Merge pull request #6239 from stenzek/d3d-videocommon-efbtotex

D3D: Use VideoCommon EFB-to-Texture Shaders

4 files changed, 42 insertions, 197 deletions

diff --git a/Source/Core/VideoBackends/D3D/PixelShaderCache.cpp b/Source/Core/VideoBackends/D3D/PixelShaderCache.cpp
index 3625cf4e0c..664ecec3a2 100644
--- a/Source/Core/VideoBackends/D3D/PixelShaderCache.cpp
+++ b/Source/Core/VideoBackends/D3D/PixelShaderCache.cpp
@@ -38,9 +38,7 @@ LinearDiskCache<PixelShaderUid, u8> g_ps_disk_cache;
 LinearDiskCache<UberShader::PixelShaderUid, u8> g_uber_ps_disk_cache;
 extern std::unique_ptr<VideoCommon::AsyncShaderCompiler> g_async_compiler;
 
-ID3D11PixelShader* s_ColorMatrixProgram[2] = {nullptr};
 ID3D11PixelShader* s_ColorCopyProgram[2] = {nullptr};
-ID3D11PixelShader* s_DepthMatrixProgram[2] = {nullptr};
 ID3D11PixelShader* s_ClearProgram = nullptr;
 ID3D11PixelShader* s_AnaglyphProgram = nullptr;
 ID3D11PixelShader* s_DepthResolveProgram = nullptr;
@@ -104,106 +102,6 @@ const char color_copy_program_code_msaa[] = {
     "ocol0 /= SAMPLES;\n"
     "}\n"};
 
