Merge pull request #22465 from hrydgard/hw-ge-blends

HW backends: Improve accuracy of blending

Common/GPU/Vulkan/VulkanQueueRunner.cpp

@@ -933,7 +933,7 @@ void VulkanQueueRunner::LogRenderPass(const VKRStep &pass, bool verbose) {
 				INFO_LOG(Log::G3D, "  BindGraphicsPipeline(%x)", (int)(intptr_t)cmd.graphics_pipeline.pipeline);
 				break;
 			case VKRRenderCommand::BLEND:
-				INFO_LOG(Log::G3D, "  BlendColor(%08x)", cmd.blendColor.color);
+				INFO_LOG(Log::G3D, "  BlendColor(%f, %f, %f, %f)", cmd.blendColor.color[0], cmd.blendColor.color[1], cmd.blendColor.color[2], cmd.blendColor.color[3]);
 				break;
 			case VKRRenderCommand::CLEAR:
 				INFO_LOG(Log::G3D, "  Clear");
@@ -1179,12 +1179,8 @@ void VulkanQueueRunner::PerformRenderPass(const VKRStep &step, VkCommandBuffer c
 		}
 
 		case VKRRenderCommand::BLEND:
-		{
-			float bc[4];
-			Uint8x4ToFloat4(bc, c.blendColor.color);
-			vkCmdSetBlendConstants(cmd, bc);
+			vkCmdSetBlendConstants(cmd, c.blendColor.color);
 			break;
-		}
 
 		case VKRRenderCommand::PUSH_CONSTANTS:
 			if (pipelineOK) {

Common/GPU/Vulkan/VulkanQueueRunner.h

@@ -99,7 +99,7 @@ struct VkRenderData {
 			uint8_t stencilRef;
 		} stencil;
 		struct {
-			uint32_t color;
+			float color[4];
 		} blendColor;
 		struct {
 			VkShaderStageFlags stages;

Common/GPU/Vulkan/VulkanRenderManager.h

@@ -396,11 +396,11 @@ public:
 		data.stencil.stencilRef = refValue;
 	}
 
-	void SetBlendFactor(uint32_t color) {
+	void SetBlendFactor(const float color[4]) {
 		_dbg_assert_(curRenderStep_ && curRenderStep_->stepType == VKRStepType::RENDER);
 		VkRenderData &data = curRenderStep_->commands.push_uninitialized();
 		data.cmd = VKRRenderCommand::BLEND;
-		data.blendColor.color = color;
+		memcpy(data.blendColor.color, color, sizeof(data.blendColor.color));
 	}
 
 	void PushConstants(VkShaderStageFlags stages, int offset, int size, void *constants) {

Common/GPU/Vulkan/thin3d_vulkan.cpp

@@ -1370,8 +1370,7 @@ void VKContext::SetViewport(const Viewport &viewport) {
 }
 
 void VKContext::SetBlendFactor(float color[4]) {
-	uint32_t col = Float4ToUint8x4(color);
-	renderManager_.SetBlendFactor(col);
+	renderManager_.SetBlendFactor(color);
 }
 
 void VKContext::SetStencilParams(uint8_t refValue, uint8_t writeMask, uint8_t compareMask) {

