diff options
| author | gdkchan <gab.dark.100@gmail.com> | 2019-04-17 20:57:08 -0300 |
|---|---|---|
| committer | jduncanator <1518948+jduncanator@users.noreply.github.com> | 2019-04-18 09:57:08 +1000 |
| commit | 6b23a2c125b9c48b5ebea92716004ef68698bb0f (patch) | |
| tree | 69332df6fbbd8e2bddc522ba682fcc5c7a69e101 /Ryujinx.Graphics/Shader/Translation/Optimizations | |
| parent | b2e88b04a85b41cc60af3485d88c90429e84a218 (diff) | |
New shader translator implementation (#654)
* Start implementing a new shader translator
* Fix shift instructions and a typo
* Small refactoring on StructuredProgram, move RemovePhis method to a separate class
* Initial geometry shader support
* Implement TLD4
* Fix -- There's no negation on FMUL32I
* Add constant folding and algebraic simplification optimizations, nits
* Some leftovers from constant folding
* Avoid cast for constant assignments
* Add a branch elimination pass, and misc small fixes
* Remove redundant branches, add expression propagation and other improvements on the code
* Small leftovers -- add missing break and continue, remove unused properties, other improvements
* Add null check to handle empty block cases on block visitor
* Add HADD2 and HMUL2 half float shader instructions
* Optimize pack/unpack sequences, some fixes related to half float instructions
* Add TXQ, TLD, TLDS and TLD4S shader texture instructions, and some support for bindless textures, some refactoring on codegen
* Fix copy paste mistake that caused RZ to be ignored on the AST instruction
* Add workaround for conditional exit, and fix half float instruction with constant buffer
* Add missing 0.0 source for TLDS.LZ variants
* Simplify the switch for TLDS.LZ
* Texture instructions related fixes
* Implement the HFMA instruction, and some misc. fixes
* Enable constant folding on UnpackHalf2x16 instructions
* Refactor HFMA to use OpCode* for opcode decoding rather than on the helper methods
* Remove the old shader translator
* Remove ShaderDeclInfo and other unused things
* Add dual vertex shader support
* Add ShaderConfig, used to pass shader type and maximum cbuffer size
* Move and rename some instruction enums
* Move texture instructions into a separate file
* Move operand GetExpression and locals management to OperandManager
* Optimize opcode decoding using a simple list and binary search
* Add missing condition for do-while on goto elimination
* Misc. fixes on texture instructions
* Simplify TLDS switch
* Address PR feedback, and a nit
Diffstat (limited to 'Ryujinx.Graphics/Shader/Translation/Optimizations')
5 files changed, 753 insertions, 0 deletions
diff --git a/Ryujinx.Graphics/Shader/Translation/Optimizations/BranchElimination.cs b/Ryujinx.Graphics/Shader/Translation/Optimizations/BranchElimination.cs new file mode 100644 index 00000000..2b0f1905 --- /dev/null +++ b/Ryujinx.Graphics/Shader/Translation/Optimizations/BranchElimination.cs @@ -0,0 +1,64 @@ +using Ryujinx.Graphics.Shader.IntermediateRepresentation; +using System; + +namespace Ryujinx.Graphics.Shader.Translation.Optimizations +{ + static class BranchElimination + { + public static bool Eliminate(BasicBlock block) + { + if (block.HasBranch && IsRedundantBranch((Operation)block.GetLastOp(), Next(block))) + { + block.Branch = null; + + return true; + } + + return false; + } + + private static bool IsRedundantBranch(Operation current, BasicBlock nextBlock) + { + //Here we check that: + //- The current block ends with a branch. + //- The next block only contains a branch. + //- The branch on the next block is unconditional. + //- Both branches are jumping to the same location. + //In this case, the branch on the current block can be removed, + //as the next block is going to jump to the same place anyway. + if (nextBlock == null) + { + return false; + } + + if (!