Merge pull request #22898 from hrydgard/depal-pixel-mapped

Depal: Skip the bilinear filter on exactly pixel-mapped draws

GPU/Common/ShaderUniforms.cpp

@@ -180,7 +180,14 @@ void BaseUpdateUniforms(UB_VS_FS_Base *ub, uint64_t dirtyUniforms, bool useBuffe
 	}
 }
 
+void DirtyDepalIfPixelMappedChanged(bool pixelMapped) {
+	if (gstate_c.shaderDepalMode != ShaderDepalMode::OFF && pixelMapped != gstate_c.depalPixelMapped) {
+		gstate_c.Dirty(DIRTY_DEPAL);
+	}
+}
+
 uint32_t PackDepalBits(bool pixelMapped) {
+	gstate_c.depalPixelMapped = pixelMapped;
 	const int indexMask = gstate.getClutIndexMask();
 	const int indexShift = gstate.getClutIndexShift();
 	const int indexOffset = gstate.getClutIndexStartPos() >> 4;

GPU/Common/ShaderUniforms.h

@@ -110,6 +110,8 @@ void LightUpdateUniforms(UB_VS_Lights *ub, uint64_t dirtyUniforms);
 
 uint32_t PackLightControlBits();
 uint32_t PackDepalBits(bool pixelMapped);
+// Call before reading the dirty uniforms: the depal uniform depends on pixelMapped, which changes per draw.
+void DirtyDepalIfPixelMappedChanged(bool pixelMapped);
 
 void UpdateFogCoef(const GEState &state, float fogCoef[2]);
 

GPU/Common/SoftwareTransformCommon.cpp

@@ -1051,26 +1051,37 @@ static bool ExpandRectangles(int vertexCount, int &numDecodedVerts, int vertsSiz
 		vScale /= gstate_c.curTextureHeight;
 	}
 
+	// Look for pixel mapping, two ways at once:
+	// - pixelMapped (with smart 2D filtering): each rect is 1:1 in at least one direction. Turns the texture filter
+	//   off and the sprite border fix with it. This one can change what's on screen.
+	// - exactlyPixelMapped: each rect is 1:1 in both directions at a whole-texel offset, so every pixel center lands
+	//   on a texel center and bilinear gives the nearest texel anyway. Shader depal filters by hand, with four
+	//   texel and four palette lookups per pixel, a lot for a slow GPU on a full-screen pass (DBZ Tenkaichi Tag
+	//   Team has two per frame), so for it this turns the filter off without changing anything. Positions are only
+	//   in pixels in through mode.
 	bool pixelMapped = g_Config.bSmart2DTexFiltering && !gstate_c.textureIsVideo;
-	if (pixelMapped) {
-		// Check for pixel mapping. If we find pixel mapping, abandon spriteBorderFix.
-		for (int i = 0; i < vertexCount; i += 2) {
-			const TransformedVertex &transVtxTL = transformed[indsIn[i + 0]];
-			const TransformedVertex &transVtxBR = transformed[indsIn[i + 1]];
-
-			if (pixelMapped) {
-				float dx = transVtxBR.x - transVtxTL.x;
-				float dy = transVtxBR.y - transVtxTL.y;
-				float du = transVtxBR.u - transVtxTL.u;
-				float dv = transVtxBR.v - transVtxTL.v;
-
-				// NOTE: We will accept it as pixel mapped if only one dimension is stretched. This fixes dialog frames in FFI.
-				// Though, there could be false positives in other games due to this. Let's see if it is a problem...
-				if (dx <= 0 || dy <= 0 || (dx != du && dy != dv)) {
-					pixelMapped = false;
-					break;
-				}
-			}
+	bool exactlyPixelMapped = throughmode && gstate_c.shaderDepalMode != ShaderDepalMode::OFF;
+	for (int i = 0; i < vertexCount && (pixelMapped || exactlyPixelMapped); i += 2) {
+		const TransformedVertex &transVtxTL = transformed[indsIn[i + 0]];
+		const TransformedVertex &transVtxBR = transformed[indsIn[i + 1]];
+		const float dx = transVtxBR.x - transVtxTL.x;
+		const float dy = transVtxBR.y - transVtxTL.y;
+		const float du = transVtxBR.u - transVtxTL.u;
+		const float dv = transVtxBR.v - transVtxTL.v;
+		if (dx <= 0 || dy <= 0) {
+			pixelMapped = false;
+			exactlyPixelMapped = false;
+			break;
+		}
+		// NOTE: We will accept it as pixel mapped if only one dimension is stretched. This fixes dialog frames in FFI.
+		// Though, there could be false positives in other games due to this. Let's see if it is a problem...
+		if (dx != du && dy != dv) {
+			pixelMapped = false;
+		}
+		const float offsetU = transVtxTL.u - transVtxTL.x;
+		const float offsetV = transVtxTL.v - transVtxTL.y;
+		if (dx != du || dy != dv || offsetU != floorf(offsetU) || offsetV != floorf(offsetV)) {
+			exactlyPixelMapped = false;
 		}
 	}
 
