Merge pull request #13906 from unknownbrackets/skinning

GPU: Correct shader gen with weights as floats
hrydgard · Jan 10, 2021 · 5880ee4 · 5880ee4
2 parents 291f9fd + a73c15b
commit 5880ee4
Show file tree

Hide file tree

Showing 15 changed files with 28 additions and 28 deletions.
diff --git a/GPU/Common/ShaderId.cpp b/GPU/Common/ShaderId.cpp
@@ -61,7 +61,7 @@ std::string VertexShaderDesc(const VShaderID &id) {
 	return desc.str();
 }
 
-void ComputeVertexShaderID(VShaderID *id_out, u32 vertType, bool useHWTransform, bool useHWTessellation) {
+void ComputeVertexShaderID(VShaderID *id_out, u32 vertType, bool useHWTransform, bool useHWTessellation, bool weightsAsFloat) {
 	bool isModeThrough = (vertType & GE_VTYPE_THROUGH) != 0;
 	bool doTexture = gstate.isTextureMapEnabled() && !gstate.isModeClear();
 	bool doShadeMapping = doTexture && (gstate.getUVGenMode() == GE_TEXMAP_ENVIRONMENT_MAP);
@@ -112,7 +112,7 @@ void ComputeVertexShaderID(VShaderID *id_out, u32 vertType, bool useHWTransform,
 			id.SetBits(VS_BIT_BONES, 3, TranslateNumBones(vertTypeGetNumBoneWeights(vertType)) - 1);
 			// 2 bits. We should probably send in the weight scalefactor as a uniform instead,
 			// or simply preconvert all weights to floats.
-			id.SetBits(VS_BIT_WEIGHT_FMTSCALE, 2, (vertType & GE_VTYPE_WEIGHT_MASK) >> GE_VTYPE_WEIGHT_SHIFT);
+			id.SetBits(VS_BIT_WEIGHT_FMTSCALE, 2, weightsAsFloat ? 0 : (vertType & GE_VTYPE_WEIGHT_MASK) >> GE_VTYPE_WEIGHT_SHIFT);
 		}
 
 		if (gstate.isLightingEnabled()) {

diff --git a/GPU/Common/ShaderId.h b/GPU/Common/ShaderId.h
@@ -225,7 +225,7 @@ class Bugs;
 }
 
 
-void ComputeVertexShaderID(VShaderID *id, uint32_t vertexType, bool useHWTransform, bool useHWTessellation);
+void ComputeVertexShaderID(VShaderID *id, uint32_t vertexType, bool useHWTransform, bool useHWTessellation, bool weightsAsFloat);
 // Generates a compact string that describes the shader. Useful in a list to get an overview
 // of the current flora of shaders.
 std::string VertexShaderDesc(const VShaderID &id);

diff --git a/GPU/Common/VertexShaderGenerator.cpp b/GPU/Common/VertexShaderGenerator.cpp
@@ -822,17 +822,17 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 			// Uncomment this to screw up bone shaders to check the vertex shader software fallback
 			// WRITE(p, "THIS SHOULD ERROR! #error");
 			if (numBoneWeights == 1 && ShaderLanguageIsOpenGL(compat.shaderLanguage))
-				WRITE(p, "  %s skinMatrix = w1 * u_bone0", boneMatrix);
+				WRITE(p, "  %s skinMatrix = mul(w1, u_bone0)", boneMatrix);
 			else
-				WRITE(p, "  %s skinMatrix = w1.x * u_bone0", boneMatrix);
+				WRITE(p, "  %s skinMatrix = mul(w1.x, u_bone0)", boneMatrix);
 			for (int i = 1; i < numBoneWeights; i++) {
 				const char *weightAttr = boneWeightAttr[i];
 				// workaround for "cant do .x of scalar" issue.
 				if (ShaderLanguageIsOpenGL(compat.shaderLanguage)) {
 					if (numBoneWeights == 1 && i == 0) weightAttr = "w1";
 					if (numBoneWeights == 5 && i == 4) weightAttr = "w2";
 				}
-				WRITE(p, " + %s * u_bone%i", weightAttr, i);
+				WRITE(p, " + mul(%s, u_bone%i)", weightAttr, i);
 			}
 
