Cleanup with SkAlphaMulQ_SSE2()
Related nanobench results:
before:
10M 18 7.03µs 7.31µs 7.38µs 8.46µs 6% ▂▁▂▂▂▃▄▁█▁ 8888 bitmaprect_80_filter_identity
10M 43 6.96µs 6.97µs 6.99µs 7.19µs 1% ▁▂▁▁▁▁▁█▁▁ 8888 bitmaprect_80_nofilter_identity
10M 14 35.7µs 35.8µs 35.9µs 36.3µs 1% ▃▂▁▂▁█▂▁▁▁ 8888 bitmap_BGRA_8888_update_scale_bilerp
10M 16 35.5µs 35.6µs 35.7µs 36.3µs 1% █▅▂▁▁▁▃▂▁▁ 8888 bitmap_BGRA_8888_update_volatile_scale_bilerp
10M 16 35.4µs 35.4µs 35.5µs 36.8µs 1% ▂▁█▁▁▁▁▂▁▁ 8888 bitmap_BGRA_8888_scale_bilerp
10M 25 16.4µs 16.6µs 16.7µs 17.4µs 2% ▂▁▁▂▁▁▁▅▅█ 8888 bitmap_Index_8
10M 15 37.9µs 38µs 38µs 38.4µs 0% ▄▆▂▁▁▁█▂▁▁ 8888 bitmap_RGB_565
10M 33 11.1µs 11.1µs 11.1µs 11.2µs 0% ▆▂█▂▂▂▁▁▂▁ 8888 bitmap_BGRA_8888_scale
after:
10M 9 7.04µs 7.06µs 7.1µs 7.32µs 1% █▅▂▁▁▂▁▁▁▁ 8888 bitmaprect_80_filter_identity
10M 18 7.01µs 7.02µs 7.05µs 7.25µs 1% █▂▁▁▁▁▁▁▁▁ 8888 bitmaprect_80_nofilter_identity
10M 5 33.9µs 34µs 34.1µs 34.5µs 1% █▃▂▂▁▁▁▅▃▂ 8888 bitmap_BGRA_8888_update_scale_bilerp
10M 7 35.5µs 35.5µs 35.6µs 36.3µs 1% ▃▂▂▁▂▁▂▁█▂ 8888 bitmap_BGRA_8888_update_volatile_scale_bilerp
10M 7 35.5µs 35.5µs 35.7µs 36.8µs 1% ▂▁▁▁▁▁▁▁▁█ 8888 bitmap_BGRA_8888_scale_bilerp
10M 11 16.4µs 16.4µs 16.4µs 16.6µs 0% █▂▁▁▂▁▁▁▂▁ 8888 bitmap_Index_8
10M 7 37.3µs 37.4µs 38.4µs 47.8µs 9% ▁▁▁▁▁▁▁▁▁█ 8888 bitmap_RGB_565
10M 33 11µs 11µs 11.1µs 11.2µs 1% ▄█▅▃▂▁▁▁▁▁ 8888 bitmap_BGRA_8888_scale
BUG=skia:
Review URL: https://codereview.chromium.org/755573002
diff --git a/src/opts/SkBlitRow_opts_SSE2.cpp b/src/opts/SkBlitRow_opts_SSE2.cpp
index cc68b70..17cd20a 100644
--- a/src/opts/SkBlitRow_opts_SSE2.cpp
+++ b/src/opts/SkBlitRow_opts_SSE2.cpp
@@ -38,57 +38,14 @@
const __m128i *s = reinterpret_cast<const __m128i*>(src);
__m128i *d = reinterpret_cast<__m128i*>(dst);
- __m128i rb_mask = _mm_set1_epi32(0x00FF00FF);
- __m128i ag_mask = _mm_set1_epi32(0xFF00FF00);
- // Move scale factors to upper byte of word
- __m128i src_scale_wide = _mm_set1_epi16(src_scale << 8);
- __m128i dst_scale_wide = _mm_set1_epi16(dst_scale << 8);
while (count >= 4) {
// Load 4 pixels each of src and dest.
__m128i src_pixel = _mm_loadu_si128(s);
__m128i dst_pixel = _mm_load_si128(d);
- // Interleave Atom port 0/1 operations based on the execution port
- // constraints that multiply can only be executed on port 0 (while
- // boolean operations can be executed on either port 0 or port 1)
- // because GCC currently doesn't do a good job scheduling
- // instructions based on these constraints.
-
- // Get red and blue pixels into lower byte of each word.
- // (0, r, 0, b, 0, r, 0, b, 0, r, 0, b, 0, r, 0, b)
- __m128i src_rb = _mm_and_si128(rb_mask, src_pixel);
-
- // Multiply by scale.
- // (4 x (0, rs.h, 0, bs.h))
- // where rs.h stands for the higher byte of r * scale, and
- // bs.h the higher byte of b * scale.
- src_rb = _mm_mulhi_epu16(src_rb, src_scale_wide);
-
- // Get alpha and green pixels into higher byte of each word.
- // (a, 0, g, 0, a, 0, g, 0, a, 0, g, 0, a, 0, g, 0)
- __m128i src_ag = _mm_and_si128(ag_mask, src_pixel);
-
- // Multiply by scale.
