Add AVX2 version of ConvolveVertically

src/opts/SkBitmapFilter_opts.h

Index: src/opts/SkBitmapFilter_opts.h
diff --git a/src/opts/SkBitmapFilter_opts.h b/src/opts/SkBitmapFilter_opts.h
index f22b5c2368050681f6979fe08dfaeeec8405c6ca..94d97f74a80de6f1251a78d47010a6ec34a0c514 100644
--- a/src/opts/SkBitmapFilter_opts.h
+++ b/src/opts/SkBitmapFilter_opts.h
@@ -10,7 +10,9 @@
 
 #include "SkConvolver.h"
 
-#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2
+    #include <immintrin.h>

[mtklein_C, 2016/12/05 18:42:46]

I think we can just use immintrin.h for everything >= SSE2.

[xiangze.zhang, 2016/12/07 11:17:01]

Done.

+#elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
     #include <emmintrin.h>
 #elif defined(SK_ARM_HAS_NEON)
     #include <arm_neon.h>
@@ -18,7 +20,160 @@
 
 namespace SK_OPTS_NS {
 
-#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2
+
+    static SK_ALWAYS_INLINE
+    void ComputeCoefficientRow(SkConvolutionFilter1D::ConvolutionFixed filterValue, const unsigned char* sourceDataRows,

[mtklein_C, 2016/12/05 18:42:45]

Typically we name static functions with lower_case_and_underscores(), and pass
mutated values as pointers instead of references so that it's clear they're
being mutated at the call site.   In my experience this doesn't have any effect
on code generation, especially with SK_ALWAYS_INLINE.

[xiangze.zhang, 2016/12/07 11:17:01]

Done.

+                               __m256i& Y01, __m256i& Y23, __m256i& Y45, __m256i& Y67) {
+        __m256i coefs = _mm256_set1_epi16(filterValue);
+        __m256i pixels = _mm256_lddqu_si256(reinterpret_cast<const __m256i *>(sourceDataRows));

[mtklein_C, 2016/12/05 18:42:45]

I have always shied away from lddqu, instead just using loadu.  Is it important
here?  Not sure I see any difference when benching.

[xiangze.zhang, 2016/12/07 11:17:01]

Done.

+        __m256i zero = _mm256_setzero_si256();
+
+        // [16] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
+        __m256i Yt1 = _mm256_unpacklo_epi8(pixels, zero);

[mtklein_C, 2016/12/05 18:42:45]

Does Yt1 mean ymm, temporary #1?

We usually find this sort of code a lot easier to read and maintain if you give
each value some sort of meaningful name (like pixels_03_16bit) then never change
it, letting the compiler handle register allocation.  It can also help
readability if you can find ways to vertically align analogous parts of register
or operation names.

E.g.

__m256i pixels_03_16bit = _mm256_unpacklo_epi8(pixels, zero);

__m256i scaled_03_lo = _mm256_mullo_epi16(pixels_03_16bit, coefs),
        scaled_03_hi = _mm256_mulhi_epi16(pixels_03_16bit, coefs);

Y01 = _mm256_add_epi32(Y01, _mm256_unpacklo_epi16(scaled_03_lo, scaled_03_hi));
Y23 = _mm256_add_epi32(Y23, _mm256_unpackhi_epi16(scaled_03_lo, scaled_03_hi));
...

[xiangze.zhang, 2016/12/07 11:17:01]

Done.

