode(GetDepthMode());
+ }
+
+ /// <summary>
+ /// Updates polygon mode state based on current GPU state.
+ /// </summary>
+ private void UpdatePolygonMode()
+ {
+ _context.Renderer.Pipeline.SetPolygonMode(_state.State.PolygonModeFront, _state.State.PolygonModeBack);
+ }
+
+ /// <summary>
+ /// Updates host depth bias (also called polygon offset) state based on current GPU state.
+ /// </summary>
+ private void UpdateDepthBiasState()
+ {
+ var depthBias = _state.State.DepthBiasState;
+
+ float factor = _state.State.DepthBiasFactor;
+ float units = _state.State.DepthBiasUnits;
+ float clamp = _state.State.DepthBiasClamp;
+
+ PolygonModeMask enables;
+
+ enables = (depthBias.PointEnable ? PolygonModeMask.Point : 0);
+ enables |= (depthBias.LineEnable ? PolygonModeMask.Line : 0);
+ enables |= (depthBias.FillEnable ? PolygonModeMask.Fill : 0);
+
+ _pipeline.BiasEnable = enables;
+ _context.Renderer.Pipeline.SetDepthBias(enables, factor, units / 2f, clamp);
+ }
+
+ /// <summary>
+ /// Updates host stencil test state based on current GPU state.
+ /// </summary>
+ private void UpdateStencilTestState()
+ {
+ var backMasks = _state.State.StencilBackMasks;
+ var test = _state.State.StencilTestState;
+ var backTest = _state.State.StencilBackTestState;
+
+ CompareOp backFunc;
+ StencilOp backSFail;
+ StencilOp backDpPass;
+ StencilOp backDpFail;
+ int backFuncRef;
+ int backFuncMask;
+ int backMask;
+
+ if (backTest.TwoSided)
+ {
+ backFunc = backTest.BackFunc;
+ backSFail = backTest.BackSFail;
+ backDpPass = backTest.BackDpPass;
+ backDpFail = backTest.BackDpFail;
+ backFuncRef = backMasks.FuncRef;
+ backFuncMask = backMasks.FuncMask;
+ backMask = backMasks.Mask;
+ }
+ else
+ {
+ backFunc = test.FrontFunc;
+ backSFail = test.FrontSFail;
+ backDpPass = test.FrontDpPass;
+ backDpFail = test.FrontDpFail;
+ backFuncRef = test.FrontFuncRef;
+ backFuncMask = test.FrontFuncMask;
+ backMask = test.FrontMask;
+ }
+
+ StencilTestDescriptor descriptor = new StencilTestDescriptor(
+ test.Enable,
+ test.FrontFunc,
+ test.FrontSFail,
+ test.FrontDpPass,
+ test.FrontDpFail,
+ test.FrontFuncRef,
+ test.FrontFuncMask,
+ test.FrontMask,
+ backFunc,
+ backSFail,
+ backDpPass,
+ backDpFail,
+ backFuncRef,
+ backFuncMask,
+ backMask);
+
+ _pipeline.StencilTest = descriptor;
+ _context.Renderer.Pipeline.SetStencilTest(descriptor);
+ }
+
+ /// <summary>
+ /// Updates user-defined clipping based on the guest GPU state.
+ /// </summary>
+ private void UpdateUserClipState()
+ {
+ uint clipMask = _state.State.ClipDistanceEnable & _vsClipDistancesWritten;
+
+ for (int i = 0; i < Constants.TotalClipDistances; ++i)
+ {
+ _context.Renderer.Pipeline.SetUserClipDistance(i, (clipMask & (1 << i)) != 0);
+ }
+ }
+
+ /// <summary>
+ /// Updates current sampler pool address and size based on guest GPU state.
+ /// </summary>
+ private void UpdateSamplerPoolState()
+ {
+ var texturePool = _state.State.TexturePoolState;
+ var samplerPool = _state.State.SamplerPoolState;
+
+ var samplerIndex = _state.State.SamplerIndex;
+
+ int maximumId = samplerIndex == SamplerIndex.ViaHeaderIndex
+ ? texturePool.MaximumId
+ : samplerPool.MaximumId;
+
+ _channel.TextureManager.SetGraphicsSamplerPool(samplerPool.Address.Pack(), maximumId, samplerIndex);
+ }
+
+ /// <summary>
+ /// Updates current texture pool address and size based on guest GPU state.
+ /// </summary>
+ private void UpdateTexturePoolState()
+ {
+ var texturePool = _state.State.TexturePoolState;
+
+ _channel.TextureManager.SetGraphicsTexturePool(texturePool.Address.Pack(), texturePool.MaximumId);
+ _channel.TextureManager.SetGraphicsTextureBufferIndex((int)_state.State.TextureBufferIndex);
+
+ _currentSpecState.SetPoolState(GetPoolState());
+ }
+
+ /// <summary>
+ /// Updates host vertex attributes based on guest GPU state.
+ /// </summary>
+ private void UpdateVertexAttribState()
+ {
+ uint vbEnableMask = _vbEnableMask;
+
+ Span<VertexAttribDescriptor> vertexAttribs = stackalloc VertexAttribDescriptor[Constants.TotalVertexAttribs];
+
+ for (int index = 0; index < Constants.TotalVertexAttribs; index++)
+ {
+ var vertexAttrib = _state.State.VertexAttribState[index];
+
+ int bufferIndex = vertexAttrib.UnpackBufferIndex();
+
+ if ((vbEnableMask & (1u << bufferIndex)) == 0)
+ {
+ // Using a vertex buffer that doesn't exist is invalid, so let's use a dummy attribute for those cases.
+ vertexAttribs[index] = new VertexAttribDescriptor(0, 0, true, Format.R32G32B32A32Float);
+ continue;
+ }
+
+ if (!FormatTable.TryGetAttribFormat(vertexAttrib.UnpackFormat(), out Format format))
+ {
+ Logger.Debug?.Print(LogClass.Gpu, $"Invalid attribute format 0x{vertexAttrib.UnpackFormat():X}.");
+
+ format = vertexAttrib.UnpackType() switch
+ {
+ VertexAttribType.Sint => Format.R32G32B32A32Sint,
+ VertexAttribType.Uint => Format.R32G32B32A32Uint,
+ _ => Format.R32G32B32A32Float
+ };
+ }
+
+ vertexAttribs[index] = new VertexAttribDescriptor(
+ bufferIndex,
+ vertexAttrib.UnpackOffset(),
+ vertexAttrib.UnpackIsConstant(),
+ format);
+ }
+
+ _pipeline.SetVertexAttribs(vertexAttribs);
+ _context.Renderer.Pipeline.SetVertexAttribs(vertexAttribs);
+ _currentSpecState.SetAttributeTypes(ref _state.State.VertexAttribState);
+ }
+
+ /// <summary>
+ /// Updates host line width based on guest GPU state.
+ /// </summary>
+ private void UpdateLineState()
+ {
+ float width = _state.State.LineWidthSmooth;
+ bool smooth = _state.State.LineSmoothEnable;
+
+ _pipeline.LineWidth = width;
+ _context.Renderer.Pipeline.SetLineParameters(width, smooth);
+ }
+
+ /// <summary>
+ /// Updates host point size based on guest GPU state.
+ /// </summary>
+ private void UpdatePointState()
+ {
+ float size = _state.State.PointSize;
+ bool isProgramPointSize = _state.State.VertexProgramPointSize;
+ bool enablePointSprite = _state.State.PointSpriteEnable;
+
+ // TODO: Need to figure out a way to map PointCoordReplace enable bit.
+ Origin origin = (_state.State.PointCoordReplace & 4) == 0 ? Origin.LowerLeft : Origin.UpperLeft;
+
+ _context.Renderer.Pipeline.SetPointParameters(size, isProgramPointSize, enablePointSprite, origin);
+
+ _currentSpecState.SetProgramPointSizeEnable(isProgramPointSize);
+ _currentSpecState.SetPointSize(size);
+ }
+
+ /// <summary>
+ /// Updates host primitive restart based on guest GPU state.
+ /// </summary>
+ private void UpdatePrimitiveRestartState()
+ {
+ PrimitiveRestartState primitiveRestart = _state.State.PrimitiveRestartState;
+ bool enable = primitiveRestart.Enable && (_drawState.DrawIndexed || _state.State.PrimitiveRestartDrawArrays);
+
+ _pipeline.PrimitiveRestartEnable = enable;
+ _context.Renderer.Pipeline.SetPrimitiveRestart(enable, primitiveRestart.Index);
+ }
+
+ /// <summary>
+ /// Updates host index buffer binding based on guest GPU state.
+ /// </summary>
+ private void UpdateIndexBufferState()
+ {
+ var indexBuffer = _state.State.IndexBufferState;
+
+ if (_drawState.IndexCount == 0)
+ {
+ return;
+ }
+
+ ulong gpuVa = indexBuffer.Address.Pack();
+
+ // Do not use the end address to calculate the size, because
+ // the result may be much larger than the real size of the index buffer.
+ ulong size = (ulong)(_drawState.FirstIndex + _drawState.IndexCount);
+
+ switch (indexBuffer.Type)
+ {
+ case IndexType.UShort: size *= 2; break;
+ case IndexType.UInt: size *= 4; break;
+ }
+
+ _channel.BufferManager.SetIndexBuffer(gpuVa, size, indexBuffer.Type);
+ }
+
+ /// <summary>
+ /// Updates host vertex buffer bindings based on guest GPU state.
