gsdx sw: factorize common draw scanline code

Ymm inherite from Xmm so it is useless to duplicate the code

Add a parameter to alltrue to test the good register

plugins/GSdx/GSDrawScanlineCodeGenerator.cpp

@@ -122,82 +122,7 @@ GSDrawScanlineCodeGenerator::GSDrawScanlineCodeGenerator(void* param, uint64 key
 	}
 }
 
-#if _M_SSE >= 0x501
+void GSDrawScanlineCodeGenerator::modulate16(const Xmm& a, const Operand& f, uint8 shift)
-
-void GSDrawScanlineCodeGenerator::modulate16(const Ymm& a, const Operand& f, int shift)
-{
-	if(shift == 0)
-	{
-		vpmulhrsw(a, f);
-	}
-	else
-	{
-		vpsllw(a, (uint8)(shift + 1));
-		vpmulhw(a, f);
-	}
-}
-
-void GSDrawScanlineCodeGenerator::lerp16(const Ymm& a, const Ymm& b, const Ymm& f, int shift)
-{
-	vpsubw(a, b);
-	modulate16(a, f, shift);
-	vpaddw(a, b);
-}
-
-void GSDrawScanlineCodeGenerator::lerp16_4(const Ymm& a, const Ymm& b, const Ymm& f)
-{
-	vpsubw(a, b);
-	vpmullw(a, f);
-	vpsraw(a, 4);
-	vpaddw(a, b);
-}
-
-void GSDrawScanlineCodeGenerator::mix16(const Ymm& a, const Ymm& b, const Ymm& temp)
-{
-	vpblendw(a, b, 0xaa);
-}
-
-void GSDrawScanlineCodeGenerator::clamp16(const Ymm& a, const Ymm& temp)
-{
-	vpackuswb(a, a);
-	vpermq(a, a, _MM_SHUFFLE(3, 1, 2, 0)); // this sucks
-	vpmovzxbw(a, a);
-}
-
-void GSDrawScanlineCodeGenerator::alltrue()
-{
-	vpmovmskb(eax, ymm7);
-	cmp(eax, 0xffffffff);
-	je("step", T_NEAR);
-}
-
-void GSDrawScanlineCodeGenerator::blend(const Ymm& a, const Ymm& b, const Ymm& mask)
-{
-	vpand(b, mask);
-	vpandn(mask, a);
-	vpor(a, b, mask);
-}
-
-void GSDrawScanlineCodeGenerator::blendr(const Ymm& b, const Ymm& a, const Ymm& mask)
-{
-	vpand(b, mask);
-	vpandn(mask, a);
-	vpor(b, mask);
-}
-
-void GSDrawScanlineCodeGenerator::blend8(const Ymm& a, const Ymm& b)
-{
-	vpblendvb(a, a, b, xmm0);
-}
-
-void GSDrawScanlineCodeGenerator::blend8r(const Ymm& b, const Ymm& a)
-{
-	vpblendvb(b, a, b, xmm0);
-}
-
-#else
-
-void GSDrawScanlineCodeGenerator::modulate16(const Xmm& a, const Operand& f, int shift)
 {
 	if(g_cpu.has(util::Cpu::tAVX))
 	{
@@ -226,7 +151,7 @@ void GSDrawScanlineCodeGenerator::modulate16(const Xmm& a, const Operand& f, int
 	}
 }
 
-void GSDrawScanlineCodeGenerator::lerp16(const Xmm& a, const Xmm& b, const Xmm& f, int shift)
+void GSDrawScanlineCodeGenerator::lerp16(const Xmm& a, const Xmm& b, const Xmm& f, uint8 shift)
 {
 	if(g_cpu.has(util::Cpu::tAVX))
 	{
@@ -288,6 +213,15 @@ void GSDrawScanlineCodeGenerator::clamp16(const Xmm& a, const Xmm& temp)
 	if(g_cpu.has(util::Cpu::tAVX))
 	{
 		vpackuswb(a, a);
+
+#if _M_SSE >= 0x501
+		// Greg: why ?
+		if(g_cpu.has(util::Cpu::tAVX2)) {
+			ASSERT(a.isYMM());
+			vpermq(Ymm(a.getIdx()), Ymm(a.getIdx()), _MM_SHUFFLE(3, 1, 2, 0)); // this sucks
+		}
+#endif
+
 		vpmovzxbw(a, a);
 	}
 	else
@@ -306,18 +240,20 @@ void GSDrawScanlineCodeGenerator::clamp16(const Xmm& a, const Xmm& temp)
 	}
 }
 