-const char color_matrix_program_code[] = {"sampler samp0 : register(s0);\n"
-                                          "Texture2DArray Tex0 : register(t0);\n"
-                                          "uniform float4 cColMatrix[7] : register(c0);\n"
-                                          "void main(\n"
-                                          "out float4 ocol0 : SV_Target,\n"
-                                          "in float4 pos : SV_Position,\n"
-                                          "in float3 uv0 : TEXCOORD0){\n"
-                                          "float4 texcol = Tex0.Sample(samp0,uv0);\n"
-                                          "texcol = floor(texcol * cColMatrix[5])*cColMatrix[6];\n"
-                                          "ocol0 = "
-                                          "float4(dot(texcol,cColMatrix[0]),dot(texcol,cColMatrix["
-                                          "1]),dot(texcol,cColMatrix[2]),dot(texcol,cColMatrix[3]))"
-                                          " + cColMatrix[4];\n"
-                                          "}\n"};
-
-const char color_matrix_program_code_msaa[] = {
-    "#define SAMPLES %d\n"
-    "sampler samp0 : register(s0);\n"
-    "Texture2DMSArray<float4, SAMPLES> Tex0 : register(t0);\n"
-    "uniform float4 cColMatrix[7] : register(c0);\n"
-    "void main(\n"
-    "out float4 ocol0 : SV_Target,\n"
-    "in float4 pos : SV_Position,\n"
-    "in float3 uv0 : TEXCOORD0){\n"
-    "int width, height, slices, samples;\n"
-    "Tex0.GetDimensions(width, height, slices, samples);\n"
-    "float4 texcol = 0;\n"
-    "for(int i = 0; i < SAMPLES; ++i)\n"
-    "	texcol += Tex0.Load(int3(uv0.x*(width), uv0.y*(height), uv0.z), i);\n"
-    "texcol /= SAMPLES;\n"
-    "texcol = floor(texcol * cColMatrix[5])*cColMatrix[6];\n"
-    "ocol0 = "
-    "float4(dot(texcol,cColMatrix[0]),dot(texcol,cColMatrix[1]),dot(texcol,cColMatrix[2]),dot("
-    "texcol,cColMatrix[3])) + cColMatrix[4];\n"
-    "}\n"};
-
-const char depth_matrix_program[] = {"sampler samp0 : register(s0);\n"
-                                     "Texture2DArray Tex0 : register(t0);\n"
-                                     "uniform float4 cColMatrix[7] : register(c0);\n"
-                                     "void main(\n"
-                                     "out float4 ocol0 : SV_Target,\n"
-                                     " in float4 pos : SV_Position,\n"
-                                     " in float3 uv0 : TEXCOORD0){\n"
-                                     "	float4 texcol = Tex0.Sample(samp0,uv0);\n"
-                                     "	int depth = int((1.0 - texcol.x) * 16777216.0);\n"
-
-                                     // Convert to Z24 format
-                                     "	int4 workspace;\n"
-                                     "	workspace.r = (depth >> 16) & 255;\n"
-                                     "	workspace.g = (depth >> 8) & 255;\n"
-                                     "	workspace.b = depth & 255;\n"
-
-                                     // Convert to Z4 format
-                                     "	workspace.a = (depth >> 16) & 0xF0;\n"
-
-                                     // Normalize components to [0.0..1.0]
-                                     "	texcol = float4(workspace) / 255.0;\n"
-
-                                     // Apply color matrix
-                                     "	ocol0 = "
-                                     "float4(dot(texcol,cColMatrix[0]),dot(texcol,cColMatrix[1]),"
-                                     "dot(texcol,cColMatrix[2]),dot(texcol,cColMatrix[3])) + "
-                                     "cColMatrix[4];\n"
-                                     "}\n"};
-
-const char depth_matrix_program_msaa[] = {
-    "#define SAMPLES %d\n"
-    "sampler samp0 : register(s0);\n"
-    "Texture2DMSArray<float4, SAMPLES> Tex0 : register(t0);\n"
-    "uniform float4 cColMatrix[7] : register(c0);\n"
-    "void main(\n"
-    "out float4 ocol0 : SV_Target,\n"
-    " in float4 pos : SV_Position,\n"
-    " in float3 uv0 : TEXCOORD0){\n"
-    "	int width, height, slices, samples;\n"
-    "	Tex0.GetDimensions(width, height, slices, samples);\n"
-    "	float4 texcol = 0;\n"
-    "	for(int i = 0; i < SAMPLES; ++i)\n"
-    "		texcol += Tex0.Load(int3(uv0.x*(width), uv0.y*(height), uv0.z), i);\n"
-    "	texcol /= SAMPLES;\n"
-    "	int depth = int((1.0 - texcol.x) * 16777216.0);\n"
-
-    // Convert to Z24 format
-    "	int4 workspace;\n"
-    "	workspace.r = (depth >> 16) & 255;\n"
-    "	workspace.g = (depth >> 8) & 255;\n"
-    "	workspace.b = depth & 255;\n"
-
-    // Convert to Z4 format
-    "	workspace.a = (depth >> 16) & 0xF0;\n"
-
-    // Normalize components to [0.0..1.0]
-    "	texcol = float4(workspace) / 255.0;\n"
-
-    // Apply color matrix
-    "	ocol0 = "
-    "float4(dot(texcol,cColMatrix[0]),dot(texcol,cColMatrix[1]),dot(texcol,cColMatrix[2]),dot("
-    "texcol,cColMatrix[3])) + cColMatrix[4];\n"
-    "}\n"};
-
 const char depth_resolve_program[] = {
     "#define SAMPLES %d\n"
     "Texture2DMSArray<float4, SAMPLES> Tex0 : register(t0);\n"
@@ -368,49 +266,6 @@ ID3D11PixelShader* PixelShaderCache::GetColorCopyProgram(bool multisampled)
   }
 }
 
