GS/DX11: Add pixel shader based CAS.

It's much faster than uav and it's backwards compatible across all feature levels.

bin/resources/shaders/dx11/cas.hlsl

@@ -19,13 +19,16 @@
 
 cbuffer cb : register(b0)
 {
-  uint4 const0;
-  uint4 const1;
-  int2 srcOffset;
+	uint4 const0;
+	uint4 const1;
+	int2 srcOffset;
 };
 
 Texture2D InputTexture : register(t0);
-RWTexture2D<float4> OutputTexture : register(u0);
+
+#if PS_CAS == 0
+	RWTexture2D<float4> OutputTexture : register(u0);
+#endif
 
 #define A_GPU 1
 #define A_HLSL 1
@@ -34,7 +37,7 @@ RWTexture2D<float4> OutputTexture : register(u0);
 
 AF3 CasLoad(ASU2 p)
 {
-  return InputTexture.Load(int3(srcOffset, 0) + int3(p, 0)).rgb;
+	return InputTexture.Load(int3(srcOffset, 0) + int3(p, 0)).rgb;
 }
 
 // Lets you transform input from the load into a linear color space between 0 and 1. See ffx_cas.h
@@ -43,33 +46,56 @@ void CasInput(inout AF1 r, inout AF1 g, inout AF1 b) {}
 
 #include "ffx_cas.h"
 
+#if PS_CAS == 0
+
 [numthreads(64, 1, 1)]
-void main(uint3 LocalThreadId : SV_GroupThreadID, uint3 WorkGroupId : SV_GroupID)
+void cs_main(uint3 LocalThreadId : SV_GroupThreadID, uint3 WorkGroupId : SV_GroupID)
 {
-  // Do remapping of local xy in workgroup for a more PS-like swizzle pattern.
-  AU2 gxy = ARmp8x8(LocalThreadId.x) + AU2(WorkGroupId.x << 4u, WorkGroupId.y << 4u);
+	// Do remapping of local xy in workgroup for a more PS-like swizzle pattern.
+	AU2 gxy = ARmp8x8(LocalThreadId.x) + AU2(WorkGroupId.x << 4u, WorkGroupId.y << 4u);
 
 #if CAS_SHARPEN_ONLY
-  const bool sharpenOnly = true;
+	const bool sharpenOnly = true;
 #else
-  const bool sharpenOnly = false;
+	const bool sharpenOnly = false;
 #endif
 
-  // Filter.
-  AF3 c = (float3)0.0f;
+	// Filter.
+	AF3 c = (float3)0.0f;
 
-  CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
-  OutputTexture[ASU2(gxy)] = AF4(c, 1);
-  gxy.x += 8u;
+	CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
+	OutputTexture[ASU2(gxy)] = AF4(c, 1);
+	gxy.x += 8u;
 
-  CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
-  OutputTexture[ASU2(gxy)] = AF4(c, 1);
-  gxy.y += 8u;
+	CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
+	OutputTexture[ASU2(gxy)] = AF4(c, 1);
+	gxy.y += 8u;
 
-  CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
-  OutputTexture[ASU2(gxy)] = AF4(c, 1);
-  gxy.x -= 8u;
+	CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
+	OutputTexture[ASU2(gxy)] = AF4(c, 1);
+	gxy.x -= 8u;
 
-  CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
-  OutputTexture[ASU2(gxy)] = AF4(c, 1);
+	CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
+	OutputTexture[ASU2(gxy)] = AF4(c, 1);
 }
+
+#elif PS_CAS == 1
+
+float4 ps_main(float4 position : SV_Position, float2 uv : TEXCOORD0) : SV_Target
+{
+	// Pixel coordinate for the current fragment
+	AU2 gxy = AU2(position.xy);
+
+#if CAS_SHARPEN_ONLY
+	const bool sharpenOnly = true;
+#else
+	const bool sharpenOnly = false;
+#endif
+
+	AF3 c = (float3)0.0f;
+	CasFilter(c.r, c.g, c.b, gxy, const0, const1, sharpenOnly);
+
+	return AF4(c, 1.0f);
+}
+
+#endif

pcsx2/GS/Renderers/Common/GSDevice.cpp

@@ -1236,10 +1236,10 @@ void GSDevice::CAS(GSTexture*& tex, GSVector4i& src_rect, GSVector4& src_uv, con
 	if (!m_cas || m_cas->GetWidth() != dst_width || m_cas->GetHeight() != dst_height)
 	{
 		delete m_cas;
-		m_cas = CreateSurface(GSTexture::ShaderWriteTexture, dst_width, dst_height, 1, GSTexture::Format::Color);
+		m_cas = CreateSurface(PSTypeCAS() ? GSTexture::ShaderWriteTarget : GSTexture::ShaderWriteTexture, dst_width, dst_height, 1, GSTexture::Format::Color);
 		if (!m_cas)
 		{
-			Console.Error("Failed to allocate CAS RW texture.");
+			Console.Error("GS: Failed to allocate CAS RW texture.");
 			return;
 		}
 	}
@@ -1254,7 +1254,7 @@ void GSDevice::CAS(GSTexture*& tex, GSVector4i& src_rect, GSVector4& src_uv, con
 	if (!DoCAS(src_tex, m_cas, sharpen_only, consts))
 	{
 		// leave textures intact if we failed
-		Console.Warning("Applying CAS failed.");
+		Console.Warning("GS: Applying CAS failed.");
 		return;
 	}
 

pcsx2/GS/Renderers/Common/GSDevice.h

@@ -1514,6 +1514,7 @@ protected:
 
