ARM Skia NEON patches - 33 - Convolution filter

NEON version of the convolutionProcs

The bitmap_scale benchmark is now twice as fast on ARM.

Signed-off-by: Kévin PETIT <kevin.petit@arm.com>

BUG=

Committed: http://code.google.com/p/skia/source/detail?r=12154

R=djsollen@google.com, mtklein@google.com, humper@google.com, epoger@google.com

Author: kevin.petit.arm@gmail.com

Review URL: https://codereview.chromium.org/27533004

git-svn-id: http://skia.googlecode.com/svn/trunk@12166 2bbb7eff-a529-9590-31e7-b0007b416f81

2 files changed, 518 insertions, 286 deletions

diff --git a/src/opts/SkBitmapProcState_arm_neon.cpp b/src/opts/SkBitmapProcState_arm_neon.cpp
index d50707dce0..35ba462c82 100644
--- a/src/opts/SkBitmapProcState_arm_neon.cpp
+++ b/src/opts/SkBitmapProcState_arm_neon.cpp
@@ -90,3 +90,423 @@ const SkBitmapProcState::SampleProc16 gSkBitmapProcStateSample16_neon[] = {
     // Don't support A8 -> 565
     NULL, NULL, NULL, NULL
 };
+
+///////////////////////////////////////////////////////////////////////////////
+
+#include <arm_neon.h>
+#include "SkConvolver.h"
+
+// Convolves horizontally along a single row. The row data is given in
+// |srcData| and continues for the numValues() of the filter.
+void convolveHorizontally_neon(const unsigned char* srcData,
+                               const SkConvolutionFilter1D& filter,
+                               unsigned char* outRow,
+                               bool hasAlpha) {
+    // Loop over each pixel on this row in the output image.
+    int numValues = filter.numValues();
+    for (int outX = 0; outX < numValues; outX++) {
+        uint8x8_t coeff_mask0 = vcreate_u8(0x0100010001000100);
+        uint8x8_t coeff_mask1 = vcreate_u8(0x0302030203020302);
+        uint8x8_t coeff_mask2 = vcreate_u8(0x0504050405040504);
+        uint8x8_t coeff_mask3 = vcreate_u8(0x0706070607060706);
+        // Get the filter that determines the current output pixel.
+        int filterOffset, filterLength;
+        const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
+            filter.FilterForValue(outX, &filterOffset, &filterLength);
+
+        // Compute the first pixel in this row that the filter affects. It will
+        // touch |filterLength| pixels (4 bytes each) after this.
+        const unsigned char* rowToFilter = &srcData[filterOffset * 4];
+
+        // Apply the filter to the row to get the destination pixel in |accum|.
+        int32x4_t accum = vdupq_n_s32(0);
+        for (int filterX = 0; filterX < filterLength >> 2; filterX++) {
+            // Load 4 coefficients
+            int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
+            coeffs = vld1_s16(filterValues);
+            coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
+            coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
+            coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
+            coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
+
+            // Load pixels and calc
+            uint8x16_t pixels = vld1q_u8(rowToFilter);
+            int16x8_t p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));
+            int16x8_t p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels)));
+
+            int16x4_t p0_src = vget_low_s16(p01_16);
+            int16x4_t p1_src = vget_high_s16(p01_16);
+            int16x4_t p2_src = vget_low_s16(p23_16);
+            int16x4_t p3_src = vget_high_s16(p23_16);
+
+            int32x4_t p0 = vmull_s16(p0_src, coeff0);
+            int32x4_t p1 = vmull_s16(p1_src, coeff1);
+            int32x4_t p2 = vmull_s16(p2_src, coeff2);
+            int32x4_t p3 = vmull_s16(p3_src, coeff3);
+
+            accum += p0;
+            accum += p1;
+            accum += p2;
+            accum += p3;
+
+            // Advance the pointers
+            rowToFilter += 16;
+            filterValues += 4;
+        }
+        int r = filterLength & 3;
+        if (r) {
+            const uint16_t mask[4][4] = {
+                {0, 0, 0, 0},
+                {0xFFFF, 0, 0, 0},
+                {0xFFFF, 0xFFFF, 0, 0},
+                {0xFFFF, 0xFFFF, 0xFFFF, 0}
+            };
+            uint16x4_t coeffs;
+            int16x4_t coeff0, coeff1, coeff2;
+            coeffs = vld1_u16(reinterpret_cast<const uint16_t*>(filterValues));
+            coeffs &= vld1_u16(&mask[r][0]);
+            coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask0));
+            coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask1));
+            coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask2));
+
+            // Load pixels and calc
+            uint8x16_t pixels = vld1q_u8(rowToFilter);
+            int16x8_t p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));
+            int16x8_t p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels)));
+            int32x4_t p0 = vmull_s16(vget_low_s16(p01_16), coeff0);
+            int32x4_t p1 = vmull_s16(vget_high_s16(p01_16), coeff1);
+            int32x4_t p2 = vmull_s16(vget_low_s16(p23_16), coeff2);
+
+            accum += p0;
+            accum += p1;
+            accum += p2;
+        }
+
+        // Bring this value back in range. All of the filter scaling factors
+        // are in fixed point with kShiftBits bits of fractional part.
+        accum = vshrq_n_s32(accum, SkConvolutionFilter1D::kShiftBits);
+
+        // Pack and store the new pixel.
+        int16x4_t accum16 = vqmovn_s32(accum);
+        uint8x8_t accum8 = vqmovun_s16(vcombine_s16(accum16, accum16));
+        vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow), vreinterpret_u32_u8(accum8), 0);
+        outRow += 4;
+    }
+}
+
+// Does vertical convolution to produce one output row. The filter values and
+// length are given in the first two parameters. These are applied to each
+// of the rows pointed to in the |sourceDataRows| array, with each row
+// being |pixelWidth| wide.
+//
+// The output must have room for |pixelWidth * 4| bytes.
+template<bool hasAlpha>
+void convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
+                             int filterLength,
+                             unsigned char* const* sourceDataRows,
+                             int pixelWidth,
+                             unsigned char* outRow) {
+    int width = pixelWidth & ~3;
+
+    int32x4_t accum0, accum1, accum2, accum3;
+    int16x4_t coeff16;
+
+    // Output four pixels per iteration (16 bytes).
+    for (int outX = 0; outX < width; outX += 4) {
+
+        // Accumulated result for each pixel. 32 bits per RGBA channel.
+        accum0 = accum1 = accum2 = accum3 = vdupq_n_s32(0);
+
+        // Convolve with one filter coefficient per iteration.
+        for (int filterY = 0; filterY < filterLength; filterY++) {
+
+            // Duplicate the filter coefficient 4 times.
