gsdx ogl: Bypass the texture cache when the frame buffer will be sampled.

Proof of concept. It should provide a huge speedup when accurate
blending is enabled for tri-ace / Jak / R&C (shadow rendering).

See #2894
Need PR#2892

v2: Add const + comment to explain that code isn't ideal.

plugins/GSdx/Renderers/HW/GSRendererHW.h

@@ -180,4 +180,7 @@ public:
 	void InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r);
 	void InvalidateLocalMem(const GIFRegBITBLTBUF& BITBLTBUF, const GSVector4i& r, bool clut = false);
 	void Draw();
+
+	// Called by the texture cache to know if current texture is useful
+	virtual bool IsDummyTexture() const { return false;}
 };

plugins/GSdx/Renderers/HW/GSTextureCache.cpp

@@ -21,6 +21,7 @@
 
 #include "stdafx.h"
 #include "GSTextureCache.h"
+#include "GSRendererHW.h"
 #include "GSUtil.h"
 
 bool GSTextureCache::m_disable_partial_invalidation = false;
@@ -1173,6 +1174,31 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 			AttachPaletteToSource(src, psm.pal, true);
 		}
 	}
+	else if (dst && static_cast<GSRendererHW*>(m_renderer)->IsDummyTexture())
+	{
+		// This shortcut is a temporary solution. It isn't a good solution
+		// as it won't work with Channel Shuffle/Texture Shuffle pattern
+		// (we need texture cache result to detect those effects).
+		// Instead a better solution would be to defer the copy/StrechRect later
+		// in the rendering.
+		// Still this poor solution is enough for a huge speed up in a couple of games
+		//
+		// Be aware that you can't use StrechRect between BeginScene/EndScene.
+		// So it could be tricky to put in the middle of the DrawPrims
+
+		// Texture is created to keep code compatibility
+		GSTexture* dTex = m_renderer->m_dev->CreateRenderTarget(tw, th);
+
+		// Keep a trace of origin of the texture
+		src->m_texture = dTex;
+		src->m_target = true;
+		src->m_from_target = dst->m_texture;
+
+		// Even if we sample the framebuffer directly we might need the palette
+		// to handle the format conversion on GPU
+		if (psm.pal > 0)
+			AttachPaletteToSource(src, psm.pal, true);
+	}
 	else if (dst)
 	{
 		// TODO: clean up this mess

plugins/GSdx/Renderers/OpenGL/GSRendererOGL.cpp

@@ -1027,9 +1027,10 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 	// Detect framebuffer read that will need special handling
 	if ((m_context->FRAME.Block() == m_context->TEX0.TBP0) && PRIM->TME && m_sw_blending) {
 		if ((m_context->FRAME.FBMSK == 0x00FFFFFF) && (m_vt.m_primclass == GS_TRIANGLE_CLASS)) {
-			// Ratchet & Clank, Jak, Tri-Ace (Star Ocean 3) games use this pattern to compute the shadows.
-			// Alpha (multiplication) tfx is mostly equivalent to -1/+1 stencil operation.
-			GL_DBG("ERROR: Source and Target are the same! Let's sample the framebuffer");
+			// This pattern is used by several games to emulate a stencil (shadow)
+			// Ratchet & Clank, Jak do alpha integer multiplication (tfx) which is mostly equivalent to +1/-1
+			// Tri-Ace (Star Ocean 3/RadiataStories/VP2) uses a palette to handle the +1/-1
+			GL_DBG("Source and Target are the same! Let's sample the framebuffer");
 			m_ps_sel.tex_is_fb = 1;
 			m_require_full_barrier = true;
 		} else if (m_prim_overlap != PRIM_OVERLAP_NO) {
@@ -1417,3 +1418,10 @@ void GSRendererOGL::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sour
 		dev->Recycle(hdr_rt);
 	}
 }
+
+bool GSRendererOGL::IsDummyTexture() const
+{
+	// Texture is actually the frame buffer. Stencil emulation to compute shadow (Jak series/tri-ace game)
+	// Will hit the "m_ps_sel.tex_is_fb = 1" path in the draw
+	return (m_context->FRAME.Block() == m_context->TEX0.TBP0) && PRIM->TME && m_sw_blending && m_vt.m_primclass == GS_TRIANGLE_CLASS && (m_context->FRAME.FBMSK == 0x00FFFFFF);
+}

plugins/GSdx/Renderers/OpenGL/GSRendererOGL.h

@@ -90,4 +90,6 @@ class GSRendererOGL final : public GSRendererHW
 		PRIM_OVERLAP PrimitiveOverlap();
 
 		void SendDraw();
+
+		bool IsDummyTexture() const final;
 };