-ID3D11PixelShader* PixelShaderCache::GetColorMatrixProgram(bool multisampled)
-{
-  if (!multisampled || g_ActiveConfig.iMultisamples <= 1)
-  {
-    return s_ColorMatrixProgram[0];
-  }
-  else if (s_ColorMatrixProgram[1])
-  {
-    return s_ColorMatrixProgram[1];
-  }
-  else
-  {
-    // create MSAA shader for current AA mode
-    std::string buf =
-        StringFromFormat(color_matrix_program_code_msaa, g_ActiveConfig.iMultisamples);
-    s_ColorMatrixProgram[1] = D3D::CompileAndCreatePixelShader(buf);
-    CHECK(s_ColorMatrixProgram[1] != nullptr, "Create color matrix MSAA pixel shader");
-    D3D::SetDebugObjectName(s_ColorMatrixProgram[1], "color matrix MSAA pixel shader");
-    return s_ColorMatrixProgram[1];
-  }
-}
-
-ID3D11PixelShader* PixelShaderCache::GetDepthMatrixProgram(bool multisampled)
-{
-  if (!multisampled || g_ActiveConfig.iMultisamples <= 1)
-  {
-    return s_DepthMatrixProgram[0];
-  }
-  else if (s_DepthMatrixProgram[1])
-  {
-    return s_DepthMatrixProgram[1];
-  }
-  else
-  {
-    // create MSAA shader for current AA mode
-    std::string buf = StringFromFormat(depth_matrix_program_msaa, g_ActiveConfig.iMultisamples);
-    s_DepthMatrixProgram[1] = D3D::CompileAndCreatePixelShader(buf);
-    CHECK(s_DepthMatrixProgram[1] != nullptr, "Create depth matrix MSAA pixel shader");
-    D3D::SetDebugObjectName(s_DepthMatrixProgram[1], "depth matrix MSAA pixel shader");
-    return s_DepthMatrixProgram[1];
-  }
-}
-
 ID3D11PixelShader* PixelShaderCache::GetClearProgram()
 {
   return s_ClearProgram;
@@ -490,16 +345,6 @@ void PixelShaderCache::Init()
   CHECK(s_ColorCopyProgram[0] != nullptr, "Create color copy pixel shader");
   D3D::SetDebugObjectName(s_ColorCopyProgram[0], "color copy pixel shader");
 
-  // used for color conversion
-  s_ColorMatrixProgram[0] = D3D::CompileAndCreatePixelShader(color_matrix_program_code);
-  CHECK(s_ColorMatrixProgram[0] != nullptr, "Create color matrix pixel shader");
-  D3D::SetDebugObjectName(s_ColorMatrixProgram[0], "color matrix pixel shader");
-
-  // used for depth copy
-  s_DepthMatrixProgram[0] = D3D::CompileAndCreatePixelShader(depth_matrix_program);
-  CHECK(s_DepthMatrixProgram[0] != nullptr, "Create depth matrix pixel shader");
-  D3D::SetDebugObjectName(s_DepthMatrixProgram[0], "depth matrix pixel shader");
-
   Clear();
 
   SETSTAT(stats.numPixelShadersCreated, 0);
@@ -557,8 +402,6 @@ void PixelShaderCache::Clear()
 void PixelShaderCache::InvalidateMSAAShaders()
 {
   SAFE_RELEASE(s_ColorCopyProgram[1]);
-  SAFE_RELEASE(s_ColorMatrixProgram[1]);
-  SAFE_RELEASE(s_DepthMatrixProgram[1]);
   SAFE_RELEASE(s_rgb8_to_rgba6[1]);
   SAFE_RELEASE(s_rgba6_to_rgb8[1]);
   SAFE_RELEASE(s_DepthResolveProgram);
@@ -574,8 +417,6 @@ void PixelShaderCache::Shutdown()
   for (int i = 0; i < 2; ++i)
   {
     SAFE_RELEASE(s_ColorCopyProgram[i]);
-    SAFE_RELEASE(s_ColorMatrixProgram[i]);
-    SAFE_RELEASE(s_DepthMatrixProgram[i]);
     SAFE_RELEASE(s_rgba6_to_rgb8[i]);
     SAFE_RELEASE(s_rgb8_to_rgba6[i]);
   }
diff --git a/Source/Core/VideoBackends/D3D/PixelShaderCache.h b/Source/Core/VideoBackends/D3D/PixelShaderCache.h
index 18979cc8ef..9a0faed9e5 100644
--- a/Source/Core/VideoBackends/D3D/PixelShaderCache.h
+++ b/Source/Core/VideoBackends/D3D/PixelShaderCache.h
@@ -32,9 +32,7 @@ public:
 
   static ID3D11Buffer* GetConstantBuffer();
 
