Add a WebPExtractAlpha function to dsp

This is the opposite of WebPDispatchAlpha

+ Implement the SSE2 version

(cherry picked from commit cddd334050)

Conflicts:
	src/dsp/alpha_processing_sse2.c

Change-Id: I0c297309255f508c5261da8aad01f7e57f924d6c

Android.mk

@@ -33,6 +33,7 @@ LOCAL_SRC_FILES := \
     src/dec/vp8l.c \
     src/dec/webp.c \
     src/dsp/alpha_processing.c \
+    src/dsp/alpha_processing_sse2.c \
     src/dsp/cpu.c \
     src/dsp/dec.c \
     src/dsp/dec_clip_tables.c \

Makefile.vc

@@ -172,6 +172,7 @@ DEMUX_OBJS = \
 
 DSP_DEC_OBJS = \
     $(DIROBJ)\dsp\alpha_processing.obj \
+    $(DIROBJ)\dsp\alpha_processing_sse2.obj \
     $(DIROBJ)\dsp\cpu.obj \
     $(DIROBJ)\dsp\dec.obj \
     $(DIROBJ)\dsp\dec_clip_tables.obj \

makefile.unix

@@ -108,6 +108,7 @@ DEMUX_OBJS = \
 
 DSP_DEC_OBJS = \
     src/dsp/alpha_processing.o \
+    src/dsp/alpha_processing_sse2.o \
     src/dsp/cpu.o \
     src/dsp/dec.o \
     src/dsp/dec_clip_tables.o \

src/dsp/Makefile.am

@@ -38,6 +38,7 @@ libwebpdsp_avx2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
 libwebpdsp_avx2_la_CFLAGS = $(AM_CFLAGS) $(AVX2_FLAGS)
 
 libwebpdspdecode_sse2_la_SOURCES =
+libwebpdspdecode_sse2_la_SOURCES += alpha_processing_sse2.c
 libwebpdspdecode_sse2_la_SOURCES += dec_sse2.c
 libwebpdspdecode_sse2_la_SOURCES += lossless_sse2.c
 libwebpdspdecode_sse2_la_SOURCES += upsampling_sse2.c

src/dsp/alpha_processing.c

@@ -284,15 +284,46 @@ static void ApplyAlphaMultiply_16b(uint8_t* rgba4444,
 #endif
 }
 
+static int ExtractAlpha(const uint8_t* argb, int argb_stride,
+                        int width, int height,
+                        uint8_t* alpha, int alpha_stride) {
+  uint8_t alpha_mask = 0xff;
+  int i, j;
+
+  for (j = 0; j < height; ++j) {
+    for (i = 0; i < width; ++i) {
+      const uint8_t alpha_value = argb[4 * i];
+      alpha[i] = alpha_value;
+      alpha_mask &= alpha_value;
+    }
+    argb += argb_stride;
+    alpha += alpha_stride;
+  }
+  return (alpha_mask == 0xff);
+}
+
 void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int);
 void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int);
+int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
 
 //------------------------------------------------------------------------------
 // Init function
 
+extern void WebPInitAlphaProcessingSSE2(void);
+
 void WebPInitAlphaProcessing(void) {
   WebPMultARGBRow = MultARGBRow;
   WebPMultRow = MultRow;
   WebPApplyAlphaMultiply = ApplyAlphaMultiply;
   WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply_16b;
+  WebPExtractAlpha = ExtractAlpha;
+
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitAlphaProcessingSSE2();
+    }
+#endif
+  }
 }

src/dsp/alpha_processing_sse2.c

@@ -0,0 +1,77 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Utilities for processing transparent channel.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_SSE2)
+#include <emmintrin.h>
+
+//------------------------------------------------------------------------------
+
+static int ExtractAlpha(const uint8_t* argb, int argb_stride,
+                        int width, int height,
+                        uint8_t* alpha, int alpha_stride) {
+  // alpha_and stores an 'and' operation of all the alpha[] values. The final
+  // value is not 0xff if any of the alpha[] is not equal to 0xff.
+  uint32_t alpha_and = 0xff;
+  int i, j;
+  const __m128i a_mask = _mm_set1_epi32(0xffu);  // to preserve alpha
+  const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u);
+  __m128i all_alphas = all_0xff;
+
+  // We must be able to access 3 extra bytes after the last written byte
+  // 'src[4 * width - 4]', because we don't know if alpha is the first or the
+  // last byte of the quadruplet.
+  const int limit = (width - 1) & ~7;
+
+  for (j = 0; j < height; ++j) {
+    const __m128i* src = (const __m128i*)argb;
+    for (i = 0; i < limit; i += 8) {
+      // load 32 argb bytes
+      const __m128i a0 = _mm_loadu_si128(src + 0);
+      const __m128i a1 = _mm_loadu_si128(src + 1);
+      const __m128i b0 = _mm_and_si128(a0, a_mask);
+      const __m128i b1 = _mm_and_si128(a1, a_mask);
+      const __m128i c0 = _mm_packs_epi32(b0, b1);
+      const __m128i d0 = _mm_packus_epi16(c0, c0);
+      // store
+      _mm_storel_epi64((__m128i*)&alpha[i], d0);
+      // accumulate eight alpha 'and' in parallel
+      all_alphas = _mm_and_si128(all_alphas, d0);
+      src += 2;
+    }
+    for (; i < width; ++i) {
+      const uint32_t alpha_value = argb[4 * i];
+      alpha[i] = alpha_value;
+      alpha_and &= alpha_value;
+    }
+    argb += argb_stride;
+    alpha += alpha_stride;
+  }
+  // Combine the eight alpha 'and' into a 8-bit mask.
+  alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff));
+  return (alpha_and == 0xff);
+}
+
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// Init function
+
+extern void WebPInitAlphaProcessingSSE2(void);
+
+void WebPInitAlphaProcessingSSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  WebPExtractAlpha = ExtractAlpha;
+#endif
+}

src/dsp/dsp.h

@@ -257,6 +257,13 @@ extern void (*WebPApplyAlphaMultiply)(
 extern void (*WebPApplyAlphaMultiply4444)(
     uint8_t* rgba4444, int w, int h, int stride);
 
+// Extract the alpha values from 32b values in argb[] and pack them into alpha[]
+// (this is the opposite of WebPDispatchAlpha).
+// Returns true if there's only trivial 0xff alpha values.
+extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride,
+                               int width, int height,
+                               uint8_t* alpha, int alpha_stride);
+
 // Pre-Multiply operation transforms x into x * A / 255  (where x=Y,R,G or B).
 // Un-Multiply operation transforms x into x * 255 / A.