-void GSDrawScanlineCodeGenerator::alltrue()
+void GSDrawScanlineCodeGenerator::alltrue(const Xmm& test)
 {
+	uint32 mask = test.isYMM() ? 0xffffffff : 0xffff;
+
 	if(g_cpu.has(util::Cpu::tAVX))
 	{
-		vpmovmskb(eax, xmm7);
+		vpmovmskb(eax, test);
-		cmp(eax, 0xffff);
+		cmp(eax, mask);
 		je("step", T_NEAR);
 	}
 	else
 	{
-		pmovmskb(eax, xmm7);
+		pmovmskb(eax, test);
-		cmp(eax, 0xffff);
+		cmp(eax, mask);
 		je("step", T_NEAR);
 	}
 }
@@ -416,5 +352,3 @@ void GSDrawScanlineCodeGenerator::split16_2x8(const Xmm& l, const Xmm& h, const
 		psrlw(h, 8);
 	}
 }
-
-#endif

plugins/GSdx/GSDrawScanlineCodeGenerator.h

@@ -71,17 +71,6 @@ class GSDrawScanlineCodeGenerator : public GSCodeGenerator
 	void ReadTexel(int pixels, int mip_offset = 0);
 	void ReadTexel(const Ymm& dst, const Ymm& addr, uint8 i);
 
-	void modulate16(const Ymm& a, const Operand& f, int shift);
-	void lerp16(const Ymm& a, const Ymm& b, const Ymm& f, int shift);
-	void lerp16_4(const Ymm& a, const Ymm& b, const Ymm& f);
-	void mix16(const Ymm& a, const Ymm& b, const Ymm& temp);
-	void clamp16(const Ymm& a, const Ymm& temp);
-	void alltrue();
-	void blend(const Ymm& a, const Ymm& b, const Ymm& mask);
-	void blendr(const Ymm& b, const Ymm& a, const Ymm& mask);
-	void blend8(const Ymm& a, const Ymm& b);
-	void blend8r(const Ymm& b, const Ymm& a);
-
 	#else
 
 	void Generate_SSE();
@@ -138,20 +127,20 @@ class GSDrawScanlineCodeGenerator : public GSCodeGenerator
 	void ReadTexel_AVX(int pixels, int mip_offset = 0);
 	void ReadTexel_AVX(const Xmm& dst, const Xmm& addr, uint8 i);
 
-	void modulate16(const Xmm& a, const Operand& f, int shift);
+	#endif
-	void lerp16(const Xmm& a, const Xmm& b, const Xmm& f, int shift);
+
+	void modulate16(const Xmm& a, const Operand& f, uint8 shift);
+	void lerp16(const Xmm& a, const Xmm& b, const Xmm& f, uint8 shift);
 	void lerp16_4(const Xmm& a, const Xmm& b, const Xmm& f);
 	void mix16(const Xmm& a, const Xmm& b, const Xmm& temp);
 	void clamp16(const Xmm& a, const Xmm& temp);
-	void alltrue();
+	void alltrue(const Xmm& test);
 	void blend(const Xmm& a, const Xmm& b, const Xmm& mask);
 	void blendr(const Xmm& b, const Xmm& a, const Xmm& mask);
 	void blend8(const Xmm& a, const Xmm& b);
 	void blend8r(const Xmm& b, const Xmm& a);
 	void split16_2x8(const Xmm& l, const Xmm& h, const Xmm& src);
 