- // (4 x (as.h, as.l, gs.h, gs.l))
- src_ag = _mm_mulhi_epu16(src_ag, src_scale_wide);
-
- // Clear the lower byte of the a*scale and g*scale results
- // (4 x (as.h, 0, gs.h, 0))
- src_ag = _mm_and_si128(src_ag, ag_mask);
-
- // Operations the destination pixels are the same as on the
- // source pixels. See the comments above.
- __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel);
- dst_rb = _mm_mulhi_epu16(dst_rb, dst_scale_wide);
- __m128i dst_ag = _mm_and_si128(ag_mask, dst_pixel);
- dst_ag = _mm_mulhi_epu16(dst_ag, dst_scale_wide);
- dst_ag = _mm_and_si128(dst_ag, ag_mask);
-
- // Combine back into RGBA.
- // (4 x (as.h, rs.h, gs.h, bs.h))
- src_pixel = _mm_or_si128(src_rb, src_ag);
- dst_pixel = _mm_or_si128(dst_rb, dst_ag);
+ src_pixel = SkAlphaMulQ_SSE2(src_pixel, src_scale);
+ dst_pixel = SkAlphaMulQ_SSE2(dst_pixel, dst_scale);
// Add result
__m128i result = _mm_add_epi8(src_pixel, dst_pixel);
@@ -368,34 +325,12 @@
const __m128i *s = reinterpret_cast<const __m128i*>(src);
__m128i *d = reinterpret_cast<__m128i*>(dst);
- __m128i rb_mask = _mm_set1_epi32(0x00FF00FF);
- __m128i src_scale_wide = _mm_set1_epi16(scale);
__m128i color_wide = _mm_set1_epi32(color);
while (count >= 4) {
- // Load 4 pixels each of src and dest.
__m128i src_pixel = _mm_loadu_si128(s);
+ src_pixel = SkAlphaMulQ_SSE2(src_pixel, scale);
- // Get red and blue pixels into lower byte of each word.
- __m128i src_rb = _mm_and_si128(rb_mask, src_pixel);
-
- // Get alpha and green into lower byte of each word.
- __m128i src_ag = _mm_srli_epi16(src_pixel, 8);
-
- // Multiply by scale.
- src_rb = _mm_mullo_epi16(src_rb, src_scale_wide);
- src_ag = _mm_mullo_epi16(src_ag, src_scale_wide);
-
- // Divide by 256.
- src_rb = _mm_srli_epi16(src_rb, 8);
- src_ag = _mm_andnot_si128(rb_mask, src_ag);
-
- // Combine back into RGBA.
- src_pixel = _mm_or_si128(src_rb, src_ag);
-
- // Add color to result.
__m128i result = _mm_add_epi8(color_wide, src_pixel);
-
- // Store result.
_mm_store_si128(d, result);
s++;
d++;
diff --git a/src/opts/SkColor_opts_SSE2.h b/src/opts/SkColor_opts_SSE2.h
index 741d1ab..c1e41f0 100644
--- a/src/opts/SkColor_opts_SSE2.h
+++ b/src/opts/SkColor_opts_SSE2.h
@@ -42,24 +42,42 @@
return prod;
}
+static const __m128i rb_mask = _mm_set1_epi32(0x00FF00FF);
+static const __m128i ag_mask = _mm_set1_epi32(0xFF00FF00);
+
// Portable version SkAlphaMulQ is in SkColorPriv.h.
static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const __m128i& scale) {
- __m128i mask = _mm_set1_epi32(0xFF00FF);
__m128i s = _mm_or_si128(_mm_slli_epi32(scale, 16), scale);
// uint32_t rb = ((c & mask) * scale) >> 8
- __m128i rb = _mm_and_si128(mask, c);
+ __m128i rb = _mm_and_si128(rb_mask, c);
rb = _mm_mullo_epi16(rb, s);
rb = _mm_srli_epi16(rb, 8);
// uint32_t ag = ((c >> 8) & mask) * scale
__m128i ag = _mm_srli_epi16(c, 8);
- ASSERT_EQ(ag, _mm_and_si128(mask, ag)); // ag = _mm_srli_epi16(c, 8) did this for us.
+ ASSERT_EQ(ag, _mm_and_si128(rb_mask, ag)); // ag = _mm_srli_epi16(c, 8) did this for us.
ag = _mm_mullo_epi16(ag, s);
// (rb & mask) | (ag & ~mask)
- ASSERT_EQ(rb, _mm_and_si128(mask, rb)); // rb = _mm_srli_epi16(rb, 8) did this for us.
- ag = _mm_andnot_si128(mask, ag);
+ ASSERT_EQ(rb, _mm_and_si128(rb_mask, rb)); // rb = _mm_srli_epi16(rb, 8) did this for us.
+ ag = _mm_and_si128(ag_mask, ag);
+ return _mm_or_si128(rb, ag);
+}
+
+// Fast path for SkAlphaMulQ_SSE2 with a constant scale factor.
+static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const unsigned scale) {
+ __m128i s = _mm_set1_epi16(scale << 8); // Move scale factor to upper byte of word.
+
+ // With mulhi, red and blue values are already in the right place and
+ // don't need to be divided by 256.
+ __m128i rb = _mm_and_si128(rb_mask, c);
+ rb = _mm_mulhi_epu16(rb, s);
+
+ __m128i ag = _mm_and_si128(ag_mask, c);
+ ag = _mm_mulhi_epu16(ag, s); // Alpha and green values are in the higher byte of each word.
+ ag = _mm_and_si128(ag_mask, ag);
+
return _mm_or_si128(rb, ag);
}