Add support for binding the depth buffer as a color target. Fixes Kurohyo depth sorting

Common/GPU/Shader.cpp

@@ -93,7 +93,7 @@ void ShaderLanguageDesc::Init(ShaderLanguage lang) {
 			fragColor0 = "outfragment.target";
 			fragColor1 = "outfragment.target1";
 		} else {
-			fragColor0 = "target";
+			fragColor0 = "outfragment.target";
 		}
 		varying_fs = "in";
 		varying_vs = "out";

GPU/Common/DepalettizeShaderCommon.cpp

@@ -145,8 +145,9 @@ void GenerateDepalShader300(char *buffer, GEBufferFormat pixelFormat, ShaderLang
 	}
 
 	float texturePixels = 256;
-	if (clutFormat != GE_CMODE_32BIT_ABGR8888)
+	if (clutFormat != GE_CMODE_32BIT_ABGR8888) {
 		texturePixels = 512;
+	}
 
 	if (shift) {
 		WRITE(p, "  index = (int(uint(index) >> uint(%i)) & 0x%02x)", shift, mask);

GPU/Common/FragmentShaderGenerator.cpp

@@ -90,6 +90,7 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 	bool shaderDepal = id.Bit(FS_BIT_SHADER_DEPAL) && !texture3D;  // combination with texture3D not supported. Enforced elsewhere too.
 	bool bgraTexture = id.Bit(FS_BIT_BGRA_TEXTURE);
 	bool colorWriteMask = id.Bit(FS_BIT_COLOR_WRITEMASK) && compat.bitwiseOps;
+	bool colorToDepth = id.Bit(FS_BIT_COLOR_TO_DEPTH);
 
 	GEComparison alphaTestFunc = (GEComparison)id.Bits(FS_BIT_ALPHA_TEST_FUNC, 3);
 	GEComparison colorTestFunc = (GEComparison)id.Bits(FS_BIT_COLOR_TEST_FUNC, 2);
@@ -122,7 +123,7 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 	bool readFramebufferTex = readFramebuffer && !gstate_c.Supports(GPU_SUPPORTS_ANY_FRAMEBUFFER_FETCH);
 
 	bool needFragCoord = readFramebuffer || gstate_c.Supports(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT);
-	bool writeDepth = gstate_c.Supports(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT);
+	bool writeDepth = gstate_c.Supports(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT) || colorToDepth;
 
 	if (shaderDepal && !doTexture) {
 		*errorString = "depal requires a texture";
@@ -135,6 +136,11 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 	}
 
 	if (compat.shaderLanguage == ShaderLanguage::GLSL_VULKAN) {
+		if (colorToDepth) {
+			WRITE(p, "precision highp int;\n");
+			WRITE(p, "precision highp float;\n");
+		}
+
 		if (useDiscardStencilBugWorkaround && !gstate_c.Supports(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT)) {
 			WRITE(p, "layout (depth_unchanged) out float gl_FragDepth;\n");
 		}
@@ -274,9 +280,16 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 				WRITE(p, "  float depth : SV_Depth;\n");
 			}
 			WRITE(p, "};\n");
+		} else if (compat.shaderLanguage == HLSL_D3D9) {
+			WRITE(p, "struct PS_OUT {\n");
+			WRITE(p, "  vec4 target : COLOR;\n");
+			if (writeDepth) {
+				WRITE(p, "  float depth : DEPTH;\n");
+			}
+			WRITE(p, "};\n");
 		}
 	} else if (ShaderLanguageIsOpenGL(compat.shaderLanguage)) {
-		if ((shaderDepal || colorWriteMask) && gl_extensions.IsGLES) {
+		if ((shaderDepal || colorWriteMask || colorToDepth) && gl_extensions.IsGLES) {
 			WRITE(p, "precision highp int;\n");
 		}
 