+ /// </summary>
+ private void UpdateVertexBufferState()
+ {
+ IndexType indexType = _state.State.IndexBufferState.Type;
+ bool indexTypeSmall = indexType == IndexType.UByte || indexType == IndexType.UShort;
+
+ _drawState.IsAnyVbInstanced = false;
+
+ bool drawIndexed = _drawState.DrawIndexed;
+ bool drawIndirect = _drawState.DrawIndirect;
+ int drawFirstVertex = _drawState.DrawFirstVertex;
+ int drawVertexCount = _drawState.DrawVertexCount;
+ uint vbEnableMask = 0;
+
+ for (int index = 0; index < Constants.TotalVertexBuffers; index++)
+ {
+ var vertexBuffer = _state.State.VertexBufferState[index];
+
+ if (!vertexBuffer.UnpackEnable())
+ {
+ _pipeline.VertexBuffers[index] = new BufferPipelineDescriptor(false, 0, 0);
+ _channel.BufferManager.SetVertexBuffer(index, 0, 0, 0, 0);
+
+ continue;
+ }
+
+ GpuVa endAddress = _state.State.VertexBufferEndAddress[index];
+
+ ulong address = vertexBuffer.Address.Pack();
+
+ if (_channel.MemoryManager.IsMapped(address))
+ {
+ vbEnableMask |= 1u << index;
+ }
+
+ int stride = vertexBuffer.UnpackStride();
+
+ bool instanced = _state.State.VertexBufferInstanced[index];
+
+ int divisor = instanced ? vertexBuffer.Divisor : 0;
+
+ _drawState.IsAnyVbInstanced |= divisor != 0;
+
+ ulong vbSize = endAddress.Pack() - address + 1;
+ ulong size;
+
+ if (_drawState.IbStreamer.HasInlineIndexData || drawIndexed || stride == 0 || instanced)
+ {
+ // This size may be (much) larger than the real vertex buffer size.
+ // Avoid calculating it this way, unless we don't have any other option.
+
+ size = vbSize;
+
+ if (stride > 0 && indexTypeSmall && drawIndexed && !drawIndirect && !instanced)
+ {
+ // If the index type is a small integer type, then we might be still able
+ // to reduce the vertex buffer size based on the maximum possible index value.
+
+ ulong maxVertexBufferSize = indexType == IndexType.UByte ? 0x100UL : 0x10000UL;
+
+ maxVertexBufferSize += _state.State.FirstVertex;
+ maxVertexBufferSize *= (uint)stride;
+
+ size = Math.Min(size, maxVertexBufferSize);
+ }
+ }
+ else
+ {
+ // For non-indexed draws, we can guess the size from the vertex count
+ // and stride.
+
+ int firstInstance = (int)_state.State.FirstInstance;
+
+ size = Math.Min(vbSize, (ulong)((firstInstance + drawFirstVertex + drawVertexCount) * stride));
+ }
+
+ _pipeline.VertexBuffers[index] = new BufferPipelineDescriptor(_channel.MemoryManager.IsMapped(address), stride, divisor);
+ _channel.BufferManager.SetVertexBuffer(index, address, size, stride, divisor);
+ }
+
+ if (_vbEnableMask != vbEnableMask)
+ {
+ _vbEnableMask = vbEnableMask;
+ UpdateVertexAttribState();
+ }
+ }
+
+ /// <summary>
+ /// Updates host face culling and orientation based on guest GPU state.
+ /// </summary>
+ private void UpdateFaceState()
+ {
+ var yControl = _state.State.YControl;
+ var face = _state.State.FaceState;
+
+ _pipeline.CullEnable = face.CullEnable;
+ _pipeline.CullMode = face.CullFace;
+ _context.Renderer.Pipeline.SetFaceCulling(face.CullEnable, face.CullFace);
+
+ UpdateFrontFace(yControl, face.FrontFace);
+ }
+
+ /// <summary>
+ /// Updates the front face based on the current front face and the origin.
+ /// </summary>
+ /// <param name="yControl">Y control register value, where the origin is located</param>
+ /// <param name="frontFace">Front face</param>
+ private void UpdateFrontFace(YControl yControl, FrontFace frontFace)
+ {
+ bool isUpperLeftOrigin = !yControl.HasFlag(YControl.TriangleRastFlip);
+
+ if (isUpperLeftOrigin)
+ {
+ frontFace = frontFace == FrontFace.CounterClockwise ? FrontFace.Clockwise : FrontFace.CounterClockwise;
+ }
+
+ _pipeline.FrontFace = frontFace;
+ _context.Renderer.Pipeline.SetFrontFace(frontFace);
+ }
+
+ /// <summary>
+ /// Updates host render target color masks, based on guest GPU state.
+ /// This defines which color channels are written to each color buffer.
+ /// </summary>
+ private void UpdateRtColorMask()
+ {
+ bool rtColorMaskShared = _state.State.RtColorMaskShared;
+
+ Span<uint> componentMasks = stackalloc uint[Constants.TotalRenderTargets];
+
+ for (int index = 0; index < Constants.TotalRenderTargets; index++)
+ {
+ var colorMask = _state.State.RtColorMask[rtColorMaskShared ? 0 : index];
+
+ uint componentMask;
+
+ componentMask = (colorMask.UnpackRed() ? 1u : 0u);
+ componentMask |= (colorMask.UnpackGreen() ? 2u : 0u);
+ componentMask |= (colorMask.UnpackBlue() ? 4u : 0u);
+ componentMask |= (colorMask.UnpackAlpha() ? 8u : 0u);
+
+ componentMasks[index] = componentMask;
+ _pipeline.ColorWriteMask[index] = componentMask;
+ }
+
+ _context.Renderer.Pipeline.SetRenderTargetColorMasks(componentMasks);
+ }
+
+ /// <summary>
+ /// Updates host render target color buffer blending state, based on guest state.
+ /// </summary>
+ private void UpdateBlendState()
+ {
+ if (_state.State.BlendUcodeEnable != BlendUcodeEnable.Disabled)
+ {
+ if (_context.Capabilities.SupportsBlendEquationAdvanced && _blendManager.TryGetAdvancedBlend(out var blendDescriptor))
+ {
+ // Try to HLE it using advanced blend on the host if we can.
+ _context.Renderer.Pipeline.SetBlendState(blendDescriptor);
+ return;
+ }
+ else
+ {
+ // TODO: Blend emulation fallback.
+ }
+ }
+
+ bool blendIndependent = _state.State.BlendIndependent;
+ ColorF blendConstant = _state.State.BlendConstant;
+
+ bool dualSourceBlendEnabled = false;
+
+ if (blendIndependent)
+ {
+ for (int index = 0; index < Constants.TotalRenderTargets; index++)
+ {
+ bool enable = _state.State.BlendEnable[index];
+ var blend = _state.State.BlendState[index];
+
+ var descriptor = new BlendDescriptor(
+ enable,
+ blendConstant,
+ blend.ColorOp,
+ FilterBlendFactor(blend.ColorSrcFactor, index),
+ FilterBlendFactor(blend.ColorDstFactor, index),
+ blend.AlphaOp,
+ FilterBlendFactor(blend.AlphaSrcFactor, index),
+ FilterBlendFactor(blend.AlphaDstFactor, index));
+
+ if (enable &&
+ (blend.ColorSrcFactor.IsDualSource() ||
+ blend.ColorDstFactor.IsDualSource() ||
+ blend.AlphaSrcFactor.IsDualSource() ||
+ blend.AlphaDstFactor.IsDualSource()))
+ {
+ dualSourceBlendEnabled = true;
+ }
+
+ _pipeline.BlendDescriptors[index] = descriptor;
+ _context.Renderer.Pipeline.SetBlendState(index, descriptor);
+ }
+ }
+ else
+ {
+ bool enable = _state.State.BlendEnable[0];
+ var blend = _state.State.BlendStateCommon;
+
+ var descriptor = new BlendDescriptor(
+ enable,
+ blendConstant,
+ blend.ColorOp,
+ FilterBlendFactor(blend.ColorSrcFactor, 0),
+ FilterBlendFactor(blend.ColorDstFactor, 0),
+ blend.AlphaOp,
+ FilterBlendFactor(blend.AlphaSrcFactor, 0),
+ FilterBlendFactor(blend.AlphaDstFactor, 0));
+
+ if (enable &&
+ (blend.ColorSrcFactor.IsDualSource() ||
+ blend.ColorDstFactor.IsDualSource() ||
+ blend.AlphaSrcFactor.IsDualSource() ||
+ blend.AlphaDstFactor.IsDualSource()))
+ {
+ dualSourceBlendEnabled = true;
+ }
+
+ for (int index = 0; index < Constants.TotalRenderTargets; index++)
+ {
+ _pipeline.BlendDescriptors[index] = descriptor;
+ _context.Renderer.Pipeline.SetBlendState(index, descriptor);
+ }
+ }
+
+ _currentSpecState.SetDualSourceBlendEnabled(dualSourceBlendEnabled);
+ }
+
+ /// <summary>
+ /// Gets a blend factor for the color target currently.
+ /// This will return <paramref name="factor"/> unless the target format has no alpha component,
+ /// in which case it will replace destination alpha factor with a constant factor of one or zero.
+ /// </summary>
+ /// <param name="factor">Input factor</param>
+ /// <param name="index">Color target index</param>
+ /// <returns>New blend factor</returns>
+ private BlendFactor FilterBlendFactor(BlendFactor factor, int index)
+ {
+ // If any color target format without alpha is being used, we need to make sure that
+ // if blend is active, it will not use destination alpha as a factor.
+ // That is required because RGBX formats are emulated using host RGBA formats.
+
+ if (_state.State.RtColorState[index].Format.NoAlpha())
+ {
+ switch (factor)
+ {
+ case BlendFactor.DstAlpha:
+ case BlendFactor.DstAlphaGl:
+ factor = BlendFactor.One;
+ break;
+ case BlendFactor.OneMinusDstAlpha:
+ case BlendFactor.OneMinusDstAlphaGl:
+ factor = BlendFactor.Zero;
+ break;
+ }
+ }
+
+ return factor;
+ }
+
+ /// <summary>
+ /// Updates host logical operation state, based on guest state.
+ /// </summary>
+ private void UpdateLogicOpState()
+ {
+ LogicalOpState logicOpState = _state.State.LogicOpState;
+
+ _pipeline.SetLogicOpState(logicOpState.Enable, logicOpState.LogicalOp);
+ _context.Renderer.Pipeline.SetLogicOpState(logicOpState.Enable, logicOpState.LogicalOp);
+ }
+
+ /// <summary>
+ /// Updates multisample state, based on guest state.
