Add SSSE3 Optimizations for premul and swap

src/opts/SkSwizzler_opts.h

Index: src/opts/SkSwizzler_opts.h
diff --git a/src/opts/SkSwizzler_opts.h b/src/opts/SkSwizzler_opts.h
index c7edcb9ece068ec29b321b95295163e908b6f004..2d8476d11b2c54867c7c0b73b7ba0a2d85f312be 100644
--- a/src/opts/SkSwizzler_opts.h
+++ b/src/opts/SkSwizzler_opts.h
@@ -174,6 +174,121 @@ static void swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) {
     swaprb_xxxa_portable(dst, src, count);
 }
 
+#elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+
+template <bool kSwapRB>
+static void premul_xxxa_should_swaprb(uint32_t dst[], const uint32_t src[], int count) {
+    const __m128i zeros = _mm_setzero_si128();
+    const __m128i _128 = _mm_set1_epi16(128);
+    const __m128i _257 = _mm_set1_epi16(257);
+    __m128i planar;
+    if (kSwapRB) {
+        planar = _mm_setr_epi8(2,6,10,14, 1,5,9,13, 0,4,8,12, 3,7,11,15);
+    } else {
+        planar = _mm_setr_epi8(0,4,8,12, 1,5,9,13, 2,6,10,14, 3,7,11,15);
+    }
+
+    while (count >= 8) {
+        // We'll load 8 pixels into 4 registers, each holding a 16-bit component plane.
+
+        // First just load the 8 interlaced pixels.
+        __m128i lo = _mm_loadu_si128((const __m128i*) (src + 0)), // bgrabgra bgrabgra
+                hi = _mm_loadu_si128((const __m128i*) (src + 4)); // BGRABGRA BGRABGRA
+
+        // Swizzle them to 8-bit planar.
+        lo = _mm_shuffle_epi8(lo, planar);                        // bbbbgggg rrrraaaa
+        hi = _mm_shuffle_epi8(lo, planar);                        // BBBBGGGG RRRRAAAA
+        __m128i bg = _mm_unpacklo_epi32(lo, hi),                  // bbbbBBBB ggggGGGG
+                ra = _mm_unpackhi_epi32(lo, hi);                  // rrrrRRRR aaaaAAAA
+
+        // Unpack to 16-bit planar.
+        __m128i b = _mm_unpacklo_epi8(bg, zeros),                 // b_b_b_b_ B_B_B_B_
+                g = _mm_unpackhi_epi8(bg, zeros),                 // g_g_g_g_ G_G_G_G_
+                r = _mm_unpacklo_epi8(ra, zeros),                 // r_r_r_r_ R_R_R_R_
+                a = _mm_unpackhi_epi8(ra, zeros);                 // a_a_a_a_ A_A_A_A_
+
+        // Premultiply!  (x+127)/255 == ((x+128)*257)>>16 for 0 <= x <= 255*255.
+        b = _mm_mulhi_epu16(_mm_add_epi16(_mm_mullo_epi16(b, a), _128), _257);
+        g = _mm_mulhi_epu16(_mm_add_epi16(_mm_mullo_epi16(g, a), _128), _257);
+        r = _mm_mulhi_epu16(_mm_add_epi16(_mm_mullo_epi16(r, a), _128), _257);
+
+        // Repack into interlaced pixels.
+        bg = _mm_or_si128(b, _mm_slli_epi16(g, 8));               // bgbgbgbg BGBGBGBG
+        ra = _mm_or_si128(r, _mm_slli_epi16(a, 8));               // rararara RARARARA
+        lo = _mm_unpacklo_epi16(bg, ra);                          // bgrabgra bgrabgra
+        hi = _mm_unpackhi_epi16(bg, ra);                          // BRGABGRA BGRABGRA
+
+        // Store interlaced pixels.
+        _mm_storeu_si128((__m128i*) (dst + 0), lo);
+        _mm_storeu_si128((__m128i*) (dst + 4), hi);
+
+        src += 8;
+        dst += 8;
+        count -= 8;
+    }
+
+    if (count >= 4) {

[mtklein, 2016/01/19 18:28:30]

OK, now that we've got count >= 8 in shape, let's do something like this?