(nextBlock.Operations.First?.Value is Operation next)) + { + return false; + } + + if (next.Inst != Instruction.Branch) + { + return false; + } + + return current.Dest == next.Dest; + } + + private static BasicBlock Next(BasicBlock block) + { + block = block.Next; + + while (block != null && block.Operations.Count == 0) + { + if (block.HasBranch) + { + throw new InvalidOperationException("Found a bogus empty block that \"ends with a branch\"."); + } + + block = block.Next; + } + + return block; + } + } +}
\ No newline at end of file diff --git a/Ryujinx.Graphics/Shader/Translation/Optimizations/ConstantFolding.cs b/Ryujinx.Graphics/Shader/Translation/Optimizations/ConstantFolding.cs new file mode 100644 index 00000000..a2e05ef1 --- /dev/null +++ b/Ryujinx.Graphics/Shader/Translation/Optimizations/ConstantFolding.cs @@ -0,0 +1,323 @@ +using Ryujinx.Graphics.Shader.Decoders; +using Ryujinx.Graphics.Shader.IntermediateRepresentation; +using System; + +using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper; + +namespace Ryujinx.Graphics.Shader.Translation.Optimizations +{ + static class ConstantFolding + { + public static void Fold(Operation operation) + { + if (!AreAllSourcesConstant(operation)) + { + return; + } + + switch (operation.Inst) + { + case Instruction.Add: + EvaluateBinary(operation, (x, y) => x + y); + break; + + case Instruction.BitwiseAnd: + EvaluateBinary(operation, (x, y) => x & y); + break; + + case Instruction.BitwiseExclusiveOr: + EvaluateBinary(operation, (x, y) => x ^ y); + break; + + case Instruction.BitwiseNot: + EvaluateUnary(operation, (x) => ~x); + break; + + case Instruction.BitwiseOr: + EvaluateBinary(operation, (x, y) => x | y); + break; + + case Instruction.BitfieldExtractS32: + BitfieldExtractS32(operation); + break; + + case Instruction.BitfieldExtractU32: + BitfieldExtractU32(operation); + break; + + case Instruction.Clamp: + EvaluateTernary(operation, (x, y, z) => Math.Clamp(x, y, z)); + break; + + case Instruction.ClampU32: + EvaluateTernary(operation, (x, y, z) => (int)Math.Clamp((uint)x, (uint)y, (uint)z)); + break; + + case Instruction.CompareEqual: + EvaluateBinary(operation, (x, y) => x == y); + break; + + case Instruction.CompareGreater: + EvaluateBinary(operation, (x, y) => x > y); + break; + + case Instruction.CompareGreaterOrEqual: + EvaluateBinary(operation, (x, y) => x >= y); + break; + + case Instruction.CompareGreaterOrEqualU32: + EvaluateBinary(operation, (x, y) => (uint)x >= (uint)y); + break; + + case Instruction.CompareGreaterU32: + EvaluateBinary(operation, (x, y) => (uint)x > (uint)y); + break; + + case Instruction.CompareLess: + EvaluateBinary(operation, (x, y) => x < y); + break; + + case Instruction.CompareLessOrEqual: + EvaluateBinary(operation, (x, y) => x <= y); + break; + + case Instruction.CompareLessOrEqualU32: + EvaluateBinary(operation, (x, y) => (uint)x <= (uint)y); + break; + + case Instruction.CompareLessU32: + EvaluateBinary(operation, (x, y) => (uint)x < (uint)y); + break; + + case Instruction.CompareNotEqual: + EvaluateBinary(operation, (x, y) => x != y); + break; + + case Instruction.Divide: + EvaluateBinary(operation, (x, y) => y != 0 ? x / y : 0); + break; + + case Instruction.FP | Instruction.Add: + EvaluateFPBinary(operation, (x, y) => x + y); + break; + + case Instruction.FP | Instruction.Clamp: + EvaluateFPTernary(operation, (x, y, z) => Math.Clamp(x, y, z)); + break; + + case Instruction.FP | Instruction.CompareEqual: + EvaluateFPBinary(operation, (x, y) => x == y); + break; + + case Instruction.FP | Instruction.CompareGreater: + EvaluateFPBinary(operation, (x, y) => x > y); + break; + + case Instruction.FP | Instruction.CompareGreaterOrEqual: + EvaluateFPBinary(operation, (x, y) => x >= y); + break; + + case Instruction.FP | Instruction.CompareLess: + EvaluateFPBinary(operation, (x, y) => x < y); + break; + + case Instruction.FP | Instruction.CompareLessOrEqual: + EvaluateFPBinary(operation, (x, y) => x <= y); + break; + + case Instruction.FP | Instruction.CompareNotEqual: + EvaluateFPBinary(operation, (x, y) => x != y); + break; + + case Instruction.FP | Instruction.Divide: + EvaluateFPBinary(operation, (x, y) => x / y); + break; + + case Instruction.FP | Instruction.Multiply: + EvaluateFPBinary(operation, (x, y) => x * y); + break; + + case Instruction.FP | Instruction.Negate: + EvaluateFPUnary(operation, (x) => -x); + break; + + case Instruction.FP | Instruction.Subtract: + EvaluateFPBinary(operation, (x, y) => x - y); + break; + + case Instruction.IsNan: + EvaluateFPUnary(operation, (x) => float.IsNaN(x)); + break; + + case Instruction.Maximum: + EvaluateBinary(operation, (x, y) => Math.Max(x, y)); + break; + + case Instruction.MaximumU32: + EvaluateBinary(operation, (x, y) => (int)Math.Max((uint)x, (uint)y)); + break; + + case Instruction.Minimum: + EvaluateBinary(operation, (x, y) => Math.Min(x, y)); + break; + + case Instruction.MinimumU32: + EvaluateBinary(operation, (x, y) => (int)Math.Min((uint)x, (uint)y)); + break; + + case Instruction.Multiply: + EvaluateBinary(operation, (x, y) => x * y); + break; + + case Instruction.Negate: + EvaluateUnary(operation, (x) => -x); + break; + + case Instruction.ShiftLeft: + EvaluateBinary(operation, (x, y) => x << y); + break; + + case Instruction.ShiftRightS32: + EvaluateBinary(operation, (x, y) => x >> y); + break; + + case Instruction.ShiftRightU32: + EvaluateBinary(operation, (x, y) => (int)((uint)x >> y)); + break; + + case Instruction.Subtract: + EvaluateBinary(operation, (x, y) => x - y); + break; + + case Instruction.UnpackHalf2x16: + UnpackHalf2x16(operation); + break; + } + } + + private static bool AreAllSourcesConstant(Operation operation) + { + for (int index = 0; index < operation.SourcesCount; index++) + { + if (operation.GetSource(index).Type != OperandType.Constant) + { + return false; + } + } + + return true; + } + + private static void BitfieldExtractS32(Operation operation) + { + int value = GetBitfieldExtractValue(operation); + + int shift = 32 - operation.GetSource(2).Value; + + value = (value << shift) >> shift; + + operation.TurnIntoCopy(Const(value)); + } + + private static void BitfieldExtractU32(Operation operation) + { + operation.TurnIntoCopy(Const(GetBitfieldExtractValue(operation))); + } + + private static int GetBitfieldExtractValue(Operation operation) + { + int value = operation.GetSource(0).Value; + int lsb = operation.GetSource(1).Value; + int length = operation.GetSource(2).Value; + + return value.Extract(lsb, length); + } + + private static void UnpackHalf2x16(Operation operation) + { + int value = operation.GetSource(0).Value; + + value = (value >> operation.ComponentIndex * 16) & 0xffff; + + operation.TurnIntoCopy(ConstF(HalfConversion.HalfToSingle(value))); + } + + private static void FPNegate(Operation operation) + { + float value = operation.GetSource(0).AsFloat(); + + operation.TurnIntoCopy(ConstF(-value)); + } + + private static void EvaluateUnary(Operation operation, Func<int, int> op) + { + int x = operation.GetSource(0).Value; + + operation.TurnIntoCopy(Const(op(x))); + } + + private static void EvaluateFPUnary(Operation operation, Func<float, float> op) + { + float x = operation.GetSource(0).AsFloat(); + + operation.TurnIntoCopy(ConstF(op(x))); + } + + private static void EvaluateFPUnary(Operation operation, Func<float, bool> op) + { + float x = operation.GetSource(0).AsFloat(); + + operation.TurnIntoCopy(Const(op(x) ? IrConsts.True : IrConsts.False)); + } + + private static void EvaluateBinary(Operation operation, Func<int, int, int> op) + { + int x = operation.GetSource(0).Value; + int y = operation.GetSource(1).Value; + + operation.TurnIntoCopy(Const(op(x, y))); + } + + private static void EvaluateBinary(Operation operation, Func<int, int, bool> op) + { + int x = operation.GetSource(0).Value; + int y = operation.GetSource(1).Value; + + operation.TurnIntoCopy(Const(op(x, y) ? IrConsts.True : IrConsts.False)); + } + + private static void EvaluateFPBinary(Operation operation, Func<float, float, float> op) + { + float x = operation.GetSource(0).AsFloat(); + float y = operation.GetSource(1).AsFloat(); + + operation.TurnIntoCopy(ConstF(op(x, y))); + } + + private static void EvaluateFPBinary(Operation operation, Func<float, float, bool> op) + { + float x = operation.GetSource(0).AsFloat(); + float y = operation.GetSource(1).AsFloat(); + + operation.TurnIntoCopy(Const(op(x, y) ? IrConsts.True : IrConsts.False)); + } + + private static void EvaluateTernary(Operation operation, Func<int, int, int, int> op) + { + int x = operation.GetSource(0).Value; + int y = operation.GetSource(1).Value; + int z = operation.GetSource(2).Value; + + operation.TurnIntoCopy(Const(op(x, y, z))); + } + + private static void EvaluateFPTernary(Operation operation, Func<float, float, float, float> op) + { + float x = operation.GetSource(0).AsFloat(); + float y = operation.GetSource(1).AsFloat(); + float z = operation.GetSource(2).AsFloat(); + + operation.TurnIntoCopy(ConstF(op(x, y, z))); + } + } +}
\ No newline at end of file diff --git a/Ryujinx.Graphics/Shader/Translation/Optimizations/HalfConversion.cs b/Ryujinx.Graphics/Shader/Translation/Optimizations/HalfConversion.cs new file mode 100644 index 00000000..9ef35abc --- /dev/null +++ b/Ryujinx.Graphics/Shader/Translation/Optimizations/HalfConversion.cs @@ -0,0 +1,47 @@ +using System; + +namespace Ryujinx.Graphics.Shader.Translation.Optimizations +{ + static class HalfConversion + { + public static float HalfToSingle(int value) + { + int mantissa = (value >> 0) & 0x3ff; + int exponent = (value >> 10) & 0x1f; + int sign = (value >> 15) & 0x1; + + if (exponent == 0x1f) + { + //NaN or Infinity. + mantissa <<= 13; + exponent = 0xff; + } + else if (exponent != 0 || mantissa != 0 ) + { + if (exponent == 0) + { + //Denormal. + int e = -1; + int m = mantissa; + + do + { + e++; + m <<= 1; + } + while ((m & 0x400) == 0); + + mantissa = m & 0x3ff; + exponent = e; + } + + mantissa <<= 13; + exponent = 127 - 15 + exponent; + } + + int output = (sign << 31) | (exponent << 23) | mantissa; + + return BitConverter.Int32BitsToSingle(output); + } + } +}
