Merge pull request #6313 from unknownbrackets/gpu-minor

Implement in-shader blending on gles2 / gl2

GPU/GLES/FragmentShaderGenerator.cpp

@@ -283,10 +283,6 @@ bool ShouldUseShaderBlending() {
 	if (!gstate.isAlphaBlendEnabled()) {
 		return false;
 	}
-	// We can't blit on GLES2, so we don't support it.  We also want texelFetch (OpenGL 3.0+ / GLES3+.)
-	if (!gl_extensions.VersionGEThan(3, 0, 0) && !gl_extensions.GLES3) {
-		return false;
-	}
 	if (g_Config.iRenderingMode == FB_NON_BUFFERED_MODE) {
 		return false;
 	}
@@ -516,6 +512,9 @@ void GenerateFragmentShader(char *buffer) {
 		WRITE(p, "uniform sampler2D tex;\n");
 	if (!gstate.isModeClear() && ShouldUseShaderBlending()) {
 		if (!gl_extensions.NV_shader_framebuffer_fetch) {
+			if (!gl_extensions.VersionGEThan(3, 0, 0) && !gl_extensions.GLES3) {
+				WRITE(p, "uniform vec2 u_fbotexSize;\n");
+			}
 			WRITE(p, "uniform sampler2D fbotex;\n");
 		}
 		if (gstate.getBlendFuncA() == GE_SRCBLEND_FIXA) {
@@ -766,6 +765,8 @@ void GenerateFragmentShader(char *buffer) {
 			// TODO: EXT_shader_framebuffer_fetch on iOS 6, possibly others.
 			if (gl_extensions.NV_shader_framebuffer_fetch) {
 				WRITE(p, "  lowp vec4 destColor = gl_LastFragData[0];\n");
+			} else if (!gl_extensions.VersionGEThan(3, 0, 0) && !gl_extensions.GLES3) {
+				WRITE(p, "  lowp vec4 destColor = %s(fbotex, gl_FragCoord.xy * u_fbotexSize.xy);\n", texture);
 			} else {
 				WRITE(p, "  lowp vec4 destColor = texelFetch(fbotex, ivec2(gl_FragCoord.x, gl_FragCoord.y), 0);\n");
 			}

GPU/GLES/Framebuffer.cpp

@@ -1053,6 +1053,8 @@ void FramebufferManager::DoSetRenderFrameBuffer() {
 		gstate_c.curRTWidth = vfb->width;
 		gstate_c.curRTHeight = vfb->height;
 	}
+	gstate_c.curRTRenderWidth = vfb->renderWidth;
+	gstate_c.curRTRenderHeight = vfb->renderHeight;
 }
 
 void FramebufferManager::SetLineWidth() {

GPU/GLES/GLES_GPU.cpp

@@ -162,8 +162,8 @@ static const CommandTableEntry commandTable[] = {
 	{GE_CMD_STENCILTESTENABLE, FLAG_FLUSHBEFOREONCHANGE},
 	{GE_CMD_ALPHABLENDENABLE, FLAG_FLUSHBEFOREONCHANGE},
 	{GE_CMD_BLENDMODE, FLAG_FLUSHBEFOREONCHANGE},
-	{GE_CMD_BLENDFIXEDA, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTEONCHANGE, &GLES_GPU::Execute_BlendFixA},
-	{GE_CMD_BLENDFIXEDB, FLAG_FLUSHBEFOREONCHANGE | FLAG_EXECUTEONCHANGE, &GLES_GPU::Execute_BlendFixB},
+	{GE_CMD_BLENDFIXEDA, FLAG_FLUSHBEFOREONCHANGE},
+	{GE_CMD_BLENDFIXEDB, FLAG_FLUSHBEFOREONCHANGE},
 	{GE_CMD_MASKRGB, FLAG_FLUSHBEFOREONCHANGE},
 	{GE_CMD_MASKALPHA, FLAG_FLUSHBEFOREONCHANGE},
 	{GE_CMD_ZTEST, FLAG_FLUSHBEFOREONCHANGE},
@@ -1096,14 +1096,6 @@ void GLES_GPU::Execute_ColorRef(u32 op, u32 diff) {
 	shaderManager_->DirtyUniform(DIRTY_ALPHACOLORREF);
 }
 
-void GLES_GPU::Execute_BlendFixA(u32 op, u32 diff) {
-	shaderManager_->DirtyUniform(DIRTY_BLENDFIX);
-}
-
-void GLES_GPU::Execute_BlendFixB(u32 op, u32 diff) {
-	shaderManager_->DirtyUniform(DIRTY_BLENDFIX);
-}
-
 void GLES_GPU::Execute_WorldMtxNum(u32 op, u32 diff) {
 	// This is almost always followed by GE_CMD_WORLDMATRIXDATA.
 	const u32_le *src = (const u32_le *)Memory::GetPointer(currentList->pc + 4);
@@ -1634,11 +1626,7 @@ void GLES_GPU::ExecuteOpInternal(u32 op, u32 diff) {
 		break;
 
 	case GE_CMD_BLENDFIXEDA:
-		Execute_BlendFixA(op, diff);
-		break;
-
 	case GE_CMD_BLENDFIXEDB:
-		Execute_BlendFixB(op, diff);
 		break;
 
 	case GE_CMD_ALPHATESTENABLE:

GPU/GLES/GLES_GPU.h

@@ -128,8 +128,6 @@ public:
 	void Execute_AlphaTest(u32 op, u32 diff);
 	void Execute_StencilTest(u32 op, u32 diff);
 	void Execute_ColorRef(u32 op, u32 diff);
-	void Execute_BlendFixA(u32 op, u32 diff);
-	void Execute_BlendFixB(u32 op, u32 diff);
 	void Execute_WorldMtxNum(u32 op, u32 diff);
 	void Execute_WorldMtxData(u32 op, u32 diff);
 	void Execute_ViewMtxNum(u32 op, u32 diff);

GPU/GLES/ShaderManager.cpp

@@ -151,6 +151,7 @@ LinkedShader::LinkedShader(Shader *vs, Shader *fs, u32 vertType, bool useHWTrans
 	u_fbotex = glGetUniformLocation(program, "fbotex");
 	u_blendFixA = glGetUniformLocation(program, "u_blendFixA");
 	u_blendFixB = glGetUniformLocation(program, "u_blendFixB");
+	u_fbotexSize = glGetUniformLocation(program, "u_fbotexSize");
 
 	// Transform
 	u_view = glGetUniformLocation(program, "u_view");
@@ -225,7 +226,7 @@ LinkedShader::LinkedShader(Shader *vs, Shader *fs, u32 vertType, bool useHWTrans
 	if (u_view != -1) availableUniforms |= DIRTY_VIEWMATRIX;
 	if (u_texmtx != -1) availableUniforms |= DIRTY_TEXMATRIX;
 	if (u_stencilReplaceValue != -1) availableUniforms |= DIRTY_STENCILREPLACEVALUE;
-	if (u_blendFixA != -1 || u_blendFixB != -1) availableUniforms |= DIRTY_BLENDFIX;
+	if (u_blendFixA != -1 || u_blendFixB != -1 || u_fbotexSize != -1) availableUniforms |= DIRTY_SHADERBLEND;
 
 	// Looping up to numBones lets us avoid checking u_bone[i]
 	for (int i = 0; i < numBones; i++) {
@@ -551,9 +552,17 @@ void LinkedShader::UpdateUniforms(u32 vertType) {
 	}
 #endif
 
-	if (dirty & DIRTY_BLENDFIX) {
+	if (dirty & DIRTY_SHADERBLEND) {
 		SetColorUniform3(u_blendFixA, gstate.getFixA());
 		SetColorUniform3(u_blendFixB, gstate.getFixB());
+
+		const float fbotexSize[2] = {
+			1.0f / (float)gstate_c.curRTRenderWidth,
+			1.0f / (float)gstate_c.curRTRenderHeight,
+		};
+		if (u_fbotexSize != -1) {
+			glUniform2fv(u_fbotexSize, 1, fbotexSize);
+		}
 	}
 
 	// Lighting

GPU/GLES/ShaderManager.h

@@ -77,6 +77,7 @@ public:
 	int u_fbotex;
 	int u_blendFixA;
 	int u_blendFixB;
+	int u_fbotexSize;
 
 	// Fragment processing inputs
 	int u_alphacolorref;
@@ -130,7 +131,7 @@ enum
 	DIRTY_AMBIENT = (1 << 15),
 	DIRTY_MATAMBIENTALPHA = (1 << 16),
 
-	DIRTY_BLENDFIX = (1 << 17),  // (either one.)
+	DIRTY_SHADERBLEND = (1 << 17),  // Used only for in-shader blending.
 
 	DIRTY_UVSCALEOFFSET = (1 << 18),  // this will be dirtied ALL THE TIME... maybe we'll need to do "last value with this shader compares"
 	DIRTY_TEXCLAMP = (1 << 19),

GPU/GLES/StateMapping.cpp

@@ -195,8 +195,13 @@ void TransformDrawEngine::ApplyDrawState(int prim) {
 
 			glActiveTexture(GL_TEXTURE1);
 			framebufferManager_->BindFramebufferColor(NULL);
+			// If we are rendering at a higher resolution, linear is probably best for the dest color.
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 			glActiveTexture(GL_TEXTURE0);
 			fboTexBound_ = true;
+
+			shaderManager_->DirtyUniform(DIRTY_SHADERBLEND);
 		}
 		// None of the below logic is interesting, we're gonna do it entirely in the shader.
 		wantBlend = false;
@@ -204,6 +209,7 @@ void TransformDrawEngine::ApplyDrawState(int prim) {
 		glActiveTexture(GL_TEXTURE1);
 		glBindTexture(GL_TEXTURE_2D, 0);
 		glActiveTexture(GL_TEXTURE0);
+		fboTexBound_ = false;
 	}
 
 	glstate.blend.set(wantBlend);

GPU/GLES/TextureCache.cpp

@@ -969,8 +969,8 @@ void TextureCache::SetTextureFramebuffer(TexCacheEntry *entry, VirtualFramebuffe
 			glActiveTexture(GL_TEXTURE0);
 
 			framebufferManager_->BindFramebufferColor(framebuffer, true);
-			glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-			glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+			glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 			entry->status |= TexCacheEntry::STATUS_TEXPARAM_DIRTY;
 
 			glDisable(GL_BLEND);

GPU/GPUState.cpp

@@ -335,4 +335,6 @@ void GPUStateCache::DoState(PointerWrap &p) {
 
 	p.Do(curRTWidth);
 	p.Do(curRTHeight);
+
+	// curRTBufferWidth and curRTBufferHeight don't need to be saved.
 }

GPU/GPUState.h

@@ -475,6 +475,9 @@ struct GPUStateCache
 	u32 curRTWidth;
 	u32 curRTHeight;
 
+	u32 curRTRenderWidth;
+	u32 curRTRenderHeight;
+
 	u32 getRelativeAddress(u32 data) const;
 	void DoState(PointerWrap &p);
 };