Improve srcover_srgb_srgb implementation.

src/opts/SkBlend_opts.h

Index: src/opts/SkBlend_opts.h
diff --git a/src/opts/SkBlend_opts.h b/src/opts/SkBlend_opts.h
index 93946438e59610961732766f8f030dcd9de37054..2e6eff62773150439649d488dcab8b0469196867 100644
--- a/src/opts/SkBlend_opts.h
+++ b/src/opts/SkBlend_opts.h
@@ -127,27 +127,33 @@ void trivial_srcover_srgb_srgb(
                 int count = SkTMin(ndst, nsrc);
                 ndst -= count;
                 const uint32_t* src = srcStart;
-                const uint32_t* end = src + (count & ~3);
+                const uint32_t* end = dst + (count & ~3);
+                ptrdiff_t delta = src - dst;
 
-                while (src < end) {
+                while (dst < end) {
                     __m128i pixels = load(src);
                     if (_mm_testc_si128(pixels, alphaMask)) {
+                         uint32_t* start = dst;
                         do {
                             store(dst, pixels);
                             dst += 4;
-                            src += 4;
-                        } while (src < end && _mm_testc_si128(pixels = load(src), alphaMask));
+                        } while (dst < end
+                                 && _mm_testc_si128(pixels = load(dst + delta), alphaMask));
+                        src += dst - start;
                     } else if (_mm_testz_si128(pixels, alphaMask)) {
                         do {
                             dst += 4;
                             src += 4;
-                        } while (src < end && _mm_testz_si128(pixels = load(src), alphaMask));
+                        } while (dst < end
+                                 && _mm_testz_si128(pixels = load(src), alphaMask));
                     } else {
+                        uint32_t* start = dst;
                         do {
-                            srcover_srgb_srgb_4(dst, src);
+                            srcover_srgb_srgb_4(dst, dst + delta);
                             dst += 4;
-                            src += 4;
-                        } while (src < end && _mm_testnzc_si128(pixels = load(src), alphaMask));
+                        } while (dst < end
+                                 && _mm_testnzc_si128(pixels = load(dst + delta), alphaMask));
+                        src += dst - start;
                     }
                 }
 
@@ -159,32 +165,34 @@ void trivial_srcover_srgb_srgb(
         }
     #else
     // SSE2 versions
+
+        // Note: In the next three comparisons a group of 4 pixels is converted to a group of
+        // "signed" pixels because the sse2 does not have an unsigned comparison.
+        // Make it so that we can use the signed comparison operators by biasing
+        // 0x00xxxxxx to 0x80xxxxxxx which is the smallest values and biasing 0xffxxxxxx to
+        // 0x7fxxxxxx which is the largest set of values.
         static inline bool check_opaque_alphas(__m128i pixels) {
+            __m128i signedPixels = _mm_xor_si128(pixels, _mm_set1_epi32(0x80000000));
             int mask =
                 _mm_movemask_epi8(
-                    _mm_cmpeq_epi32(
-                        _mm_andnot_si128(pixels, _mm_set1_epi32(0xFF000000)),
-                        _mm_setzero_si128()));
-            return mask == 0xFFFF;
+                    _mm_cmplt_epi32(signedPixels, _mm_set1_epi32(0x7F000000)));
+            return mask == 0;
         }
 
         static inline bool check_transparent_alphas(__m128i pixels) {
+            __m128i signedPixels = _mm_xor_si128(pixels, _mm_set1_epi32(0x80000000));
             int mask =
                 _mm_movemask_epi8(
-                    _mm_cmpeq_epi32(
-                        _mm_and_si128(pixels, _mm_set1_epi32(0xFF000000)),
-                        _mm_setzero_si128()));
-            return mask == 0xFFFF;
+                    _mm_cmpgt_epi32(signedPixels, _mm_set1_epi32(0x80FFFFFF)));

[mtklein, 2016/05/24 12:15:37]