+ /// </summary>
+ private void UpdateMultisampleState()
+ {
+ bool alphaToCoverageEnable = (_state.State.MultisampleControl & 1) != 0;
+ bool alphaToOneEnable = (_state.State.MultisampleControl & 0x10) != 0;
+
+ _context.Renderer.Pipeline.SetMultisampleState(new MultisampleDescriptor(
+ alphaToCoverageEnable,
+ _state.State.AlphaToCoverageDitherEnable,
+ alphaToOneEnable));
+
+ _currentSpecState.SetAlphaToCoverageEnable(alphaToCoverageEnable, _state.State.AlphaToCoverageDitherEnable);
+ }
+
+ /// <summary>
+ /// Updates the early z flag, based on guest state.
+ /// </summary>
+ private void UpdateEarlyZState()
+ {
+ _currentSpecState.SetEarlyZForce(_state.State.EarlyZForce);
+ }
+
+ /// <summary>
+ /// Updates host shaders based on the guest GPU state.
+ /// </summary>
+ private void UpdateShaderState()
+ {
+ var shaderCache = _channel.MemoryManager.Physical.ShaderCache;
+
+ _vtgWritesRtLayer = false;
+
+ ShaderAddresses addresses = new ShaderAddresses();
+ Span<ulong> addressesSpan = addresses.AsSpan();
+
+ ulong baseAddress = _state.State.ShaderBaseAddress.Pack();
+
+ for (int index = 0; index < 6; index++)
+ {
+ var shader = _state.State.ShaderState[index];
+ if (!shader.UnpackEnable() && index != 1)
+ {
+ continue;
+ }
+
+ addressesSpan[index] = baseAddress + shader.Offset;
+ }
+
+ CachedShaderProgram gs = shaderCache.GetGraphicsShader(ref _state.State, ref _pipeline, _channel, ref _currentSpecState.GetPoolState(), ref _currentSpecState.GetGraphicsState(), addresses);
+
+ // Consume the modified flag for spec state so that it isn't checked again.
+ _currentSpecState.SetShader(gs);
+
+ _shaderSpecState = gs.SpecializationState;
+
+ byte oldVsClipDistancesWritten = _vsClipDistancesWritten;
+
+ _drawState.VsUsesInstanceId = gs.Shaders[1]?.Info.UsesInstanceId ?? false;
+ _vsUsesDrawParameters = gs.Shaders[1]?.Info.UsesDrawParameters ?? false;
+ _vsClipDistancesWritten = gs.Shaders[1]?.Info.ClipDistancesWritten ?? 0;
+
+ if (oldVsClipDistancesWritten != _vsClipDistancesWritten)
+ {
+ UpdateUserClipState();
+ }
+
+ UpdateShaderBindings(gs.Bindings);
+
+ for (int stageIndex = 0; stageIndex < Constants.ShaderStages; stageIndex++)
+ {
+ ShaderProgramInfo info = gs.Shaders[stageIndex + 1]?.Info;
+
+ if (info?.UsesRtLayer == true)
+ {
+ _vtgWritesRtLayer = true;
+ }
+
+ _currentProgramInfo[stageIndex] = info;
+ }
+
+ _context.Renderer.Pipeline.SetProgram(gs.HostProgram);
+ }
+
+ /// <summary>
+ /// Updates bindings consumed by the shader on the texture and buffer managers.
+ /// </summary>
+ /// <param name="bindings">Bindings for the active shader</param>
+ private void UpdateShaderBindings(CachedShaderBindings bindings)
+ {
+ _channel.TextureManager.SetGraphicsBindings(bindings);
+ _channel.BufferManager.SetGraphicsBufferBindings(bindings);
+ }
+
+ /// <summary>
+ /// Gets the current texture pool state.
+ /// </summary>
+ /// <returns>Texture pool state</returns>
+ private GpuChannelPoolState GetPoolState()
+ {
+ return new GpuChannelPoolState(
+ _state.State.TexturePoolState.Address.Pack(),
+ _state.State.TexturePoolState.MaximumId,
+ (int)_state.State.TextureBufferIndex);
+ }
+
+ /// <summary>
+ /// Gets the depth mode that is currently being used (zero to one or minus one to one).
+ /// </summary>
+ /// <returns>Current depth mode</returns>
+ private DepthMode GetDepthMode()
+ {
+ ref var transform = ref _state.State.ViewportTransform[0];
+ ref var extents = ref _state.State.ViewportExtents[0];
+
+ DepthMode depthMode;
+
+ if (!float.IsInfinity(extents.DepthNear) &&
+ !float.IsInfinity(extents.DepthFar) &&
+ (extents.DepthFar - extents.DepthNear) != 0)
+ {
+ // Try to guess the depth mode being used on the high level API
+ // based on current transform.
+ // It is setup like so by said APIs:
+ // If depth mode is ZeroToOne:
+ // TranslateZ = Near
+ // ScaleZ = Far - Near
+ // If depth mode is MinusOneToOne:
+ // TranslateZ = (Near + Far) / 2
+ // ScaleZ = (Far - Near) / 2
+ // DepthNear/Far are sorted such as that Near is always less than Far.
+ depthMode = extents.DepthNear != transform.TranslateZ &&
+ extents.DepthFar != transform.TranslateZ
+ ? DepthMode.MinusOneToOne
+ : DepthMode.ZeroToOne;
+ }
+ else
+ {
+ // If we can't guess from the viewport transform, then just use the depth mode register.
+ depthMode = (DepthMode)(_state.State.DepthMode & 1);
+ }
+
+ return depthMode;
+ }
+
+ /// <summary>
+ /// Forces the shaders to be rebound on the next draw.
+ /// </summary>
+ public void ForceShaderUpdate()
+ {
+ _updateTracker.ForceDirty(ShaderStateIndex);
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs
new file mode 100644
index 00000000..caeee18e
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs
@@ -0,0 +1,620 @@
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.GPFifo;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+using Ryujinx.Graphics.Gpu.Engine.Threed.Blender;
+using System;
+using System.Collections.Generic;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Represents a 3D engine class.
+ /// </summary>
+ class ThreedClass : IDeviceState
+ {
+ private readonly GpuContext _context;
+ private readonly GPFifoClass _fifoClass;
+ private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+ private readonly InlineToMemoryClass _i2mClass;
+ private readonly AdvancedBlendManager _blendManager;
+ private readonly DrawManager _drawManager;
+ private readonly SemaphoreUpdater _semaphoreUpdater;
+ private readonly ConstantBufferUpdater _cbUpdater;
+ private readonly StateUpdater _stateUpdater;
+
+ /// <summary>
+ /// Creates a new instance of the 3D engine class.
+ /// </summary>
+ /// <param name="context">GPU context</param>
+ /// <param name="channel">GPU channel</param>
+ public ThreedClass(GpuContext context, GpuChannel channel, GPFifoClass fifoClass)
+ {
+ _context = context;
+ _fifoClass = fifoClass;
+ _state = new DeviceStateWithShadow<ThreedClassState>(new Dictionary<string, RwCallback>
+ {
+ { nameof(ThreedClassState.LaunchDma), new RwCallback(LaunchDma, null) },
+ { nameof(ThreedClassState.LoadInlineData), new RwCallback(LoadInlineData, null) },
+ { nameof(ThreedClassState.SyncpointAction), new RwCallback(IncrementSyncpoint, null) },
+ { nameof(ThreedClassState.InvalidateSamplerCacheNoWfi), new RwCallback(InvalidateSamplerCacheNoWfi, null) },
+ { nameof(ThreedClassState.InvalidateTextureHeaderCacheNoWfi), new RwCallback(InvalidateTextureHeaderCacheNoWfi, null) },
+ { nameof(ThreedClassState.TextureBarrier), new RwCallback(TextureBarrier, null) },
+ { nameof(ThreedClassState.LoadBlendUcodeStart), new RwCallback(LoadBlendUcodeStart, null) },
+ { nameof(ThreedClassState.LoadBlendUcodeInstruction), new RwCallback(LoadBlendUcodeInstruction, null) },
+ { nameof(ThreedClassState.TextureBarrierTiled), new RwCallback(TextureBarrierTiled, null) },
+ { nameof(ThreedClassState.DrawTextureSrcY), new RwCallback(DrawTexture, null) },
+ { nameof(ThreedClassState.DrawVertexArrayBeginEndInstanceFirst), new RwCallback(DrawVertexArrayBeginEndInstanceFirst, null) },
+ { nameof(ThreedClassState.DrawVertexArrayBeginEndInstanceSubsequent), new RwCallback(DrawVertexArrayBeginEndInstanceSubsequent, null) },
+ { nameof(ThreedClassState.VbElementU8), new RwCallback(VbElementU8, null) },
+ { nameof(ThreedClassState.VbElementU16), new RwCallback(VbElementU16, null) },
+ { nameof(ThreedClassState.VbElementU32), new RwCallback(VbElementU32, null) },
+ { nameof(ThreedClassState.ResetCounter), new RwCallback(ResetCounter, null) },
+ { nameof(ThreedClassState.RenderEnableCondition), new RwCallback(null, Zero) },
+ { nameof(ThreedClassState.DrawEnd), new RwCallback(DrawEnd, null) },
+ { nameof(ThreedClassState.DrawBegin), new RwCallback(DrawBegin, null) },
+ { nameof(ThreedClassState.DrawIndexBuffer32BeginEndInstanceFirst), new RwCallback(DrawIndexBuffer32BeginEndInstanceFirst, null) },
+ { nameof(ThreedClassState.DrawIndexBuffer16BeginEndInstanceFirst), new RwCallback(DrawIndexBuffer16BeginEndInstanceFirst, null) },
+ { nameof(ThreedClassState.DrawIndexBuffer8BeginEndInstanceFirst), new RwCallback(DrawIndexBuffer8BeginEndInstanceFirst, null) },
+ { nameof(ThreedClassState.DrawIndexBuffer32BeginEndInstanceSubsequent), new RwCallback(DrawIndexBuffer32BeginEndInstanceSubsequent, null) },
+ { nameof(ThreedClassState.DrawIndexBuffer16BeginEndInstanceSubsequent), new RwCallback(DrawIndexBuffer16BeginEndInstanceSubsequent, null) },
+ { nameof(ThreedClassState.DrawIndexBuffer8BeginEndInstanceSubsequent), new RwCallback(DrawIndexBuffer8BeginEndInstanceSubsequent, null) },
+ { nameof(ThreedClassState.IndexBufferCount), new RwCallback(SetIndexBufferCount, null) },
+ { nameof(ThreedClassState.Clear), new RwCallback(Clear, null) },
+ { nameof(ThreedClassState.SemaphoreControl), new RwCallback(Report, null) },
+ { nameof(ThreedClassState.SetFalcon04), new RwCallback(SetFalcon04, null) },
+ { nameof(ThreedClassState.UniformBufferUpdateData), new RwCallback(ConstantBufferUpdate, null) },
+ { nameof(ThreedClassState.UniformBufferBindVertex), new RwCallback(ConstantBufferBindVertex, null) },
+ { nameof(ThreedClassState.UniformBufferBindTessControl), new RwCallback(ConstantBufferBindTessControl, null) },
+ { nameof(ThreedClassState.UniformBufferBindTessEvaluation), new RwCallback(ConstantBufferBindTessEvaluation, null) },
+ { nameof(ThreedClassState.UniformBufferBindGeometry), new RwCallback(ConstantBufferBindGeometry, null) },
+ { nameof(ThreedClassState.UniformBufferBindFragment), new RwCallback(ConstantBufferBindFragment, null) }
+ });
+
+ _i2mClass = new InlineToMemoryClass(context, channel, initializeState: false);
+
+ var spec = new SpecializationStateUpdater(context);
+ var drawState = new DrawState();
+
+ _drawManager = new DrawManager(context, channel, _state, drawState, spec);
+ _blendManager = new AdvancedBlendManager(_state);
+ _semaphoreUpdater = new SemaphoreUpdater(context, channel, _state);
+ _cbUpdater = new ConstantBufferUpdater(channel, _state);
+ _stateUpdater = new StateUpdater(context, channel, _state, drawState, _blendManager, spec);
+
+ // This defaults to "always", even without any register write.