 	/// Applies CAS and writes to the destination texture, which should be a shader writeable texture.
 	virtual bool DoCAS(GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants) = 0;
+	virtual bool PSTypeCAS() = 0;
 
 	/// Perform texture operations for ImGui
 	void UpdateImGuiTextures();

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -715,7 +715,6 @@ void GSDevice11::SetFeatures(IDXGIAdapter1* adapter)
 	                          SupportsTextureFormat(m_dev.get(), DXGI_FORMAT_BC3_UNORM);
 
 	m_features.bptc_textures = SupportsTextureFormat(m_dev.get(), DXGI_FORMAT_BC7_UNORM);
-	m_features.cas_sharpening = (m_feature_level >= D3D_FEATURE_LEVEL_11_0);
 	m_features.test_and_sample_depth = (m_feature_level >= D3D_FEATURE_LEVEL_11_0);
 	m_features.depth_feedback = m_features.multidraw_fb_copy && GSConfig.DepthFeedbackMode == GSDepthFeedbackMode::Depth;
 	m_features.aa1 = GSConfig.HWAA1 && m_features.vs_expand && m_features.feedback_loops();
@@ -2273,7 +2272,7 @@ void GSDevice11::SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, u8
 bool GSDevice11::CreateCASShaders()
 {
 	CD3D11_BUFFER_DESC desc(NUM_CAS_CONSTANTS * sizeof(u32), D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DEFAULT);
-	HRESULT hr = m_dev->CreateBuffer(&desc, nullptr, m_cas.cb.put());
+	const HRESULT hr = m_dev->CreateBuffer(&desc, nullptr, m_cas.cb.put());
 	if (FAILED(hr))
 		return false;
 
@@ -2281,15 +2280,14 @@ bool GSDevice11::CreateCASShaders()
 	if (!cas_source.has_value() || !GetCASShaderSource(&cas_source.value()))
 		return false;
 
-	static constexpr D3D_SHADER_MACRO sharpen_only_macros[] = {
-		{"CAS_SHARPEN_ONLY", "1"},
-		{nullptr, nullptr}};
+	static constexpr D3D_SHADER_MACRO sharpen_macros[] = {{"CAS_SHARPEN_ONLY", "1"}, {"PS_CAS", "1"}, {nullptr, nullptr}};
+	static constexpr D3D_SHADER_MACRO upscale_macros[] = {{"PS_CAS", "1"}, {nullptr, nullptr}};
 
-	m_cas.cs_sharpen = m_shader_cache.GetComputeShader(m_dev.get(), cas_source.value(), sharpen_only_macros, "main");
-	m_cas.cs_upscale = m_shader_cache.GetComputeShader(m_dev.get(), cas_source.value(), nullptr, "main");
-	if (!m_cas.cs_sharpen || !m_cas.cs_upscale)
+	m_cas.ps_sharpen = m_shader_cache.GetPixelShader(m_dev.get(), cas_source.value(), sharpen_macros, "ps_main");
+	m_cas.ps_upscale = m_shader_cache.GetPixelShader(m_dev.get(), cas_source.value(), upscale_macros, "ps_main");
+	if (!m_cas.ps_sharpen || !m_cas.ps_upscale)
 	{
-		Console.Error("D3D11: Failed to create CAS compute shaders.");
+		Console.Error("D3D11: Failed to create CAS pixel shaders.");
 		return false;
 	}
 
@@ -2298,27 +2296,14 @@ bool GSDevice11::CreateCASShaders()
 
 bool GSDevice11::DoCAS(GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants)
 {
-	g_perfmon.Put(GSPerfMon::TextureCopies, 1);
+	const GSVector2i s = dTex->GetSize();
 
-	static constexpr int threadGroupWorkRegionDim = 16;
-	const int dispatchX = (dTex->GetWidth() + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim;
-	const int dispatchY = (dTex->GetHeight() + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim;
+	const GSVector4 sRect(0, 0, 1, 1);
+	const GSVector4 dRect(0, 0, s.x, s.y);
 