Can't we trim the xor here by exploiting the constraint that these pixels are
premultiplied?  Pixels with transparent alphas must be exactly 0x00000000.

    return 0xffff == _mm_movemask_epi8(_mm_cmpeq_epi32(pixels, 
                                                       _mm_setzero_si128()));

That oughta be something like
    pcmpeqd   <zero>,   <pixels>
    pmovmskb  <pixels>, <reg>
    cmpl      $0xffff,  <reg>

+            return mask == 0;
         }
 
         static inline bool check_partial_alphas(__m128i pixels) {
-            __m128i alphas = _mm_and_si128(pixels, _mm_set1_epi32(0xFF000000));
-            int mask =
-                _mm_movemask_epi8(
-                    _mm_cmpeq_epi8(
-                        _mm_srai_epi32(alphas, 8),
-                        alphas));
-            return mask == 0xFFFF;
+            __m128i signedPixels = _mm_xor_si128(pixels, _mm_set1_epi32(0x80000000));
+            __m128i opaque       = _mm_cmplt_epi32(signedPixels, _mm_set1_epi32(0x7F000000));

[mtklein, 2016/05/24 13:00:13]

I think we can make this logic clearer.

To start, I think it'd help readability to change `opaque` and `transparent` to
`not_opaque` and `not_transparent`.

From there, it looks like the return value is
    !( !opaque ^ !transparent )

That expression is unusual enough it'd probably benefit from a truth table:

    opaque, transparent  ~~~> result
    1, 1  ~~~> (impossible input)
    0, 1  ~~~> !( !0 ^ !1 ) -->  !( 1 ^ 0)  --> !(1) --> 0
    1, 0  ~~~> !( !1 ^ !0 ) --> 0 by symmetry
    0, 0  ~~~> !( !0 ^ !0 ) --> ! ( 1 ^ 1 ) -->  !(0) --> 1

+            __m128i transparent  = _mm_cmpgt_epi32(signedPixels, _mm_set1_epi32(0x80FFFFFF));
+            int mask             = _mm_movemask_epi8(_mm_xor_si128(opaque, transparent));
+            return mask == 0;
         }
 
         void srcover_srgb_srgb(
@@ -193,30 +201,33 @@ void trivial_srcover_srgb_srgb(
                 int count = SkTMin(ndst, nsrc);
                 ndst -= count;
                 const uint32_t* src = srcStart;
-                const uint32_t* end = src + (count & ~3);
+                const uint32_t* end = dst + (count & ~3);
+                const ptrdiff_t delta = src - dst;
 
                 __m128i pixels = load(src);
                 do {
                     if (check_opaque_alphas(pixels)) {
+                        uint32_t* start = dst;
                         do {
                             store(dst, pixels);
                             dst += 4;
-                            src += 4;
-                        } while (src < end && check_opaque_alphas(pixels = load(src)));
+                        } while (dst < end && check_opaque_alphas((pixels = load(dst + delta))));
+                        src += dst - start;
                     } else if (check_transparent_alphas(pixels)) {
-                        const uint32_t* start = src;
+                        const uint32_t* start = dst;
                         do {
-                            src += 4;
-                        } while (src < end && check_transparent_alphas(pixels = load(src)));
-                        dst += src - start;
+                            dst += 4;
+                        } while (dst < end && check_transparent_alphas(pixels = load(dst + delta)));
+                        src += dst - start;
                     } else {
+                        const uint32_t* start = dst;
                         do {
-                            srcover_srgb_srgb_4(dst, src);
+                            srcover_srgb_srgb_4(dst, dst + delta);
                             dst += 4;
-                            src += 4;
-                        } while (src < end && check_partial_alphas(pixels = load(src)));
+                        } while (dst < end && check_partial_alphas(pixels = load(dst + delta)));
+                        src += dst - start;
                     }
-                } while (src < end);
+                } while (dst < end);
 
                 count = count & 3;
                 while (count-- > 0) {