Core/CmdLine.cpp

@@ -205,6 +205,7 @@ static const CommandLineParam g_autoParams[] = {
 	{POFF(screenshotFilename), CmdParamType::String, "screenshot", '\0', "Compare rendered output against a reference screenshot FILE", CmdLineMode::Headless},
 	{POFF(screenshotFilenameSave), CmdParamType::String, "screenshot-save", '\0', "Save rendered screenshot to specified path (PNG if the path ends in .png, BMP otherwise)", CmdLineMode::Headless},
 	{POFF(replayEnd), CmdParamType::Int, "replay-end", '\0', "When replaying a GE frame dump, only draw primitives up to this one (1-based)", CmdLineMode::Headless},
+	{POFF(dumpCompat), CmdParamType::Bool, "dump-compat", '\0', "Apply the game's compat.ini settings when replaying a GE frame dump (off by default)", CmdLineMode::Headless},
 	{POFF(depthFilenameSave), CmdParamType::String, "depth-save", '\0', "Save the depth buffer to FILE along with the screenshot (u32 width, u32 height, u16 depths; float depths if FILE ends in .f32)", CmdLineMode::Headless},
 	{POFF(screenshotRenderTarget), CmdParamType::Bool, "screenshot-render-target", '\0', "Take screenshots of the current render target instead of the displayed framebuffer", CmdLineMode::Headless},
 	{POFF(screenshotFilenameDiff), CmdParamType::String, "screenshot-diff", '\0', "Save a visual comparison image to FILE when comparing screenshots", CmdLineMode::Headless},

Core/CmdLine.h

@@ -178,6 +178,8 @@ struct CommandLineOptions {
 	std::optional<std::string> screenshotFilenameSave;
 	std::optional<std::string> depthFilenameSave;
 	std::optional<int> replayEnd;
+	// Headless: apply the game's compatibility settings when replaying a GE frame dump.
+	std::optional<bool> dumpCompat;
 	std::optional<bool> screenshotRenderTarget;
 	std::optional<std::string> screenshotFilenameDiff;
 	// Headless: preserve the alpha channel when saving PNG screenshots.

Core/CoreParameter.h

@@ -75,6 +75,8 @@ struct CoreParameter {
 	bool startBreak = false;
 	std::string *collectDebugOutput = nullptr;
 	bool headLess = false;   // Try to avoid messageboxes etc
+	// Headless leaves out compat settings when replaying frame dumps, unless this is set.
+	bool dumpCompat = false;
 
 	// Internal PSP rendering resolution and scale factor.
 	int renderScaleFactor = 1;

Core/System.cpp

@@ -498,7 +498,7 @@ static bool CPU_Init(FileLoader *fileLoader, IdentifiedFileType type, std::strin
 	// Homebrew get fake disc IDs assigned to the global paramSFO, so they shouldn't clash with real games.
 
 	std::string discId = g_paramSFO.GetDiscID();
-	if (type == IdentifiedFileType::PPSSPP_GE_DUMP && g_CoreParameter.headLess) {
+	if (type == IdentifiedFileType::PPSSPP_GE_DUMP && g_CoreParameter.headLess && !g_CoreParameter.dumpCompat) {
 		// Headless replays frame dumps to compare them with references and with the PSP, which a game's
 		// workarounds would skew: DarkStalkers' present hack replaces the game's final upscale.
 		g_CoreParameter.compat.Clear();

GPU/Common/FragmentShaderGenerator.cpp

@@ -1008,7 +1008,21 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 				return false;
 			}
 
-			WRITE(p, "  v.rgb = v.rgb * %s;\n", srcFactor);
+			if (replaceBlend == REPLACE_BLEND_PRE_SRC && replaceBlendFuncA == GE_SRCBLEND_FIXA && IsGEExactFixedBlend(replaceBlendFuncB, replaceBlendEq)) {
+				// The GE's product of a color and a factor is floor((2c+1)(2f+1)/1024), and the hardware blend multiplies
+				// the dst by (2B+1)/512 (see ConvertBlendState). Lowering the src term by 0.5 - (2B+1)/1024 turns the
+				// rounding of the sum into the GE's floor, except where that would go below zero.
+				// The color is truncated too: the GE's bilinear filter truncates (4-bit weights, see the software
+				// Sampler), and rounding the GPU's finer filtered value instead makes each texel about a level
+				// brighter. That doubles God Eater Burst's faint bloom, which feeds back into itself pass after pass.
+				// The 0.01 keeps an unfiltered texel's exact integer from flooring a level down on float error.
+				WRITE(p, "  vec3 fixA2 = floor(u_blendFixA * 510.0 + 0.5) + 1.0;\n");
+				WRITE(p, "  vec3 fixB2 = floor(u_blendFixB * 510.0 + 0.5) + 1.0;\n");
+				WRITE(p, "  v.rgb = floor((floor(v.rgb * 255.0 + 0.01) * 2.0 + 1.0) * fixA2 / 1024.0);\n");
+				WRITE(p, "  v.rgb = (v.rgb - 0.5 + fixB2 / 1024.0) / 255.0;\n");
+			} else {
+				WRITE(p, "  v.rgb = v.rgb * %s;\n", srcFactor);
+			}
 		}
 