+            // [16] cj cj cj cj
+            coeff16 = vdup_n_s16(filterValues[filterY]);
+
+            // Load four pixels (16 bytes) together.
+            // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
+            uint8x16_t src8 = vld1q_u8(&sourceDataRows[filterY][outX << 2]);
+
+            int16x8_t src16_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(src8)));
+            int16x8_t src16_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(src8)));
+            int16x4_t src16_0 = vget_low_s16(src16_01);
+            int16x4_t src16_1 = vget_high_s16(src16_01);
+            int16x4_t src16_2 = vget_low_s16(src16_23);
+            int16x4_t src16_3 = vget_high_s16(src16_23);
+
+            accum0 += vmull_s16(src16_0, coeff16);
+            accum1 += vmull_s16(src16_1, coeff16);
+            accum2 += vmull_s16(src16_2, coeff16);
+            accum3 += vmull_s16(src16_3, coeff16);
+        }
+
+        // Shift right for fixed point implementation.
+        accum0 = vshrq_n_s32(accum0, SkConvolutionFilter1D::kShiftBits);
+        accum1 = vshrq_n_s32(accum1, SkConvolutionFilter1D::kShiftBits);
+        accum2 = vshrq_n_s32(accum2, SkConvolutionFilter1D::kShiftBits);
+        accum3 = vshrq_n_s32(accum3, SkConvolutionFilter1D::kShiftBits);
+
+        // Packing 32 bits |accum| to 16 bits per channel (signed saturation).
+        // [16] a1 b1 g1 r1 a0 b0 g0 r0
+        int16x8_t accum16_0 = vcombine_s16(vqmovn_s32(accum0), vqmovn_s32(accum1));
+        // [16] a3 b3 g3 r3 a2 b2 g2 r2
+        int16x8_t accum16_1 = vcombine_s16(vqmovn_s32(accum2), vqmovn_s32(accum3));
+
+        // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
+        // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
+        uint8x16_t accum8 = vcombine_u8(vqmovun_s16(accum16_0), vqmovun_s16(accum16_1));
+
+        if (hasAlpha) {
+            // Compute the max(ri, gi, bi) for each pixel.
+            // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
+            uint8x16_t a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 8));
+            // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
+            uint8x16_t b = vmaxq_u8(a, accum8); // Max of r and g
+            // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
+            a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 16));
+            // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
+            b = vmaxq_u8(a, b); // Max of r and g and b.
+            // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
+            b = vreinterpretq_u8_u32(vshlq_n_u32(vreinterpretq_u32_u8(b), 24));
+
+            // Make sure the value of alpha channel is always larger than maximum
+            // value of color channels.
+            accum8 = vmaxq_u8(b, accum8);
+        } else {
+            // Set value of alpha channels to 0xFF.
+            accum8 = vreinterpretq_u8_u32(vreinterpretq_u32_u8(accum8) | vdupq_n_u32(0xFF000000));
+        }
+
+        // Store the convolution result (16 bytes) and advance the pixel pointers.
+        vst1q_u8(outRow, accum8);
+        outRow += 16;
+    }
+
+    // Process the leftovers when the width of the output is not divisible
+    // by 4, that is at most 3 pixels.
+    int r = pixelWidth & 3;
+    if (r) {
+
+        accum0 = accum1 = accum2 = vdupq_n_s32(0);
+
+        for (int filterY = 0; filterY < filterLength; ++filterY) {
+            coeff16 = vdup_n_s16(filterValues[filterY]);
+
+            // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
+            uint8x16_t src8 = vld1q_u8(&sourceDataRows[filterY][width << 2]);
+
+            int16x8_t src16_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(src8)));
+            int16x8_t src16_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(src8)));
+            int16x4_t src16_0 = vget_low_s16(src16_01);
+            int16x4_t src16_1 = vget_high_s16(src16_01);
+            int16x4_t src16_2 = vget_low_s16(src16_23);
+
+            accum0 += vmull_s16(src16_0, coeff16);
+            accum1 += vmull_s16(src16_1, coeff16);
+            accum2 += vmull_s16(src16_2, coeff16);
+        }
+
+        accum0 = vshrq_n_s32(accum0, SkConvolutionFilter1D::kShiftBits);
+        accum1 = vshrq_n_s32(accum1, SkConvolutionFilter1D::kShiftBits);
+        accum2 = vshrq_n_s32(accum2, SkConvolutionFilter1D::kShiftBits);
+
+        int16x8_t accum16_0 = vcombine_s16(vqmovn_s32(accum0), vqmovn_s32(accum1));
+        int16x8_t accum16_1 = vcombine_s16(vqmovn_s32(accum2), vqmovn_s32(accum2));
+
+        uint8x16_t accum8 = vcombine_u8(vqmovun_s16(accum16_0), vqmovun_s16(accum16_1));
+
+        if (hasAlpha) {
+            // Compute the max(ri, gi, bi) for each pixel.
+            // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
+            uint8x16_t a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 8));
+            // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
+            uint8x16_t b = vmaxq_u8(a, accum8); // Max of r and g
+            // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
+            a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 16));
+            // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
+            b = vmaxq_u8(a, b); // Max of r and g and b.
+            // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
+            b = vreinterpretq_u8_u32(vshlq_n_u32(vreinterpretq_u32_u8(b), 24));
+
+            // Make sure the value of alpha channel is always larger than maximum
+            // value of color channels.
+            accum8 = vmaxq_u8(b, accum8);
+        } else {
+            // Set value of alpha channels to 0xFF.
+            accum8 = vreinterpretq_u8_u32(vreinterpretq_u32_u8(accum8) | vdupq_n_u32(0xFF000000));
+        }
+
+        switch(r) {
+        case 1:
+            vst1q_lane_u32(reinterpret_cast<uint32_t*>(outRow), vreinterpretq_u32_u8(accum8), 0);
+            break;
+        case 2:
+            vst1_u32(reinterpret_cast<uint32_t*>(outRow),
+                     vreinterpret_u32_u8(vget_low_u8(accum8)));
+            break;
+        case 3:
+            vst1_u32(reinterpret_cast<uint32_t*>(outRow),
+                     vreinterpret_u32_u8(vget_low_u8(accum8)));
+            vst1q_lane_u32(reinterpret_cast<uint32_t*>(outRow+8), vreinterpretq_u32_u8(accum8), 2);
+            break;
+        }
+    }
+}
+
+void convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
+                             int filterLength,
+                             unsigned char* const* sourceDataRows,
+                             int pixelWidth,
+                             unsigned char* outRow,
+                             bool sourceHasAlpha) {
+    if (sourceHasAlpha) {
+        convolveVertically_neon<true>(filterValues, filterLength,
+                                      sourceDataRows, pixelWidth,
