From 2ec7fcecb7d1f0bc8f943a3f7cb4d2e215bc4e76 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Fri, 19 Nov 2021 03:17:02 +0100
Subject: Vulkan: implement D24S8 <-> RGBA8 convertions.

---
 src/video_core/host_shaders/CMakeLists.txt          |  2 ++
 .../host_shaders/convert_abgr8_to_d24s8.frag        | 17 +++++++++++++++++
 .../host_shaders/convert_d24s8_to_abgr8.frag        | 21 +++++++++++++++++++++
 3 files changed, 40 insertions(+)
 create mode 100644 src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
 create mode 100644 src/video_core/host_shaders/convert_d24s8_to_abgr8.frag

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index d779a967a..fd3e41434 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -10,6 +10,8 @@ set(SHADER_FILES
     astc_decoder.comp
     block_linear_unswizzle_2d.comp
     block_linear_unswizzle_3d.comp
+    convert_abgr8_to_d24s8.frag
+    convert_d24s8_to_abgr8.frag
     convert_depth_to_float.frag
     convert_float_to_depth.frag
     full_screen_triangle.vert
diff --git a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
new file mode 100644
index 000000000..f7657e50a
--- /dev/null
+++ b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
@@ -0,0 +1,17 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450
+// #extension GL_ARB_shader_stencil_export : require
+
+layout(binding = 0) uniform sampler2D color_texture;
+
+void main() {
+    ivec2 coord = ivec2(gl_FragCoord.xy);
+    uvec4 color = uvec4(texelFetch(color_texture, coord, 0).rgba * (exp2(8) - 1.0f));
+    uint depth_unorm = (color.r << 16) | (color.g << 8) | color.b;
+
+    gl_FragDepth = float(depth_unorm) / (exp2(24.0) - 1.0f);
+    // gl_FragStencilRefARB = int(color.a);
+}
diff --git a/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
new file mode 100644
index 000000000..ff3bf8209
--- /dev/null
+++ b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
@@ -0,0 +1,21 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450
+
+layout(binding = 0) uniform sampler2D depth_tex;
+layout(binding = 1) uniform isampler2D stencil_tex;
+
+layout(location = 0) out vec4 color;
+
+void main() {
+    ivec2 coord = ivec2(gl_FragCoord.xy);
+    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
+    uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
+
+    color.r = float(depth >> 16) / (exp2(8) - 1.0);
+    color.g = float((depth >> 8) & 0x00FF) / (exp2(8) - 1.0);
+    color.b = float(depth & 0x00FF) / (exp2(8) - 1.0);
+    color.a = float(stencil) / (exp2(8) - 1.0);
+}
-- 
cgit v1.2.3


From b805c7bf058c6da04620cf75880509bdf6d5986c Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Fri, 19 Nov 2021 06:27:44 +0100
Subject: TextureCache: Implement additional D24S8 convertions.

---
 src/video_core/host_shaders/CMakeLists.txt          |  2 ++
 .../host_shaders/convert_d24s8_to_b10g11r11.frag    | 21 +++++++++++++++++++++
 .../host_shaders/convert_d24s8_to_r16g16.frag       | 21 +++++++++++++++++++++
 3 files changed, 44 insertions(+)
 create mode 100644 src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
 create mode 100644 src/video_core/host_shaders/convert_d24s8_to_r16g16.frag

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index fd3e41434..87042195a 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -12,6 +12,8 @@ set(SHADER_FILES
     block_linear_unswizzle_3d.comp
     convert_abgr8_to_d24s8.frag
     convert_d24s8_to_abgr8.frag
+    convert_d24s8_to_b10g11r11.frag
+    convert_d24s8_to_r16g16.frag
     convert_depth_to_float.frag
     convert_float_to_depth.frag
     full_screen_triangle.vert
diff --git a/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag b/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
new file mode 100644
index 000000000..c743d3a13
--- /dev/null
+++ b/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
@@ -0,0 +1,21 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450
+
+layout(binding = 0) uniform sampler2D depth_tex;
+layout(binding = 1) uniform isampler2D stencil_tex;
+
+layout(location = 0) out vec4 color;
+
+void main() {
+    ivec2 coord = ivec2(gl_FragCoord.xy);
+    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
+    uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
+
+    color.b = float(depth >> 22) / (exp2(10) - 1.0);
+    color.g = float((depth >> 11) & 0x00FF) / (exp2(11) - 1.0);
+    color.r = float(depth & 0x00FF) / (exp2(11) - 1.0);
+    color.a = 1.0f;
+}
diff --git a/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag b/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
new file mode 100644
index 000000000..2a9443d3d
--- /dev/null
+++ b/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
@@ -0,0 +1,21 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450
+
+layout(binding = 0) uniform sampler2D depth_tex;
+layout(binding = 1) uniform isampler2D stencil_tex;
+
+layout(location = 0) out vec4 color;
+
+void main() {
+    ivec2 coord = ivec2(gl_FragCoord.xy);
+    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
+    uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
+
+    color.r = float(depth >> 16) / (exp2(16) - 1.0);
+    color.g = float((depth >> 16) & 0x00FF) / (exp2(16) - 1.0);
+    color.b = 0.0f;
+    color.a = 1.0f;
+}
-- 
cgit v1.2.3