 			WRITE(p, ";\n");

diff --git a/GPU/D3D11/DrawEngineD3D11.cpp b/GPU/D3D11/DrawEngineD3D11.cpp
@@ -523,7 +523,7 @@ void DrawEngineD3D11::DoFlush() {
 
 		D3D11VertexShader *vshader;
 		D3D11FragmentShader *fshader;
-		shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, useHWTransform, useHWTessellation_);
+		shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, useHWTransform, useHWTessellation_, decOptions_.expandAllWeightsToFloat);
 		ID3D11InputLayout *inputLayout = SetupDecFmtForDraw(vshader, dec_->GetDecVtxFmt(), dec_->VertexType());
 		context_->PSSetShader(fshader->GetShader(), nullptr, 0);
 		context_->VSSetShader(vshader->GetShader(), nullptr, 0);
@@ -609,7 +609,7 @@ void DrawEngineD3D11::DoFlush() {
 
 			D3D11VertexShader *vshader;
 			D3D11FragmentShader *fshader;
-			shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, false, false);
+			shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, false, false, decOptions_.expandAllWeightsToFloat);
 			context_->PSSetShader(fshader->GetShader(), nullptr, 0);
 			context_->VSSetShader(vshader->GetShader(), nullptr, 0);
 			shaderManager_->UpdateUniforms(framebufferManager_->UseBufferedRendering());

diff --git a/GPU/D3D11/ShaderManagerD3D11.cpp b/GPU/D3D11/ShaderManagerD3D11.cpp
@@ -177,13 +177,13 @@ void ShaderManagerD3D11::BindUniforms() {
 	context_->PSSetConstantBuffers(0, 1, ps_cbs);
 }
 
-void ShaderManagerD3D11::GetShaders(int prim, u32 vertType, D3D11VertexShader **vshader, D3D11FragmentShader **fshader, bool useHWTransform, bool useHWTessellation) {
+void ShaderManagerD3D11::GetShaders(int prim, u32 vertType, D3D11VertexShader **vshader, D3D11FragmentShader **fshader, bool useHWTransform, bool useHWTessellation, bool weightsAsFloat) {
 	VShaderID VSID;
 	FShaderID FSID;
 
 	if (gstate_c.IsDirty(DIRTY_VERTEXSHADER_STATE)) {
 		gstate_c.Clean(DIRTY_VERTEXSHADER_STATE);
-		ComputeVertexShaderID(&VSID, vertType, useHWTransform, useHWTessellation);
+		ComputeVertexShaderID(&VSID, vertType, useHWTransform, useHWTessellation, weightsAsFloat);
 	} else {
 		VSID = lastVSID_;
 	}

diff --git a/GPU/D3D11/ShaderManagerD3D11.h b/GPU/D3D11/ShaderManagerD3D11.h
@@ -85,7 +85,7 @@ class ShaderManagerD3D11 : public ShaderManagerCommon {
 	ShaderManagerD3D11(Draw::DrawContext *draw, ID3D11Device *device, ID3D11DeviceContext *context, D3D_FEATURE_LEVEL featureLevel);
 	~ShaderManagerD3D11();
 
-	void GetShaders(int prim, u32 vertType, D3D11VertexShader **vshader, D3D11FragmentShader **fshader, bool useHWTransform, bool useHWTessellation);
+	void GetShaders(int prim, u32 vertType, D3D11VertexShader **vshader, D3D11FragmentShader **fshader, bool useHWTransform, bool useHWTessellation, bool weightsAsFloat);
 	void ClearShaders();
 	void DirtyLastShader() override;
 

diff --git a/GPU/Directx9/DrawEngineDX9.cpp b/GPU/Directx9/DrawEngineDX9.cpp
@@ -324,7 +324,7 @@ void DrawEngineDX9::DoFlush() {
 	GEPrimitiveType prim = prevPrim_;
 	ApplyDrawState(prim);
 
