Accelerate convolution functions using NEON

Adds Arm NEON versions of convolve_4_rows_horizontally,
convolve_horizontally and convolve_vertically.

This NEON convolution code was originally added here: [1]. Then, at the
introduction of GPU raster, the code (then in SkBitmapFilter_opts.h) was
removed [2].

The re-introduction of this code gives a significant performance boost
to Arm platforms which do not have GPU raster enabled - currently all
Arm Chromebooks and old Android devices.

Similar SIMD-accelerated convolution code exists in src/skia/ext for
Intel and MIPS platforms which do not have GPU raster enabled. This
commit simply adds an Arm NEON counterpart.

[1] https://codereview.chromium.org/27533004
[2] https://skia-review.googlesource.com/c/skia/+/24644

Change-Id: Ied7d65c0f1e2fd3cb54dee3ff70dcccb8eafcb8a
Reviewed-on: https://chromium-review.googlesource.com/911789
Commit-Queue: Mike Klein <mtklein@chromium.org>
Reviewed-by: Mike Klein <mtklein@chromium.org>
Cr-Commit-Position: refs/heads/master@{#539143}

skia/BUILD.gn

@@ -281,6 +281,11 @@ component("skia") {
       "ext/convolver_mips_dspr2.cc",
       "ext/convolver_mips_dspr2.h",
     ]
+  } else if (current_cpu == "arm" || current_cpu == "arm64") {
+    sources += [
+      "ext/convolver_neon.cc",
+      "ext/convolver_neon.h",
+    ]
   }
 
   # The imported Skia gni source paths are made absolute by gn.

skia/ext/convolver.cc

@@ -8,6 +8,7 @@
 #include "skia/ext/convolver.h"
 #include "skia/ext/convolver_SSE2.h"
 #include "skia/ext/convolver_mips_dspr2.h"
+#include "skia/ext/convolver_neon.h"
 #include "third_party/skia/include/core/SkSize.h"
 #include "third_party/skia/include/core/SkTypes.h"
 
@@ -369,6 +370,11 @@ void SetupSIMD(ConvolveProcs *procs) {
   procs->extra_horizontal_reads = 3;
   procs->convolve_vertically = &ConvolveVertically_mips_dspr2;
   procs->convolve_horizontally = &ConvolveHorizontally_mips_dspr2;
+#elif defined SIMD_NEON
+  procs->extra_horizontal_reads = 3;
+  procs->convolve_vertically = &ConvolveVertically_Neon;
+  procs->convolve_4rows_horizontally = &Convolve4RowsHorizontally_Neon;
+  procs->convolve_horizontally = &ConvolveHorizontally_Neon;
 #endif
 }
 

skia/ext/convolver.h

@@ -25,6 +25,12 @@
     defined(__mips_dsp) && (__mips_dsp_rev >= 2)
 #define SIMD_MIPS_DSPR2 1
 #endif
+
+#if defined(ARCH_CPU_ARM_FAMILY) && \
+    (defined(__ARM_NEON__) || defined(__ARM_NEON))
+#define SIMD_NEON 1
+#endif
+
 // avoid confusion with Mac OS X's math library (Carbon)
 #if defined(__APPLE__)
 #undef FloatToFixed

skia/ext/convolver_neon.h

+// Copyright 2018 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.
+
+#ifndef SKIA_EXT_CONVOLVER_NEON_H_
+#define SKIA_EXT_CONVOLVER_NEON_H_
+
+#include "skia/ext/convolver.h"
+
+namespace skia {
+
+void ConvolveHorizontally_Neon(const unsigned char* src_data,
+                               const ConvolutionFilter1D& filter,
+                               unsigned char* out_row,
+                               bool has_alpha);
+
+void Convolve4RowsHorizontally_Neon(const unsigned char* src_data[4],
+                                    const ConvolutionFilter1D& filter,
+                                    unsigned char* out_row[4]);
+
+void ConvolveVertically_Neon(const ConvolutionFilter1D::Fixed* filter_values,
+                             int filter_length,
+                             unsigned char* const* source_data_rows,
+                             int pixel_width,
+                             unsigned char* out_row,
+                             bool has_alpha);
+
+}  // namespace skia
+
+#endif  // SKIA_EXT_CONVOLVER_NEON_H_