+                                      outRow);
+    } else {
+        convolveVertically_neon<false>(filterValues, filterLength,
+                                       sourceDataRows, pixelWidth,
+                                       outRow);
+    }
+}
+
+// Convolves horizontally along four rows. The row data is given in
+// |src_data| and continues for the num_values() of the filter.
+// The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
+// refer to that function for detailed comments.
+void convolve4RowsHorizontally_neon(const unsigned char* srcData[4],
+                                    const SkConvolutionFilter1D& filter,
+                                    unsigned char* outRow[4]) {
+
+    uint8x8_t coeff_mask0 = vcreate_u8(0x0100010001000100);
+    uint8x8_t coeff_mask1 = vcreate_u8(0x0302030203020302);
+    uint8x8_t coeff_mask2 = vcreate_u8(0x0504050405040504);
+    uint8x8_t coeff_mask3 = vcreate_u8(0x0706070607060706);
+    int num_values = filter.numValues();
+
+    int filterOffset, filterLength;
+    // |mask| will be used to decimate all extra filter coefficients that are
+    // loaded by SIMD when |filter_length| is not divisible by 4.
+    // mask[0] is not used in following algorithm.
+    const uint16_t mask[4][4] = {
+        {0, 0, 0, 0},
+        {0xFFFF, 0, 0, 0},
+        {0xFFFF, 0xFFFF, 0, 0},
+        {0xFFFF, 0xFFFF, 0xFFFF, 0}
+    };
+
+    // Output one pixel each iteration, calculating all channels (RGBA) together.
+    for (int outX = 0; outX < num_values; outX++) {
+
+        const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
+        filter.FilterForValue(outX, &filterOffset, &filterLength);
+
+        // four pixels in a column per iteration.
+        int32x4_t accum0 = vdupq_n_s32(0);
+        int32x4_t accum1 = vdupq_n_s32(0);
+        int32x4_t accum2 = vdupq_n_s32(0);
+        int32x4_t accum3 = vdupq_n_s32(0);
+
+        int start = (filterOffset<<2);
+
+        // We will load and accumulate with four coefficients per iteration.
+        for (int filter_x = 0; filter_x < (filterLength >> 2); filter_x++) {
+            int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
+
+            coeffs = vld1_s16(filterValues);
+            coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
+            coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
+            coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
+            coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
+
+            uint8x16_t pixels;
+            int16x8_t p01_16, p23_16;
+            int32x4_t p0, p1, p2, p3;
+
+
+#define ITERATION(src, accum)                                       \
+    pixels = vld1q_u8(src);                                         \
+    p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));  \
+    p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels))); \
+    p0 = vmull_s16(vget_low_s16(p01_16), coeff0);                   \
+    p1 = vmull_s16(vget_high_s16(p01_16), coeff1);                  \
+    p2 = vmull_s16(vget_low_s16(p23_16), coeff2);                   \
+    p3 = vmull_s16(vget_high_s16(p23_16), coeff3);                  \
+    accum += p0;                                                    \
+    accum += p1;                                                    \
+    accum += p2;                                                    \
+    accum += p3
+
+            ITERATION(srcData[0] + start, accum0);
+            ITERATION(srcData[1] + start, accum1);
+            ITERATION(srcData[2] + start, accum2);
+            ITERATION(srcData[3] + start, accum3);
+
+            start += 16;
+            filterValues += 4;
+        }
+
+        int r = filterLength & 3;
+        if (r) {
+            int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
+            coeffs = vld1_s16(filterValues);
+            coeffs &= vreinterpret_s16_u16(vld1_u16(&mask[r][0]));
+            coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
+            coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
+            coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
+            coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
+
+            uint8x16_t pixels;
+            int16x8_t p01_16, p23_16;
+            int32x4_t p0, p1, p2, p3;
+
+            ITERATION(srcData[0] + start, accum0);
+            ITERATION(srcData[1] + start, accum1);
+            ITERATION(srcData[2] + start, accum2);
+            ITERATION(srcData[3] + start, accum3);
+        }
+
+        int16x4_t accum16;
+        uint8x8_t res0, res1, res2, res3;
+
+#define PACK_RESULT(accum, res)                                         \
+        accum = vshrq_n_s32(accum, SkConvolutionFilter1D::kShiftBits);  \
+        accum16 = vqmovn_s32(accum);                                    \
+        res = vqmovun_s16(vcombine_s16(accum16, accum16));
+
+        PACK_RESULT(accum0, res0);
+        PACK_RESULT(accum1, res1);
+        PACK_RESULT(accum2, res2);
+        PACK_RESULT(accum3, res3);
+
+        vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[0]), vreinterpret_u32_u8(res0), 0);
+        vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[1]), vreinterpret_u32_u8(res1), 0);
+        vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[2]), vreinterpret_u32_u8(res2), 0);
+        vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[3]), vreinterpret_u32_u8(res3), 0);
+        outRow[0] += 4;
+        outRow[1] += 4;
+        outRow[2] += 4;
+        outRow[3] += 4;
+    }
+}
+
+void applySIMDPadding_neon(SkConvolutionFilter1D *filter) {
+    // Padding |paddingCount| of more dummy coefficients after the coefficients
+    // of last filter to prevent SIMD instructions which load 8 or 16 bytes
+    // together to access invalid memory areas. We are not trying to align the
+    // coefficients right now due to the opaqueness of <vector> implementation.
+    // This has to be done after all |AddFilter| calls.
+    for (int i = 0; i < 8; ++i) {
+        filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0));
+    }
+}
+
+void platformConvolutionProcs_arm_neon(SkConvolutionProcs* procs) {
+    procs->fExtraHorizontalReads = 3;
+    procs->fConvolveVertically = &convolveVertically_neon;
+    procs->fConvolve4RowsHorizontally = &convolve4RowsHorizontally_neon;
+    procs->fConvolveHorizontally = &convolveHorizontally_neon;
+    procs->fApplySIMDPadding = &applySIMDPadding_neon;
+}
+
diff --git a/src/opts/SkBitmapProcState_opts_arm.cpp b/src/opts/SkBitmapProcState_opts_arm.cpp
index badb0f4d3b..96fbebd4e1 100644
--- a/src/opts/SkBitmapProcState_opts_arm.cpp
+++ b/src/opts/SkBitmapProcState_opts_arm.cpp
@@ -50,54 +50,54 @@ void SI8_D16_nofilter_DX_arm(const SkBitmapProcState& s,
         const uint16_t* SK_RESTRICT xx = (const uint16_t*)(xy + 1);
 