\ No newline at end of file diff --git a/Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs b/Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs new file mode 100644 index 00000000..88118e3a --- /dev/null +++ b/Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs @@ -0,0 +1,172 @@ +using Ryujinx.Graphics.Shader.IntermediateRepresentation; +using System.Collections.Generic; +using System.Linq; + +namespace Ryujinx.Graphics.Shader.Translation.Optimizations +{ + static class Optimizer + { + public static void Optimize(BasicBlock[] blocks) + { + bool modified; + + do + { + modified = false; + + for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++) + { + BasicBlock block = blocks[blkIndex]; + + LinkedListNode<INode> node = block.Operations.First; + + while (node != null) + { + LinkedListNode<INode> nextNode = node.Next; + + bool isUnused = IsUnused(node.Value); + + if (!(node.Value is Operation operation) || isUnused) + { + if (isUnused) + { + RemoveNode(block, node); + + modified = true; + } + + node = nextNode; + + continue; + } + + ConstantFolding.Fold(operation); + + Simplification.Simplify(operation); + + if (DestIsLocalVar(operation)) + { + if (operation.Inst == Instruction.Copy) + { + PropagateCopy(operation); + + RemoveNode(block, node); + + modified = true; + } + else if (operation.Inst == Instruction.PackHalf2x16 && PropagatePack(operation)) + { + if (operation.Dest.UseOps.Count == 0) + { + RemoveNode(block, node); + } + + modified = true; + } + } + + node = nextNode; + } + + if (BranchElimination.Eliminate(block)) + { + RemoveNode(block, block.Operations.Last); + + modified = true; + } + } + } + while (modified); + } + + private static void PropagateCopy(Operation copyOp) + { + //Propagate copy source operand to all uses of + //the destination operand. + Operand dest = copyOp.Dest; + Operand src = copyOp.GetSource(0); + + INode[] uses = dest.UseOps.ToArray(); + + foreach (INode useNode in uses) + { + for (int index = 0; index < useNode.SourcesCount; index++) + { + if (useNode.GetSource(index) == dest) + { + useNode.SetSource(index, src); + } + } + } + } + + private static bool PropagatePack(Operation packOp) + { + //Propagate pack source operands to uses by unpack + //instruction. The source depends on the unpack instruction. + bool modified = false; + + Operand dest = packOp.Dest; + Operand src0 = packOp.GetSource(0); + Operand src1 = packOp.GetSource(1); + + INode[] uses = dest.UseOps.ToArray(); + + foreach (INode useNode in uses) + { + if (!(useNode is Operation operation) || operation.Inst != Instruction.UnpackHalf2x16) + { + continue; + } + + if (operation.GetSource(0) == dest) + { + operation.TurnIntoCopy(operation.ComponentIndex == 1 ? src1 : src0); + + modified = true; + } + } + + return modified; + } + + private static void RemoveNode(BasicBlock block, LinkedListNode<INode> llNode) + { + //Remove a node from the nodes list, and also remove itself + //from all the use lists on the operands that this node uses. + block.Operations.Remove(llNode); + + Queue<INode> nodes = new Queue<INode>(); + + nodes.Enqueue(llNode.Value); + + while (nodes.TryDequeue(out INode node)) + { + for (int index = 0; index < node.SourcesCount; index++) + { + Operand src = node.GetSource(index); + + if (src.Type != OperandType.LocalVariable) + { + continue; + } + + if (src.UseOps.Remove(node) && src.UseOps.Count == 0) + { + nodes.Enqueue(src.AsgOp); + } + } + } + } + + private static bool IsUnused(INode node) + { + return DestIsLocalVar(node) && node.Dest.UseOps.Count == 0; + } + + private static bool DestIsLocalVar(INode node) + { + return node.Dest != null && node.Dest.Type == OperandType.LocalVariable; + } + } +}