-  static ID3D11PixelShader* GetColorMatrixProgram(bool multisampled);
   static ID3D11PixelShader* GetColorCopyProgram(bool multisampled);
-  static ID3D11PixelShader* GetDepthMatrixProgram(bool multisampled);
   static ID3D11PixelShader* GetClearProgram();
   static ID3D11PixelShader* GetAnaglyphProgram();
   static ID3D11PixelShader* GetDepthResolveProgram();
diff --git a/Source/Core/VideoBackends/D3D/TextureCache.cpp b/Source/Core/VideoBackends/D3D/TextureCache.cpp
index a7eab7c041..cca8448421 100644
--- a/Source/Core/VideoBackends/D3D/TextureCache.cpp
+++ b/Source/Core/VideoBackends/D3D/TextureCache.cpp
@@ -29,8 +29,6 @@
 
 namespace DX11
 {
-static const size_t MAX_COPY_BUFFERS = 32;
-static ID3D11Buffer* s_efbcopycbuf[MAX_COPY_BUFFERS] = {0};
 static std::unique_ptr<PSTextureEncoder> g_encoder;
 
 void TextureCache::CopyEFB(u8* dst, const EFBCopyParams& params, u32 native_width,
@@ -207,17 +205,16 @@ TextureCache::TextureCache()
 
 TextureCache::~TextureCache()
 {
-  for (unsigned int k = 0; k < MAX_COPY_BUFFERS; ++k)
-    SAFE_RELEASE(s_efbcopycbuf[k]);
-
   g_encoder->Shutdown();
   g_encoder.reset();
 
   SAFE_RELEASE(palette_buf);
   SAFE_RELEASE(palette_buf_srv);
   SAFE_RELEASE(palette_uniform);
-  for (ID3D11PixelShader*& shader : palette_pixel_shader)
+  for (auto*& shader : palette_pixel_shader)
     SAFE_RELEASE(shader);
+  for (auto& iter : m_efb_to_tex_pixel_shaders)
+    SAFE_RELEASE(iter.second);
 }
 
 void TextureCache::CopyEFBToCacheEntry(TCacheEntry* entry, bool is_depth_copy,
@@ -227,20 +224,25 @@ void TextureCache::CopyEFBToCacheEntry(TCacheEntry* entry, bool is_depth_copy,
 {
   auto* destination_texture = static_cast<DXTexture*>(entry->texture.get());
 
-  // When copying at half size, in multisampled mode, resolve the color/depth buffer first.
-  // This is because multisampled texture reads go through Load, not Sample, and the linear
-  // filter is ignored.
-  bool multisampled = (g_ActiveConfig.iMultisamples > 1);
-  ID3D11ShaderResourceView* efbTexSRV = is_depth_copy ?
-                                            FramebufferManager::GetEFBDepthTexture()->GetSRV() :
-                                            FramebufferManager::GetEFBColorTexture()->GetSRV();
-  if (multisampled && scale_by_half)
+  bool multisampled = g_ActiveConfig.iMultisamples > 1;
+  ID3D11ShaderResourceView* efb_tex_srv;
+  if (multisampled)
+  {
+    efb_tex_srv = is_depth_copy ? FramebufferManager::GetResolvedEFBDepthTexture()->GetSRV() :
+                                  FramebufferManager::GetResolvedEFBColorTexture()->GetSRV();
+  }
+  else
   {
-    multisampled = false;
-    efbTexSRV = is_depth_copy ? FramebufferManager::GetResolvedEFBDepthTexture()->GetSRV() :
-                                FramebufferManager::GetResolvedEFBColorTexture()->GetSRV();
+    efb_tex_srv = is_depth_copy ? FramebufferManager::GetEFBDepthTexture()->GetSRV() :
+                                  FramebufferManager::GetEFBColorTexture()->GetSRV();
   }
 
+  auto uid = TextureConversionShaderGen::GetShaderUid(dst_format, is_depth_copy, is_intensity,
+                                                      scale_by_half);
+  ID3D11PixelShader* pixel_shader = GetEFBToTexPixelShader(uid);
+  if (!pixel_shader)
+    return;
+
   g_renderer->ResetAPIState();
 
   // stretch picture with increased internal resolution
@@ -249,20 +251,6 @@ void TextureCache::CopyEFBToCacheEntry(TCacheEntry* entry, bool is_depth_copy,
                       static_cast<float>(destination_texture->GetConfig().height));
   D3D::context->RSSetViewports(1, &vp);
 