         asm volatile (
-                      "cmp        %[count8], #0                   \n\t"   // compare loop counter with 0
-                      "beq        2f                              \n\t"   // if loop counter == 0, exit
-                      "1:                                             \n\t"
-                      "ldmia      %[xx]!, {r5, r7, r9, r11}       \n\t"   // load ptrs to pixels 0-7
-                      "subs       %[count8], %[count8], #1        \n\t"   // decrement loop counter
-                      "uxth       r4, r5                          \n\t"   // extract ptr 0
-                      "mov        r5, r5, lsr #16                 \n\t"   // extract ptr 1
-                      "uxth       r6, r7                          \n\t"   // extract ptr 2
-                      "mov        r7, r7, lsr #16                 \n\t"   // extract ptr 3
-                      "ldrb       r4, [%[srcAddr], r4]            \n\t"   // load pixel 0 from image
-                      "uxth       r8, r9                          \n\t"   // extract ptr 4
-                      "ldrb       r5, [%[srcAddr], r5]            \n\t"   // load pixel 1 from image
-                      "mov        r9, r9, lsr #16                 \n\t"   // extract ptr 5
-                      "ldrb       r6, [%[srcAddr], r6]            \n\t"   // load pixel 2 from image
-                      "uxth       r10, r11                        \n\t"   // extract ptr 6
-                      "ldrb       r7, [%[srcAddr], r7]            \n\t"   // load pixel 3 from image
-                      "mov        r11, r11, lsr #16               \n\t"   // extract ptr 7
-                      "ldrb       r8, [%[srcAddr], r8]            \n\t"   // load pixel 4 from image
-                      "add        r4, r4, r4                      \n\t"   // double pixel 0 for RGB565 lookup
-                      "ldrb       r9, [%[srcAddr], r9]            \n\t"   // load pixel 5 from image
-                      "add        r5, r5, r5                      \n\t"   // double pixel 1 for RGB565 lookup
-                      "ldrb       r10, [%[srcAddr], r10]          \n\t"   // load pixel 6 from image
-                      "add        r6, r6, r6                      \n\t"   // double pixel 2 for RGB565 lookup
-                      "ldrb       r11, [%[srcAddr], r11]          \n\t"   // load pixel 7 from image
-                      "add        r7, r7, r7                      \n\t"   // double pixel 3 for RGB565 lookup
-                      "ldrh       r4, [%[table], r4]              \n\t"   // load pixel 0 RGB565 from colmap
-                      "add        r8, r8, r8                      \n\t"   // double pixel 4 for RGB565 lookup
-                      "ldrh       r5, [%[table], r5]              \n\t"   // load pixel 1 RGB565 from colmap
-                      "add        r9, r9, r9                      \n\t"   // double pixel 5 for RGB565 lookup
-                      "ldrh       r6, [%[table], r6]              \n\t"   // load pixel 2 RGB565 from colmap
-                      "add        r10, r10, r10                   \n\t"   // double pixel 6 for RGB565 lookup
-                      "ldrh       r7, [%[table], r7]              \n\t"   // load pixel 3 RGB565 from colmap
-                      "add        r11, r11, r11                   \n\t"   // double pixel 7 for RGB565 lookup
-                      "ldrh       r8, [%[table], r8]              \n\t"   // load pixel 4 RGB565 from colmap
-                      "ldrh       r9, [%[table], r9]              \n\t"   // load pixel 5 RGB565 from colmap
-                      "ldrh       r10, [%[table], r10]            \n\t"   // load pixel 6 RGB565 from colmap
-                      "ldrh       r11, [%[table], r11]            \n\t"   // load pixel 7 RGB565 from colmap
-                      "pkhbt      r5, r4, r5, lsl #16             \n\t"   // pack pixels 0 and 1
-                      "pkhbt      r6, r6, r7, lsl #16             \n\t"   // pack pixels 2 and 3
-                      "pkhbt      r8, r8, r9, lsl #16             \n\t"   // pack pixels 4 and 5
-                      "pkhbt      r10, r10, r11, lsl #16          \n\t"   // pack pixels 6 and 7
-                      "stmia      %[colors]!, {r5, r6, r8, r10}   \n\t"   // store last 8 pixels
-                      "bgt        1b                              \n\t"   // loop if counter > 0
-                      "2:                                             \n\t"
