feat: FX 支持 USE 指令

Blinue · Jan 7, 2025 · bae66e7 · bae66e7
1 parent e5bbb25
commit bae66e7
Show file tree

Hide file tree

Showing 19 changed files with 129 additions and 92 deletions.
diff --git a/src/Effects/ACNet.hlsl b/src/Effects/ACNet.hlsl
@@ -1,10 +1,9 @@
 // ACNet
 // 移植自 https://github.com/TianZerL/ACNetGLSL/blob/f20a6b6b7327f4caf588b06c6b21f18e40dae1ce/glsl/ACNet.glsl
 
-
 //!MAGPIE EFFECT
 //!VERSION 4
-//!USE_FP16
+//!USE FP16
 
 #include "StubDefs.hlsli"
 
@@ -40,7 +39,6 @@ Texture2D tex3;
 //!FORMAT R16G16B16A16_FLOAT
 Texture2D tex4;
 
-
 //!SAMPLER
 //!FILTER POINT
 SamplerState sam;
@@ -49,9 +47,12 @@ SamplerState sam;
 //!FILTER LINEAR
 SamplerState sam1;
 
-
 //!COMMON
 
+#ifdef MP_DEBUG
+#pragma warning(disable: 4714)	// X4714: sum of temp registers and indexable temp registers times 256 threads exceeds the recommended total 16384.  Performance may be reduced
+#endif
+
 #define RELU(x) max(x, 0)
 
 MF4 MulAdd(MF3 x, MF3x4 y, MF4 a) {
@@ -2542,15 +2543,15 @@ void Pass9(uint2 blockStart, uint3 threadId) {
 			uint2 destPos = gxy + uint2(i, j);
 
 			uint index = j * 2 + i;
-			MF luma = clamp(
+			MF luma = saturate(
 				target1.x * kernelsL10[0 + index] +
 				target1.y * kernelsL10[4 + index] +
 				target1.z * kernelsL10[8 + index] +
 				target1.w * kernelsL10[12 + index] +
 				target2.x * kernelsL10[16 + index] +
 				target2.y * kernelsL10[20 + index] +
 				target2.z * kernelsL10[24 + index] +
-				target2.w * kernelsL10[28 + index], 0.0, 1.0);
+				target2.w * kernelsL10[28 + index]);
 
 			MF2 originUV = mul(rgb2uv, INPUT.SampleLevel(sam1, (destPos + 0.5f) * outputPt, 0).rgb);
 			OUTPUT[destPos] = MF4(mul(yuv2rgb, MF3(luma, originUV)), 1);

diff --git a/src/Effects/Anime4K/Anime4K_Restore_L.hlsl b/src/Effects/Anime4K/Anime4K_Restore_L.hlsl
@@ -3,7 +3,7 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME Anime4K_Restore_2
-//!USE_FP16
+//!USE FP16
 
 #include "..\StubDefs.hlsli"
 

diff --git a/src/Effects/Anime4K/Anime4K_Restore_Soft_UL.hlsl b/src/Effects/Anime4K/Anime4K_Restore_Soft_UL.hlsl
@@ -4,7 +4,7 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME Anime4K_Restore_Soft_4
-//!USE_FP16
+//!USE FP16
 
 #include "..\StubDefs.hlsli"
 

diff --git a/src/Effects/Anime4K/Anime4K_Restore_UL.hlsl b/src/Effects/Anime4K/Anime4K_Restore_UL.hlsl
@@ -4,7 +4,7 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME Anime4K_Restore_4
-//!USE_FP16
+//!USE FP16
 
 #include "..\StubDefs.hlsli"
 

diff --git a/src/Effects/Anime4K/Anime4K_Upscale_Denoise_UL.hlsl b/src/Effects/Anime4K/Anime4K_Upscale_Denoise_UL.hlsl
@@ -4,7 +4,7 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME Anime4K_Upscale_Denoise_3
-//!USE_FP16
+//!USE FP16
 