 		if (replaceBlend == REPLACE_BLEND_READ_FRAMEBUFFER) {

GPU/Common/GPUStateUtils.cpp

@@ -19,6 +19,7 @@
 #include <algorithm>
 #include <limits>
 
+#include "Common/Data/Convert/SmallDataConvert.h"
 #include "Core/ConfigValues.h"
 #include "Core/System.h"
 #include "Core/Config.h"
@@ -262,6 +263,16 @@ StencilValueType ReplaceAlphaWithStencilType() {
 	return STENCIL_VALUE_KEEP;
 }
 
+void GenericBlendState::blendColorToFloat4(float out[4]) const {
+	Uint8x4ToFloat4(out, blendColor);
+	if (blendColorGE) {
+		for (int i = 0; i < 3; i++) {
+			const int c = (blendColor >> (i * 8)) & 0xFF;
+			out[i] = (2 * c + 1) / 512.0f;
+		}
+	}
+}
+
 ReplaceBlendType ReplaceBlendWithShader(GEBufferFormat bufferFormat) {
 	if (gstate_c.blueToAlpha) {
 		return REPLACE_BLEND_BLUE_TO_ALPHA;
@@ -324,6 +335,10 @@ ReplaceBlendType ReplaceBlendWithShader(GEBufferFormat bufferFormat) {
 			// Using a copy isn't accurate either, though, when there's overlap.
 			if (gstate_c.Use(GPU_USE_FRAMEBUFFER_FETCH))
 				return REPLACE_BLEND_READ_FRAMEBUFFER;
+			// Through mode is mostly full screen passes that don't overlap themselves, so a copy is safe there.
+			// Super Stardust composites its bloom at alpha 0xFF, doubling the dst, which 2x alpha clamps to 1x.
+			if (gstate.isModeThrough())
+				return REPLACE_BLEND_READ_FRAMEBUFFER;
 			return REPLACE_BLEND_PRE_SRC_2X_ALPHA;
 
 		case GE_DSTBLEND_DOUBLEINVSRCALPHA:
@@ -450,6 +465,11 @@ ReplaceBlendType ReplaceBlendWithShader(GEBufferFormat bufferFormat) {
 			if (gstate.getFixA() == 0xFFFFFF && gstate.getFixB() == 0x000000) {
 				// Some games specify this. Some GPUs may prefer blending off entirely.
 				return REPLACE_BLEND_NO;
+			} else if (IsGEExactFixedBlend(funcB, gstate.getBlendEq())) {
+				// 0 and 0xFFFFFF are exact as blend factors. Other fixed colors need the GE's math, see ConvertBlendState.
+				const bool trivialA = gstate.getFixA() == 0xFFFFFF || gstate.getFixA() == 0x000000;
+				const bool trivialB = gstate.getFixB() == 0xFFFFFF || gstate.getFixB() == 0x000000;
+				return trivialA && trivialB ? REPLACE_BLEND_STANDARD : REPLACE_BLEND_PRE_SRC;
 			} else if (gstate.getFixA() == 0xFFFFFF || gstate.getFixA() == 0x000000 || gstate.getFixB() == 0xFFFFFF || gstate.getFixB() == 0x000000) {
 				// We can represent this with standard factors.
 				return REPLACE_BLEND_STANDARD;
@@ -488,8 +508,12 @@ ReplaceBlendType ReplaceBlendWithShader(GEBufferFormat bufferFormat) {
 				// We will just hope that doubling alpha for the dst factor will not clamp too badly.
 				if (gstate_c.Use(GPU_USE_FRAMEBUFFER_FETCH))
 					return REPLACE_BLEND_READ_FRAMEBUFFER;
+				// Through mode is mostly full screen passes that don't overlap themselves, so a copy is safe there.
+				// MotorStorm brightens its frame with DSTCOLOR + DOUBLESRCALPHA at alpha 0xFE, which is up to 3x the dst,
+				// and doubling alpha in the shader clamps that to 2x.
+				if (gstate.isModeThrough())
+					return REPLACE_BLEND_READ_FRAMEBUFFER;
 				// Hm, this is similar to the L.A. Rush case above. This will not be accurate.
-				// Wonder in which games we encounter this? One example is MotorStorm.
 				return REPLACE_BLEND_2X_ALPHA;
 			}
 