-                      : [xx] "+r" (xx), [count8] "+r" (count8), [colors] "+r" (colors)
-                      : [table] "r" (table), [srcAddr] "r" (srcAddr)
-                      : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
-                      );
+            "cmp        %[count8], #0                  \n\t"  // compare loop counter with 0
+            "beq        2f                             \n\t"  // if loop counter == 0, exit
+            "1:                                        \n\t"
+            "ldmia      %[xx]!, {r5, r7, r9, r11}      \n\t"  // load ptrs to pixels 0-7
+            "subs       %[count8], %[count8], #1       \n\t"  // decrement loop counter
+            "uxth       r4, r5                         \n\t"  // extract ptr 0
+            "mov        r5, r5, lsr #16                \n\t"  // extract ptr 1
+            "uxth       r6, r7                         \n\t"  // extract ptr 2
+            "mov        r7, r7, lsr #16                \n\t"  // extract ptr 3
+            "ldrb       r4, [%[srcAddr], r4]           \n\t"  // load pixel 0 from image
+            "uxth       r8, r9                         \n\t"  // extract ptr 4
+            "ldrb       r5, [%[srcAddr], r5]           \n\t"  // load pixel 1 from image
+            "mov        r9, r9, lsr #16                \n\t"  // extract ptr 5
+            "ldrb       r6, [%[srcAddr], r6]           \n\t"  // load pixel 2 from image
+            "uxth       r10, r11                       \n\t"  // extract ptr 6
+            "ldrb       r7, [%[srcAddr], r7]           \n\t"  // load pixel 3 from image
+            "mov        r11, r11, lsr #16              \n\t"  // extract ptr 7
+            "ldrb       r8, [%[srcAddr], r8]           \n\t"  // load pixel 4 from image
+            "add        r4, r4, r4                     \n\t"  // double pixel 0 for RGB565 lookup
+            "ldrb       r9, [%[srcAddr], r9]           \n\t"  // load pixel 5 from image
+            "add        r5, r5, r5                     \n\t"  // double pixel 1 for RGB565 lookup
+            "ldrb       r10, [%[srcAddr], r10]         \n\t"  // load pixel 6 from image
+            "add        r6, r6, r6                     \n\t"  // double pixel 2 for RGB565 lookup
+            "ldrb       r11, [%[srcAddr], r11]         \n\t"  // load pixel 7 from image
+            "add        r7, r7, r7                     \n\t"  // double pixel 3 for RGB565 lookup
+            "ldrh       r4, [%[table], r4]             \n\t"  // load pixel 0 RGB565 from colmap
+            "add        r8, r8, r8                     \n\t"  // double pixel 4 for RGB565 lookup
+            "ldrh       r5, [%[table], r5]             \n\t"  // load pixel 1 RGB565 from colmap
+            "add        r9, r9, r9                     \n\t"  // double pixel 5 for RGB565 lookup
+            "ldrh       r6, [%[table], r6]             \n\t"  // load pixel 2 RGB565 from colmap
+            "add        r10, r10, r10                  \n\t"  // double pixel 6 for RGB565 lookup
+            "ldrh       r7, [%[table], r7]             \n\t"  // load pixel 3 RGB565 from colmap
+            "add        r11, r11, r11                  \n\t"  // double pixel 7 for RGB565 lookup
+            "ldrh       r8, [%[table], r8]             \n\t"  // load pixel 4 RGB565 from colmap
+            "ldrh       r9, [%[table], r9]             \n\t"  // load pixel 5 RGB565 from colmap
+            "ldrh       r10, [%[table], r10]           \n\t"  // load pixel 6 RGB565 from colmap
+            "ldrh       r11, [%[table], r11]           \n\t"  // load pixel 7 RGB565 from colmap
+            "pkhbt      r5, r4, r5, lsl #16            \n\t"  // pack pixels 0 and 1
+            "pkhbt      r6, r6, r7, lsl #16            \n\t"  // pack pixels 2 and 3
+            "pkhbt      r8, r8, r9, lsl #16            \n\t"  // pack pixels 4 and 5
+            "pkhbt      r10, r10, r11, lsl #16         \n\t"  // pack pixels 6 and 7
+            "stmia      %[colors]!, {r5, r6, r8, r10}  \n\t"  // store last 8 pixels
+            "bgt        1b                             \n\t"  // loop if counter > 0
+            "2:                                        \n\t"
+            : [xx] "+r" (xx), [count8] "+r" (count8), [colors] "+r" (colors)
+            : [table] "r" (table), [srcAddr] "r" (srcAddr)
+            : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
+        );
 