 #include "..\StubDefs.hlsli"
 

diff --git a/src/Effects/Anime4K/Anime4K_Upscale_L.hlsl b/src/Effects/Anime4K/Anime4K_Upscale_L.hlsl
@@ -4,11 +4,10 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME Anime4K_Upscale_1
-//!USE_FP16
+//!USE FP16, MulAdd
 
 #include "..\StubDefs.hlsli"
 
-
 //!TEXTURE
 Texture2D INPUT;
 

diff --git a/src/Effects/Anime4K/Anime4K_Upscale_VL.hlsl b/src/Effects/Anime4K/Anime4K_Upscale_VL.hlsl
@@ -4,7 +4,7 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME Anime4K_Upscale_2
-//!USE_FP16
+//!USE FP16
 
 #include "..\StubDefs.hlsli"
 

diff --git a/src/Effects/CRT/CRT_Geom.hlsl b/src/Effects/CRT/CRT_Geom.hlsl
@@ -25,7 +25,7 @@
 
 //!MAGPIE EFFECT
 //!VERSION 4
-//!USE_DYNAMIC
+//!USE Dynamic
 
 
 //!PARAMETER

diff --git a/src/Effects/CuNNy/CuNNy-16x16C-NVL.hlsl b/src/Effects/CuNNy/CuNNy-16x16C-NVL.hlsl
@@ -17,7 +17,7 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME CuNNy-D16N16
-//!USE_FP16
+//!USE FP16
 
 #include "..\StubDefs.hlsli"
 

diff --git a/src/Effects/CuNNy/CuNNy-6x8C-NVL.hlsl b/src/Effects/CuNNy/CuNNy-6x8C-NVL.hlsl
@@ -17,7 +17,7 @@
 //!MAGPIE EFFECT
 //!VERSION 4
 //!SORT_NAME CuNNy-D08N06
-//!USE_FP16
+//!USE FP16
 
 
 //!TEXTURE

diff --git a/src/Effects/FSRCNNX/FSRCNNX.hlsl b/src/Effects/FSRCNNX/FSRCNNX.hlsl
@@ -4,7 +4,7 @@
 
 //!MAGPIE EFFECT
 //!VERSION 4
-//!USE_FP16
+//!USE FP16
 
 
 //!TEXTURE

diff --git a/src/Magpie.Core/DirectXHelper.cpp b/src/Magpie.Core/DirectXHelper.cpp
@@ -26,7 +26,7 @@ bool DirectXHelper::CompileComputeShader(
 	flags |= D3DCOMPILE_SKIP_OPTIMIZATION | D3DCOMPILE_DEBUG;
 #else
 	flags |= D3DCOMPILE_OPTIMIZATION_LEVEL3;
-#endif // _DEBUG
+#endif
 
 	std::unique_ptr<D3D_SHADER_MACRO[]> mc(new D3D_SHADER_MACRO[macros.size() + 1]);
 	for (UINT i = 0; i < macros.size(); ++i) {

diff --git a/src/Magpie.Core/EffectCacheManager.cpp b/src/Magpie.Core/EffectCacheManager.cpp
@@ -64,7 +64,7 @@ void serialize(Archive& ar, EffectSamplerDesc& o) {
 
 template<typename Archive>
 void serialize(Archive& ar, EffectPassDesc& o) {
-	ar& o.cso& o.inputs& o.outputs& o.numThreads[0] & o.numThreads[1] & o.numThreads[2] & o.blockSize& o.desc& o.isPSStyle;
+	ar& o.cso& o.inputs& o.outputs& o.numThreads[0] & o.numThreads[1] & o.numThreads[2] & o.blockSize& o.desc& o.flags;
 }
 
 template<typename Archive>
@@ -76,7 +76,7 @@ static constexpr uint32_t MAX_CACHE_COUNT = 127;
 
 // 缓存版本
 // 当缓存文件结构有更改时更新它，使旧缓存失效
-static constexpr uint32_t EFFECT_CACHE_VERSION = 14;
+static constexpr uint32_t EFFECT_CACHE_VERSION = 15;
 
 
 static std::wstring GetLinearEffectName(std::wstring_view effectName) {