From 1d5e6a51d7f66cf089d541a009c84c373fd5c6ab Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Fri, 19 Nov 2021 23:22:44 +0100
Subject: TextureCache: Add B10G11R11 to D24S8 converter.

---
 src/video_core/host_shaders/CMakeLists.txt            |  1 +
 .../host_shaders/convert_b10g11r11_to_d24s8.frag      | 19 +++++++++++++++++++
 2 files changed, 20 insertions(+)
 create mode 100644 src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index 87042195a..a2e046f12 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -11,6 +11,7 @@ set(SHADER_FILES
     block_linear_unswizzle_2d.comp
     block_linear_unswizzle_3d.comp
     convert_abgr8_to_d24s8.frag
+    convert_b10g11r11_to_d24s8.frag
     convert_d24s8_to_abgr8.frag
     convert_d24s8_to_b10g11r11.frag
     convert_d24s8_to_r16g16.frag
diff --git a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
new file mode 100644
index 000000000..b7358c15c
--- /dev/null
+++ b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
@@ -0,0 +1,19 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450
+// #extension GL_ARB_shader_stencil_export : require
+
+layout(binding = 0) uniform sampler2D color_texture;
+
+void main() {
+    ivec2 coord = ivec2(gl_FragCoord.xy);
+    vec4 color = texelFetch(color_texture, coord, 0).rgba;
+    uint depth_stencil_unorm = (uint(color.b * (exp2(10) - 1.0f)) << 22)
+                      | (uint(color.g * (exp2(11) - 1.0f)) << 11)
+                      | (uint(color.r * (exp2(11) - 1.0f)));
+
+    gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
+    // gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
+}
-- 
cgit v1.2.3


From e02cff2f69f9a90777f87f85f290f83fc04c16ec Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sat, 20 Nov 2021 00:02:12 +0100
Subject: TextureCache: Add R16G16 to D24S8 converter.

---
 src/video_core/host_shaders/CMakeLists.txt             |  1 +
 .../host_shaders/convert_r16g16_to_d24s8.frag          | 18 ++++++++++++++++++
 2 files changed, 19 insertions(+)
 create mode 100644 src/video_core/host_shaders/convert_r16g16_to_d24s8.frag

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index a2e046f12..1c91999d7 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -17,6 +17,7 @@ set(SHADER_FILES
     convert_d24s8_to_r16g16.frag
     convert_depth_to_float.frag
     convert_float_to_depth.frag
+    convert_r16g16_to_d24s8.frag
     full_screen_triangle.vert
     fxaa.frag
     fxaa.vert
diff --git a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
new file mode 100644
index 000000000..7b1b914f6
--- /dev/null
+++ b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
@@ -0,0 +1,18 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450
+// #extension GL_ARB_shader_stencil_export : require
+
+layout(binding = 0) uniform sampler2D color_texture;
+
+void main() {
+    ivec2 coord = ivec2(gl_FragCoord.xy);
+    vec4 color = texelFetch(color_texture, coord, 0).rgba;
+    uint depth_stencil_unorm = (uint(color.r * (exp2(16) - 1.0f)) << 16)
+                      | (uint(color.g * (exp2(16) - 1.0f)) << 16);
+
+    gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
+    // gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
+}
-- 
cgit v1.2.3


From 4ca6e9a9e23b96d259cc1f1ba50b9464a8fa12e8 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sat, 20 Nov 2021 06:17:01 +0100
Subject: TextureCache: Assure full conversions on depth/stencil write shaders.