         for (i = (count & 7); i > 0; --i) {
             src = srcAddr[*xx++]; *colors++ = table[src];
@@ -136,52 +136,52 @@ void SI8_opaque_D32_nofilter_DX_arm(const SkBitmapProcState& s,
         const uint16_t* xx = (const uint16_t*)(xy + 1);
 
         asm volatile (
-                      "subs       %[count], %[count], #8          \n\t"   // decrement count by 8, set flags
-                      "blt        2f                              \n\t"   // if count < 0, branch to singles
-                      "1:                                             \n\t"   // eights loop
-                      "ldmia      %[xx]!, {r5, r7, r9, r11}       \n\t"   // load ptrs to pixels 0-7
-                      "uxth       r4, r5                          \n\t"   // extract ptr 0
-                      "mov        r5, r5, lsr #16                 \n\t"   // extract ptr 1
-                      "uxth       r6, r7                          \n\t"   // extract ptr 2
-                      "mov        r7, r7, lsr #16                 \n\t"   // extract ptr 3
-                      "ldrb       r4, [%[srcAddr], r4]            \n\t"   // load pixel 0 from image
-                      "uxth       r8, r9                          \n\t"   // extract ptr 4
-                      "ldrb       r5, [%[srcAddr], r5]            \n\t"   // load pixel 1 from image
-                      "mov        r9, r9, lsr #16                 \n\t"   // extract ptr 5
-                      "ldrb       r6, [%[srcAddr], r6]            \n\t"   // load pixel 2 from image
-                      "uxth       r10, r11                        \n\t"   // extract ptr 6
-                      "ldrb       r7, [%[srcAddr], r7]            \n\t"   // load pixel 3 from image
-                      "mov        r11, r11, lsr #16               \n\t"   // extract ptr 7
-                      "ldrb       r8, [%[srcAddr], r8]            \n\t"   // load pixel 4 from image
-                      "ldrb       r9, [%[srcAddr], r9]            \n\t"   // load pixel 5 from image
-                      "ldrb       r10, [%[srcAddr], r10]          \n\t"   // load pixel 6 from image
-                      "ldrb       r11, [%[srcAddr], r11]          \n\t"   // load pixel 7 from image
-                      "ldr        r4, [%[table], r4, lsl #2]      \n\t"   // load pixel 0 SkPMColor from colmap
-                      "ldr        r5, [%[table], r5, lsl #2]      \n\t"   // load pixel 1 SkPMColor from colmap
-                      "ldr        r6, [%[table], r6, lsl #2]      \n\t"   // load pixel 2 SkPMColor from colmap
-                      "ldr        r7, [%[table], r7, lsl #2]      \n\t"   // load pixel 3 SkPMColor from colmap
-                      "ldr        r8, [%[table], r8, lsl #2]      \n\t"   // load pixel 4 SkPMColor from colmap
-                      "ldr        r9, [%[table], r9, lsl #2]      \n\t"   // load pixel 5 SkPMColor from colmap
-                      "ldr        r10, [%[table], r10, lsl #2]    \n\t"   // load pixel 6 SkPMColor from colmap
-                      "ldr        r11, [%[table], r11, lsl #2]    \n\t"   // load pixel 7 SkPMColor from colmap
-                      "subs       %[count], %[count], #8          \n\t"   // decrement loop counter
-                      "stmia      %[colors]!, {r4-r11}            \n\t"   // store 8 pixels
-                      "bge        1b                              \n\t"   // loop if counter >= 0
-                      "2:                                             \n\t"
-                      "adds       %[count], %[count], #8          \n\t"   // fix up counter, set flags
-                      "beq        4f                              \n\t"   // if count == 0, branch to exit
-                      "3:                                             \n\t"   // singles loop
-                      "ldrh       r4, [%[xx]], #2                 \n\t"   // load pixel ptr
-                      "subs       %[count], %[count], #1          \n\t"   // decrement loop counter
-                      "ldrb       r5, [%[srcAddr], r4]            \n\t"   // load pixel from image
-                      "ldr        r6, [%[table], r5, lsl #2]      \n\t"   // load SkPMColor from colmap
-                      "str        r6, [%[colors]], #4             \n\t"   // store pixel, update ptr
-                      "bne        3b                              \n\t"   // loop if counter != 0
-                      "4:                                             \n\t"   // exit
-                      : [xx] "+r" (xx), [count] "+r" (count), [colors] "+r" (colors)
-                      : [table] "r" (table), [srcAddr] "r" (srcAddr)
-                      : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
-                      );
+            "subs       %[count], %[count], #8        \n\t"   // decrement count by 8, set flags
+            "blt        2f                            \n\t"   // if count < 0, branch to singles
+            "1:                                       \n\t"   // eights loop
+            "ldmia      %[xx]!, {r5, r7, r9, r11}     \n\t"   // load ptrs to pixels 0-7
+            "uxth       r4, r5                        \n\t"   // extract ptr 0
+            "mov        r5, r5, lsr #16               \n\t"   // extract ptr 1
+            "uxth       r6, r7                        \n\t"   // extract ptr 2
+            "mov        r7, r7, lsr #16               \n\t"   // extract ptr 3
+            "ldrb       r4, [%[srcAddr], r4]          \n\t"   // load pixel 0 from image
+            "uxth       r8, r9                        \n\t"   // extract ptr 4
+            "ldrb       r5, [%[srcAddr], r5]          \n\t"   // load pixel 1 from image
+            "mov        r9, r9, lsr #16               \n\t"   // extract ptr 5
+            "ldrb       r6, [%[srcAddr], r6]          \n\t"   // load pixel 2 from image
+            "uxth       r10, r11                      \n\t"   // extract ptr 6
+            "ldrb       r7, [%[srcAddr], r7]          \n\t"   // load pixel 3 from image
+            "mov        r11, r11, lsr #16             \n\t"   // extract ptr 7
+            "ldrb       r8, [%[srcAddr], r8]          \n\t"   // load pixel 4 from image
+            "ldrb       r9, [%[srcAddr], r9]          \n\t"   // load pixel 5 from image
+            "ldrb       r10, [%[srcAddr], r10]        \n\t"   // load pixel 6 from image
+            "ldrb       r11, [%[srcAddr], r11]        \n\t"   // load pixel 7 from image
+            "ldr        r4, [%[table], r4, lsl #2]    \n\t"   // load pixel 0 SkPMColor from colmap
+            "ldr        r5, [%[table], r5, lsl #2]    \n\t"   // load pixel 1 SkPMColor from colmap
+            "ldr        r6, [%[table], r6, lsl #2]    \n\t"   // load pixel 2 SkPMColor from colmap
+            "ldr        r7, [%[table], r7, lsl #2]    \n\t"   // load pixel 3 SkPMColor from colmap
+            "ldr        r8, [%[table], r8, lsl #2]    \n\t"   // load pixel 4 SkPMColor from colmap
+            "ldr        r9, [%[table], r9, lsl #2]    \n\t"   // load pixel 5 SkPMColor from colmap
+            "ldr        r10, [%[table], r10, lsl #2]  \n\t"   // load pixel 6 SkPMColor from colmap
+            "ldr        r11, [%[table], r11, lsl #2]  \n\t"   // load pixel 7 SkPMColor from colmap
+            "subs       %[count], %[count], #8        \n\t"   // decrement loop counter
+            "stmia      %[colors]!, {r4-r11}          \n\t"   // store 8 pixels
+            "bge        1b                            \n\t"   // loop if counter >= 0
+            "2:                                       \n\t"
+            "adds       %[count], %[count], #8        \n\t"   // fix up counter, set flags
+            "beq        4f                            \n\t"   // if count == 0, branch to exit
+            "3:                                       \n\t"   // singles loop
+            "ldrh       r4, [%[xx]], #2               \n\t"   // load pixel ptr
+            "subs       %[count], %[count], #1        \n\t"   // decrement loop counter
+            "ldrb       r5, [%[srcAddr], r4]          \n\t"   // load pixel from image
+            "ldr        r6, [%[table], r5, lsl #2]    \n\t"   // load SkPMColor from colmap
+            "str        r6, [%[colors]], #4           \n\t"   // store pixel, update ptr
+            "bne        3b                            \n\t"   // loop if counter != 0
+            "4:                                       \n\t"   // exit
+            : [xx] "+r" (xx), [count] "+r" (count), [colors] "+r" (colors)
+            : [table] "r" (table), [srcAddr] "r" (srcAddr)
+            : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
+        );
     }
 