@@ -1095,6 +1106,10 @@ static bool ExpandRectangles(int vertexCount, int &numDecodedVerts, int vertsSiz
 		}
 	}
 
+	// The sprite border fix moves the UVs off the texel centers.
+	if (spriteBorderFixL != 0.0f || spriteBorderFixR != 0.0f || spriteBorderFixT != 0.0f || spriteBorderFixB != 0.0f) {
+		exactlyPixelMapped = false;
+	}
 
 	// Like the software renderer's ProcessRect: rectangles aren't clipped, one with a vertex behind the camera
 	// or outside the screen range is dropped whole. With depth clamp, a vertex past the near plane is exempt
@@ -1213,7 +1228,7 @@ static bool ExpandRectangles(int vertexCount, int &numDecodedVerts, int vertsSiz
 	}
 	numDecodedVerts = outVerts;
 	inds = newInds;
-	*pixelMappedExactly = pixelMapped;
+	*pixelMappedExactly = pixelMapped || exactlyPixelMapped;
 	*drawIndexCount = indsOut - newInds;
 	return true;
 }

GPU/D3D11/ShaderManagerD3D11.cpp

@@ -149,6 +149,7 @@ void ShaderManagerD3D11::ClearShaders() {
 }
 
 uint64_t ShaderManagerD3D11::UpdateUniforms(bool useBufferedRendering, bool pixelMapped) {
+	DirtyDepalIfPixelMappedChanged(pixelMapped);
 	uint64_t dirty = gstate_c.GetDirtyUniforms();
 	if (dirty != 0) {
 		D3D11_MAPPED_SUBRESOURCE map;

GPU/GLES/ShaderManagerGLES.cpp

@@ -717,6 +717,7 @@ Shader *ShaderManagerGLES::ApplyVertexShader(bool useHWTransform, u32 vertexType
 }
 
 LinkedShader *ShaderManagerGLES::ApplyFragmentShader(VShaderID VSID, Shader *vs, const ComputedPipelineState &pipelineState, ClipInfoFlags clipInfoFlags, bool pixelMapped) {
+	DirtyDepalIfPixelMappedChanged(pixelMapped);
 	uint64_t dirty = gstate_c.GetDirtyUniforms();
 	if (dirty) {
 		if (lastShader_)

GPU/GPUState.h

@@ -682,6 +682,8 @@ public:
 
 	ShaderDepalMode shaderDepalMode;
 	GEBufferFormat depalTextureFormat;
+	// The pixelMapped that the depal uniform was last packed with (it turns bilinear depal off).
+	bool depalPixelMapped;
 
 	void SetShaderDepal(ShaderDepalMode mode, GEBufferFormat texFormat = {}) {
 		if (mode != shaderDepalMode || (mode != ShaderDepalMode::OFF && texFormat != depalTextureFormat)) {

GPU/Metal/ShaderManagerMetal.mm

@@ -86,6 +86,7 @@ void ShaderManagerMetal::ClearShaders() {
 }
 
 uint64_t ShaderManagerMetal::UpdateUniforms(bool useBufferedRendering, bool pixelMapped) {
+	DirtyDepalIfPixelMappedChanged(pixelMapped);
 	uint64_t dirty = gstate_c.GetDirtyUniforms();
 	if (dirty != 0) {
 		if (dirty & DIRTY_BASE_UNIFORMS)

GPU/Vulkan/ShaderManagerVulkan.cpp

@@ -218,6 +218,7 @@ void ShaderManagerVulkan::ClearShaders() {
 }
 
 uint64_t ShaderManagerVulkan::UpdateUniforms(bool useBufferedRendering, bool pixelMapped) {
+	DirtyDepalIfPixelMappedChanged(pixelMapped);
 	uint64_t dirty = gstate_c.GetDirtyUniforms();
 	if (dirty != 0) {
 		if (dirty & DIRTY_BASE_UNIFORMS)