-	#endif
-
 public:
 	GSDrawScanlineCodeGenerator(void* param, uint64 key, void* code, size_t maxsize);
 

plugins/GSdx/GSDrawScanlineCodeGenerator.x64.avx.cpp

@@ -727,7 +727,7 @@ void GSDrawScanlineCodeGenerator::TestZ_AVX(const Xmm& temp1, const Xmm& temp2)
 			break;
 		}
 
-		alltrue();
+		alltrue(_test);
 	}
 }
 
@@ -1337,7 +1337,7 @@ void GSDrawScanlineCodeGenerator::TestAlpha_AVX()
 	case AFAIL_KEEP:
 		// test |= t;
 		vpor(_test, xmm1);
-		alltrue();
+		alltrue(_test);
 		break;
 
 	case AFAIL_FB_ONLY:
@@ -1509,7 +1509,7 @@ void GSDrawScanlineCodeGenerator::TestDestAlpha_AVX()
 
 	vpor(_test, xmm1);
 
-	alltrue();
+	alltrue(_test);
 }
 
 void GSDrawScanlineCodeGenerator::WriteMask_AVX()

plugins/GSdx/GSDrawScanlineCodeGenerator.x86.avx.cpp

@@ -689,7 +689,7 @@ void GSDrawScanlineCodeGenerator::TestZ_AVX(const Xmm& temp1, const Xmm& temp2)
 			break;
 		}
 
-		alltrue();
+		alltrue(xmm7);
 	}
 }
 
@@ -2130,7 +2130,7 @@ void GSDrawScanlineCodeGenerator::TestAlpha_AVX()
 	case AFAIL_KEEP:
 		// test |= t;
 		vpor(xmm7, xmm1);
-		alltrue();
+		alltrue(xmm7);
 		break;
 
 	case AFAIL_FB_ONLY:
@@ -2313,7 +2313,7 @@ void GSDrawScanlineCodeGenerator::TestDestAlpha_AVX()
 
 	vpor(xmm7, xmm1);
 
-	alltrue();
+	alltrue(xmm7);
 }
 
 void GSDrawScanlineCodeGenerator::WriteMask_AVX()

plugins/GSdx/GSDrawScanlineCodeGenerator.x86.avx2.cpp

@@ -691,7 +691,7 @@ void GSDrawScanlineCodeGenerator::TestZ(const Ymm& temp1, const Ymm& temp2)
 			break;
 		}
 
-		alltrue();
+		alltrue(ymm7);
 	}
 }
 
@@ -2118,7 +2118,7 @@ void GSDrawScanlineCodeGenerator::TestAlpha()
 	case AFAIL_KEEP:
 		// test |= t;
 		vpor(ymm7, ymm1);
-		alltrue();
+		alltrue(ymm7);
 		break;
 
 	case AFAIL_FB_ONLY:
@@ -2309,7 +2309,7 @@ void GSDrawScanlineCodeGenerator::TestDestAlpha()
 
 	vpor(ymm7, ymm1);
 
-	alltrue();
+	alltrue(ymm7);
 }
 
 void GSDrawScanlineCodeGenerator::WriteMask()

plugins/GSdx/GSDrawScanlineCodeGenerator.x86.cpp

@@ -694,7 +694,7 @@ void GSDrawScanlineCodeGenerator::TestZ_SSE(const Xmm& temp1, const Xmm& temp2)
 			break;
 		}
 
-		alltrue();
+		alltrue(xmm7);
 	}
 }
 
@@ -2162,7 +2162,7 @@ void GSDrawScanlineCodeGenerator::TestAlpha_SSE()
 	case AFAIL_KEEP:
 		// test |= t;
 		por(xmm7, xmm1);
-		alltrue();
+		alltrue(xmm7);
 		break;
 
 	case AFAIL_FB_ONLY:
@@ -2344,7 +2344,7 @@ void GSDrawScanlineCodeGenerator::TestDestAlpha_SSE()
 
 	por(xmm7, xmm1);
 
-	alltrue();
+	alltrue(xmm7);
 }
 
 void GSDrawScanlineCodeGenerator::WriteMask_SSE()