@@ -441,8 +454,12 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 			WRITE(p, "  float gl_FragDepth;\n");
 		}
 	} else if (compat.shaderLanguage == HLSL_D3D9) {
-		WRITE(p, "vec4 main( PS_IN In ) : COLOR {\n");
+		WRITE(p, "PS_OUT main( PS_IN In ) {\n");
+		WRITE(p, "  PS_OUT outfragment;\n");
 		WRITE(p, "  vec4 target;\n");
+		if (colorToDepth) {
+			WRITE(p, "  float gl_FragDepth;\n");
+		}
 	} else {
 		WRITE(p, "void main() {\n");
 	}
@@ -1034,7 +1051,9 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 
 	// Final color computed - apply color write mask.
 	// TODO: Maybe optimize to only do math on the affected channels?
-	// Or .. meh.
+	// Or .. meh. That would require more shader bits. Though we could
+	// of course optimize for the common mask 0xF00000, though again, blue-to-alpha
+	// does a better job with that.
 	if (colorWriteMask) {
 		WRITE(p, "  highp uint v32 = packUnorm4x8(%s);\n", compat.fragColor0);
 		WRITE(p, "  highp uint d32 = packUnorm4x8(destColor);\n");
@@ -1047,6 +1066,22 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 		WRITE(p, "  %s = vec4(0.0, 0.0, 0.0, %s.z);  // blue to alpha\n", compat.fragColor0, compat.fragColor0);
 	}
 
+	if (colorToDepth) {
+		DepthScaleFactors factors = GetDepthScaleFactors();
+
+		if (compat.bitwiseOps) {
+			WRITE(p, "  highp float depthValue = float(int(%s.x * 31.99) | (int(%s.y * 63.99) << 5) | (int(%s.z * 31.99) << 11)) / 65535.0;\n", "v", "v", "v"); // compat.fragColor0, compat.fragColor0, compat.fragColor0);
+		} else {
+			// D3D9-compatible alternative
+			WRITE(p, "  highp float depthValue = (floor(%s.x * 31.99) + floor(%s.y * 63.99) * 32.0 + floor(%s.z * 31.99) * 2048.0) / 65535.0;\n", "v", "v", "v"); // compat.fragColor0, compat.fragColor0, compat.fragColor0);
+		}
+		if (factors.scale != 1.0 || factors.offset != 0.0) {
+			WRITE(p, "  gl_FragDepth = (depthValue / %f) + %f;\n", factors.scale / 65535.0f, factors.offset);
+		} else {
+			WRITE(p, "  gl_FragDepth = depthValue;\n");
+		}
+	}
+
 	if (gstate_c.Supports(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT)) {
 		const double scale = DepthSliceFactor() * 65535.0;
 
@@ -1060,7 +1095,7 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 				WRITE(p, "  z = floor(z * %f) * (1.0 / %f);\n", scale, scale);
 			}
 		} else {
-			WRITE(p, "  z = (1.0/65535.0) * floor(z * 65535.0);\n");
+			WRITE(p, "  z = (1.0 / 65535.0) * floor(z * 65535.0);\n");
 		}
 		WRITE(p, "  gl_FragDepth = z;\n");
 	} else if (useDiscardStencilBugWorkaround) {
@@ -1071,13 +1106,11 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 		WRITE(p, "  gl_FragDepth = gl_FragCoord.z;\n");
 	}
 
-	if (compat.shaderLanguage == HLSL_D3D11) {
+	if (compat.shaderLanguage == HLSL_D3D11 || compat.shaderLanguage == HLSL_D3D9) {
 		if (writeDepth) {
 			WRITE(p, "  outfragment.depth = gl_FragDepth;\n");
 		}
 		WRITE(p, "  return outfragment;\n");
-	} else if (compat.shaderLanguage == HLSL_D3D9) {
-		WRITE(p, "  return target;\n");
 	}
 
 	WRITE(p, "}\n");

GPU/Common/FramebufferManagerCommon.cpp

@@ -275,14 +275,22 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 	gstate_c.SetCurRTOffset(0, 0);
 	bool vfbFormatChanged = false;
 
+	if (params.fb_address == params.z_address) {
+		// Most likely Z will not be used in this pass, as that would wreak havoc (undefined behavior for sure)
+		// We probably don't need to do anything about that, but let's log it.
+		WARN_LOG_ONCE(color_equal_z, G3D, "Framebuffer bound with color addr == z addr, likely will not use Z in this pass: %08x", params.fb_address);
+	}
+
+	FramebufferRenderMode mode = FB_MODE_NORMAL;
+
 	// Find a matching framebuffer
 	VirtualFramebuffer *vfb = nullptr;
 	for (size_t i = 0; i < vfbs_.size(); ++i) {
 		VirtualFramebuffer *v = vfbs_[i];
 
 		const u32 bpp = v->format == GE_FORMAT_8888 ? 4 : 2;
 