+ // Reads just return 0, regardless of what was set there.
+ _state.State.RenderEnableCondition = Condition.Always;
+ }
+
+ /// <summary>
+ /// Reads data from the class registers.
+ /// </summary>
+ /// <param name="offset">Register byte offset</param>
+ /// <returns>Data at the specified offset</returns>
+ [MethodImpl(MethodImplOptions.AggressiveInlining)]
+ public int Read(int offset) => _state.Read(offset);
+
+ /// <summary>
+ /// Writes data to the class registers.
+ /// </summary>
+ /// <param name="offset">Register byte offset</param>
+ /// <param name="data">Data to be written</param>
+ [MethodImpl(MethodImplOptions.AggressiveInlining)]
+ public void Write(int offset, int data)
+ {
+ _state.WriteWithRedundancyCheck(offset, data, out bool valueChanged);
+
+ if (valueChanged)
+ {
+ _stateUpdater.SetDirty(offset);
+ }
+ }
+
+ /// <summary>
+ /// Sets the shadow ram control value of all sub-channels.
+ /// </summary>
+ /// <param name="control">New shadow ram control value</param>
+ public void SetShadowRamControl(int control)
+ {
+ _state.State.SetMmeShadowRamControl = (uint)control;
+ }
+
+ /// <summary>
+ /// Updates current host state for all registers modified since the last call to this method.
+ /// </summary>
+ public void UpdateState()
+ {
+ _fifoClass.CreatePendingSyncs();
+ _cbUpdater.FlushUboDirty();
+ _stateUpdater.Update();
+ }
+
+ /// <summary>
+ /// Updates current host state for all registers modified since the last call to this method.
+ /// </summary>
+ /// <param name="mask">Mask where each bit set indicates that the respective state group index should be checked</param>
+ public void UpdateState(ulong mask)
+ {
+ _stateUpdater.Update(mask);
+ }
+
+ /// <summary>
+ /// Updates render targets (color and depth-stencil buffers) based on current render target state.
+ /// </summary>
+ /// <param name="updateFlags">Flags indicating which render targets should be updated and how</param>
+ /// <param name="singleUse">If this is not -1, it indicates that only the given indexed target will be used.</param>
+ public void UpdateRenderTargetState(RenderTargetUpdateFlags updateFlags, int singleUse = -1)
+ {
+ _stateUpdater.UpdateRenderTargetState(updateFlags, singleUse);
+ }
+
+ /// <summary>
+ /// Updates scissor based on current render target state.
+ /// </summary>
+ public void UpdateScissorState()
+ {
+ _stateUpdater.UpdateScissorState();
+ }
+
+ /// <summary>
+ /// Marks the entire state as dirty, forcing a full host state update before the next draw.
+ /// </summary>
+ public void ForceStateDirty()
+ {
+ _drawManager.ForceStateDirty();
+ _stateUpdater.SetAllDirty();
+ }
+
+ /// <summary>
+ /// Marks the specified register offset as dirty, forcing the associated state to update on the next draw.
+ /// </summary>
+ /// <param name="offset">Register offset</param>
+ public void ForceStateDirty(int offset)
+ {
+ _stateUpdater.SetDirty(offset);
+ }
+
+ /// <summary>
+ /// Forces the shaders to be rebound on the next draw.
+ /// </summary>
+ public void ForceShaderUpdate()
+ {
+ _stateUpdater.ForceShaderUpdate();
+ }
+
+ /// <summary>
+ /// Create any syncs from WaitForIdle command that are currently pending.
+ /// </summary>
+ public void CreatePendingSyncs()
+ {
+ _fifoClass.CreatePendingSyncs();
+ }
+
+ /// <summary>
+ /// Flushes any queued UBO updates.
+ /// </summary>
+ public void FlushUboDirty()
+ {
+ _cbUpdater.FlushUboDirty();
+ }
+
+ /// <summary>
+ /// Perform any deferred draws.
+ /// </summary>
+ public void PerformDeferredDraws()
+ {
+ _drawManager.PerformDeferredDraws();
+ }
+
+ /// <summary>
+ /// Updates the currently bound constant buffer.
+ /// </summary>
+ /// <param name="data">Data to be written to the buffer</param>
+ public void ConstantBufferUpdate(ReadOnlySpan<int> data)
+ {
+ _cbUpdater.Update(data);
+ }
+
+ /// <summary>
+ /// Launches the Inline-to-Memory DMA copy operation.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void LaunchDma(int argument)
+ {
+ _i2mClass.LaunchDma(ref Unsafe.As<ThreedClassState, InlineToMemoryClassState>(ref _state.State), argument);
+ }
+
+ /// <summary>
+ /// Pushes a block of data to the Inline-to-Memory engine.
+ /// </summary>
+ /// <param name="data">Data to push</param>
+ public void LoadInlineData(ReadOnlySpan<int> data)
+ {
+ _i2mClass.LoadInlineData(data);
+ }
+
+ /// <summary>
+ /// Pushes a word of data to the Inline-to-Memory engine.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void LoadInlineData(int argument)
+ {
+ _i2mClass.LoadInlineData(argument);
+ }
+
+ /// <summary>
+ /// Performs an incrementation on a syncpoint.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ public void IncrementSyncpoint(int argument)
+ {
+ uint syncpointId = (uint)argument & 0xFFFF;
+
+ _context.AdvanceSequence();
+ _context.CreateHostSyncIfNeeded(true, true);
+ _context.Renderer.UpdateCounters(); // Poll the query counters, the game may want an updated result.
+ _context.Synchronization.IncrementSyncpoint(syncpointId);
+ }
+
+ /// <summary>
+ /// Invalidates the cache with the sampler descriptors from the sampler pool.
+ /// </summary>
+ /// <param name="argument">Method call argument (unused)</param>
+ private void InvalidateSamplerCacheNoWfi(int argument)
+ {
+ _context.AdvanceSequence();
+ }
+
+ /// <summary>
+ /// Invalidates the cache with the texture descriptors from the texture pool.
+ /// </summary>
+ /// <param name="argument">Method call argument (unused)</param>
+ private void InvalidateTextureHeaderCacheNoWfi(int argument)
+ {
+ _context.AdvanceSequence();
+ }
+
+ /// <summary>
+ /// Issues a texture barrier.
+ /// This waits until previous texture writes from the GPU to finish, before
+ /// performing new operations with said textures.
+ /// </summary>
+ /// <param name="argument">Method call argument (unused)</param>
+ private void TextureBarrier(int argument)
+ {
+ _context.Renderer.Pipeline.TextureBarrier();
+ }
+
+ /// <summary>
+ /// Sets the start offset of the blend microcode in memory.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void LoadBlendUcodeStart(int argument)
+ {
+ _blendManager.LoadBlendUcodeStart(argument);
+ }
+
+ /// <summary>
+ /// Pushes one word of blend microcode.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void LoadBlendUcodeInstruction(int argument)
+ {
+ _blendManager.LoadBlendUcodeInstruction(argument);
+ }
+
+ /// <summary>
+ /// Issues a texture barrier.
+ /// This waits until previous texture writes from the GPU to finish, before
+ /// performing new operations with said textures.
+ /// This performs a per-tile wait, it is only valid if both the previous write
+ /// and current access has the same access patterns.
+ /// This may be faster than the regular barrier on tile-based rasterizers.
+ /// </summary>
+ /// <param name="argument">Method call argument (unused)</param>
+ private void TextureBarrierTiled(int argument)
+ {
+ _context.Renderer.Pipeline.TextureBarrierTiled();
+ }
+
+ /// <summary>
+ /// Draws a texture, without needing to specify shader programs.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawTexture(int argument)
+ {
+ _drawManager.DrawTexture(this, argument);
+ }
+
+ /// <summary>
+ /// Performs a non-indexed draw with the specified topology, index and count.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawVertexArrayBeginEndInstanceFirst(int argument)
+ {
+ _drawManager.DrawVertexArrayBeginEndInstanceFirst(this, argument);
+ }
+
+ /// <summary>
+ /// Performs a non-indexed draw with the specified topology, index and count,
+ /// while incrementing the current instance.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawVertexArrayBeginEndInstanceSubsequent(int argument)
+ {
+ _drawManager.DrawVertexArrayBeginEndInstanceSubsequent(this, argument);
+ }
+
+ /// <summary>
+ /// Pushes four 8-bit index buffer elements.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void VbElementU8(int argument)
+ {
+ _drawManager.VbElementU8(argument);
+ }
+
+ /// <summary>
+ /// Pushes two 16-bit index buffer elements.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void VbElementU16(int argument)
+ {
+ _drawManager.VbElementU16(argument);
+ }
+
+ /// <summary>
+ /// Pushes one 32-bit index buffer element.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void VbElementU32(int argument)
+ {
+ _drawManager.VbElementU32(argument);
+ }
+
+ /// <summary>
+ /// Resets the value of an internal GPU counter back to zero.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void ResetCounter(int argument)
+ {
+ _semaphoreUpdater.ResetCounter(argument);
+ }
+
+ /// <summary>
+ /// Finishes the draw call.
