NEON-ize RGBA to RGB code

third_party/WebKit/Source/platform/image-encoders/JPEGImageEncoder.cpp

 /*
  * Copyright (c) 2010, Google Inc. All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are
  * met:
  *
  *     * Redistributions of source code must retain the above copyright
  * notice, this list of conditions and the following disclaimer.
  *     * Redistributions in binary form must reproduce the above
@@ skipping 15 matching lines @@
  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
 #include "platform/image-encoders/JPEGImageEncoder.h"
 
 #include "SkColorPriv.h"
 #include "platform/geometry/IntSize.h"
 #include "platform/graphics/ImageBuffer.h"
+#include "platform/image-encoders/RGBAtoRGB.h"
 #include "wtf/CurrentTime.h"
 #include "wtf/PtrUtil.h"
 #include <memory>
 
 extern "C" {
 #include <setjmp.h>
 #include <stdio.h>  // jpeglib.h needs stdio.h FILE
 #include "jpeglib.h"
 }
 
 namespace blink {
 
+void RGBAtoRGBScalar(const unsigned char* pixels,
+                     unsigned pixelCount,
+                     unsigned char* output) {
+  // Per <canvas> spec, composite the input image pixels source-over on black.
+  for (; pixelCount-- > 0; pixels += 4) {
+    unsigned char alpha = pixels[3];
+    if (alpha != 255) {
+      *output++ = SkMulDiv255Round(pixels[0], alpha);
+      *output++ = SkMulDiv255Round(pixels[1], alpha);
+      *output++ = SkMulDiv255Round(pixels[2], alpha);
+    } else {
+      *output++ = pixels[0];
+      *output++ = pixels[1];
+      *output++ = pixels[2];
+    }
+  }
+}
+
+// TODO(cavalcantii): use regular macro, see https://crbug.com/673067.
+#ifdef __ARM_NEON__
+
+// TODO(cavalcantii): this came straight from SkSwizzler_opts.h, I wonder
+// how to include it?

[msarett, 2016/12/18 18:48:37]

Sorry for being unclear, there's no plan to make the Skia code public.  I was
suggesting looking at it just to try out the implementation, not to make you
worry about code sharing.

You can remove this TODO :).

+inline uint8x8_t div255Round(uint16x8_t x) {
+  return vraddhn_u16(x, vrshrq_n_u16(x, 8));
+}
+
+inline uint8x8_t scale(uint8x8_t x, uint8x8_t y) {
+  return div255Round(vmull_u8(x, y));
+}
+
+void RGBAtoRGBNeon(const unsigned char* input,
+                   const unsigned pixelCount,
+                   unsigned char* output) {
+  const unsigned pixelsPerLoad = 16;
+  const unsigned rgbaStep = pixelsPerLoad * 4, rgbStep = pixelsPerLoad * 3;
+  // Input registers.
+  uint8x16x4_t rgba;
+  // Output registers.
+  uint8x16x3_t rgb;
+  // Intermediate registers.
+  uint8x8_t low, high;
+  uint8x16_t result;
+  unsigned counter;
+
+  auto transformColor = [&](size_t channel) {
+    // Extracts the low/high part of the 128 bits and scale using Skia code.
+    low = vget_low_u8(rgba.val[channel]);
+    high = vget_high_u8(rgba.val[channel]);
+    low = scale(low, vget_low_u8(rgba.val[3]));
+    high = scale(high, vget_high_u8(rgba.val[3]));
+    result = vcombine_u8(low, high);
+    // Write back the channel to a 128 bits register.
+    rgb.val[channel] = result;
+  };
+
+  for (counter = 0; counter + pixelsPerLoad <= pixelCount;
+       counter += pixelsPerLoad) {
+    // Reads 16 pixels at once, each color channel in a different
+    // 128 bits register.
+    rgba = vld4q_u8(input);
+
+    transformColor(0);
+    transformColor(1);
+    transformColor(2);
+
+    // Write back (interleaved) results to output.
+    vst3q_u8(output, rgb);
+
+    // Advance to next elements (could be avoided loading register with
+    // increment after i.e. "vld4 {vector}, [r1]!").
+    input += rgbaStep;
+    output += rgbStep;
+  }
+
+  // Handle the tail elements.
+  unsigned remaining = pixelCount;
+  remaining -= counter;
+  if (remaining != 0) {
+    RGBAtoRGBScalar(input, remaining, output);
+  }
+}
+#endif
+
 struct JPEGOutputBuffer : public jpeg_destination_mgr {
   DISALLOW_NEW();
   Vector<unsigned char>* output;
   Vector<unsigned char> buffer;
 };
 
 class JPEGImageEncoderStateImpl final : public JPEGImageEncoderState {
  public:
   JPEGImageEncoderStateImpl() {}
   ~JPEGImageEncoderStateImpl() override {
@@ skipping 30 matching lines @@
   JPEGOutputBuffer* out = static_cast<JPEGOutputBuffer*>(cinfo->dest);
   const size_t size = out->buffer.size() - out->free_in_buffer;
   out->output->append(out->buffer.data(), size);
 }
 
 static void handleError(j_common_ptr common) {
   jmp_buf* jumpBufferPtr = static_cast<jmp_buf*>(common->client_data);
   longjmp(*jumpBufferPtr, -1);
 }
 
-static void RGBAtoRGB(const unsigned char* pixels,
-                      unsigned pixelCount,
-                      unsigned char* output) {
-  // Per <canvas> spec, composite the input image pixels source-over on black.
-
-  for (; pixelCount-- > 0; pixels += 4) {
-    unsigned char alpha = pixels[3];
-    if (alpha != 255) {
-      *output++ = SkMulDiv255Round(pixels[0], alpha);
-      *output++ = SkMulDiv255Round(pixels[1], alpha);
-      *output++ = SkMulDiv255Round(pixels[2], alpha);
-    } else {
-      *output++ = pixels[0];
-      *output++ = pixels[1];
-      *output++ = pixels[2];
-    }
-  }
-}
-
 static void disableSubsamplingForHighQuality(jpeg_compress_struct* cinfo,
                                              int quality) {
   if (quality < 100)
     return;
 
   for (int i = 0; i < MAX_COMPONENTS; ++i) {
     cinfo->comp_info[i].h_samp_factor = 1;
     cinfo->comp_info[i].v_samp_factor = 1;
   }
 }
@@ skipping 123 matching lines @@
   std::unique_ptr<JPEGImageEncoderState> encoderState =
       JPEGImageEncoderState::create(imageData.size(), quality, output);
   if (!encoderState)
     return false;
 
   return JPEGImageEncoder::encodeWithPreInitializedState(
       std::move(encoderState), imageData.pixels());
 }
 
 }  // namespace blink