-	VSShader *vshader = shaderManager_->ApplyShader(CanUseHardwareTransform(prim), useHWTessellation_, lastVType_);
+	VSShader *vshader = shaderManager_->ApplyShader(CanUseHardwareTransform(prim), useHWTessellation_, lastVType_, decOptions_.expandAllWeightsToFloat);
 
 	if (vshader->UseHWTransform()) {
 		LPDIRECT3DVERTEXBUFFER9 vb_ = NULL;
@@ -505,7 +505,7 @@ void DrawEngineDX9::DoFlush() {
 		}
 
 		ApplyDrawStateLate();
-		vshader = shaderManager_->ApplyShader(CanUseHardwareTransform(prim), useHWTessellation_, lastVType_);
+		vshader = shaderManager_->ApplyShader(CanUseHardwareTransform(prim), useHWTessellation_, lastVType_, decOptions_.expandAllWeightsToFloat);
 		IDirect3DVertexDeclaration9 *pHardwareVertexDecl = SetupDecFmtForDraw(vshader, dec_->GetDecVtxFmt(), dec_->VertexType());
 
 		if (pHardwareVertexDecl) {
@@ -568,7 +568,7 @@ void DrawEngineDX9::DoFlush() {
 			framebufferManager_->SetSafeSize(result.safeWidth, result.safeHeight);
 
 		ApplyDrawStateLate();
-		vshader = shaderManager_->ApplyShader(false, false, lastVType_);
+		vshader = shaderManager_->ApplyShader(false, false, lastVType_, decOptions_.expandAllWeightsToFloat);
 
 		if (result.action == SW_DRAW_PRIMITIVES) {
 			if (result.setStencil) {

diff --git a/GPU/Directx9/ShaderManagerDX9.cpp b/GPU/Directx9/ShaderManagerDX9.cpp
@@ -541,15 +541,15 @@ void ShaderManagerDX9::DirtyLastShader() { // disables vertex arrays
 	lastPShader_ = nullptr;
 }
 
-VSShader *ShaderManagerDX9::ApplyShader(bool useHWTransform, bool useHWTessellation, u32 vertType) {
+VSShader *ShaderManagerDX9::ApplyShader(bool useHWTransform, bool useHWTessellation, u32 vertType, bool weightsAsFloat) {
 	// Always use software for flat shading to fix the provoking index.
 	bool tess = gstate_c.submitType == SubmitType::HW_BEZIER || gstate_c.submitType == SubmitType::HW_SPLINE;
 	useHWTransform = useHWTransform && (tess || gstate.getShadeMode() != GE_SHADE_FLAT);
 
 	VShaderID VSID;
 	if (gstate_c.IsDirty(DIRTY_VERTEXSHADER_STATE)) {
 		gstate_c.Clean(DIRTY_VERTEXSHADER_STATE);
-		ComputeVertexShaderID(&VSID, vertType, useHWTransform, useHWTessellation);
+		ComputeVertexShaderID(&VSID, vertType, useHWTransform, useHWTessellation, weightsAsFloat);
 	} else {
 		VSID = lastVSID_;
 	}
@@ -598,7 +598,7 @@ VSShader *ShaderManagerDX9::ApplyShader(bool useHWTransform, bool useHWTessellat
 			}
 			delete vs;
 
-			ComputeVertexShaderID(&VSID, vertType, false, false);
+			ComputeVertexShaderID(&VSID, vertType, false, false, weightsAsFloat);
 