@@ -1068,7 +1092,19 @@ static void ConvertBlendState(GenericBlendState &blendState, FBReadSetting useFB
 		glBlendFuncB = toDualSource(glBlendFuncB);
 	}
 
-	if (blendFuncA == GE_SRCBLEND_FIXA || blendFuncB == GE_DSTBLEND_FIXB) {
+	if (usePreSrc && replaceBlend == REPLACE_BLEND_PRE_SRC && blendFuncA == GE_SRCBLEND_FIXA && IsGEExactFixedBlend(blendFuncB, blendFuncEq)) {
+		// The shader computed the src term like the GE and lowered it so that the hardware's rounding of the
+		// sum lands on the GE's floor of it, given the dst factor (2B+1)/512 the GE multiplies with.
+		if (fixB == 0 || fixB == 0xFFFFFF) {
+			glBlendFuncB = fixB == 0 ? BlendFactor::ZERO : BlendFactor::ONE;
+			if (constantAlphaGL == BlendFactor::CONSTANT_ALPHA) {
+				blendState.defaultBlendColor(constantAlpha);
+			}
+		} else {
+			glBlendFuncB = BlendFactor::CONSTANT_COLOR;
+			blendState.setBlendColorGE(fixB, constantAlpha);
+		}
+	} else if (blendFuncA == GE_SRCBLEND_FIXA || blendFuncB == GE_DSTBLEND_FIXB) {
 		if (glBlendFuncA == BlendFactor::INVALID && glBlendFuncB != BlendFactor::INVALID) {
 			// Can use blendcolor trivially.
 			blendState.setBlendColor(fixA, constantAlpha);

GPU/Common/GPUStateUtils.h

@@ -68,6 +68,11 @@ int ClearModeChannelMask();
 ReplaceAlphaType ReplaceAlphaWithStencil(ReplaceBlendType replaceBlend);
 ReplaceBlendType ReplaceBlendWithShader(GEBufferFormat bufferFormat);
 
+// Fixed color blends (with FIXA as the src factor) that REPLACE_BLEND_PRE_SRC computes like the GE.
+inline bool IsGEExactFixedBlend(GEBlendDstFactor funcB, GEBlendMode eq) {
+	return funcB == GE_DSTBLEND_FIXB && eq == GE_BLENDMODE_MUL_AND_ADD;
+}
+
 // This is for the fallback path if real logic ops are not available.
 SimulateLogicOpType SimulateLogicOpShaderTypeIfNeeded();
 
@@ -143,6 +148,8 @@ struct GenericBlendState {
 
 	bool useBlendColor;
 	u32 blendColor;
+	// The color channels are a GE fixed blend factor, applied as (2c+1)/512 like the GE does.
+	bool blendColorGE;
 
 	void setFactors(BlendFactor srcC, BlendFactor dstC, BlendFactor srcA, BlendFactor dstA) {
 		srcColor = srcC;
@@ -157,11 +164,16 @@ struct GenericBlendState {
 	void setBlendColor(uint32_t color, uint8_t alpha) {
 		blendColor = color | ((uint32_t)alpha << 24);
 		useBlendColor = true;
+		blendColorGE = false;
+	}
+	void setBlendColorGE(uint32_t color, uint8_t alpha) {
+		setBlendColor(color, alpha);
+		blendColorGE = true;
 	}
 	void defaultBlendColor(uint8_t alpha) {
-		blendColor = 0xFFFFFF | ((uint32_t)alpha << 24);
-		useBlendColor = true;
+		setBlendColor(0xFFFFFF, alpha);
 	}
+	void blendColorToFloat4(float out[4]) const;
 
 	void Log();
 };

