Xfermode: SSE2 implementation of exclusion_modeproc

src/opts/SkXfermode_opts_SSE2.cpp

 #include "SkColorPriv.h"
 #include "SkColor_opts_SSE2.h"
 #include "SkMathPriv.h"
 #include "SkXfermode.h"
 #include "SkXfermode_opts_SSE2.h"
 #include "SkXfermode_proccoeff.h"
 
 ////////////////////////////////////////////////////////////////////////////////
 // 4 pixels SSE2 version functions
 ////////////////////////////////////////////////////////////////////////////////
@@ skipping 66 matching lines @@
     __m128i g = blendfunc_multiply_byte_SSE2(sg, dg, sa, da);
 
 
     __m128i sb = SkGetPackedB32_SSE2(src);
     __m128i db = SkGetPackedB32_SSE2(dst);
     __m128i b = blendfunc_multiply_byte_SSE2(sb, db, sa, da);
 
     return SkPackARGB32_SSE2(a, r, g, b);
 }
 
+static inline __m128i exclusion_byte_SSE2(const __m128i& sc, const __m128i& dc,
+                                          const __m128i&, __m128i&) {
+    __m128i tmp1 = _mm_mullo_epi16(_mm_set1_epi32(255), sc); // 255 * sc
+    __m128i tmp2 = _mm_mullo_epi16(_mm_set1_epi32(255), dc); // 255 * dc
+    tmp1 = _mm_add_epi32(tmp1, tmp2);
+    tmp2 = _mm_mullo_epi16(sc, dc);                          // sc * dc
+    tmp2 = _mm_slli_epi32(tmp2, 1);                          // 2 * sc * dc
+
+    __m128i r = _mm_sub_epi32(tmp1, tmp2);
+    return clamp_div255round_SSE2(r);
+}
+
+static __m128i exclusion_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
+    __m128i sa = SkGetPackedA32_SSE2(src);
+    __m128i da = SkGetPackedA32_SSE2(dst);
+
+    __m128i a = srcover_byte_SSE2(sa, da);
+    __m128i r = exclusion_byte_SSE2(SkGetPackedR32_SSE2(src),
+                                    SkGetPackedR32_SSE2(dst), sa, da);
+    __m128i g = exclusion_byte_SSE2(SkGetPackedG32_SSE2(src),
+                                    SkGetPackedG32_SSE2(dst), sa, da);
+    __m128i b = exclusion_byte_SSE2(SkGetPackedB32_SSE2(src),
+                                    SkGetPackedB32_SSE2(dst), sa, da);
+    return SkPackARGB32_SSE2(a, r, g, b);
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 typedef __m128i (*SkXfermodeProcSIMD)(const __m128i& src, const __m128i& dst);
 
 extern SkXfermodeProcSIMD gSSE2XfermodeProcs[];
 
 SkSSE2ProcCoeffXfermode::SkSSE2ProcCoeffXfermode(SkReadBuffer& buffer)
     : INHERITED(buffer) {
     fProcSIMD = reinterpret_cast<void*>(gSSE2XfermodeProcs[this->getMode()]);
 }
@@ skipping 143 matching lines @@
     NULL, // kScreen_Mode
 
     NULL, // kOverlay_Mode
     NULL, // kDarken_Mode
     NULL, // kLighten_Mode
     NULL, // kColorDodge_Mode
     NULL, // kColorBurn_Mode
     NULL, // kHardLight_Mode
     NULL, // kSoftLight_Mode
     NULL, // kDifference_Mode
-    NULL, // kExclusion_Mode
+    exclusion_modeproc_SSE2,
     multiply_modeproc_SSE2,
 
     NULL, // kHue_Mode
     NULL, // kSaturation_Mode
     NULL, // kColor_Mode
     NULL, // kLuminosity_Mode
 };
 
 SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_SSE2(const ProcCoeff& rec,
                                                          SkXfermode::Mode mode) {
     void* procSIMD = reinterpret_cast<void*>(gSSE2XfermodeProcs[mode]);
 
     if (procSIMD != NULL) {
         return SkNEW_ARGS(SkSSE2ProcCoeffXfermode, (rec, mode, procSIMD));
     }
     return NULL;
 }