+        // [32] cja3 cjb3 cjg3 cjr3 cja2 cjb2 cjg2 cjr2
+        __m256i Yt2 = _mm256_mulhi_epi16(Yt1, coefs);
+        // [32] cja1 cjb1 cjg1 cjr1 cja0 cjb0 cjg0 cjr0
+        Yt1 = _mm256_mullo_epi16(Yt1, coefs);
+
+        // pixel 0 and 1
+        __m256i Yt3 = _mm256_unpacklo_epi16(Yt1, Yt2);
+        Y01 = _mm256_add_epi32(Y01, Yt3);
+        // pixel 2 and 3
+        Yt3 = _mm256_unpackhi_epi16(Yt1, Yt2);
+        Y23 = _mm256_add_epi32(Y23, Yt3);
+
+        // [16] a7 b7 g7 r7 a6 b6 g6 r6 a5 b5 g5 r5 a4 b4 g4 r4
+        Yt1 = _mm256_unpackhi_epi8(pixels, zero);
+        // [32] cja7 cjb7 cjg7 cjr7 cja6 cjb6 cjg6 cjr6
+        Yt2 = _mm256_mulhi_epi16(Yt1, coefs);
+        // [32] cja5 cjb5 cjg5 cjr5 cja4 cjb4 cjg4 cjr4
+        Yt1 = _mm256_mullo_epi16(Yt1, coefs);
+
+        // pixel 4 and 5
+        Yt3 = _mm256_unpacklo_epi16(Yt1, Yt2);
+        Y45 = _mm256_add_epi32(Y45, Yt3);
+        // pixel 6 and 7
+        Yt3 = _mm256_unpackhi_epi16(Yt1, Yt2);
+        Y67 = _mm256_add_epi32(Y67, Yt3);
+    }
+
+    template<bool hasAlpha>
+    void ConvolveVertically(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
+                            int filterLength,
+                            unsigned char* const * sourceDataRows,
+                            int pixelWidth,
+                            unsigned char* outRow) {
+
+        int outX, filterY;
+        int width = pixelWidth & ~7;
+        int length = filterLength & ~3;
+
+        __m256i Yt0, Yt1, Yt2;
+
+        // Output eight pixels per iteration (32 bytes).
+        for (outX = 0; outX < width; outX += 8) {
+            // Accumulated result for each pixel. 32 bits per RGBA channel.
+            __m256i Y01 = _mm256_setzero_si256();
+            __m256i Y23 = _mm256_setzero_si256();
+            __m256i Y45 = _mm256_setzero_si256();
+            __m256i Y67 = _mm256_setzero_si256();
+
+            // Convolve with 4 filter coefficient per iteration.
+            for (int filterY = 0; filterY < length; filterY += 4) {
+                ComputeCoefficientRow(filterValues[filterY], sourceDataRows[filterY] + outX * 4, Y01, Y23, Y45, Y67);

[mtklein_C, 2016/12/05 18:42:46]

This might line up neater with a couple "+ 0".

[xiangze.zhang, 2016/12/07 11:17:01]

Done.

+                ComputeCoefficientRow(filterValues[filterY + 1], sourceDataRows[filterY + 1] + outX * 4, Y01, Y23, Y45, Y67);
+                ComputeCoefficientRow(filterValues[filterY + 2], sourceDataRows[filterY + 2] + outX * 4, Y01, Y23, Y45, Y67);
+                ComputeCoefficientRow(filterValues[filterY + 3], sourceDataRows[filterY + 3] + outX * 4, Y01, Y23, Y45, Y67);
+            }
+            for (filterY = length; filterY < filterLength; filterY++) {
+                ComputeCoefficientRow(filterValues[filterY], sourceDataRows[filterY] + outX * 4, Y01, Y23, Y45, Y67);
+            }
+
+            // Shift right for fixed point implementation.
+            Y01 = _mm256_srai_epi32(Y01, SkConvolutionFilter1D::kShiftBits);
+            Y23 = _mm256_srai_epi32(Y23, SkConvolutionFilter1D::kShiftBits);
+            Y45 = _mm256_srai_epi32(Y45, SkConvolutionFilter1D::kShiftBits);
+            Y67 = _mm256_srai_epi32(Y67, SkConvolutionFilter1D::kShiftBits);
+
+            // Packing 32 bits |accum| to 16 bits per channel (signed saturation).
+            // [16] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
+            Yt0 = _mm256_packs_epi32(Y01, Y23);
+
+            // Packing 32 bits |accum| to 16 bits per channel (signed saturation).
+            // [16] a7 b7 g7 r7 a6 b6 g6 r6 a5 b5 g5 r5 a4 b4 g4 r4
+            Yt1 = _mm256_packs_epi32(Y45, Y67);
+
+            // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
+            //  [8] a7 b7 g7 r7 a6 b6 g6 r6 a5 b5 g5 r5 a4 b4 g4 r4 a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
+            Yt0 = _mm256_packus_epi16(Yt0, Yt1);
+
+            if (hasAlpha) {
+                // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
+                Yt1 = _mm256_srli_epi32(Yt0, 8);
+                // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
+                Yt2 = _mm256_max_epu8(Yt1, Yt0);  // Max of r and g.
+                // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
+                Yt1 = _mm256_srli_epi32(Yt0, 16);
+                // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
+                Yt2 = _mm256_max_epu8(Yt1, Yt2);  // Max of r and g and b.
+                // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
+                Yt2 = _mm256_slli_epi32(Yt2, 24);
+
+                // Make sure the value of alpha channel is always larger than maximum

[mtklein_C, 2016/12/05 18:42:46]

This comment might be better moved up just under if (hasAlpha) {

[xiangze.zhang, 2016/12/07 11:17:01]

Done.