---
 src/video_core/host_shaders/convert_abgr8_to_d24s8.frag     | 4 ++--
 src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag | 4 ++--
 src/video_core/host_shaders/convert_r16g16_to_d24s8.frag    | 4 ++--
 3 files changed, 6 insertions(+), 6 deletions(-)

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
index f7657e50a..4e4ab6a26 100644
--- a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
+++ b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
@@ -3,7 +3,7 @@
 // Refer to the license.txt file included.
 
 #version 450
-// #extension GL_ARB_shader_stencil_export : require
+#extension GL_ARB_shader_stencil_export : require
 
 layout(binding = 0) uniform sampler2D color_texture;
 
@@ -13,5 +13,5 @@ void main() {
     uint depth_unorm = (color.r << 16) | (color.g << 8) | color.b;
 
     gl_FragDepth = float(depth_unorm) / (exp2(24.0) - 1.0f);
-    // gl_FragStencilRefARB = int(color.a);
+    gl_FragStencilRefARB = int(color.a);
 }
diff --git a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
index b7358c15c..2999a84cf 100644
--- a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
+++ b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
@@ -3,7 +3,7 @@
 // Refer to the license.txt file included.
 
 #version 450
-// #extension GL_ARB_shader_stencil_export : require
+#extension GL_ARB_shader_stencil_export : require
 
 layout(binding = 0) uniform sampler2D color_texture;
 
@@ -15,5 +15,5 @@ void main() {
                       | (uint(color.r * (exp2(11) - 1.0f)));
 
     gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
-    // gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
+    gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
 }
diff --git a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
index 7b1b914f6..3df70575e 100644
--- a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
+++ b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
@@ -3,7 +3,7 @@
 // Refer to the license.txt file included.
 
 #version 450
-// #extension GL_ARB_shader_stencil_export : require
+#extension GL_ARB_shader_stencil_export : require
 
 layout(binding = 0) uniform sampler2D color_texture;
 
@@ -14,5 +14,5 @@ void main() {
                       | (uint(color.g * (exp2(16) - 1.0f)) << 16);
 
     gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
-    // gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
+    gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
 }
-- 
cgit v1.2.3


From b96caf200d047b81554c3839c7a6a7c35b251944 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 21 Nov 2021 20:52:39 +0100
Subject: HostShaders: Fix D24S8 convertion shaders.

---
 .../host_shaders/convert_abgr8_to_d24s8.frag          |  7 ++++---
 .../host_shaders/convert_b10g11r11_to_d24s8.frag      | 18 +++++++++++++-----
 .../host_shaders/convert_d24s8_to_abgr8.frag          | 10 ++++++----
 .../host_shaders/convert_d24s8_to_b10g11r11.frag      | 19 +++++++++++++++----
 .../host_shaders/convert_d24s8_to_r16g16.frag         |  7 ++++---
 .../host_shaders/convert_r16g16_to_d24s8.frag         |  9 +++++----
 6 files changed, 47 insertions(+), 23 deletions(-)

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
index 4e4ab6a26..d51397a0c 100644
--- a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
+++ b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
@@ -10,8 +10,9 @@ layout(binding = 0) uniform sampler2D color_texture;
 void main() {
     ivec2 coord = ivec2(gl_FragCoord.xy);
     uvec4 color = uvec4(texelFetch(color_texture, coord, 0).rgba * (exp2(8) - 1.0f));
-    uint depth_unorm = (color.r << 16) | (color.g << 8) | color.b;
+    uvec4 bytes = color << uvec4(24, 16, 8, 0);
+    uint depth_stencil_unorm = bytes.x | bytes.y | bytes.z | bytes.w;
 
-    gl_FragDepth = float(depth_unorm) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(color.a);
+    gl_FragDepth = float(depth_stencil_unorm & 0x00FFFFFFu) / (exp2(24.0) - 1.0f);
+    gl_FragStencilRefARB = int(depth_stencil_unorm >> 24);
 }
diff --git a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
index 2999a84cf..11bdd861d 100644
--- a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
+++ b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
@@ -7,13 +7,21 @@
 
 layout(binding = 0) uniform sampler2D color_texture;
 
+uint conv_from_float(float value_f, uint mantissa_bits) {
+    uint value = floatBitsToInt(value_f);
+    uint exp = (value >> 23) & 0x1Fu;
+    uint mantissa_shift = 32u - mantissa_bits;
+    uint mantissa = (value << 9u) >> mantissa_shift;
+    return (exp << mantissa_bits) | mantissa;
+}
+
 void main() {
     ivec2 coord = ivec2(gl_FragCoord.xy);
     vec4 color = texelFetch(color_texture, coord, 0).rgba;
-    uint depth_stencil_unorm = (uint(color.b * (exp2(10) - 1.0f)) << 22)
-                      | (uint(color.g * (exp2(11) - 1.0f)) << 11)
-                      | (uint(color.r * (exp2(11) - 1.0f)));
+    uint depth_stencil_unorm = (conv_from_float(color.r, 6u) << 21)
+                      | (conv_from_float(color.g, 6u) << 10)
+                      | conv_from_float(color.b, 5u);
 
