gsdx-ogl: optimize Cs * As + Cd and Cs * Af + Cd blending

Basically the code does the alpha multiplication in the shader therefore
the blend unit only does a pure addition. This way the multiplication is
accurate and accurate_blending doesn't requires a costly barrier.

This code also avoid variable duplication to make the code more separated.
Hopefully blending can be done in a separated function

It is preliminary work to support fast color clipping with HDR

v2: fix assertion compilation failure

v3: fix regression in not accurate mode

v3: Cs * As/Af is not an accumulation

Those cases don't need the Cd addition and were already optimized anyway

Fix a regression on GoW2

plugins/GSdx/GSDeviceOGL.cpp

@@ -666,6 +666,7 @@ GLuint GSDeviceOGL::CompilePS(PSSelector sel)
 		+ format("#define PS_SHUFFLE %d\n", sel.shuffle)
 		+ format("#define PS_READ_BA %d\n", sel.read_ba)
 		+ format("#define PS_FBMASK %d\n", sel.fbmask)
+		+ format("#define PS_BLEND_ACCU %d\n", sel.blend_accu)
 		;
 
 	return m_shader->Compile("tfx.glsl", "ps_main", GL_FRAGMENT_SHADER, tfx_fs_all_glsl, macro);

plugins/GSdx/GSDeviceOGL.h

@@ -355,6 +355,7 @@ class GSDeviceOGL : public GSDevice
 				uint32 blend_b:2;
 				uint32 blend_c:2;
 				uint32 blend_d:2;
+				uint32 blend_accu:1;
 				uint32 dfmt:2;
 
 				uint32 _free2:21;
@@ -461,8 +462,10 @@ class GSDeviceOGL : public GSDevice
 				uint32 c:2;
 				uint32 d:2;
 				uint32 negative:1;
+				uint32 accu:1;
+				uint32 ps:1;
 
-				uint32 _free:22;
+				uint32 _free:20;
 			};
 
 			struct
@@ -470,15 +473,16 @@ class GSDeviceOGL : public GSDevice
 				uint32 _abe:1;
 				uint32 abcd:8;
 				uint32 _negative:1;
+				uint32 _accu:1;
+				uint32 _ps:1;
 
-				uint32 _free2:22;
+				uint32 _free2:20;
 			};
 
 			uint32 key;
 		};
 
-		// FIXME is the & useful ?
+		operator uint32() {return key;}
-		operator uint32() {return key & 0x3ff;}
 
 		OMBlendSelector() : key(0) {}
 
@@ -666,7 +670,7 @@ class GSDeviceOGL : public GSDevice
 	void SetupPS(PSSelector sel);
 	void SetupCB(const VSConstantBuffer* vs_cb, const PSConstantBuffer* ps_cb);
 	void SetupSampler(PSSamplerSelector ssel);
-	void SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, float afix, bool sw_blending =  false);
+	void SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, float afix);
 	GLuint GetSamplerID(PSSamplerSelector ssel);
 	GLuint GetPaletteSamplerID();
 

plugins/GSdx/GSRendererOGL.cpp

@@ -452,8 +452,6 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 
 	const GIFRegALPHA& ALPHA = context->ALPHA;
 	float afix = (float)context->ALPHA.FIX / 0x80;
-	bool sw_blending = false;
-	bool colclip_wrap = false;
 
