Depal: Get depth scale factors dynamically

GPU/Common/DepalettizeShaderCommon.cpp

@@ -51,13 +51,10 @@ void GenerateDepalShader300(ShaderWriter &writer, const DepalConfig &config) {
 	writer.C("  vec2 texcoord = v_texcoord;\n");
 
 	// Implement the swizzle we need to simulate, if a game uses 8888 framebuffers and any other mode than "6" to access depth textures.
-	// This implements the "2" mode swizzle (it fixes up the Y direction but not X. See comments on issue #15898)
+	// This implements the "2" mode swizzle (it fixes up the Y direction but not X. See comments on issue #15898, Tantalus games)
 	// NOTE: This swizzle can be made to work with any power-of-2 resolution scaleFactor by shifting
 	// the bits around, but not sure how to handle 3x scaling. For now this is 1x-only (rough edges at higher resolutions).
 	if (config.bufferFormat == GE_FORMAT_DEPTH16) {
-		DepthScaleFactors factors = GetDepthScaleFactors(gstate_c.UseFlags());
-		writer.ConstFloat("z_scale", factors.ScaleU16());
-		writer.ConstFloat("z_offset", factors.Offset());
 		if (config.depthUpperBits == 0x2) {
 			writer.C(R"(
   int x = int((texcoord.x / scaleFactor) * texSize.x);

GPU/Common/Draw2D.cpp

@@ -39,23 +39,28 @@ static const SamplerDef samplers[1] = {
 	{ 0, "tex", SamplerFlags::ARRAY_ON_VULKAN },
 };
 
-const UniformDef g_draw2Duniforms[2] = {
+const UniformDef g_draw2Duniforms[4] = {
 	{ "vec2", "texSize", 0 },
 	{ "float", "scaleFactor", 1},
+	{ "float", "z_scale", 2 },
+	{ "float", "z_offset", 3 },
 };
 
 struct Draw2DUB {
 	float texSizeX;
 	float texSizeY;
 	float scaleFactor;
+	float zScale;
+	float zOffset;
 };
 
 const UniformBufferDesc draw2DUBDesc{ sizeof(Draw2DUB), {
 	{ "texSize", -1, 0, UniformType::FLOAT2, 0 },
 	{ "scaleFactor", -1, 1, UniformType::FLOAT1, 8 },
+	{ "z_scale", -1, 1, UniformType::FLOAT1, 12 },
+	{ "z_offset", -1, 1, UniformType::FLOAT1, 16 },
 } };
 
-
 Draw2DPipelineInfo GenerateDraw2DCopyColorFs(ShaderWriter &writer) {
 	writer.DeclareSamplers(samplers);
 	writer.BeginFSMain(Slice<UniformDef>::empty(), varyings);
@@ -95,8 +100,8 @@ Draw2DPipelineInfo GenerateDraw2DCopyDepthFs(ShaderWriter &writer) {
 	writer.EndFSMain("outColor");
 
 	return Draw2DPipelineInfo{
-		"draw2d_copy_depth",
+		"draw2d_copy_r16_to_depth",
-		RASTER_DEPTH,
+		RASTER_DEPTH,  // Unused in this case, I think.
 		RASTER_DEPTH,
 	};
 }
@@ -319,6 +324,10 @@ void Draw2D::DrawStrip2D(Draw::Texture *tex, const Draw2DVertex *verts, int vert
 	ub.texSizeY = tex ? tex->Height() : texH;
 	ub.scaleFactor = (float)scaleFactor;
 
+	DepthScaleFactors zScaleFactors = GetDepthScaleFactors(gstate_c.UseFlags());
+	ub.zScale = zScaleFactors.Scale();
+	ub.zOffset = zScaleFactors.Offset();
+
 	draw_->BindPipeline(pipeline->pipeline);
 	draw_->UpdateDynamicUniformBuffer(&ub, sizeof(ub));
 
@@ -356,7 +365,7 @@ Draw2DPipeline *FramebufferManagerCommon::Get2DPipeline(Draw2DShader shader) {
 		pipeline = draw2DPipelineColorRect2Lin_;
 		break;
 
-	case DRAW2D_COPY_DEPTH:
+	case DRAW2D_COPY_R16_TO_DEPTH:
 		if (!draw_->GetDeviceCaps().fragmentShaderDepthWriteSupported) {
 			// Can't do it
 			return nullptr;