     s.fBitmap->getColorTable()->unlockColors();
@@ -222,201 +222,13 @@ void SkBitmapProcState::platformProcs() {
     }
 }
 
-/////////////////////////////////////
-
-/* FUNCTIONS BELOW ARE SCALAR STUBS INTENDED FOR ARM DEVELOPERS TO REPLACE */
-
-/////////////////////////////////////
-
-
-static inline unsigned char ClampTo8(int a) {
-    if (static_cast<unsigned>(a) < 256) {
-        return a;  // Avoid the extra check in the common case.
-    }
-    if (a < 0) {
-        return 0;
-    }
-    return 255;
-}
-
-// Convolves horizontally along a single row. The row data is given in
-// |srcData| and continues for the numValues() of the filter.
-void convolveHorizontally_arm(const unsigned char* srcData,
-                              const SkConvolutionFilter1D& filter,
-                              unsigned char* outRow,
-                              bool hasAlpha) {
-    // Loop over each pixel on this row in the output image.
-    int numValues = filter.numValues();
-    for (int outX = 0; outX < numValues; outX++) {
-        // Get the filter that determines the current output pixel.
-        int filterOffset, filterLength;
-        const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
-            filter.FilterForValue(outX, &filterOffset, &filterLength);
-
-        // Compute the first pixel in this row that the filter affects. It will
-        // touch |filterLength| pixels (4 bytes each) after this.
-        const unsigned char* rowToFilter = &srcData[filterOffset * 4];
-
-        // Apply the filter to the row to get the destination pixel in |accum|.
-        int accum[4] = {0};
-        for (int filterX = 0; filterX < filterLength; filterX++) {
-            SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterX];
-            accum[0] += curFilter * rowToFilter[filterX * 4 + 0];
-            accum[1] += curFilter * rowToFilter[filterX * 4 + 1];
-            accum[2] += curFilter * rowToFilter[filterX * 4 + 2];
-            if (hasAlpha) {
-                accum[3] += curFilter * rowToFilter[filterX * 4 + 3];
-            }
-        }
-
-        // Bring this value back in range. All of the filter scaling factors
-        // are in fixed point with kShiftBits bits of fractional part.
-        accum[0] >>= SkConvolutionFilter1D::kShiftBits;
-        accum[1] >>= SkConvolutionFilter1D::kShiftBits;
-        accum[2] >>= SkConvolutionFilter1D::kShiftBits;
-        if (hasAlpha) {
-            accum[3] >>= SkConvolutionFilter1D::kShiftBits;
-        }
-
-        // Store the new pixel.
-        outRow[outX * 4 + 0] = ClampTo8(accum[0]);
-        outRow[outX * 4 + 1] = ClampTo8(accum[1]);
-        outRow[outX * 4 + 2] = ClampTo8(accum[2]);
-        if (hasAlpha) {
-            outRow[outX * 4 + 3] = ClampTo8(accum[3]);
-        }
-    }
-}
-
-// Does vertical convolution to produce one output row. The filter values and
-// length are given in the first two parameters. These are applied to each
-// of the rows pointed to in the |sourceDataRows| array, with each row
-// being |pixelWidth| wide.
-//
-// The output must have room for |pixelWidth * 4| bytes.
-template<bool hasAlpha>
-    void convolveVertically_arm(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
-                            int filterLength,
-                            unsigned char* const* sourceDataRows,
-                            int pixelWidth,
-                            unsigned char* outRow) {
-        // We go through each column in the output and do a vertical convolution,
-        // generating one output pixel each time.
-        for (int outX = 0; outX < pixelWidth; outX++) {
-            // Compute the number of bytes over in each row that the current column
-            // we're convolving starts at. The pixel will cover the next 4 bytes.
-            int byteOffset = outX * 4;
-
-            // Apply the filter to one column of pixels.
-            int accum[4] = {0};
-            for (int filterY = 0; filterY < filterLength; filterY++) {
-                SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterY];
-                accum[0] += curFilter * sourceDataRows[filterY][byteOffset + 0];
-                accum[1] += curFilter * sourceDataRows[filterY][byteOffset + 1];
-                accum[2] += curFilter * sourceDataRows[filterY][byteOffset + 2];
-                if (hasAlpha) {
-                    accum[3] += curFilter * sourceDataRows[filterY][byteOffset + 3];
-                }
-            }
-
-            // Bring this value back in range. All of the filter scaling factors
-            // are in fixed point with kShiftBits bits of precision.
-            accum[0] >>= SkConvolutionFilter1D::kShiftBits;
-            accum[1] >>= SkConvolutionFilter1D::kShiftBits;
-            accum[2] >>= SkConvolutionFilter1D::kShiftBits;
-            if (hasAlpha) {
-                accum[3] >>= SkConvolutionFilter1D::kShiftBits;
-            }
-
-            // Store the new pixel.
-            outRow[byteOffset + 0] = ClampTo8(accum[0]);
-            outRow[byteOffset + 1] = ClampTo8(accum[1]);
-            outRow[byteOffset + 2] = ClampTo8(accum[2]);
-            if (hasAlpha) {
-                unsigned char alpha = ClampTo8(accum[3]);
-
-                // Make sure the alpha channel doesn't come out smaller than any of the
-                // color channels. We use premultipled alpha channels, so this should
-                // never happen, but rounding errors will cause this from time to time.
-                // These "impossible" colors will cause overflows (and hence random pixel
-                // values) when the resulting bitmap is drawn to the screen.
-                //
-                // We only need to do this when generating the final output row (here).
-                int maxColorChannel = SkTMax(outRow[byteOffset + 0],
-                                               SkTMax(outRow[byteOffset + 1],
-                                                      outRow[byteOffset + 2]));
-                if (alpha < maxColorChannel) {
-                    outRow[byteOffset + 3] = maxColorChannel;
-                } else {
-                    outRow[byteOffset + 3] = alpha;
-                }
-            } else {
-                // No alpha channel, the image is opaque.
-                outRow[byteOffset + 3] = 0xff;
-            }
-        }
-    }
-
-void convolveVertically_arm(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
-                            int filterLength,
-                            unsigned char* const* sourceDataRows,
-                            int pixelWidth,
-                            unsigned char* outRow,
-                            bool sourceHasAlpha) {
-    if (sourceHasAlpha) {
-        convolveVertically_arm<true>(filterValues, filterLength,
-                                     sourceDataRows, pixelWidth,
-                                     outRow);
-    } else {
-        convolveVertically_arm<false>(filterValues, filterLength,
-                                      sourceDataRows, pixelWidth,
-                                      outRow);
-    }
-}
-
-// Convolves horizontally along four rows. The row data is given in
-// |src_data| and continues for the num_values() of the filter.
-// The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
-// refer to that function for detailed comments.
-void convolve4RowsHorizontally_arm(const unsigned char* src_data[4],
-                                   const SkConvolutionFilter1D& filter,
-                                   unsigned char* out_row[4]) {
-}
-
-///////////////////////////
-
-/* STOP REWRITING FUNCTIONS HERE, BUT DON'T FORGET TO EDIT THE
-   PLATFORM CONVOLUTION PROCS BELOW */
+///////////////////////////////////////////////////////////////////////////////
 