\ No newline at end of file diff --git a/Ryujinx.Graphics/Shader/Translation/Optimizations/Simplification.cs b/Ryujinx.Graphics/Shader/Translation/Optimizations/Simplification.cs new file mode 100644 index 00000000..56b1543f --- /dev/null +++ b/Ryujinx.Graphics/Shader/Translation/Optimizations/Simplification.cs @@ -0,0 +1,147 @@ +using Ryujinx.Graphics.Shader.IntermediateRepresentation; + +using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper; + +namespace Ryujinx.Graphics.Shader.Translation.Optimizations +{ + static class Simplification + { + private const int AllOnes = ~0; + + public static void Simplify(Operation operation) + { + switch (operation.Inst) + { + case Instruction.Add: + case Instruction.BitwiseExclusiveOr: + TryEliminateBinaryOpComutative(operation, 0); + break; + + case Instruction.BitwiseAnd: + TryEliminateBitwiseAnd(operation); + break; + + case Instruction.BitwiseOr: + TryEliminateBitwiseOr(operation); + break; + + case Instruction.ConditionalSelect: + TryEliminateConditionalSelect(operation); + break; + + case Instruction.Divide: + TryEliminateBinaryOpY(operation, 1); + break; + + case Instruction.Multiply: + TryEliminateBinaryOpComutative(operation, 1); + break; + + case Instruction.ShiftLeft: + case Instruction.ShiftRightS32: + case Instruction.ShiftRightU32: + case Instruction.Subtract: + TryEliminateBinaryOpY(operation, 0); + break; + } + } + + private static void TryEliminateBitwiseAnd(Operation operation) + { + //Try to recognize and optimize those 3 patterns (in order): + //x & 0xFFFFFFFF == x, 0xFFFFFFFF & y == y, + //x & 0x00000000 == 0x00000000, 0x00000000 & y == 0x00000000 + Operand x = operation.GetSource(0); + Operand y = operation.GetSource(1); + + if (IsConstEqual(x, AllOnes)) + { + operation.TurnIntoCopy(y); + } + else if (IsConstEqual(y, AllOnes)) + { + operation.TurnIntoCopy(x); + } + else if (IsConstEqual(x, 0) || IsConstEqual(y, 0)) + { + operation.TurnIntoCopy(Const(0)); + } + } + + private static void TryEliminateBitwiseOr(Operation operation) + { + //Try to recognize and optimize those 3 patterns (in order): + //x | 0x00000000 == x, 0x00000000 | y == y, + //x | 0xFFFFFFFF == 0xFFFFFFFF, 0xFFFFFFFF | y == 0xFFFFFFFF + Operand x = operation.GetSource(0); + Operand y = operation.GetSource(1); + + if (IsConstEqual(x, 0)) + { + operation.TurnIntoCopy(y); + } + else if (IsConstEqual(y, 0)) + { + operation.TurnIntoCopy(x); + } + else if (IsConstEqual(x, AllOnes) || IsConstEqual(y, AllOnes)) + { + operation.TurnIntoCopy(Const(AllOnes)); + } + } + + private static void TryEliminateBinaryOpY(Operation operation, int comparand) + { + Operand x = operation.GetSource(0); + Operand y = operation.GetSource(1); + + if (IsConstEqual(y, comparand)) + { + operation.TurnIntoCopy(x); + } + } + + private static void TryEliminateBinaryOpComutative(Operation operation, int comparand) + { + Operand x = operation.GetSource(0); + Operand y = operation.GetSource(1); + + if (IsConstEqual(x, comparand)) + { + operation.TurnIntoCopy(y); + } + else if (IsConstEqual(y, comparand)) + { + operation.TurnIntoCopy(x); + } + } + + private static void TryEliminateConditionalSelect(Operation operation) + { + Operand cond = operation.GetSource(0); + + if (cond.Type != OperandType.Constant) + { + return; + } + + //The condition is constant, we can turn it into a copy, and select + //the source based on the condition value. + int srcIndex = cond.Value != 0 ? 1 : 2; + + Operand source = operation.GetSource(srcIndex); + + operation.TurnIntoCopy(source); + } + + private static bool IsConstEqual(Operand operand, int comparand) + { + if (operand.Type != OperandType.Constant) + { + return false; + } + + return operand.Value == comparand; + } + } +}
\ No newline at end of file |