-  // set transformation
-  if (nullptr == s_efbcopycbuf[cbuf_id])
-  {
-    const D3D11_BUFFER_DESC cbdesc =
-        CD3D11_BUFFER_DESC(28 * sizeof(float), D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DEFAULT);
-    D3D11_SUBRESOURCE_DATA data;
-    data.pSysMem = colmat;
-    HRESULT hr = D3D::device->CreateBuffer(&cbdesc, &data, &s_efbcopycbuf[cbuf_id]);
-    CHECK(SUCCEEDED(hr), "Create efb copy constant buffer %d", cbuf_id);
-    D3D::SetDebugObjectName(s_efbcopycbuf[cbuf_id],
-                            "a constant buffer used in TextureCache::CopyRenderTargetToTexture");
-  }
-  D3D::stateman->SetPixelConstants(s_efbcopycbuf[cbuf_id]);
-
   const TargetRectangle targetSource = g_renderer->ConvertEFBRectangle(src_rect);
   // TODO: try targetSource.asRECT();
   const D3D11_RECT sourcerect =
@@ -284,13 +272,24 @@ void TextureCache::CopyEFBToCacheEntry(TCacheEntry* entry, bool is_depth_copy,
 
   // Create texture copy
   D3D::drawShadedTexQuad(
-      efbTexSRV, &sourcerect, g_renderer->GetTargetWidth(), g_renderer->GetTargetHeight(),
-      is_depth_copy ? PixelShaderCache::GetDepthMatrixProgram(multisampled) :
-                      PixelShaderCache::GetColorMatrixProgram(multisampled),
-      VertexShaderCache::GetSimpleVertexShader(), VertexShaderCache::GetSimpleInputLayout(),
-      GeometryShaderCache::GetCopyGeometryShader());
+      efb_tex_srv, &sourcerect, g_renderer->GetTargetWidth(), g_renderer->GetTargetHeight(),
+      pixel_shader, VertexShaderCache::GetSimpleVertexShader(),
+      VertexShaderCache::GetSimpleInputLayout(), GeometryShaderCache::GetCopyGeometryShader());
 
   FramebufferManager::BindEFBRenderTarget();
   g_renderer->RestoreAPIState();
 }
+
+ID3D11PixelShader*
+TextureCache::GetEFBToTexPixelShader(const TextureConversionShaderGen::TCShaderUid& uid)
+{
+  auto iter = m_efb_to_tex_pixel_shaders.find(uid);
+  if (iter != m_efb_to_tex_pixel_shaders.end())
+    return iter->second;
+
+  ShaderCode code = TextureConversionShaderGen::GenerateShader(APIType::D3D, uid.GetUidData());
+  ID3D11PixelShader* shader = D3D::CompileAndCreatePixelShader(code.GetBuffer());
+  m_efb_to_tex_pixel_shaders.emplace(uid, shader);
+  return shader;
+}
 }
diff --git a/Source/Core/VideoBackends/D3D/TextureCache.h b/Source/Core/VideoBackends/D3D/TextureCache.h
index d650c04d4c..3f48ddb6f7 100644
--- a/Source/Core/VideoBackends/D3D/TextureCache.h
+++ b/Source/Core/VideoBackends/D3D/TextureCache.h
@@ -4,8 +4,11 @@
 
 #pragma once
 
+#include <map>
+
 #include "VideoBackends/D3D/D3DTexture.h"
 #include "VideoCommon/TextureCacheBase.h"
+#include "VideoCommon/TextureConverterShaderGen.h"
 
 class AbstractTexture;
 struct TextureConfig;
@@ -39,9 +42,13 @@ private:
 
   bool CompileShaders() override { return true; }
   void DeleteShaders() override {}
+  ID3D11PixelShader* GetEFBToTexPixelShader(const TextureConversionShaderGen::TCShaderUid& uid);
+
   ID3D11Buffer* palette_buf;
   ID3D11ShaderResourceView* palette_buf_srv;
   ID3D11Buffer* palette_uniform;
   ID3D11PixelShader* palette_pixel_shader[3];
+
+  std::map<TextureConversionShaderGen::TCShaderUid, ID3D11PixelShader*> m_efb_to_tex_pixel_shaders;
 };
 }