 			// TODO: Look for existing shader with the appropriate ID, use that instead of generating a new one - however, need to make sure
 			// that that shader ID is not used when computing the linked shader ID below, because then IDs won't match

diff --git a/GPU/Directx9/ShaderManagerDX9.h b/GPU/Directx9/ShaderManagerDX9.h
@@ -80,7 +80,7 @@ class ShaderManagerDX9 : public ShaderManagerCommon {
 	~ShaderManagerDX9();
 
 	void ClearCache(bool deleteThem);  // TODO: deleteThem currently not respected
-	VSShader *ApplyShader(bool useHWTransform, bool useHWTessellation, u32 vertType);
+	VSShader *ApplyShader(bool useHWTransform, bool useHWTessellation, u32 vertType, bool weightsAsFloat);
 	void DirtyShader();
 	void DirtyLastShader() override;
 

diff --git a/GPU/GLES/DrawEngineGLES.cpp b/GPU/GLES/DrawEngineGLES.cpp
@@ -333,7 +333,7 @@ void DrawEngineGLES::DoFlush() {
 	GEPrimitiveType prim = prevPrim_;
 
 	VShaderID vsid;
-	Shader *vshader = shaderManager_->ApplyVertexShader(CanUseHardwareTransform(prim), useHWTessellation_, lastVType_, &vsid);
+	Shader *vshader = shaderManager_->ApplyVertexShader(CanUseHardwareTransform(prim), useHWTessellation_, lastVType_, decOptions_.expandAllWeightsToFloat, &vsid);
 
 	GLRBuffer *vertexBuffer = nullptr;
 	GLRBuffer *indexBuffer = nullptr;

diff --git a/GPU/GLES/ShaderManagerGLES.cpp b/GPU/GLES/ShaderManagerGLES.cpp
@@ -656,10 +656,10 @@ Shader *ShaderManagerGLES::CompileVertexShader(VShaderID VSID) {
 	return new Shader(render_, codeBuffer_, desc, GL_VERTEX_SHADER, useHWTransform, attrMask, uniformMask);
 }
 
-Shader *ShaderManagerGLES::ApplyVertexShader(bool useHWTransform, bool useHWTessellation, u32 vertType, VShaderID *VSID) {
+Shader *ShaderManagerGLES::ApplyVertexShader(bool useHWTransform, bool useHWTessellation, u32 vertType, bool weightsAsFloat, VShaderID *VSID) {
 	if (gstate_c.IsDirty(DIRTY_VERTEXSHADER_STATE)) {
 		gstate_c.Clean(DIRTY_VERTEXSHADER_STATE);
-		ComputeVertexShaderID(VSID, vertType, useHWTransform, useHWTessellation);
+		ComputeVertexShaderID(VSID, vertType, useHWTransform, useHWTessellation, weightsAsFloat);
 	} else {
 		*VSID = lastVSID_;
 	}
@@ -690,7 +690,7 @@ Shader *ShaderManagerGLES::ApplyVertexShader(bool useHWTransform, bool useHWTess
 
 			// Can still work with software transform.
 			VShaderID vsidTemp;
-			ComputeVertexShaderID(&vsidTemp, vertType, false, false);
+			ComputeVertexShaderID(&vsidTemp, vertType, false, false, weightsAsFloat);
 			vs = CompileVertexShader(vsidTemp);
 		}
 

diff --git a/GPU/GLES/ShaderManagerGLES.h b/GPU/GLES/ShaderManagerGLES.h
@@ -146,7 +146,7 @@ class ShaderManagerGLES : public ShaderManagerCommon {
 
 	// This is the old ApplyShader split into two parts, because of annoying information dependencies.
 	// If you call ApplyVertexShader, you MUST call ApplyFragmentShader soon afterwards.
-	Shader *ApplyVertexShader(bool useHWTransform, bool useHWTessellation, u32 vertType, VShaderID *VSID);
+	Shader *ApplyVertexShader(bool useHWTransform, bool useHWTessellation, u32 vertType, bool weightsAsFloat, VShaderID *VSID);
 	LinkedShader *ApplyFragmentShader(VShaderID VSID, Shader *vs, u32 vertType, bool useBufferedRendering);
 
 	void DeviceLost();

diff --git a/GPU/Vulkan/DrawEngineVulkan.cpp b/GPU/Vulkan/DrawEngineVulkan.cpp
@@ -818,7 +818,7 @@ void DrawEngineVulkan::DoFlush() {
 				ConvertStateToVulkanKey(*framebufferManager_, shaderManager_, prim, pipelineKey_, dynState_);
 			}
 
