I have found a more efficient way of detecting 1 and 0 alpha in SSE2. In addition, I found a stall on an execution unit for the lea instruction and rearranged to code to avoid that.

Before
 1,362.01 LinearSrcOvericonstrip.pngVSkOptsSSE41
 2,132.54 LinearSrcOvericonstrip.pngVSkOptsDefault
 1,717.77 LinearSrcOvericonstrip.pngVSkOptsNonSimdCore
 3,525.14 LinearSrcOvericonstrip.pngVSkOptsTrivial
11,181.78 LinearSrcOvericonstrip.pngVSkOptsBruteForce
   644.77 LinearSrcOvermandrill_512.pngVSkOptsSSE41
   682.51 LinearSrcOvermandrill_512.pngVSkOptsDefault
 1,169.65 LinearSrcOvermandrill_512.pngVSkOptsNonSimdCore
 2,486.45 LinearSrcOvermandrill_512.pngVSkOptsTrivial
11,635.94 LinearSrcOvermandrill_512.pngVSkOptsBruteForce
   217.76 LinearSrcOverplane.pngVSkOptsSSE41
   437.09 LinearSrcOverplane.pngVSkOptsDefault
   275.91 LinearSrcOverplane.pngVSkOptsNonSimdCore
   481.70 LinearSrcOverplane.pngVSkOptsTrivial
 1,504.66 LinearSrcOverplane.pngVSkOptsBruteForce
   323.90 LinearSrcOverbaby_tux.pngVSkOptsSSE41
   497.49 LinearSrcOverbaby_tux.pngVSkOptsDefault
   456.08 LinearSrcOverbaby_tux.pngVSkOptsNonSimdCore
   786.46 LinearSrcOverbaby_tux.pngVSkOptsTrivial
 2,554.65 LinearSrcOverbaby_tux.pngVSkOptsBruteForce
   484.83 LinearSrcOveryellow_rose.pngVSkOptsSSE41
   821.86 LinearSrcOveryellow_rose.pngVSkOptsDefault
   655.37 LinearSrcOveryellow_rose.pngVSkOptsNonSimdCore
 1,323.80 LinearSrcOveryellow_rose.pngVSkOptsTrivial
 5,802.61 LinearSrcOveryellow_rose.pngVSkOptsBruteForce

After changes to sse2 and sse4.1
  1,343.12 LinearSrcOvericonstrip.pngVSkOptsSSE41
  1,441.17 LinearSrcOvericonstrip.pngVSkOptsDefault
  1,679.97 LinearSrcOvericonstrip.pngVSkOptsNonSimdCore
  3,481.05 LinearSrcOvericonstrip.pngVSkOptsTrivial
 10,979.99 LinearSrcOvericonstrip.pngVSkOptsBruteForce
    574.17 LinearSrcOvermandrill_512.pngVSkOptsSSE41
    641.40 LinearSrcOvermandrill_512.pngVSkOptsDefault
  1,169.44 LinearSrcOvermandrill_512.pngVSkOptsNonSimdCore
  2,359.84 LinearSrcOvermandrill_512.pngVSkOptsTrivial
 12,106.02 LinearSrcOvermandrill_512.pngVSkOptsBruteForce
    209.95 LinearSrcOverplane.pngVSkOptsSSE41
    249.12 LinearSrcOverplane.pngVSkOptsDefault
    270.36 LinearSrcOverplane.pngVSkOptsNonSimdCore
    466.30 LinearSrcOverplane.pngVSkOptsTrivial
  1,431.14 LinearSrcOverplane.pngVSkOptsBruteForce
    309.70 LinearSrcOverbaby_tux.pngVSkOptsSSE41
    354.86 LinearSrcOverbaby_tux.pngVSkOptsDefault
    442.69 LinearSrcOverbaby_tux.pngVSkOptsNonSimdCore
    764.12 LinearSrcOverbaby_tux.pngVSkOptsTrivial
  2,756.16 LinearSrcOverbaby_tux.pngVSkOptsBruteForce
    457.70 LinearSrcOveryellow_rose.pngVSkOptsSSE41
    500.50 LinearSrcOveryellow_rose.pngVSkOptsDefault
    677.84 LinearSrcOveryellow_rose.pngVSkOptsNonSimdCore
  1,301.50 LinearSrcOveryellow_rose.pngVSkOptsTrivial
  5,786.40 LinearSrcOveryellow_rose.pngVSkOptsBruteForce

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1998373002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review-Url: https://codereview.chromium.org/1998373002

1 files changed, 45 insertions, 34 deletions

diff --git a/src/opts/SkBlend_opts.h b/src/opts/SkBlend_opts.h
index 93946438e5..2e6eff6277 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)));
+            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));
+            __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) {