+ /// This draws geometry on the bound buffers based on the current GPU state.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawEnd(int argument)
+ {
+ _drawManager.DrawEnd(this, argument);
+ }
+
+ /// <summary>
+ /// Starts draw.
+ /// This sets primitive type and instanced draw parameters.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawBegin(int argument)
+ {
+ _drawManager.DrawBegin(argument);
+ }
+
+ /// <summary>
+ /// Sets the index buffer count.
+ /// This also sets internal state that indicates that the next draw is an indexed draw.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void SetIndexBufferCount(int argument)
+ {
+ _drawManager.SetIndexBufferCount(argument);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 8-bit index buffer elements.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawIndexBuffer8BeginEndInstanceFirst(int argument)
+ {
+ _drawManager.DrawIndexBuffer8BeginEndInstanceFirst(this, argument);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 16-bit index buffer elements.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawIndexBuffer16BeginEndInstanceFirst(int argument)
+ {
+ _drawManager.DrawIndexBuffer16BeginEndInstanceFirst(this, argument);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 32-bit index buffer elements.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawIndexBuffer32BeginEndInstanceFirst(int argument)
+ {
+ _drawManager.DrawIndexBuffer32BeginEndInstanceFirst(this, argument);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 8-bit index buffer elements,
+ /// while also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawIndexBuffer8BeginEndInstanceSubsequent(int argument)
+ {
+ _drawManager.DrawIndexBuffer8BeginEndInstanceSubsequent(this, argument);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 16-bit index buffer elements,
+ /// while also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawIndexBuffer16BeginEndInstanceSubsequent(int argument)
+ {
+ _drawManager.DrawIndexBuffer16BeginEndInstanceSubsequent(this, argument);
+ }
+
+ /// <summary>
+ /// Performs a indexed draw with 32-bit index buffer elements,
+ /// while also pre-incrementing the current instance value.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void DrawIndexBuffer32BeginEndInstanceSubsequent(int argument)
+ {
+ _drawManager.DrawIndexBuffer32BeginEndInstanceSubsequent(this, argument);
+ }
+
+ /// <summary>
+ /// Clears the current color and depth-stencil buffers.
+ /// Which buffers should be cleared is also specified on the argument.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void Clear(int argument)
+ {
+ _drawManager.Clear(this, argument);
+ }
+
+ /// <summary>
+ /// Writes a GPU counter to guest memory.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void Report(int argument)
+ {
+ _semaphoreUpdater.Report(argument);
+ }
+
+ /// <summary>
+ /// Performs high-level emulation of Falcon microcode function number "4".
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void SetFalcon04(int argument)
+ {
+ _state.State.SetMmeShadowScratch[0] = 1;
+ }
+
+ /// <summary>
+ /// Updates the uniform buffer data with inline data.
+ /// </summary>
+ /// <param name="argument">New uniform buffer data word</param>
+ private void ConstantBufferUpdate(int argument)
+ {
+ _cbUpdater.Update(argument);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the vertex shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void ConstantBufferBindVertex(int argument)
+ {
+ _cbUpdater.BindVertex(argument);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the tessellation control shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void ConstantBufferBindTessControl(int argument)
+ {
+ _cbUpdater.BindTessControl(argument);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the tessellation evaluation shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void ConstantBufferBindTessEvaluation(int argument)
+ {
+ _cbUpdater.BindTessEvaluation(argument);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the geometry shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void ConstantBufferBindGeometry(int argument)
+ {
+ _cbUpdater.BindGeometry(argument);
+ }
+
+ /// <summary>
+ /// Binds a uniform buffer for the fragment shader stage.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ private void ConstantBufferBindFragment(int argument)
+ {
+ _cbUpdater.BindFragment(argument);
+ }
+
+ /// <summary>
+ /// Generic register read function that just returns 0.
+ /// </summary>
+ /// <returns>Zero</returns>
+ private static int Zero()
+ {
+ return 0;
+ }
+
+ /// <summary>
+ /// Performs a indexed or non-indexed draw.
+ /// </summary>
+ /// <param name="topology">Primitive topology</param>
+ /// <param name="count">Index count for indexed draws, vertex count for non-indexed draws</param>
+ /// <param name="instanceCount">Instance count</param>
+ /// <param name="firstIndex">First index on the index buffer for indexed draws, ignored for non-indexed draws</param>
+ /// <param name="firstVertex">First vertex on the vertex buffer</param>
+ /// <param name="firstInstance">First instance</param>
+ /// <param name="indexed">True if the draw is indexed, false otherwise</param>
+ public void Draw(
+ PrimitiveTopology topology,
+ int count,
+ int instanceCount,
+ int firstIndex,
+ int firstVertex,
+ int firstInstance,
+ bool indexed)
+ {
+ _drawManager.Draw(this, topology, count, instanceCount, firstIndex, firstVertex, firstInstance, indexed);
+ }
+
+ /// <summary>
+ /// Performs a indirect draw, with parameters from a GPU buffer.
+ /// </summary>
+ /// <param name="topology">Primitive topology</param>
+ /// <param name="indirectBufferAddress">Address of the buffer with the draw parameters, such as count, first index, etc</param>
+ /// <param name="parameterBufferAddress">Address of the buffer with the draw count</param>
+ /// <param name="maxDrawCount">Maximum number of draws that can be made</param>
+ /// <param name="stride">Distance in bytes between each entry on the data pointed to by <paramref name="indirectBufferAddress"/></param>
+ /// <param name="indexCount">Maximum number of indices that the draw can consume</param>
+ /// <param name="drawType">Type of the indirect draw, which can be indexed or non-indexed, with or without a draw count</param>
+ public void DrawIndirect(
+ PrimitiveTopology topology,
+ ulong indirectBufferAddress,
+ ulong parameterBufferAddress,
+ int maxDrawCount,
+ int stride,
+ int indexCount,
+ IndirectDrawType drawType)
+ {
+ _drawManager.DrawIndirect(this, topology, indirectBufferAddress, parameterBufferAddress, maxDrawCount, stride, indexCount, drawType);
+ }
+
+ /// <summary>
+ /// Clears the current color and depth-stencil buffers.
+ /// Which buffers should be cleared can also specified with the arguments.
+ /// </summary>
+ /// <param name="argument">Method call argument</param>
+ /// <param name="layerCount">For array and 3D textures, indicates how many layers should be cleared</param>
+ public void Clear(int argument, int layerCount)
+ {
+ _drawManager.Clear(this, argument, layerCount);
+ }
+ }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs
new file mode 100644
index 00000000..8f26f38f
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs
@@ -0,0 +1,1048 @@
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Image;
+using Ryujinx.Graphics.Shader;
+using System;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+ /// <summary>
+ /// Shader stage name.
+ /// </summary>
+ enum ShaderType
+ {
+ Vertex,
+ TessellationControl,
+ TessellationEvaluation,
+ Geometry,
+ Fragment
+ }
+
+ /// <summary>
+ /// Tessellation mode.
+ /// </summary>
+ struct TessMode
+ {
+#pragma warning disable CS0649
+ public uint Packed;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Unpacks the tessellation abstract patch type.
+ /// </summary>
+ /// <returns>Abtract patch type</returns>
+ public TessPatchType UnpackPatchType()
+ {
+ return (TessPatchType)(Packed & 3);
+ }
+
+ /// <summary>
+ /// Unpacks the spacing between tessellated vertices of the patch.
+ /// </summary>
+ /// <returns>Spacing between tessellated vertices</returns>
+ public TessSpacing UnpackSpacing()
+ {
+ return (TessSpacing)((Packed >> 4) & 3);
+ }
+
+ /// <summary>
+ /// Unpacks the primitive winding order.
+ /// </summary>
+ /// <returns>True if clockwise, false if counter-clockwise</returns>
+ public bool UnpackCw()
+ {
+ return (Packed & (1 << 8)) != 0;
+ }
+ }
+
+ /// <summary>
+ /// Transform feedback buffer state.
+ /// </summary>
+ struct TfBufferState
+ {
+#pragma warning disable CS0649
+ public Boolean32 Enable;
+ public GpuVa Address;
+ public int Size;
+ public int Offset;
+ public uint Padding0;
+ public uint Padding1;
+ public uint Padding2;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Transform feedback state.
+ /// </summary>
+ struct TfState
+ {
+#pragma warning disable CS0649
+ public int BufferIndex;
+ public int VaryingsCount;
+ public int Stride;
+ public uint Padding;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Render target color buffer state.
+ /// </summary>
+ struct RtColorState
+ {
+#pragma warning disable CS0649
+ public GpuVa Address;
+ public int WidthOrStride;
+ public int Height;
+ public ColorFormat Format;
+ public MemoryLayout MemoryLayout;
+ public int Depth;
+ public int LayerSize;
+ public int BaseLayer;
+ public int Unknown0x24;
+ public int Padding0;
+ public int Padding1;
+ public int Padding2;
+ public int Padding3;
+ public int Padding4;
+ public int Padding5;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Viewport transform parameters, for viewport transformation.
+ /// </summary>
+ struct ViewportTransform
+ {
+#pragma warning disable CS0649
+ public float ScaleX;
+ public float ScaleY;
+ public float ScaleZ;
+ public float TranslateX;
+ public float TranslateY;
+ public float TranslateZ;
+ public uint Swizzle;
+ public uint SubpixelPrecisionBias;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Unpacks viewport swizzle of the position X component.