-		if (v->fb_address == params.fb_address) {
+		if (params.fb_address == v->fb_address) {
 			vfb = v;
 			// Update fb stride in case it changed
 			if (vfb->fb_stride != params.fb_stride) {
@@ -311,6 +319,19 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 				vfb->height = drawing_height;
 			}
 			break;
+		} else if (params.fb_address == v->z_address && params.fmt != GE_FORMAT_8888 && params.fb_stride == v->z_stride) {
+			// Looks like the game might be intending to use color to write directly to a Z buffer.
+			// This is seen in Kuroyou 2.
+
+			// Ignore this in this loop, BUT, we do a lookup in the depth tracking afterwards to
+			// make sure we get the latest one.
+			WARN_LOG_ONCE(color_matches_z, G3D, "Color framebuffer bound at %08x with likely intent to write explicit Z values using color. fmt = %s", params.fb_address, GeBufferFormatToString(params.fmt));
+			// Seems impractical to use the other 16-bit formats for this due to the limited control over alpha,
+			// so we'll simply only support 565.
+			if (params.fmt == GE_FORMAT_565) {
+				mode = FB_MODE_COLOR_TO_DEPTH;
+				break;
+			}
 		} else if (v->fb_stride == params.fb_stride && v->format == params.fmt) {
 			u32 v_fb_first_line_end_ptr = v->fb_address + v->fb_stride * 4;  // This should be * bpp, but leaving like this until after 1.13 to be safe. The God of War games use this for shadows.
 			u32 v_fb_end_ptr = v->fb_address + v->fb_stride * v->height * bpp;
@@ -345,6 +366,25 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 		}
 	}
 
+	if (mode == FB_MODE_COLOR_TO_DEPTH) {
+		// Lookup in the depth tracking to find which VFB has the latest version of this Z buffer.
+		// Then bind it in color-to-depth mode.
+		//
+		// We are going to do this by having a special render mode where we take color and move to
+		// depth in the fragment shader, and set color writes to off.
+		//
+		// We'll need a special fragment shader flag to convert color to depth.
+
+		for (auto &depth : this->trackedDepthBuffers_) {
+			if (depth->z_address == params.fb_address && depth->z_stride == params.fb_stride) {
+				// Found the matching depth buffer. Use this vfb.
+				vfb = depth->vfb;
+			}
+		}
+	}
+
+	gstate_c.SetFramebufferRenderMode(mode);
+
 	if (vfb) {
 		if ((drawing_width != vfb->bufferWidth || drawing_height != vfb->bufferHeight)) {
 			// Even if it's not newly wrong, if this is larger we need to resize up.

GPU/Common/FramebufferManagerCommon.h

@@ -224,7 +224,7 @@ class FramebufferManagerCommon {
 	void SetDisplayFramebuffer(u32 framebuf, u32 stride, GEBufferFormat format);
 	void DestroyFramebuf(VirtualFramebuffer *v);
 
-	VirtualFramebuffer *DoSetRenderFrameBuffer(const FramebufferHeuristicParams &params, u32 skipDrawReason);	
+	VirtualFramebuffer *DoSetRenderFrameBuffer(const FramebufferHeuristicParams &params, u32 skipDrawReason);
 	VirtualFramebuffer *SetRenderFrameBuffer(bool framebufChanged, int skipDrawReason) {
 		// Inlining this part since it's so frequent.
 		if (!framebufChanged && currentRenderVfb_) {

GPU/Common/GPUStateUtils.cpp

@@ -1018,6 +1018,16 @@ void ConvertMaskState(GenericMaskState &maskState, bool allowFramebufferRead) {
 		return;
 	}
 