 	if (!IsOpaque() && rt)
 	{
@@ -471,7 +469,6 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 			{
 				// this works because with PABE alpha blending is on when alpha >= 0x80, but since the pixel shader
 				// cannot output anything over 0x80 (== 1.0) blending with 0x80 or turning it off gives the same result
-
 				om_bsel.abe = 0;
 			}
 			else
@@ -489,24 +486,34 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 		int blend_sel  = ((om_bsel.a * 3 + om_bsel.b) * 3 + om_bsel.c) * 3 + om_bsel.d;
 		int blend_flag = GSDeviceOGL::m_blendMapD3D9[blend_sel].bogus;
 		// SW Blend is (nearly) free. Let's use it.
-		int free_blend = m_sw_blending && ((blend_flag & NO_BAR) || (m_prim_overlap == PRIM_OVERLAP_NO));
+		bool free_blend = (blend_flag & NO_BAR) || (m_prim_overlap == PRIM_OVERLAP_NO);
+		// We really need SW blending for this one, barely used
+		bool impossible_blend = (blend_flag & A_MAX);
+		// Do the multiplication in shader for blending accumulation: Cs*As + Cd or Cs*Af + Cd
+		ps_sel.blend_accu = m_sw_blending && ALPHA.A == 0 && ALPHA.B == 2 && ALPHA.C != 1 && ALPHA.D == 1;
+		om_bsel.accu = ps_sel.blend_accu;
+
+		bool sw_blending_base = m_sw_blending && (free_blend || impossible_blend /*|| ps_sel.blend_accu*/);
 
 		// Color clip
 		bool acc_colclip_wrap = false;
 		if (env.COLCLAMP.CLAMP == 0) {
-			colclip_wrap = !tex && PRIM->PRIM != GS_POINTLIST;
+			// Not supported yet in colclip
-			acc_colclip_wrap =  (m_sw_blending >= ACC_BLEND_CCLIP || free_blend);
+			om_bsel.accu = ps_sel.blend_accu = 0;
+
+			acc_colclip_wrap =  (m_sw_blending >= ACC_BLEND_CCLIP || sw_blending_base);
 			if (acc_colclip_wrap) {
-				colclip_wrap = false;
 				ps_sel.colclip = 3;
 				GL_INS("COLCLIP SW ENABLED (blending is %d/%d/%d/%d)", ALPHA.A, ALPHA.B, ALPHA.C, ALPHA.D);
-			} else if (colclip_wrap) {
+			} else if (!PRIM->TME && PRIM->PRIM != GS_POINTLIST) {
+				// Standard (inaccurate) colclip
 				ps_sel.colclip = 1;
 				GL_INS("COLCLIP ENABLED (blending is %d/%d/%d/%d)", ALPHA.A, ALPHA.B, ALPHA.C, ALPHA.D);
 			}
+		} else {
+			sw_blending_base |= m_sw_blending && ps_sel.blend_accu;
 		}
 
-		bool impossible_blend = m_sw_blending && (blend_flag & A_MAX);
 		bool all_blend_sw;
 		switch (m_sw_blending) {
 			case ACC_BLEND_ULTRA:	all_blend_sw = true; break;
@@ -516,8 +523,8 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 			default:				all_blend_sw = false; break;
 		}
 
-		sw_blending = free_blend // Free case
+		bool sw_blending = sw_blending_base // Free case or Impossible blend
-			|| impossible_blend || all_blend_sw // Impossible blend or all
+			|| all_blend_sw // all blend
 			|| acc_colclip_wrap // accurate colclip
 			|| ps_sel.fbmask; // accurate fbmask
 
@@ -536,6 +543,7 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 #endif
 		if (sw_blending && om_bsel.abe) {
 			// select a shader that support blending
+			om_bsel.ps = 1;
 			ps_sel.blend_a = om_bsel.a;
 			ps_sel.blend_b = om_bsel.b;
 			ps_sel.blend_c = om_bsel.c;
@@ -549,18 +557,18 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 			}
 
 			// No need to flush for every primitive
-			require_barrier |= !(blend_flag & NO_BAR);
+			require_barrier |= !(blend_flag & NO_BAR) && !ps_sel.blend_accu;
 		} else {
 			ps_sel.clr1 = om_bsel.IsCLR1();
-		}
+			if (ps_sel.dfmt == 1 && ALPHA.C == 1) {
-	}
-
-	if (ps_sel.dfmt == 1) {
-		if (ALPHA.C == 1 && !sw_blending) {
 				// 24 bits doesn't have an alpha channel so use 1.0f fix factor as equivalent
 				om_bsel.c = 2;
 				afix = 1.0f;
 			}
+		}
+	}
+
+	if (ps_sel.dfmt == 1) {
 		// Disable writing of the alpha channel
 		om_csel.wa = 0;
 	}
@@ -864,7 +872,7 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 	GL_POP();
 
 	dev->OMSetColorMaskState(om_csel);
-	dev->SetupOM(om_dssel, om_bsel, afix, sw_blending);
+	dev->SetupOM(om_dssel, om_bsel, afix);
 
 	dev->SetupCB(&vs_cb, &ps_cb);
 