+                // value of color channels.
+                Yt0 = _mm256_max_epu8(Yt2, Yt0);
+            } else {
+                __m256i mask = _mm256_set1_epi32(0xff000000);
+                Yt0 = _mm256_or_si256(Yt0, mask);
+            }
+
+            // Store the convolution result (32 bytes) and advance the pixel pointers.
+            _mm256_storeu_si256(reinterpret_cast<__m256i *>(outRow), Yt0);
+            outRow += 32;
+        }
+
+        if (pixelWidth & 7) {
+            __m256i Y01 = _mm256_setzero_si256();
+            __m256i Y23 = _mm256_setzero_si256();
+            __m256i Y45 = _mm256_setzero_si256();
+            __m256i Y67 = _mm256_setzero_si256();
+
+            for (int filterY = 0; filterY < filterLength; filterY++) {
+                ComputeCoefficientRow(filterValues[filterY], sourceDataRows[filterY] + outX * 4, Y01, Y23, Y45, Y67);
+            }
+
+            Y01 = _mm256_srai_epi32(Y01, SkConvolutionFilter1D::kShiftBits);
+            Y23 = _mm256_srai_epi32(Y23, SkConvolutionFilter1D::kShiftBits);
+            Y45 = _mm256_srai_epi32(Y45, SkConvolutionFilter1D::kShiftBits);
+            Y67 = _mm256_srai_epi32(Y67, SkConvolutionFilter1D::kShiftBits);
+
+            Yt0 = _mm256_packs_epi32(Y01, Y23);
+            Yt1 = _mm256_packs_epi32(Y45, Y67);
+            Yt0 = _mm256_packus_epi16(Yt0, Yt1);
+
+            if (hasAlpha) {
+                Yt1 = _mm256_srli_epi32(Yt0, 8);
+                Yt2 = _mm256_max_epu8(Yt1, Yt0);  // Max of r and g.

[mtklein_C, 2016/12/05 18:42:45]

This is the sort of code that using more variables with better names can help
make readable.  E.g.

if (hasAlpha) {
  // If alpha is less than r,g, or b, set it to their max.
  auto max_rg  = _mm256_max_epu8(pixels, _mm256_srli_epi32(pixels,  8)),
       max_rgb = _mm256_max_epu8(max_rg, _mm256_srli_epi32(pixels, 16));
  pixels = _mm256_max_epu8(pixels, _mm256_slli_epi32(max_rgb, 24));
} else {
  // Force opaque.
  pixels = _mm256_or_si256(pixels, _mm256_set1_epi32(0xff000000));
}

[xiangze.zhang, 2016/12/07 11:17:01]

Done.

+                Yt1 = _mm256_srli_epi32(Yt0, 16);
+                Yt2 = _mm256_max_epu8(Yt1, Yt2);  // Max of r and g and b.
+                Yt2 = _mm256_slli_epi32(Yt2, 24);
+                Yt0 = _mm256_max_epu8(Yt2, Yt0);
+            } else {
+                __m256i mask = _mm256_set1_epi32(0xff000000);
+                Yt0 = _mm256_or_si256(Yt0, mask);
+            }
+
+            for (int i = width; i < pixelWidth; i++) {

[mtklein_C, 2016/12/05 18:42:46]

I'd have expected a call to _mm256_maskstore_epi32()... is this way faster?

[xiangze.zhang, 2016/12/07 11:17:01]

I tried something like this:
    __m256i mask = _mm256_setr_epi32(0xffffffff, 0, 0, 0, 0, 0, 0, 0);
    for (int i = outX + 1; i < pixelWidth; i++) {
        __m256i shift = _mm256_setr_epi32(0, 0, 1, 2, 3, 4, 5, 6);
        mask = _mm256_permutevar8x32_epi32(mask, shift);
    }
    _mm256_maskstore_epi32(reinterpret_cast<int*>(outRow), mask, accum);

And nanobench time becomes a little longer.
It is weird that bitmap_scale_filter_64_256 time also change from 306us to
314us, which should have no pixels left to go through this code path.

+                *(reinterpret_cast<int*>(outRow)) = _mm_cvtsi128_si32(_mm256_castsi256_si128(Yt0));
+                __m256i rotate = _mm256_setr_epi32(1, 2, 3, 4, 5, 6, 7, 0);
+                Yt0 = _mm256_permutevar8x32_epi32(Yt0, rotate);
+                outRow += 4;
+            }
+        }
+    }
+
+#elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
 
     static SK_ALWAYS_INLINE void AccumRemainder(const unsigned char* pixelsLeft,
             const SkConvolutionFilter1D::ConvolutionFixed* filterValues, __m128i& accum, int r) {