-    gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
+    gl_FragDepth = float(depth_stencil_unorm & 0x00FFFFFFu) / (exp2(24.0) - 1.0f);
+    gl_FragStencilRefARB = int(depth_stencil_unorm >> 24);
 }
diff --git a/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
index ff3bf8209..47f9c1abc 100644
--- a/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
+++ b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
@@ -14,8 +14,10 @@ void main() {
     uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
     uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
 
-    color.r = float(depth >> 16) / (exp2(8) - 1.0);
-    color.g = float((depth >> 8) & 0x00FF) / (exp2(8) - 1.0);
-    color.b = float(depth & 0x00FF) / (exp2(8) - 1.0);
-    color.a = float(stencil) / (exp2(8) - 1.0);
+    highp uint depth_val =
+        uint(textureLod(depth_tex, coord, 0).r * (exp2(32.0) - 1.0));
+    lowp uint stencil_val = textureLod(stencil_tex, coord, 0).r;
+    highp uvec4 components =
+        uvec4(stencil_val, (uvec3(depth_val) >> uvec3(24u, 16u, 8u)) & 0x000000FFu);
+    color = vec4(components) / (exp2(8.0) - 1.0);
 }
diff --git a/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag b/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
index c743d3a13..c2d935fcd 100644
--- a/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
+++ b/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
@@ -9,13 +9,24 @@ layout(binding = 1) uniform isampler2D stencil_tex;
 
 layout(location = 0) out vec4 color;
 
+float conv_to_float(uint value, uint mantissa_bits) {
+    uint exp = (value >> mantissa_bits) & 0x1Fu;
+    uint mantissa_shift = 32u - mantissa_bits;
+    uint mantissa = (value << mantissa_shift) >> mantissa_shift;
+    return uintBitsToFloat((exp << 23) | (mantissa << (23 - mantissa_bits)));
+}
+
 void main() {
     ivec2 coord = ivec2(gl_FragCoord.xy);
-    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
+    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(32.0) - 1.0f));
     uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
+    uint depth_stencil = (stencil << 24) | (depth >> 8);
+    uint red_int = (depth_stencil >> 21) & 0x07FF;
+    uint green_int = (depth_stencil >> 10) & 0x07FF;
+    uint blue_int = depth_stencil & 0x03FF;
 
-    color.b = float(depth >> 22) / (exp2(10) - 1.0);
-    color.g = float((depth >> 11) & 0x00FF) / (exp2(11) - 1.0);
-    color.r = float(depth & 0x00FF) / (exp2(11) - 1.0);
+    color.r = conv_to_float(red_int, 6u);
+    color.g = conv_to_float(green_int, 6u);
+    color.b = conv_to_float(blue_int, 5u);
     color.a = 1.0f;
 }
diff --git a/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag b/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
index 2a9443d3d..c48a7ac66 100644
--- a/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
+++ b/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
@@ -11,11 +11,12 @@ layout(location = 0) out vec4 color;
 
 void main() {
     ivec2 coord = ivec2(gl_FragCoord.xy);
-    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(24.0) - 1.0f));
+    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(32.0) - 1.0f));
     uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
+    uint depth_stencil = (stencil << 24) | (depth >> 8);
 
-    color.r = float(depth >> 16) / (exp2(16) - 1.0);
-    color.g = float((depth >> 16) & 0x00FF) / (exp2(16) - 1.0);
+    color.r = float(depth_stencil & 0x0000FFFFu) / (exp2(16) - 1.0);
+    color.g = float(depth_stencil >> 16) / (exp2(16) - 1.0);
     color.b = 0.0f;
     color.a = 1.0f;
 }
diff --git a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
index 3df70575e..beb2d1284 100644
--- a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
+++ b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
@@ -10,9 +10,10 @@ layout(binding = 0) uniform sampler2D color_texture;
 void main() {
     ivec2 coord = ivec2(gl_FragCoord.xy);
     vec4 color = texelFetch(color_texture, coord, 0).rgba;
-    uint depth_stencil_unorm = (uint(color.r * (exp2(16) - 1.0f)) << 16)
-                      | (uint(color.g * (exp2(16) - 1.0f)) << 16);
+    uvec2 bytes = uvec2(color.rg * (exp2(16) - 1.0f)) << uvec2(0, 16);
+    uint depth_stencil_unorm =
+        uint(color.r * (exp2(16) - 1.0f)) | (uint(color.g * (exp2(16) - 1.0f)) << 16);
 