+	if (gstate_c.renderMode == FB_MODE_COLOR_TO_DEPTH) {
+		// Suppress color writes entirely in this mode.
+		maskState.applyFramebufferRead = false;
+		maskState.rgba[0] = false;
+		maskState.rgba[1] = false;
+		maskState.rgba[2] = false;
+		maskState.rgba[3] = false;
+		return;
+	}
+
 	// Invert to convert masks from the PSP's format where 1 is don't draw to PC where 1 is draw.
 	uint32_t colorMask = ~((gstate.pmskc & 0xFFFFFF) | (gstate.pmska << 24));
 

GPU/Common/GPUStateUtils.h

@@ -81,14 +81,17 @@ struct ViewportAndScissor {
 void ConvertViewportAndScissor(bool useBufferedRendering, float renderWidth, float renderHeight, int bufferWidth, int bufferHeight, ViewportAndScissor &out);
 float ToScaledDepthFromIntegerScale(float z);
 
-// Use like this: (z - offset) * scale
 struct DepthScaleFactors {
 	float offset;
 	float scale;
 
 	float Apply(float z) const {
 		return (z - offset) * scale;
 	}
+
+	float ApplyInverse(float z) const {
+		return (z / scale) + offset;
+	}
 };
 DepthScaleFactors GetDepthScaleFactors();
 

GPU/Common/ShaderId.cpp

@@ -7,6 +7,7 @@
 #include "Core/Config.h"
 
 #include "GPU/ge_constants.h"
+#include "GPU/GPU.h"
 #include "GPU/GPUState.h"
 #include "GPU/Common/GPUStateUtils.h"
 #include "GPU/Common/ShaderId.h"
@@ -239,6 +240,8 @@ std::string FragmentShaderDesc(const FShaderID &id) {
 	if (id.Bit(FS_BIT_COLOR_AGAINST_ZERO)) desc << "ColorTest0 " << alphaTestFuncs[id.Bits(FS_BIT_COLOR_TEST_FUNC, 2)] << " ";  // first 4 match;
 	else if (id.Bit(FS_BIT_COLOR_TEST)) desc << "ColorTest " << alphaTestFuncs[id.Bits(FS_BIT_COLOR_TEST_FUNC, 2)] << " ";  // first 4 match
 
+	if (id.Bit(FS_BIT_COLOR_TO_DEPTH)) desc << "ColorToDepth ";
+
 	return desc.str();
 }
 
@@ -261,6 +264,7 @@ void ComputeFragmentShaderID(FShaderID *id_out, const Draw::Bugs &bugs) {
 		bool doFlatShading = gstate.getShadeMode() == GE_SHADE_FLAT;
 		bool useShaderDepal = gstate_c.useShaderDepal;
 		bool colorWriteMask = IsColorWriteMaskComplex(gstate_c.allowFramebufferRead);
+		bool colorToDepth = gstate_c.renderMode == FramebufferRenderMode::FB_MODE_COLOR_TO_DEPTH;
 
 		// Note how we here recompute some of the work already done in state mapping.
 		// Not ideal! At least we share the code.
@@ -292,6 +296,8 @@ void ComputeFragmentShaderID(FShaderID *id_out, const Draw::Bugs &bugs) {
 			id.SetBit(FS_BIT_3D_TEXTURE, gstate_c.curTextureIs3D);
 		}
 
+		id.SetBit(FS_BIT_COLOR_TO_DEPTH, colorToDepth);
+
 		id.SetBit(FS_BIT_LMODE, lmode);
 		if (enableAlphaTest) {
 			// 5 bits total.

GPU/Common/ShaderId.h

@@ -94,6 +94,7 @@ enum FShaderBit : uint8_t {
 	FS_BIT_NO_DEPTH_CANNOT_DISCARD_STENCIL = 49,
 	FS_BIT_COLOR_WRITEMASK = 50,
 	FS_BIT_3D_TEXTURE = 51,
+	FS_BIT_COLOR_TO_DEPTH = 52,
 };
 
 static inline FShaderBit operator +(FShaderBit bit, int i) {

GPU/Common/TextureCacheCommon.cpp

@@ -261,6 +261,10 @@ SamplerCacheKey TextureCacheCommon::GetSamplingParams(int maxLevel, const TexCac
 		}
 	}
 