GPU/D3D11/StateMappingD3D11.cpp

@@ -192,7 +192,7 @@ void DrawEngineD3D11::ApplyDrawState(int prim) {
 				}
 				dynState_.useBlendColor = blendState.useBlendColor;
 				if (blendState.useBlendColor) {
-					dynState_.blendColor = blendState.blendColor;
+					blendState.blendColorToFloat4(dynState_.blendColor);
 				}
 			} else {
 				keys_.blend.blendEnable = false;
@@ -437,8 +437,7 @@ void DrawEngineD3D11::ApplyDrawStateLate(bool applyStencilRef, uint8_t stencilRe
 	}
 	if (gstate_c.IsDirty(DIRTY_BLEND_STATE)) {
 		// Need to do this AFTER ApplyTexture because the process of depalettization can ruin the blend state.
-		float blendColor[4];
-		Uint8x4ToFloat4(blendColor, dynState_.blendColor);
+		const float *blendColor = dynState_.blendColor;
 		if (device1_) {
 			context1_->OMSetBlendState(blendState1_, blendColor, 0xFFFFFFFF);
 		} else {

GPU/D3D11/StateMappingD3D11.h

@@ -65,7 +65,7 @@ struct D3D11StateKeys {
 struct D3D11DynamicState {
 	int topology;
 	bool useBlendColor;
-	uint32_t blendColor;
+	float blendColor[4];
 	bool useStencil;
 	uint8_t stencilRef;
 	Draw::Viewport viewport;

GPU/GLES/StateMappingGLES.cpp

@@ -180,9 +180,8 @@ void DrawEngineGLES::ApplyDrawState(int prim) {
 					gstate_c.Dirty(DIRTY_SHADERBLEND);
 				}
 				if (blendState.useBlendColor) {
-					uint32_t color = blendState.blendColor;
 					float col[4];
-					Uint8x4ToFloat4(col, color);
+					blendState.blendColorToFloat4(col);
 					renderManager->SetBlendFactor(col);
 				}
 			}

GPU/Vulkan/StateMappingVulkan.cpp

@@ -173,7 +173,7 @@ void DrawEngineVulkan::ConvertStateToVulkanKey(FramebufferManagerVulkan &fbManag
 				}
 				dynState.useBlendColor = blendState.useBlendColor;
 				if (blendState.useBlendColor) {
-					dynState.blendColor = blendState.blendColor;
+					blendState.blendColorToFloat4(dynState.blendColor);
 				}
 			} else {
 				key.blendEnable = false;

GPU/Vulkan/StateMappingVulkan.h

@@ -15,7 +15,7 @@ struct VulkanDynamicState {
 	VkViewport viewport;
 	ScissorRect scissor;
 	bool useBlendColor;
-	uint32_t blendColor;
+	float blendColor[4];
 	bool useStencil;
 	uint8_t stencilRef;
 	uint8_t stencilWriteMask;

headless/Headless.cpp

@@ -1070,6 +1070,7 @@ int main(int argc, const char* argv[]) {
 	coreParameter.mountRoot = mountRoot.empty() ? Path() : Path(mountRoot);
 	coreParameter.startBreak = false;
 	coreParameter.headLess = true;
+	coreParameter.dumpCompat = cmdLineOptions.dumpCompat.value_or(false);
 	coreParameter.loadGameConfigs = false;
 	coreParameter.renderScaleFactor = cmdLineOptions.resolutionScale.value_or(1);
 	coreParameter.renderWidth = 480 * coreParameter.renderScaleFactor;