-	ID3D11ShaderResourceView* srvs[1] = {*static_cast<GSTexture11*>(sTex)};
-	ID3D11UnorderedAccessView* uavs[1] = {*static_cast<GSTexture11*>(dTex)};
-	OMSetRenderTargets(nullptr, nullptr, nullptr);
 	m_ctx->UpdateSubresource(m_cas.cb.get(), 0, nullptr, constants.data(), 0, 0);
-	m_ctx->CSSetConstantBuffers(0, 1, m_cas.cb.addressof());
-	m_ctx->CSSetShader(sharpen_only ? m_cas.cs_sharpen.get() : m_cas.cs_upscale.get(), nullptr, 0);
-	m_ctx->CSSetShaderResources(0, std::size(srvs), srvs);
-	m_ctx->CSSetUnorderedAccessViews(0, std::size(uavs), uavs, nullptr);
-	m_ctx->Dispatch(dispatchX, dispatchY, 1);
 
-	// clear bindings out to prevent hazards
-	uavs[0] = nullptr;
-	srvs[0] = nullptr;
-	m_ctx->CSSetShaderResources(0, std::size(srvs), srvs);
-	m_ctx->CSSetUnorderedAccessViews(0, std::size(uavs), uavs, nullptr);
+	DoStretchRect(sTex, sRect, dTex, dRect, sharpen_only ? m_cas.ps_sharpen.get() : m_cas.ps_upscale.get(), m_cas.cb.get(), Biln);
 
 	return true;
 }

pcsx2/GS/Renderers/DX11/GSDevice11.h

@@ -120,6 +120,7 @@ private:
 
 	bool CreateCASShaders();
 	bool DoCAS(GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants) override;
+	bool PSTypeCAS() override { return true; }
 
 	bool CreateImGuiResources();
 	void RenderImGui();
@@ -271,8 +272,8 @@ private:
 	struct
 	{
 		wil::com_ptr_nothrow<ID3D11Buffer> cb;
-		wil::com_ptr_nothrow<ID3D11ComputeShader> cs_upscale;
-		wil::com_ptr_nothrow<ID3D11ComputeShader> cs_sharpen;
+		wil::com_ptr_nothrow<ID3D11PixelShader> ps_upscale;
+		wil::com_ptr_nothrow<ID3D11PixelShader> ps_sharpen;
 	} m_cas;
 
 	struct

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -2311,10 +2311,11 @@ bool GSDevice12::CompileCASPipelines()
 	if (!cas_source.has_value() || !GetCASShaderSource(&cas_source.value()))
 		return false;
 
-	static constexpr D3D_SHADER_MACRO sharpen_only_macros[] = {{"CAS_SHARPEN_ONLY", "1"}, {nullptr, nullptr}};
+	static constexpr D3D_SHADER_MACRO upscale_macros[] = {{"PS_CAS", "0"}, {nullptr, nullptr}};
+	static constexpr D3D_SHADER_MACRO sharpen_macros[] = {{"CAS_SHARPEN_ONLY", "1"}, {"PS_CAS", "0"}, {nullptr, nullptr}};
 
-	const ComPtr<ID3DBlob> cs_upscale(m_shader_cache.GetComputeShader(cas_source.value(), nullptr, "main"));
-	const ComPtr<ID3DBlob> cs_sharpen(m_shader_cache.GetComputeShader(cas_source.value(), sharpen_only_macros, "main"));
+	const ComPtr<ID3DBlob> cs_upscale(m_shader_cache.GetComputeShader(cas_source.value(), nullptr, "cs_main"));
+	const ComPtr<ID3DBlob> cs_sharpen(m_shader_cache.GetComputeShader(cas_source.value(), sharpen_macros, "cs_main"));
 	if (!cs_upscale || !cs_sharpen)
 		return false;
 

pcsx2/GS/Renderers/DX12/GSDevice12.h

@@ -460,6 +460,7 @@ private:
 
 	bool DoCAS(
 		GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants) final;
+	bool PSTypeCAS() final { return false; }
 
 	bool GetSampler(D3D12DescriptorHandle* cpu_handle, GSHWDrawConfig::SamplerSelector ss);
 	void ClearSamplerCache() final;

pcsx2/GS/Renderers/Metal/GSDeviceMTL.h

@@ -394,6 +394,7 @@ public:
 	void DoShadeBoost(GSTexture* sTex, GSTexture* dTex, const float params[4]) override;
 
 	bool DoCAS(GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants) override;
+	bool PSTypeCAS() override { return false; }
 
 	MRCOwned<id<MTLFunction>> LoadShader(NSString* name);
 	MRCOwned<id<MTLRenderPipelineState>> MakePipeline(MTLRenderPipelineDescriptor* desc, id<MTLFunction> vertex, id<MTLFunction> fragment, NSString* name);

pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.h

@@ -300,6 +300,7 @@ private:
 
 	bool CreateCASPrograms();
 	bool DoCAS(GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants) override;
+	bool PSTypeCAS() override { return false; }
 
 	bool CreateImGuiProgram();
 	void RenderImGui();

pcsx2/GS/Renderers/Vulkan/GSDeviceVK.h

@@ -494,6 +494,7 @@ private:
 
 	bool DoCAS(
 		GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants) final;
+	bool PSTypeCAS() final { return false; }
 
 	VkSampler GetSampler(GSHWDrawConfig::SamplerSelector ss);
 	void ClearSamplerCache() final;