+ /// </summary>
+ /// <returns>Swizzle enum value</returns>
+ public ViewportSwizzle UnpackSwizzleX()
+ {
+ return (ViewportSwizzle)(Swizzle & 7);
+ }
+
+ /// <summary>
+ /// Unpacks viewport swizzle of the position Y component.
+ /// </summary>
+ /// <returns>Swizzle enum value</returns>
+ public ViewportSwizzle UnpackSwizzleY()
+ {
+ return (ViewportSwizzle)((Swizzle >> 4) & 7);
+ }
+
+ /// <summary>
+ /// Unpacks viewport swizzle of the position Z component.
+ /// </summary>
+ /// <returns>Swizzle enum value</returns>
+ public ViewportSwizzle UnpackSwizzleZ()
+ {
+ return (ViewportSwizzle)((Swizzle >> 8) & 7);
+ }
+
+ /// <summary>
+ /// Unpacks viewport swizzle of the position W component.
+ /// </summary>
+ /// <returns>Swizzle enum value</returns>
+ public ViewportSwizzle UnpackSwizzleW()
+ {
+ return (ViewportSwizzle)((Swizzle >> 12) & 7);
+ }
+ }
+
+ /// <summary>
+ /// Viewport extents for viewport clipping, also includes depth range.
+ /// </summary>
+ struct ViewportExtents
+ {
+#pragma warning disable CS0649
+ public ushort X;
+ public ushort Width;
+ public ushort Y;
+ public ushort Height;
+ public float DepthNear;
+ public float DepthFar;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Draw state for non-indexed draws.
+ /// </summary>
+ struct VertexBufferDrawState
+ {
+#pragma warning disable CS0649
+ public int First;
+ public int Count;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Color buffer clear color.
+ /// </summary>
+ struct ClearColors
+ {
+#pragma warning disable CS0649
+ public float Red;
+ public float Green;
+ public float Blue;
+ public float Alpha;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Depth bias (also called polygon offset) parameters.
+ /// </summary>
+ struct DepthBiasState
+ {
+#pragma warning disable CS0649
+ public Boolean32 PointEnable;
+ public Boolean32 LineEnable;
+ public Boolean32 FillEnable;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Indicates whenever the blend microcode processes RGB and alpha components.
+ /// </summary>
+ enum BlendUcodeEnable
+ {
+ Disabled = 0,
+ EnableRGB = 1,
+ EnableAlpha = 2,
+ EnableRGBA = 3
+ }
+
+ /// <summary>
+ /// Scissor state.
+ /// </summary>
+ struct ScissorState
+ {
+#pragma warning disable CS0649
+ public Boolean32 Enable;
+ public ushort X1;
+ public ushort X2;
+ public ushort Y1;
+ public ushort Y2;
+ public uint Padding;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Stencil test masks for back tests.
+ /// </summary>
+ struct StencilBackMasks
+ {
+#pragma warning disable CS0649
+ public int FuncRef;
+ public int Mask;
+ public int FuncMask;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Render target depth-stencil buffer state.
+ /// </summary>
+ struct RtDepthStencilState
+ {
+#pragma warning disable CS0649
+ public GpuVa Address;
+ public ZetaFormat Format;
+ public MemoryLayout MemoryLayout;
+ public int LayerSize;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Screen scissor state.
+ /// </summary>
+ struct ScreenScissorState
+ {
+#pragma warning disable CS0649
+ public ushort X;
+ public ushort Width;
+ public ushort Y;
+ public ushort Height;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Vertex attribute vector and component size.
+ /// </summary>
+ enum VertexAttribSize
+ {
+ Size32x4 = 1,
+ Size32x3 = 2,
+ Size16x4 = 3,
+ Size32x2 = 4,
+ Size16x3 = 5,
+ Size8x4 = 0xa,
+ Size16x2 = 0xf,
+ Size32 = 0x12,
+ Size8x3 = 0x13,
+ Size8x2 = 0x18,
+ Size16 = 0x1b,
+ Size8 = 0x1d,
+ Rgb10A2 = 0x30,
+ Rg11B10 = 0x31
+ }
+
+ /// <summary>
+ /// Vertex attribute component type.
+ /// </summary>
+ enum VertexAttribType
+ {
+ Snorm = 1,
+ Unorm = 2,
+ Sint = 3,
+ Uint = 4,
+ Uscaled = 5,
+ Sscaled = 6,
+ Float = 7
+ }
+
+ /// <summary>
+ /// Vertex buffer attribute state.
+ /// </summary>
+ struct VertexAttribState
+ {
+#pragma warning disable CS0649
+ public uint Attribute;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Unpacks the index of the vertex buffer this attribute belongs to.
+ /// </summary>
+ /// <returns>Vertex buffer index</returns>
+ public int UnpackBufferIndex()
+ {
+ return (int)(Attribute & 0x1f);
+ }
+
+ /// <summary>
+ /// Unpacks the attribute constant flag.
+ /// </summary>
+ /// <returns>True if the attribute is constant, false otherwise</returns>
+ public bool UnpackIsConstant()
+ {
+ return (Attribute & 0x40) != 0;
+ }
+
+ /// <summary>
+ /// Unpacks the offset, in bytes, of the attribute on the vertex buffer.
+ /// </summary>
+ /// <returns>Attribute offset in bytes</returns>
+ public int UnpackOffset()
+ {
+ return (int)((Attribute >> 7) & 0x3fff);
+ }
+
+ /// <summary>
+ /// Unpacks the Maxwell attribute format integer.
+ /// </summary>
+ /// <returns>Attribute format integer</returns>
+ public uint UnpackFormat()
+ {
+ return Attribute & 0x3fe00000;
+ }
+
+ /// <summary>
+ /// Unpacks the Maxwell attribute size.
+ /// </summary>
+ /// <returns>Attribute size</returns>
+ public VertexAttribSize UnpackSize()
+ {
+ return (VertexAttribSize)((Attribute >> 21) & 0x3f);
+ }
+
+ /// <summary>
+ /// Unpacks the Maxwell attribute component type.
+ /// </summary>
+ /// <returns>Attribute component type</returns>
+ public VertexAttribType UnpackType()
+ {
+ return (VertexAttribType)((Attribute >> 27) & 7);
+ }
+ }
+
+ /// <summary>
+ /// Render target draw buffers control.
+ /// </summary>
+ struct RtControl
+ {
+#pragma warning disable CS0649
+ public uint Packed;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Unpacks the number of active draw buffers.
+ /// </summary>
+ /// <returns>Number of active draw buffers</returns>
+ public int UnpackCount()
+ {
+ return (int)(Packed & 0xf);
+ }
+
+ /// <summary>
+ /// Unpacks the color attachment index for a given draw buffer.
+ /// </summary>
+ /// <param name="index">Index of the draw buffer</param>
+ /// <returns>Attachment index</returns>
+ public int UnpackPermutationIndex(int index)
+ {
+ return (int)((Packed >> (4 + index * 3)) & 7);
+ }
+ }
+
+ /// <summary>
+ /// 3D, 2D or 1D texture size.
+ /// </summary>
+ struct Size3D
+ {
+#pragma warning disable CS0649
+ public int Width;
+ public int Height;
+ public int Depth;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Stencil front test state and masks.
+ /// </summary>
+ struct StencilTestState
+ {
+#pragma warning disable CS0649
+ public Boolean32 Enable;
+ public StencilOp FrontSFail;
+ public StencilOp FrontDpFail;
+ public StencilOp FrontDpPass;
+ public CompareOp FrontFunc;
+ public int FrontFuncRef;
+ public int FrontFuncMask;
+ public int FrontMask;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Screen Y control register.
+ /// </summary>
+ [Flags]
+ enum YControl
+ {
+ NegateY = 1 << 0,
+ TriangleRastFlip = 1 << 4
+ }
+
+ /// <summary>
+ /// RGB color components packed as 16-bit float values.
+ /// </summary>
+ struct RgbHalf
+ {
+#pragma warning disable CS0649
+ public uint R;
+ public uint G;
+ public uint B;
+ public uint Padding;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Unpacks the red color component as a 16-bit float value.
+ /// </summary>
+ /// <returns>The component value</returns>
+ public Half UnpackR()
+ {
+ ushort value = (ushort)R;
+ return Unsafe.As<ushort, Half>(ref value);
+ }
+
+ /// <summary>
+ /// Unpacks the green color component as a 16-bit float value.
+ /// </summary>
+ /// <returns>The component value</returns>
+ public Half UnpackG()
+ {
+ ushort value = (ushort)G;
+ return Unsafe.As<ushort, Half>(ref value);
+ }
+
+ /// <summary>
+ /// Unpacks the blue color component as a 16-bit float value.
+ /// </summary>
+ /// <returns>The component value</returns>
+ public Half UnpackB()
+ {
+ ushort value = (ushort)B;
+ return Unsafe.As<ushort, Half>(ref value);
+ }
+ }
+
+ /// <summary>
+ /// Condition for conditional rendering.
+ /// </summary>
+ enum Condition
+ {
+ Never,
+ Always,
+ ResultNonZero,
+ Equal,
+ NotEqual
+ }
+
+ /// <summary>
+ /// Texture or sampler pool state.
+ /// </summary>
+ struct PoolState
+ {
+#pragma warning disable CS0649
+ public GpuVa Address;
+ public int MaximumId;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Stencil back test state.
+ /// </summary>
+ struct StencilBackTestState
+ {
+#pragma warning disable CS0649
+ public Boolean32 TwoSided;
+ public StencilOp BackSFail;
+ public StencilOp BackDpFail;
+ public StencilOp BackDpPass;
+ public CompareOp BackFunc;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Primitive restart state.
+ /// </summary>
+ struct PrimitiveRestartState
+ {
+#pragma warning disable CS0649
+ public Boolean32 Enable;
+ public int Index;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// GPU index buffer state.
+ /// This is used on indexed draws.
+ /// </summary>
+ struct IndexBufferState
+ {
+#pragma warning disable CS0649
+ public GpuVa Address;
+ public GpuVa EndAddress;
+ public IndexType Type;
+ public int First;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Face culling and orientation parameters.
