Make SSE2/Neon convolution functions not to read extra bytes

src/opts/SkBitmapProcState_arm_neon.cpp

 /*
  * Copyright 2012 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkBitmapProcState.h"
 #include "SkBitmapProcState_filter.h"
 #include "SkColorPriv.h"
@@ skipping 124 matching lines @@
             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;
+#define ACCUM_REMAINDER(src, accum) {                                                      \
+            int remainder[4] = {0};                                                        \
+            const unsigned char* pixels_left = src + (filterOffset + filterLength - r) * 4;\
+            for (int i = 0; i < r; i++) {                                                  \
+                SkConvolutionFilter1D::ConvolutionFixed coeff = filterValues[i];           \
+                remainder[0] += coeff * pixels_left[i * 4 + 0];                            \
+                remainder[1] += coeff * pixels_left[i * 4 + 1];                            \
+                remainder[2] += coeff * pixels_left[i * 4 + 2];                            \
+                remainder[3] += coeff * pixels_left[i * 4 + 3];                            \
+            }                                                                              \
+            int32x4_t t{remainder[0], remainder[1], remainder[2], remainder[3]};           \
+            accum += t; }
+            ACCUM_REMAINDER(srcData, accum);
         }
 
         // 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);
@@ skipping 187 matching lines @@
                                     unsigned char* outRow[4],
                                     size_t outRowBytes) {
 
     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);
@@ skipping 34 matching lines @@
             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);
+            ACCUM_REMAINDER(srcData[0], accum0);
+            ACCUM_REMAINDER(srcData[1], accum1);
+            ACCUM_REMAINDER(srcData[2], accum2);
+            ACCUM_REMAINDER(srcData[3], 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;
 }