auto premul8 = [](__m128i* lo, __m128i* hi) {
    // Swizzle them to 8-bit planar.
    *lo = _mm_shuffle_epi8(*lo, planar);
    ...
    *lo = _mm_unpacklo_epi16(bg, ra);
    *hi = _mm_unpackhi_epi16(bg, ra);
};

while (n >= 8) {
    __m128i lo = _mm_loadu_si128(...),
            hi = _mm_loadu_si128(....);
    premul8(&lo, &hi);
    _mm_storeu_si128(..., lo);
    _mm_storeu_si128(..., hi);
    ...
}
if (n >= 4) {
    __m128i lo = _mm_loadu_si128(...);
            hi = _mm_setzero_si128();
    premul8(&lo, &hi);
    _mm_storeu_si128(..., lo);
    ...
}
handle n <= 3

[msarett, 2016/01/19 19:17:43]

Done.

+        // First just load 4 interlaced pixels.
+        __m128i lo = _mm_loadu_si128((const __m128i*) src);       // bgrabgra bgrabgra
+
+        // Swizzle them to 8-bit planar.
+        lo = _mm_shuffle_epi8(lo, planar);                        // bbbbgggg rrrraaaa
+        __m128i bg = _mm_unpacklo_epi32(lo, zeros),               // bbbb____ gggg____
+                ra = _mm_unpackhi_epi32(lo, zeros);               // rrrr____ aaaa____
+
+        // Unpack to 16-bit planar.
+        __m128i b = _mm_unpacklo_epi8(bg, zeros),                 // b_b_b_b_ ________
+                g = _mm_unpackhi_epi8(bg, zeros),                 // g_g_g_g_ ________
+                r = _mm_unpacklo_epi8(ra, zeros),                 // r_r_r_r_ ________
+                a = _mm_unpackhi_epi8(ra, zeros);                 // a_a_a_a_ ________
+
+        // Premultiply!  (x+127)/255 == ((x+128)*257)>>16 for 0 <= x <= 255*255.
+        b = _mm_mulhi_epu16(_mm_add_epi16(_mm_mullo_epi16(b, a), _128), _257);
+        g = _mm_mulhi_epu16(_mm_add_epi16(_mm_mullo_epi16(g, a), _128), _257);
+        r = _mm_mulhi_epu16(_mm_add_epi16(_mm_mullo_epi16(r, a), _128), _257);
+
+        // Repack into interlaced pixels.
+        bg = _mm_or_si128(b, _mm_slli_epi16(g, 8));               // bgbgbgbg ________
+        ra = _mm_or_si128(r, _mm_slli_epi16(a, 8));               // rararara ________
+        lo = _mm_unpacklo_epi16(bg, ra);                          // bgrabgra bgrabgra
+
+        // Store interlaced pixels.
+        _mm_storeu_si128((__m128i*) dst, lo);
+
+        src += 4;
+        dst += 4;
+        count -= 4;
+    }
+
+    // Call portable code to finish up the tail of [0,4) pixels.
+    auto proc = kSwapRB ? premul_swaprb_xxxa_portable : premul_xxxa_portable;
+    proc(dst, src, count);
+}
+
+static void premul_xxxa(uint32_t dst[], const uint32_t src[], int count) {
+    premul_xxxa_should_swaprb<false>(dst, src, count);
+}
+
+static void premul_swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) {
+    premul_xxxa_should_swaprb<true>(dst, src, count);
+}
+
+static void swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) {
+    const __m128i swapRB = _mm_setr_epi8(2,1,0,3, 6,5,4,7, 10,9,8,11, 14,13,12,15);
+
+    while (count >= 4) {
+        __m128i bgra = _mm_loadu_si128((const __m128i*) src);
+        __m128i rgba = _mm_shuffle_epi8(bgra, swapRB);
+        _mm_storeu_si128((__m128i*) dst, rgba);
+
+        src += 4;
+        dst += 4;
+        count -= 4;
+    }
+
+    swaprb_xxxa_portable(dst, src, count);
+}
+
 #else
 
 static void premul_xxxa(uint32_t dst[], const uint32_t src[], int count) {