-    gl_FragDepth = float(depth_stencil_unorm >> 8) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(depth_stencil_unorm & 0x00FF);
+    gl_FragDepth = float(depth_stencil_unorm & 0x00FFFFFFu) / (exp2(24.0) - 1.0f);
+    gl_FragStencilRefARB = int(depth_stencil_unorm >> 24);
 }
-- 
cgit v1.2.3


From 853284943901560081f6ff992b6c04b7c33f0d21 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Mon, 22 Nov 2021 00:00:01 +0100
Subject: TextureCache: Simplify blitting of D24S8 formats and fix bugs.

---
 src/video_core/host_shaders/CMakeLists.txt         |  4 ---
 .../host_shaders/convert_b10g11r11_to_d24s8.frag   | 27 ------------------
 .../host_shaders/convert_d24s8_to_b10g11r11.frag   | 32 ----------------------
 .../host_shaders/convert_d24s8_to_r16g16.frag      | 22 ---------------
 .../host_shaders/convert_r16g16_to_d24s8.frag      | 19 -------------
 5 files changed, 104 deletions(-)
 delete mode 100644 src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
 delete mode 100644 src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
 delete mode 100644 src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
 delete mode 100644 src/video_core/host_shaders/convert_r16g16_to_d24s8.frag

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/CMakeLists.txt b/src/video_core/host_shaders/CMakeLists.txt
index 1c91999d7..fd3e41434 100644
--- a/src/video_core/host_shaders/CMakeLists.txt
+++ b/src/video_core/host_shaders/CMakeLists.txt
@@ -11,13 +11,9 @@ set(SHADER_FILES
     block_linear_unswizzle_2d.comp
     block_linear_unswizzle_3d.comp
     convert_abgr8_to_d24s8.frag
-    convert_b10g11r11_to_d24s8.frag
     convert_d24s8_to_abgr8.frag
-    convert_d24s8_to_b10g11r11.frag
-    convert_d24s8_to_r16g16.frag
     convert_depth_to_float.frag
     convert_float_to_depth.frag
-    convert_r16g16_to_d24s8.frag
     full_screen_triangle.vert
     fxaa.frag
     fxaa.vert
diff --git a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag b/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
deleted file mode 100644
index 11bdd861d..000000000
--- a/src/video_core/host_shaders/convert_b10g11r11_to_d24s8.frag
+++ /dev/null
@@ -1,27 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-#extension GL_ARB_shader_stencil_export : require
-
-layout(binding = 0) uniform sampler2D color_texture;
-
-uint conv_from_float(float value_f, uint mantissa_bits) {
-    uint value = floatBitsToInt(value_f);
-    uint exp = (value >> 23) & 0x1Fu;
-    uint mantissa_shift = 32u - mantissa_bits;
-    uint mantissa = (value << 9u) >> mantissa_shift;
-    return (exp << mantissa_bits) | mantissa;
-}
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    vec4 color = texelFetch(color_texture, coord, 0).rgba;
-    uint depth_stencil_unorm = (conv_from_float(color.r, 6u) << 21)
-                      | (conv_from_float(color.g, 6u) << 10)
-                      | conv_from_float(color.b, 5u);
-
-    gl_FragDepth = float(depth_stencil_unorm & 0x00FFFFFFu) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(depth_stencil_unorm >> 24);
-}
diff --git a/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag b/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
deleted file mode 100644
index c2d935fcd..000000000
--- a/src/video_core/host_shaders/convert_d24s8_to_b10g11r11.frag
+++ /dev/null
@@ -1,32 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-
-layout(binding = 0) uniform sampler2D depth_tex;
-layout(binding = 1) uniform isampler2D stencil_tex;
-
-layout(location = 0) out vec4 color;
-
-float conv_to_float(uint value, uint mantissa_bits) {
-    uint exp = (value >> mantissa_bits) & 0x1Fu;
-    uint mantissa_shift = 32u - mantissa_bits;
-    uint mantissa = (value << mantissa_shift) >> mantissa_shift;
-    return uintBitsToFloat((exp << 23) | (mantissa << (23 - mantissa_bits)));
-}
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(32.0) - 1.0f));
-    uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
-    uint depth_stencil = (stencil << 24) | (depth >> 8);
-    uint red_int = (depth_stencil >> 21) & 0x07FF;
-    uint green_int = (depth_stencil >> 10) & 0x07FF;
-    uint blue_int = depth_stencil & 0x03FF;
-
-    color.r = conv_to_float(red_int, 6u);
-    color.g = conv_to_float(green_int, 6u);
-    color.b = conv_to_float(blue_int, 5u);
-    color.a = 1.0f;
-}
diff --git a/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag b/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
deleted file mode 100644
index c48a7ac66..000000000
--- a/src/video_core/host_shaders/convert_d24s8_to_r16g16.frag
+++ /dev/null
@@ -1,22 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-
-layout(binding = 0) uniform sampler2D depth_tex;
-layout(binding = 1) uniform isampler2D stencil_tex;
-
-layout(location = 0) out vec4 color;
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    uint depth = uint(textureLod(depth_tex, coord, 0).r * (exp2(32.0) - 1.0f));
-    uint stencil = uint(textureLod(stencil_tex, coord, 0).r);
-    uint depth_stencil = (stencil << 24) | (depth >> 8);
-
-    color.r = float(depth_stencil & 0x0000FFFFu) / (exp2(16) - 1.0);
-    color.g = float(depth_stencil >> 16) / (exp2(16) - 1.0);
-    color.b = 0.0f;
-    color.a = 1.0f;
-}
diff --git a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag b/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
deleted file mode 100644
index beb2d1284..000000000
--- a/src/video_core/host_shaders/convert_r16g16_to_d24s8.frag
+++ /dev/null
@@ -1,19 +0,0 @@
-// Copyright 2021 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#version 450
-#extension GL_ARB_shader_stencil_export : require
-
-layout(binding = 0) uniform sampler2D color_texture;
-
-void main() {
-    ivec2 coord = ivec2(gl_FragCoord.xy);
-    vec4 color = texelFetch(color_texture, coord, 0).rgba;
-    uvec2 bytes = uvec2(color.rg * (exp2(16) - 1.0f)) << uvec2(0, 16);
-    uint depth_stencil_unorm =
-        uint(color.r * (exp2(16) - 1.0f)) | (uint(color.g * (exp2(16) - 1.0f)) << 16);
-
-    gl_FragDepth = float(depth_stencil_unorm & 0x00FFFFFFu) / (exp2(24.0) - 1.0f);
-    gl_FragStencilRefARB = int(depth_stencil_unorm >> 24);
-}
-- 
cgit v1.2.3