-///////////////////////////
+extern void platformConvolutionProcs_arm_neon(SkConvolutionProcs* procs);
 
-void applySIMDPadding_arm(SkConvolutionFilter1D *filter) {
-    // Padding |paddingCount| of more dummy coefficients after the coefficients
-    // of last filter to prevent SIMD instructions which load 8 or 16 bytes
-    // together to access invalid memory areas. We are not trying to align the
-    // coefficients right now due to the opaqueness of <vector> implementation.
-    // This has to be done after all |AddFilter| calls.
-    for (int i = 0; i < 8; ++i) {
-        filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0));
-    }
+void platformConvolutionProcs_arm(SkConvolutionProcs* procs) {
 }
 
 void SkBitmapProcState::platformConvolutionProcs(SkConvolutionProcs* procs) {
-    if (sk_cpu_arm_has_neon()) {
-        procs->fExtraHorizontalReads = 3;
-        procs->fConvolveVertically = &convolveVertically_arm;
-
-        // next line is commented out because the four-row convolution function above is
-        // just a no-op.  Please see the comment above its definition, and the SSE implementation
-        // in SkBitmapProcState_opts_SSE2.cpp for guidance on its semantics.
-        // leaving it as NULL will just cause the convolution system to not attempt
-        // to operate on four rows at once, which is correct but not performance-optimal.
-
-        // procs->fConvolve4RowsHorizontally = &convolve4RowsHorizontally_arm;
-
-        procs->fConvolve4RowsHorizontally = NULL;
-
-        procs->fConvolveHorizontally = &convolveHorizontally_arm;
-        procs->fApplySIMDPadding = &applySIMDPadding_arm;
-    }
+    SK_ARM_NEON_WRAP(platformConvolutionProcs_arm)(procs);
 }