+	if (gstate_c.renderMode == FB_MODE_COLOR_TO_DEPTH) {
+		forceFiltering = TEX_FILTER_FORCE_NEAREST;
+	}
+
 	switch (forceFiltering) {
 	case TEX_FILTER_AUTO:
 		break;
@@ -2131,8 +2135,14 @@ bool TextureCacheCommon::PrepareBuildTexture(BuildTexturePlan &plan, TexCacheEnt
 	}
 
 	// Don't scale the PPGe texture.
-	if (entry->addr > 0x05000000 && entry->addr < PSP_GetKernelMemoryEnd())
+	if (entry->addr > 0x05000000 && entry->addr < PSP_GetKernelMemoryEnd()) {
+		plan.scaleFactor = 1;
+	}
+
+	// Don't upscale textures in color-to-depth mode.
+	if (gstate_c.renderMode == FB_MODE_COLOR_TO_DEPTH) {
 		plan.scaleFactor = 1;
+	}
 
 	if ((entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) != 0 && plan.scaleFactor != 1 && plan.slowScaler) {
 		// Remember for later that we /wanted/ to scale this texture.

GPU/D3D11/StateMappingD3D11.cpp

@@ -283,7 +283,14 @@ void DrawEngineD3D11::ApplyDrawState(int prim) {
 		GenericStencilFuncState stencilState;
 		ConvertStencilFuncState(stencilState);
 
-		if (gstate.isModeClear()) {
+		if (gstate_c.renderMode == FB_MODE_COLOR_TO_DEPTH) {
+			// Enforce plain depth writing.
+			keys_.depthStencil.value = 0;
+			keys_.depthStencil.depthTestEnable = true;
+			keys_.depthStencil.depthWriteEnable = true;
+			keys_.depthStencil.stencilTestEnable = false;
+			keys_.depthStencil.depthCompareOp = D3D11_COMPARISON_ALWAYS;
+		} else if (gstate.isModeClear()) {
 			keys_.depthStencil.value = 0;
 			keys_.depthStencil.depthTestEnable = true;
 			keys_.depthStencil.depthCompareOp = D3D11_COMPARISON_ALWAYS;

GPU/Directx9/StateMappingDX9.cpp

@@ -197,7 +197,14 @@ void DrawEngineDX9::ApplyDrawState(int prim) {
 		ConvertStencilFuncState(stencilState);
 
 		// Set Stencil/Depth
-		if (gstate.isModeClear()) {
+
+		if (gstate_c.renderMode == FB_MODE_COLOR_TO_DEPTH) {
+			// Enforce plain depth writing.
+			dxstate.depthTest.enable();
+			dxstate.depthFunc.set(D3DCMP_ALWAYS);
+			dxstate.depthWrite.set(true);
+			dxstate.stencilTest.disable();
+		} else if (gstate.isModeClear()) {
 			// Depth Test
 			dxstate.depthTest.enable();
 			dxstate.depthFunc.set(D3DCMP_ALWAYS);

GPU/GLES/StateMappingGLES.cpp

@@ -250,7 +250,11 @@ void DrawEngineGLES::ApplyDrawState(int prim) {
 		GenericStencilFuncState stencilState;
 		ConvertStencilFuncState(stencilState);
 
-		if (gstate.isModeClear()) {
+		if (gstate_c.renderMode == FB_MODE_COLOR_TO_DEPTH) {
+			// Enforce plain depth writing.
+			renderManager->SetStencilDisabled();
+			renderManager->SetDepth(true, true, GL_ALWAYS);
+		} else if (gstate.isModeClear()) {
 			// Depth Test
 			if (gstate.isClearModeDepthMask()) {
 				framebufferManager_->SetDepthUpdated();

GPU/GPU.h

@@ -24,6 +24,11 @@ class GPUInterface;
 class GPUDebugInterface;
 class GraphicsContext;
 
+enum FramebufferRenderMode {
+	FB_MODE_NORMAL = 0,
+	FB_MODE_COLOR_TO_DEPTH = 1,
+};
+
 enum SkipDrawReasonFlags {
 	SKIPDRAW_SKIPFRAME = 1,
 	SKIPDRAW_NON_DISPLAYED_FB = 2,   // Skip drawing to FBO:s that have not been displayed.