@@ -907,9 +915,9 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 	{
 		SendDraw(require_barrier);
 
-		if (colclip_wrap)
+		if (ps_sel.colclip == 1)
 		{
-			ASSERT(!sw_blending);
+			ASSERT(!om_bsel.ps);
 			GL_PUSH("COLCLIP");
 			GSDeviceOGL::OMBlendSelector om_bselneg(om_bsel);
 			GSDeviceOGL::PSSelector ps_selneg(ps_sel);
@@ -963,13 +971,13 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 			om_csel.wa = a;
 
 			dev->OMSetColorMaskState(om_csel);
-			dev->SetupOM(om_dssel, om_bsel, afix, sw_blending);
+			dev->SetupOM(om_dssel, om_bsel, afix);
 
 			SendDraw(require_barrier);
 
-			if (colclip_wrap)
+			if (ps_sel.colclip == 1)
 			{
-				ASSERT(!sw_blending);
+				ASSERT(!om_bsel.ps);
 				GL_PUSH("COLCLIP");
 				GSDeviceOGL::OMBlendSelector om_bselneg(om_bsel);
 				GSDeviceOGL::PSSelector ps_selneg(ps_sel);

plugins/GSdx/GSTextureFXOGL.cpp

@@ -104,6 +104,9 @@ GSBlendStateOGL* GSDeviceOGL::CreateBlend(OMBlendSelector bsel, float afix)
 	{
 		int i = ((bsel.a * 3 + bsel.b) * 3 + bsel.c) * 3 + bsel.d;
 
+		if (bsel.accu)
+			bs->SetRGB(GL_FUNC_ADD, GL_ONE, GL_ONE);
+		else
 			bs->SetRGB(m_blendMapD3D9[i].op, m_blendMapD3D9[i].src, m_blendMapD3D9[i].dst);
 
 		if (m_blendMapD3D9[i].bogus & A_MAX) {
@@ -187,13 +190,13 @@ GLuint GSDeviceOGL::GetPaletteSamplerID()
 	return m_palette_ss;
 }
 
-void GSDeviceOGL::SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, float afix, bool sw_blending)
+void GSDeviceOGL::SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, float afix)
 {
 	GSDepthStencilOGL* dss = m_om_dss[dssel];
 
 	OMSetDepthStencilState(dss, 1);
 
-	if (sw_blending) {
+	if (bsel.ps && !bsel.accu) {
 		if (GLState::blend) {
 			GLState::blend = false;
 			glDisable(GL_BLEND);

plugins/GSdx/res/glsl/tfx_fs.glsl

@@ -453,6 +453,9 @@ void ps_blend(inout vec4 Color, float As)
 
 #if PS_BLEND_A == PS_BLEND_B
     Color.rgb = D;
+#elif PS_BLEND_ACCU == 1
+	// The D addition will be done in the blending unit
+	Color.rgb = trunc(A * C);
 #else
     Color.rgb = trunc((A - B) * C + D);
 #endif

plugins/GSdx/res/glsl_source.h

@@ -1316,6 +1316,9 @@ static const char* tfx_fs_all_glsl =
 	"\n"
 	"#if PS_BLEND_A == PS_BLEND_B\n"
 	"    Color.rgb = D;\n"
+	"#elif PS_BLEND_ACCU == 1\n"
+	"	// The D addition will be done in the blending unit\n"
+	"	Color.rgb = trunc(A * C);\n"
 	"#else\n"
 	"    Color.rgb = trunc((A - B) * C + D);\n"
 	"#endif\n"