-			shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, true, useHWTessellation_);  // usehwtransform
+			shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, true, useHWTessellation_, decOptions_.expandAllWeightsToFloat);  // usehwtransform
 			_dbg_assert_msg_(vshader->UseHWTransform(), "Bad vshader");
 
 			Draw::NativeObject object = framebufferManager_->UseBufferedRendering() ? Draw::NativeObject::FRAMEBUFFER_RENDERPASS : Draw::NativeObject::BACKBUFFER_RENDERPASS;
@@ -939,7 +939,7 @@ void DrawEngineVulkan::DoFlush() {
 					sampler = nullSampler_;
 			}
 			if (!lastPipeline_ || gstate_c.IsDirty(DIRTY_BLEND_STATE | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_RASTER_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_VERTEXSHADER_STATE | DIRTY_FRAGMENTSHADER_STATE) || prim != lastPrim_) {
-				shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, false, false);  // usehwtransform
+				shaderManager_->GetShaders(prim, lastVType_, &vshader, &fshader, false, false, decOptions_.expandAllWeightsToFloat);  // usehwtransform
 				_dbg_assert_msg_(!vshader->UseHWTransform(), "Bad vshader");
 				if (prim != lastPrim_ || gstate_c.IsDirty(DIRTY_BLEND_STATE | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_RASTER_STATE | DIRTY_DEPTHSTENCIL_STATE)) {
 					ConvertStateToVulkanKey(*framebufferManager_, shaderManager_, prim, pipelineKey_, dynState_);

diff --git a/GPU/Vulkan/ShaderManagerVulkan.cpp b/GPU/Vulkan/ShaderManagerVulkan.cpp
@@ -228,11 +228,11 @@ uint64_t ShaderManagerVulkan::UpdateUniforms(bool useBufferedRendering) {
 	return dirty;
 }
 
-void ShaderManagerVulkan::GetShaders(int prim, u32 vertType, VulkanVertexShader **vshader, VulkanFragmentShader **fshader, bool useHWTransform, bool useHWTessellation) {
+void ShaderManagerVulkan::GetShaders(int prim, u32 vertType, VulkanVertexShader **vshader, VulkanFragmentShader **fshader, bool useHWTransform, bool useHWTessellation, bool weightsAsFloat) {
 	VShaderID VSID;
 	if (gstate_c.IsDirty(DIRTY_VERTEXSHADER_STATE)) {
 		gstate_c.Clean(DIRTY_VERTEXSHADER_STATE);
-		ComputeVertexShaderID(&VSID, vertType, useHWTransform, useHWTessellation);
+		ComputeVertexShaderID(&VSID, vertType, useHWTransform, useHWTessellation, weightsAsFloat);
 	} else {
 		VSID = lastVSID_;
 	}

diff --git a/GPU/Vulkan/ShaderManagerVulkan.h b/GPU/Vulkan/ShaderManagerVulkan.h
@@ -88,7 +88,7 @@ class ShaderManagerVulkan : public ShaderManagerCommon {
 
 	void DeviceRestore(VulkanContext *vulkan, Draw::DrawContext *draw);
 
-	void GetShaders(int prim, u32 vertType, VulkanVertexShader **vshader, VulkanFragmentShader **fshader, bool useHWTransform, bool useHWTessellation);
+	void GetShaders(int prim, u32 vertType, VulkanVertexShader **vshader, VulkanFragmentShader **fshader, bool useHWTransform, bool useHWTessellation, bool weightsAsFloat);
 	void ClearShaders();
 	void DirtyShader();
 	void DirtyLastShader() override;