GPU/Common/Draw2D.h

@@ -13,7 +13,7 @@ struct Draw2DVertex {
 
 enum Draw2DShader {
 	DRAW2D_COPY_COLOR,
-	DRAW2D_COPY_DEPTH,
+	DRAW2D_COPY_R16_TO_DEPTH,
 	DRAW2D_565_TO_DEPTH,
 	DRAW2D_565_TO_DEPTH_DESWIZZLE,
 	DRAW2D_COPY_COLOR_RECT2LIN,
@@ -21,7 +21,7 @@ enum Draw2DShader {
 
 inline RasterChannel Draw2DSourceChannel(Draw2DShader shader) {
 	switch (shader) {
-	case DRAW2D_COPY_DEPTH:
+	case DRAW2D_COPY_R16_TO_DEPTH:
 		return RASTER_DEPTH;
 	case DRAW2D_COPY_COLOR:
 	case DRAW2D_565_TO_DEPTH:
@@ -38,7 +38,7 @@ struct Draw2DPipelineInfo {
 	Slice<SamplerDef> samplers;
 };
 
-extern const UniformDef g_draw2Duniforms[2];
+extern const UniformDef g_draw2Duniforms[4];
 
 struct Draw2DPipeline {
 	Draw::Pipeline *pipeline;

GPU/Common/FramebufferManagerCommon.cpp

@@ -974,7 +974,7 @@ void FramebufferManagerCommon::BlitFramebufferDepth(VirtualFramebuffer *src, Vir
 
 	// Some GPUs can copy depth but only if stencil gets to come along for the ride. We only want to use this if there is no blit functionality.
 	if (useRaster) {
-		BlitUsingRaster(src->fbo, 0, 0, w, h, dst->fbo, 0, 0, w, h, false, dst->renderScaleFactor, Get2DPipeline(Draw2DShader::DRAW2D_COPY_DEPTH), "BlitDepthRaster");
+		BlitUsingRaster(src->fbo, 0, 0, w, h, dst->fbo, 0, 0, w, h, false, dst->renderScaleFactor, Get2DPipeline(Draw2DShader::DRAW2D_COPY_R16_TO_DEPTH), "BlitDepthRaster");
 	} else if (useCopy) {
 		draw_->CopyFramebufferImage(src->fbo, 0, 0, 0, 0, dst->fbo, 0, 0, 0, 0, w, h, 1, Draw::FB_DEPTH_BIT, "CopyFramebufferDepth");
 		RebindFramebuffer("After BlitFramebufferDepth");
@@ -3036,7 +3036,7 @@ void FramebufferManagerCommon::DrawActiveTexture(float x, float y, float w, floa
 	// Rearrange to strip form.
 	std::swap(coord[2], coord[3]);
 
-	draw2D_.DrawStrip2D(nullptr, coord, 4, (flags & DRAWTEX_LINEAR) != 0, Get2DPipeline((flags & DRAWTEX_DEPTH) ? DRAW2D_COPY_DEPTH : DRAW2D_COPY_COLOR));
+	draw2D_.DrawStrip2D(nullptr, coord, 4, (flags & DRAWTEX_LINEAR) != 0, Get2DPipeline((flags & DRAWTEX_DEPTH) ? DRAW2D_COPY_R16_TO_DEPTH : DRAW2D_COPY_COLOR));
 
 	gstate_c.Dirty(DIRTY_ALL_RENDER_STATE);
 }
@@ -3135,7 +3135,7 @@ void FramebufferManagerCommon::BlitFramebuffer(VirtualFramebuffer *dst, int dstX
 		draw_->BlitFramebuffer(src->fbo, srcX1, srcY1, srcX2, srcY2, dst->fbo, dstX1, dstY1, dstX2, dstY2,
 			channel == RASTER_COLOR ? Draw::FB_COLOR_BIT : Draw::FB_DEPTH_BIT, Draw::FB_BLIT_NEAREST, tag);
 	} else {
-		Draw2DPipeline *pipeline = Get2DPipeline(channel == RASTER_COLOR ? DRAW2D_COPY_COLOR : DRAW2D_COPY_DEPTH);
+		Draw2DPipeline *pipeline = Get2DPipeline(channel == RASTER_COLOR ? DRAW2D_COPY_COLOR : DRAW2D_COPY_R16_TO_DEPTH);
 		Draw::Framebuffer *srcFBO = src->fbo;
 		if (src == dst) {
 			Draw::Framebuffer *tempFBO = GetTempFBO(TempFBO::BLIT, src->renderWidth, src->renderHeight);