SSE4.1 implementation of some alpha-processing functions

DispatchAlpha* functions are hard to speed up, compared to SSE2.
ExtractAlpha sees a ~15% speed-up though.

Change-Id: I8715c2defecbc832f469eed7e6ffd012146b52de

Android.mk

@@ -38,6 +38,7 @@ dsp_dec_srcs := \
     src/dsp/alpha_processing.c \
     src/dsp/alpha_processing_mips_dsp_r2.c \
     src/dsp/alpha_processing_sse2.c \
+    src/dsp/alpha_processing_sse41.c \
     src/dsp/argb.c \
     src/dsp/argb_mips_dsp_r2.c \
     src/dsp/argb_sse2.c \

Makefile.vc

@@ -189,6 +189,7 @@ DSP_DEC_OBJS = \
     $(DIROBJ)\dsp\alpha_processing.obj \
     $(DIROBJ)\dsp\alpha_processing_mips_dsp_r2.obj \
     $(DIROBJ)\dsp\alpha_processing_sse2.obj \
+    $(DIROBJ)\dsp\alpha_processing_sse41.obj \
     $(DIROBJ)\dsp\cpu.obj \
     $(DIROBJ)\dsp\dec.obj \
     $(DIROBJ)\dsp\dec_clip_tables.obj \

makefile.unix

@@ -129,6 +129,7 @@ DSP_DEC_OBJS = \
     src/dsp/alpha_processing.o \
     src/dsp/alpha_processing_mips_dsp_r2.o \
     src/dsp/alpha_processing_sse2.o \
+    src/dsp/alpha_processing_sse41.o \
     src/dsp/cpu.o \
     src/dsp/dec.o \
     src/dsp/dec_clip_tables.o \

src/dsp/Makefile.am

@@ -59,6 +59,7 @@ libwebpdsp_avx2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
 libwebpdsp_avx2_la_CFLAGS = $(AM_CFLAGS) $(AVX2_FLAGS)
 
 libwebpdspdecode_sse41_la_SOURCES =
+libwebpdspdecode_sse41_la_SOURCES += alpha_processing_sse41.c
 libwebpdspdecode_sse41_la_SOURCES += dec_sse41.c
 libwebpdspdecode_sse41_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
 libwebpdspdecode_sse41_la_CFLAGS = $(AM_CFLAGS) $(SSE41_FLAGS)

src/dsp/alpha_processing.c

@@ -345,6 +345,7 @@ int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
 
 extern void WebPInitAlphaProcessingMIPSdspR2(void);
 extern void WebPInitAlphaProcessingSSE2(void);
+extern void WebPInitAlphaProcessingSSE41(void);
 
 static volatile VP8CPUInfo alpha_processing_last_cpuinfo_used =
     (VP8CPUInfo)&alpha_processing_last_cpuinfo_used;
@@ -365,6 +366,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
       WebPInitAlphaProcessingSSE2();
+#if defined(WEBP_USE_SSE41)
+      if (VP8GetCPUInfo(kSSE4_1)) {
+        WebPInitAlphaProcessingSSE41();
+      }
+#endif
     }
 #endif
 #if defined(WEBP_USE_MIPS_DSP_R2)

src/dsp/alpha_processing_sse41.c

@@ -0,0 +1,92 @@
+// Copyright 2015 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, SSE4.1 variant.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_SSE41)
+
+#include <smmintrin.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 all_0xff = _mm_set1_epi32(~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) & ~15;
+  const __m128i kCstAlpha0 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1,
+                                          -1, -1, -1, -1, 12, 8, 4, 0);
+  const __m128i kCstAlpha1 = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1,
+                                          12, 8, 4, 0, -1, -1, -1, -1);
+  const __m128i kCstAlpha2 = _mm_set_epi8(-1, -1, -1, -1, 12, 8, 4, 0,
+                                          -1, -1, -1, -1, -1, -1, -1, -1);
+  const __m128i kCstAlpha3 = _mm_set_epi8(12, 8, 4, 0, -1, -1, -1, -1,
+                                          -1, -1, -1, -1, -1, -1, -1, -1);
+  for (j = 0; j < height; ++j) {
+    const __m128i* src = (const __m128i*)argb;
+    for (i = 0; i < limit; i += 16) {
+      // load 64 argb bytes
+      const __m128i a0 = _mm_loadu_si128(src + 0);
+      const __m128i a1 = _mm_loadu_si128(src + 1);
+      const __m128i a2 = _mm_loadu_si128(src + 2);
+      const __m128i a3 = _mm_loadu_si128(src + 3);
+      const __m128i b0 = _mm_shuffle_epi8(a0, kCstAlpha0);
+      const __m128i b1 = _mm_shuffle_epi8(a1, kCstAlpha1);
+      const __m128i b2 = _mm_shuffle_epi8(a2, kCstAlpha2);
+      const __m128i b3 = _mm_shuffle_epi8(a3, kCstAlpha3);
+      const __m128i c0 = _mm_or_si128(b0, b1);
+      const __m128i c1 = _mm_or_si128(b2, b3);
+      const __m128i d0 = _mm_or_si128(c0, c1);
+      // store
+      _mm_storeu_si128((__m128i*)&alpha[i], d0);
+      // accumulate sixteen alpha 'and' in parallel
+      all_alphas = _mm_and_si128(all_alphas, d0);
+      src += 4;
+    }
+    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 sixteen alpha 'and' into an 8-bit mask.
+  alpha_and |= 0xff00u;  // pretend the upper bits [8..15] were tested ok.
+  alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff));
+  return (alpha_and == 0xffffu);
+}
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void WebPInitAlphaProcessingSSE41(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE41(void) {
+  WebPExtractAlpha = ExtractAlpha;
+}
+
+#else  // !WEBP_USE_SSE41
+
+WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingSSE41)
+
+#endif  // WEBP_USE_SSE41