+ /// </summary>
+ struct FaceState
+ {
+#pragma warning disable CS0649
+ public Boolean32 CullEnable;
+ public FrontFace FrontFace;
+ public Face CullFace;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// View volume clip control.
+ /// </summary>
+ [Flags]
+ enum ViewVolumeClipControl
+ {
+ ForceDepthRangeZeroToOne = 1 << 0,
+ DepthClampDisabled = 1 << 11
+ }
+
+ /// <summary>
+ /// Logical operation state.
+ /// </summary>
+ struct LogicalOpState
+ {
+#pragma warning disable CS0649
+ public Boolean32 Enable;
+ public LogicalOp LogicalOp;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Render target color buffer mask.
+ /// This defines which color channels are written to the color buffer.
+ /// </summary>
+ struct RtColorMask
+ {
+#pragma warning disable CS0649
+ public uint Packed;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Unpacks red channel enable.
+ /// </summary>
+ /// <returns>True to write the new red channel color, false to keep the old value</returns>
+ public bool UnpackRed()
+ {
+ return (Packed & 0x1) != 0;
+ }
+
+ /// <summary>
+ /// Unpacks green channel enable.
+ /// </summary>
+ /// <returns>True to write the new green channel color, false to keep the old value</returns>
+ public bool UnpackGreen()
+ {
+ return (Packed & 0x10) != 0;
+ }
+
+ /// <summary>
+ /// Unpacks blue channel enable.
+ /// </summary>
+ /// <returns>True to write the new blue channel color, false to keep the old value</returns>
+ public bool UnpackBlue()
+ {
+ return (Packed & 0x100) != 0;
+ }
+
+ /// <summary>
+ /// Unpacks alpha channel enable.
+ /// </summary>
+ /// <returns>True to write the new alpha channel color, false to keep the old value</returns>
+ public bool UnpackAlpha()
+ {
+ return (Packed & 0x1000) != 0;
+ }
+ }
+
+ /// <summary>
+ /// Vertex buffer state.
+ /// </summary>
+ struct VertexBufferState
+ {
+#pragma warning disable CS0649
+ public uint Control;
+ public GpuVa Address;
+ public int Divisor;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Vertex buffer stride, defined as the number of bytes occupied by each vertex in memory.
+ /// </summary>
+ /// <returns>Vertex buffer stride</returns>
+ public int UnpackStride()
+ {
+ return (int)(Control & 0xfff);
+ }
+
+ /// <summary>
+ /// Vertex buffer enable.
+ /// </summary>
+ /// <returns>True if the vertex buffer is enabled, false otherwise</returns>
+ public bool UnpackEnable()
+ {
+ return (Control & (1 << 12)) != 0;
+ }
+ }
+
+ /// <summary>
+ /// Color buffer blending parameters, shared by all color buffers.
+ /// </summary>
+ struct BlendStateCommon
+ {
+#pragma warning disable CS0649
+ public Boolean32 SeparateAlpha;
+ public BlendOp ColorOp;
+ public BlendFactor ColorSrcFactor;
+ public BlendFactor ColorDstFactor;
+ public BlendOp AlphaOp;
+ public BlendFactor AlphaSrcFactor;
+ public uint Unknown0x1354;
+ public BlendFactor AlphaDstFactor;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Color buffer blending parameters.
+ /// </summary>
+ struct BlendState
+ {
+#pragma warning disable CS0649
+ public Boolean32 SeparateAlpha;
+ public BlendOp ColorOp;
+ public BlendFactor ColorSrcFactor;
+ public BlendFactor ColorDstFactor;
+ public BlendOp AlphaOp;
+ public BlendFactor AlphaSrcFactor;
+ public BlendFactor AlphaDstFactor;
+ public uint Padding;
+#pragma warning restore CS0649
+ }
+
+ /// <summary>
+ /// Graphics shader stage state.
+ /// </summary>
+ struct ShaderState
+ {
+#pragma warning disable CS0649
+ public uint Control;
+ public uint Offset;
+ public uint Unknown0x8;
+ public int MaxRegisters;
+ public ShaderType Type;
+ public uint Unknown0x14;
+ public uint Unknown0x18;
+ public uint Unknown0x1c;
+ public uint Unknown0x20;
+ public uint Unknown0x24;
+ public uint Unknown0x28;
+ public uint Unknown0x2c;
+ public uint Unknown0x30;
+ public uint Unknown0x34;
+ public uint Unknown0x38;
+ public uint Unknown0x3c;
+#pragma warning restore CS0649
+
+ /// <summary>
+ /// Unpacks shader enable information.
+ /// Must be ignored for vertex shaders, those are always enabled.
+ /// </summary>
+ /// <returns>True if the stage is enabled, false otherwise</returns>
+ public bool UnpackEnable()
+ {
+ return (Control & 1) != 0;
+ }
+ }
+
+ /// <summary>
+ /// Uniform buffer state for the uniform buffer currently being modified.
+ /// </summary>
+ struct UniformBufferState
+ {
+#pragma warning disable CS0649
+ public int Size;
+ public GpuVa Address;
+ public int Offset;
+#pragma warning restore CS0649
+ }
+
+ unsafe struct ThreedClassState : IShadowState
+ {
+#pragma warning disable CS0649
+ public uint SetObject;
+ public int SetObjectClassId => (int)((SetObject >> 0) & 0xFFFF);
+ public int SetObjectEngineId => (int)((SetObject >> 16) & 0x1F);
+ public fixed uint Reserved04[63];
+ public uint NoOperation;
+ public uint SetNotifyA;
+ public int SetNotifyAAddressUpper => (int)((SetNotifyA >> 0) & 0xFF);
+ public uint SetNotifyB;
+ public uint Notify;
+ public NotifyType NotifyType => (NotifyType)(Notify);
+ public uint WaitForIdle;
+ public uint LoadMmeInstructionRamPointer;
+ public uint LoadMmeInstructionRam;
+ public uint LoadMmeStartAddressRamPointer;
+ public uint LoadMmeStartAddressRam;
+ public uint SetMmeShadowRamControl;
+ public SetMmeShadowRamControlMode SetMmeShadowRamControlMode => (SetMmeShadowRamControlMode)((SetMmeShadowRamControl >> 0) & 0x3);
+ public fixed uint Reserved128[2];
+ public uint SetGlobalRenderEnableA;
+ public int SetGlobalRenderEnableAOffsetUpper => (int)((SetGlobalRenderEnableA >> 0) & 0xFF);
+ public uint SetGlobalRenderEnableB;
+ public uint SetGlobalRenderEnableC;
+ public int SetGlobalRenderEnableCMode => (int)((SetGlobalRenderEnableC >> 0) & 0x7);
+ public uint SendGoIdle;
+ public uint PmTrigger;
+ public uint PmTriggerWfi;
+ public fixed uint Reserved148[2];
+ public uint SetInstrumentationMethodHeader;
+ public uint SetInstrumentationMethodData;
+ public fixed uint Reserved158[10];
+ public uint LineLengthIn;
+ public uint LineCount;
+ public uint OffsetOutUpper;
+ public int OffsetOutUpperValue => (int)((OffsetOutUpper >> 0) & 0xFF);
+ public uint OffsetOut;
+ public uint PitchOut;
+ public uint SetDstBlockSize;
+ public SetDstBlockSizeWidth SetDstBlockSizeWidth => (SetDstBlockSizeWidth)((SetDstBlockSize >> 0) & 0xF);
+ public SetDstBlockSizeHeight SetDstBlockSizeHeight => (SetDstBlockSizeHeight)((SetDstBlockSize >> 4) & 0xF);
+ public SetDstBlockSizeDepth SetDstBlockSizeDepth => (SetDstBlockSizeDepth)((SetDstBlockSize >> 8) & 0xF);
+ public uint SetDstWidth;
+ public uint SetDstHeight;
+ public uint SetDstDepth;
+ public uint SetDstLayer;
+ public uint SetDstOriginBytesX;
+ public int SetDstOriginBytesXV => (int)((SetDstOriginBytesX >> 0) & 0xFFFFF);
+ public uint SetDstOriginSamplesY;
+ public int SetDstOriginSamplesYV => (int)((SetDstOriginSamplesY >> 0) & 0xFFFF);
+ public uint LaunchDma;
+ public LaunchDmaDstMemoryLayout LaunchDmaDstMemoryLayout => (LaunchDmaDstMemoryLayout)((LaunchDma >> 0) & 0x1);
+ public LaunchDmaCompletionType LaunchDmaCompletionType => (LaunchDmaCompletionType)((LaunchDma >> 4) & 0x3);
+ public LaunchDmaInterruptType LaunchDmaInterruptType => (LaunchDmaInterruptType)((LaunchDma >> 8) & 0x3);
+ public LaunchDmaSemaphoreStructSize LaunchDmaSemaphoreStructSize => (LaunchDmaSemaphoreStructSize)((LaunchDma >> 12) & 0x1);
+ public bool LaunchDmaReductionEnable => (LaunchDma & 0x2) != 0;
+ public LaunchDmaReductionOp LaunchDmaReductionOp => (LaunchDmaReductionOp)((LaunchDma >> 13) & 0x7);
+ public LaunchDmaReductionFormat LaunchDmaReductionFormat => (LaunchDmaReductionFormat)((LaunchDma >> 2) & 0x3);
+ public bool LaunchDmaSysmembarDisable => (LaunchDma & 0x40) != 0;
+ public uint LoadInlineData;
+ public fixed uint Reserved1B8[22];
+ public Boolean32 EarlyZForce;
+ public fixed uint Reserved214[45];
+ public uint SyncpointAction;
+ public fixed uint Reserved2CC[10];
+ public uint BlendUcodeNormalizedDst;
+ public fixed uint Reserved2F8[10];
+ public TessMode TessMode;
+ public Array4<float> TessOuterLevel;
+ public Array2<float> TessInnerLevel;
+ public fixed uint Reserved33C[16];
+ public Boolean32 RasterizeEnable;
+ public Array4<TfBufferState> TfBufferState;
+ public fixed uint Reserved400[192];
