Add AVX2 version of ConvolveVertically

src/core/SkConvolver.cpp

 // Copyright (c) 2011 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "SkConvolver.h"
 #include "SkOpts.h"
 #include "SkTArray.h"
 
 namespace {
     // Stores a list of rows in a circular buffer. The usage is you write into it
@@ skipping 174 matching lines @@
         filterY.FilterForValue(0, &filterOffset, &filterLength);
     int nextXRow = filterOffset;
 
     // We loop over each row in the input doing a horizontal convolution. This
     // will result in a horizontally convolved image. We write the results into
     // a circular buffer of convolved rows and do vertical convolution as rows
     // are available. This prevents us from having to store the entire
     // intermediate image and helps cache coherency.
     // We will need four extra rows to allow horizontal convolution could be done
     // simultaneously. We also pad each row in row buffer to be aligned-up to
-    // 16 bytes.
+    // 32 bytes.
     // TODO(jiesun): We do not use aligned load from row buffer in vertical
     // convolution pass yet. Somehow Windows does not like it.
-    int rowBufferWidth = (filterX.numValues() + 15) & ~0xF;
+    int rowBufferWidth = (filterX.numValues() + 31) & ~0x1F;
     int rowBufferHeight = maxYFilterSize +
                           (SkOpts::convolve_4_rows_horizontally != nullptr ? 4 : 0);
 
     // check for too-big allocation requests : crbug.com/528628
     {
         int64_t size = sk_64_mul(rowBufferWidth, rowBufferHeight);
         // need some limit, to avoid over-committing success from malloc, but then
         // crashing when we try to actually use the memory.
         // 100meg seems big enough to allow "normal" zoom factors and image sizes through
         // while avoiding the crash seen by the bug (crbug.com/528628)
@@ skipping 54 matching lines @@
         unsigned char* const* firstRowForFilter =
             &rowsToConvolve[filterOffset - firstRowInCircularBuffer];
 
         SkOpts::convolve_vertically(filterValues, filterLength,
                                     firstRowForFilter,
                                     filterX.numValues(), curOutputRow,
                                     sourceHasAlpha);
     }
     return true;
 }