From 1e474fb9d1c5e35c2bb8822d58556ef3358e66e9 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Mon, 22 Nov 2021 00:21:42 +0100
Subject: Texture Cache: Correct conversion shaders.

---
 src/video_core/host_shaders/convert_abgr8_to_d24s8.frag | 2 +-
 src/video_core/host_shaders/convert_d24s8_to_abgr8.frag | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
index d51397a0c..ea055ddad 100644
--- a/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
+++ b/src/video_core/host_shaders/convert_abgr8_to_d24s8.frag
@@ -9,7 +9,7 @@ layout(binding = 0) uniform sampler2D color_texture;
 
 void main() {
     ivec2 coord = ivec2(gl_FragCoord.xy);
-    uvec4 color = uvec4(texelFetch(color_texture, coord, 0).rgba * (exp2(8) - 1.0f));
+    uvec4 color = uvec4(texelFetch(color_texture, coord, 0).abgr * (exp2(8) - 1.0f));
     uvec4 bytes = color << uvec4(24, 16, 8, 0);
     uint depth_stencil_unorm = bytes.x | bytes.y | bytes.z | bytes.w;
 
diff --git a/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
index 47f9c1abc..94368fb59 100644
--- a/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
+++ b/src/video_core/host_shaders/convert_d24s8_to_abgr8.frag
@@ -19,5 +19,5 @@ void main() {
     lowp uint stencil_val = textureLod(stencil_tex, coord, 0).r;
     highp uvec4 components =
         uvec4(stencil_val, (uvec3(depth_val) >> uvec3(24u, 16u, 8u)) & 0x000000FFu);
-    color = vec4(components) / (exp2(8.0) - 1.0);
+    color.abgr = vec4(components) / (exp2(8.0) - 1.0);
 }
-- 
cgit v1.2.3