+ public Array4<TfState> TfState;
+ public fixed uint Reserved740[1];
+ public Boolean32 TfEnable;
+ public fixed uint Reserved748[46];
+ public Array8<RtColorState> RtColorState;
+ public Array16<ViewportTransform> ViewportTransform;
+ public Array16<ViewportExtents> ViewportExtents;
+ public fixed uint ReservedD00[29];
+ public VertexBufferDrawState VertexBufferDrawState;
+ public uint DepthMode;
+ public ClearColors ClearColors;
+ public float ClearDepthValue;
+ public fixed uint ReservedD94[3];
+ public uint ClearStencilValue;
+ public fixed uint ReservedDA4[2];
+ public PolygonMode PolygonModeFront;
+ public PolygonMode PolygonModeBack;
+ public Boolean32 PolygonSmoothEnable;
+ public fixed uint ReservedDB8[2];
+ public DepthBiasState DepthBiasState;
+ public int PatchVertices;
+ public BlendUcodeEnable BlendUcodeEnable;
+ public uint BlendUcodeSize;
+ public fixed uint ReservedDD8[2];
+ public uint TextureBarrier;
+ public uint WatchdogTimer;
+ public Boolean32 PrimitiveRestartDrawArrays;
+ public uint ReservedDEC;
+ public uint LoadBlendUcodeStart;
+ public uint LoadBlendUcodeInstruction;
+ public fixed uint ReservedDF8[2];
+ public Array16<ScissorState> ScissorState;
+ public fixed uint ReservedF00[21];
+ public StencilBackMasks StencilBackMasks;
+ public fixed uint ReservedF60[5];
+ public uint InvalidateTextures;
+ public fixed uint ReservedF78[1];
+ public uint TextureBarrierTiled;
+ public fixed uint ReservedF80[4];
+ public Boolean32 RtColorMaskShared;
+ public fixed uint ReservedF94[19];
+ public RtDepthStencilState RtDepthStencilState;
+ public ScreenScissorState ScreenScissorState;
+ public fixed uint ReservedFFC[33];
+ public int DrawTextureDstX;
+ public int DrawTextureDstY;
+ public int DrawTextureDstWidth;
+ public int DrawTextureDstHeight;
+ public long DrawTextureDuDx;
+ public long DrawTextureDvDy;
+ public int DrawTextureSamplerId;
+ public int DrawTextureTextureId;
+ public int DrawTextureSrcX;
+ public int DrawTextureSrcY;
+ public fixed uint Reserved10B0[18];
+ public uint ClearFlags;
+ public fixed uint Reserved10FC[25];
+ public Array32<VertexAttribState> VertexAttribState;
+ public fixed uint Reserved11E0[13];
+ public uint DrawVertexArrayBeginEndInstanceFirst;
+ public uint DrawVertexArrayBeginEndInstanceSubsequent;
+ public RtControl RtControl;
+ public fixed uint Reserved1220[2];
+ public Size3D RtDepthStencilSize;
+ public SamplerIndex SamplerIndex;
+ public fixed uint Reserved1238[37];
+ public Boolean32 DepthTestEnable;
+ public fixed uint Reserved12D0[4];
+ public Boolean32 AlphaToCoverageDitherEnable;
+ public Boolean32 BlendIndependent;
+ public Boolean32 DepthWriteEnable;
+ public Boolean32 AlphaTestEnable;
+ public fixed uint Reserved12F0[5];
+ public uint VbElementU8;
+ public uint Reserved1308;
+ public CompareOp DepthTestFunc;
+ public float AlphaTestRef;
+ public CompareOp AlphaTestFunc;
+ public uint Reserved1318;
+ public ColorF BlendConstant;
+ public fixed uint Reserved132C[4];
+ public BlendStateCommon BlendStateCommon;
+ public Boolean32 BlendEnableCommon;
+ public Array8<Boolean32> BlendEnable;
+ public StencilTestState StencilTestState;
+ public fixed uint Reserved13A0[3];
+ public YControl YControl;
+ public float LineWidthSmooth;
+ public float LineWidthAliased;
+ public fixed uint Reserved13B8[27];
+ public uint InvalidateSamplerCacheNoWfi;
+ public uint InvalidateTextureHeaderCacheNoWfi;
+ public fixed uint Reserved142C[2];
+ public uint FirstVertex;
+ public uint FirstInstance;
+ public fixed uint Reserved143C[17];
+ public Array8<RgbHalf> BlendUcodeConstants;
+ public fixed uint Reserved1500[4];
+ public uint ClipDistanceEnable;
+ public uint Reserved1514;
+ public float PointSize;
+ public uint Reserved151C;
+ public Boolean32 PointSpriteEnable;
+ public fixed uint Reserved1524[3];
+ public uint ResetCounter;
+ public Boolean32 MultisampleEnable;
+ public Boolean32 RtDepthStencilEnable;
+ public uint MultisampleControl;
+ public fixed uint Reserved1540[4];
+ public GpuVa RenderEnableAddress;
+ public Condition RenderEnableCondition;
+ public PoolState SamplerPoolState;
+ public uint Reserved1568;
+ public float DepthBiasFactor;
+ public Boolean32 LineSmoothEnable;
+ public PoolState TexturePoolState;
+ public fixed uint Reserved1580[5];
+ public StencilBackTestState StencilBackTestState;
+ public fixed uint Reserved15A8[5];
+ public float DepthBiasUnits;
+ public fixed uint Reserved15C0[4];
+ public TextureMsaaMode RtMsaaMode;
+ public fixed uint Reserved15D4[5];
+ public uint VbElementU32;
+ public uint Reserved15EC;
+ public uint VbElementU16;
+ public fixed uint Reserved15F4[4];
+ public uint PointCoordReplace;
+ public GpuVa ShaderBaseAddress;
+ public uint Reserved1610;
+ public uint DrawEnd;
+ public uint DrawBegin;
+ public fixed uint Reserved161C[10];
+ public PrimitiveRestartState PrimitiveRestartState;
+ public fixed uint Reserved164C[95];
+ public IndexBufferState IndexBufferState;
+ public uint IndexBufferCount;
+ public uint DrawIndexBuffer32BeginEndInstanceFirst;
+ public uint DrawIndexBuffer16BeginEndInstanceFirst;
+ public uint DrawIndexBuffer8BeginEndInstanceFirst;
+ public uint DrawIndexBuffer32BeginEndInstanceSubsequent;
+ public uint DrawIndexBuffer16BeginEndInstanceSubsequent;
+ public uint DrawIndexBuffer8BeginEndInstanceSubsequent;
+ public fixed uint Reserved17FC[32];
+ public float DepthBiasClamp;
+ public Array16<Boolean32> VertexBufferInstanced;
+ public fixed uint Reserved18C0[20];
+ public Boolean32 VertexProgramPointSize;
+ public uint Reserved1914;
+ public FaceState FaceState;
+ public fixed uint Reserved1924[2];
+ public uint ViewportTransformEnable;
+ public fixed uint Reserved1930[3];
+ public ViewVolumeClipControl ViewVolumeClipControl;
+ public fixed uint Reserved1940[2];
+ public Boolean32 PrimitiveTypeOverrideEnable;
+ public fixed uint Reserved194C[9];
+ public PrimitiveTypeOverride PrimitiveTypeOverride;
+ public fixed uint Reserved1974[20];
+ public LogicalOpState LogicOpState;
+ public uint Reserved19CC;
+ public uint Clear;
+ public fixed uint Reserved19D4[11];
+ public Array8<RtColorMask> RtColorMask;
+ public fixed uint Reserved1A20[56];
+ public GpuVa SemaphoreAddress;
+ public int SemaphorePayload;
+ public uint SemaphoreControl;
+ public fixed uint Reserved1B10[60];
+ public Array16<VertexBufferState> VertexBufferState;
+ public fixed uint Reserved1D00[64];
+ public Array8<BlendState> BlendState;
+ public Array16<GpuVa> VertexBufferEndAddress;
+ public fixed uint Reserved1F80[32];
+ public Array6<ShaderState> ShaderState;
+ public fixed uint Reserved2180[96];
+ public uint SetFalcon00;
+ public uint SetFalcon01;
+ public uint SetFalcon02;
+ public uint SetFalcon03;
+ public uint SetFalcon04;
+ public uint SetFalcon05;
+ public uint SetFalcon06;
+ public uint SetFalcon07;
+ public uint SetFalcon08;
+ public uint SetFalcon09;
+ public uint SetFalcon10;
+ public uint SetFalcon11;
+ public uint SetFalcon12;
+ public uint SetFalcon13;
+ public uint SetFalcon14;
+ public uint SetFalcon15;
+ public uint SetFalcon16;
+ public uint SetFalcon17;
+ public uint SetFalcon18;
+ public uint SetFalcon19;
+ public uint SetFalcon20;
+ public uint SetFalcon21;
+ public uint SetFalcon22;
+ public uint SetFalcon23;
+ public uint SetFalcon24;
+ public uint SetFalcon25;
+ public uint SetFalcon26;
+ public uint SetFalcon27;
+ public uint SetFalcon28;
+ public uint SetFalcon29;
+ public uint SetFalcon30;
+ public uint SetFalcon31;
+ public UniformBufferState UniformBufferState;
+ public Array16<uint> UniformBufferUpdateData;
+ public fixed uint Reserved23D0[16];
+ public uint UniformBufferBindVertex;
+ public fixed uint Reserved2414[7];
+ public uint UniformBufferBindTessControl;
+ public fixed uint Reserved2434[7];
+ public uint UniformBufferBindTessEvaluation;
+ public fixed uint Reserved2454[7];
+ public uint UniformBufferBindGeometry;
+ public fixed uint Reserved2474[7];
+ public uint UniformBufferBindFragment;
+ public fixed uint Reserved2494[93];
+ public uint TextureBufferIndex;
+ public fixed uint Reserved260C[125];
+ public Array4<Array32<uint>> TfVaryingLocations;
+ public fixed uint Reserved2A00[640];
+ public MmeShadowScratch SetMmeShadowScratch;
+#pragma warning restore CS0649
+ }
+}