Bug 1551088 - Part 3. Add SSSE3 and AVX2 implementations of unpacking. r=lsalzman

These variants perform significantly faster than the C implementations
according to local testing and that in treeherder. Image decoding is as
much as 40% faster in the most simple cases (solid green PNG image).

Differential Revision: https://phabricator.services.mozilla.com/D46446

--HG--
extra : moz-landing-system : lando

gfx/2d/Swizzle.cpp

@@ -151,6 +151,20 @@ void SwizzleRow_SSE2(const uint8_t*, uint8_t*, int32_t);
         SwizzleRow_SSE2<ShouldSwapRB(aSrcFormat, aDstFormat), \
                         ShouldForceOpaque(aSrcFormat, aDstFormat)>)
 
+template <bool aSwapRB>
+void UnpackRowRGB24_SSSE3(const uint8_t*, uint8_t*, int32_t);
+
+#define UNPACK_ROW_RGB_SSSE3(aDstFormat)             \
+  FORMAT_CASE_ROW(SurfaceFormat::R8G8B8, aDstFormat, \
+                  UnpackRowRGB24_SSSE3<ShouldSwapRB(SurfaceFormat::R8G8B8, aDstFormat)>)
+
+template <bool aSwapRB>
+void UnpackRowRGB24_AVX2(const uint8_t*, uint8_t*, int32_t);
+
+#define UNPACK_ROW_RGB_AVX2(aDstFormat)              \
+  FORMAT_CASE_ROW(SurfaceFormat::R8G8B8, aDstFormat, \
+                  UnpackRowRGB24_AVX2<ShouldSwapRB(SurfaceFormat::R8G8B8, aDstFormat)>)
+
 #endif
 
 #ifdef USE_NEON
@@ -985,6 +999,24 @@ bool SwizzleData(const uint8_t* aSrc, int32_t aSrcStride,
 
 SwizzleRowFn SwizzleRow(SurfaceFormat aSrcFormat, SurfaceFormat aDstFormat) {
 #ifdef USE_SSE2
+  if (mozilla::supports_avx2()) switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
+      UNPACK_ROW_RGB_AVX2(SurfaceFormat::R8G8B8X8)
+      UNPACK_ROW_RGB_AVX2(SurfaceFormat::R8G8B8A8)
+      UNPACK_ROW_RGB_AVX2(SurfaceFormat::B8G8R8X8)
+      UNPACK_ROW_RGB_AVX2(SurfaceFormat::B8G8R8A8)
+      default:
+        break;
+    }
+
+  if (mozilla::supports_ssse3()) switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
+      UNPACK_ROW_RGB_SSSE3(SurfaceFormat::R8G8B8X8)
+      UNPACK_ROW_RGB_SSSE3(SurfaceFormat::R8G8B8A8)
+      UNPACK_ROW_RGB_SSSE3(SurfaceFormat::B8G8R8X8)
+      UNPACK_ROW_RGB_SSSE3(SurfaceFormat::B8G8R8A8)
+      default:
+        break;
+    }
+
   if (mozilla::supports_sse2()) switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
       SWIZZLE_ROW_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
       SWIZZLE_ROW_SSE2(SurfaceFormat::B8G8R8X8, SurfaceFormat::R8G8B8X8)

gfx/2d/SwizzleAVX2.cpp

@@ -0,0 +1,85 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "Swizzle.h"
+
+#include <immintrin.h>
+#include <tmmintrin.h>
+
+namespace mozilla {
+namespace gfx {
+
+template <bool aSwapRB>
+void UnpackRowRGB24_SSSE3(const uint8_t* aSrc, uint8_t* aDst, int32_t aLength);
+
+template <bool aSwapRB>
+void UnpackRowRGB24_AVX2(const uint8_t* aSrc, uint8_t* aDst, int32_t aLength) {
+  // Because this implementation will read an additional 8 bytes of data that
+  // is ignored and masked over, we cannot use the accelerated version for the
+  // last 1-10 pixels (3-30 bytes remaining) to guarantee we don't access memory
+  // outside the buffer (we read in 32 byte chunks).
+  if (aLength < 11) {
+    UnpackRowRGB24_SSSE3<aSwapRB>(aSrc, aDst, aLength);
+    return;
+  }
+
+  // Because we are expanding, we can only process the data back to front in
+  // case we are performing this in place.
+  int32_t alignedRow = (aLength - 4) & ~7;
+  int32_t remainder = aLength - alignedRow;
+
+  const uint8_t* src = aSrc + alignedRow * 3;
+  uint8_t* dst = aDst + alignedRow * 4;
+
+  // Handle any 3-10 remaining pixels.
+  UnpackRowRGB24_SSSE3<aSwapRB>(src, dst, remainder);
+
+  // Used to shuffle the two final 32-bit words which we ignore into the last
+  // 32-bit word of each 128-bit lane, such that
+  //   RGBR GBRG BRGB RGBR GBRG BRGB RGBR GBRG
+  //   BRGB RGBR GBRG BRGB ZZZZ ZZZZ ZZZZ ZZZZ
+  // becomes
+  //   RGBR GBRG BRGB RGBR GBRG BRGB ZZZZ ZZZZ
+  //   RGBR GBRG BRGB RGBR GBRG BRGB ZZZZ ZZZZ
+  const __m256i discardMask = _mm256_set_epi32(7, 5, 4, 3, 6, 2, 1, 0);
+
+  // Used to shuffle 8-bit words within a 128-bit lane, such that we transform
+  //   RGBR GBRG BRGB RGBR GBRG BRGB ZZZZ ZZZZ
+  // into
+  //   RGBZ RGBZ RGBZ RGBZ RGBZ RGBZ RGBZ RGBZ
+  // or
+  //   BGRZ BGRZ BGRZ BGRZ BGRZ BGRZ BGRZ BGRZ
+  const __m256i colorMask =
+      aSwapRB ? _mm256_set_epi8(15, 9, 10, 11, 14, 6, 7, 8, 13, 3, 4, 5, 12, 0,
+                                1, 2, 15, 9, 10, 11, 14, 6, 7, 8, 13, 3, 4, 5,
+                                12, 0, 1, 2)
+              : _mm256_set_epi8(15, 11, 10, 9, 14, 8, 7, 6, 13, 5, 4, 3, 12, 2,
+                                1, 0, 15, 11, 10, 9, 14, 8, 7, 6, 13, 5, 4, 3,
+                                12, 2, 1, 0);
+
+  // Used to transform RGBZ/BGRZ to RGBX/BGRX, or force the alpha opaque.
+  const __m256i alphaMask = _mm256_set1_epi32(0xFF000000);
+
+  // Process all 8-pixel chunks as one vector.
+  src -= 8 * 3;
+  dst -= 8 * 4;
+  while (src >= aSrc) {
+    __m256i px = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(src));
+    px = _mm256_permutevar8x32_epi32(px, discardMask);
+    px = _mm256_shuffle_epi8(px, colorMask);
+    px = _mm256_or_si256(px, alphaMask);
+    _mm256_storeu_si256(reinterpret_cast<__m256i*>(dst), px);
+    src -= 8 * 3;
+    dst -= 8 * 4;
+  }
+}
+
+// Force instantiation of swizzle variants here.
+template void UnpackRowRGB24_AVX2<false>(const uint8_t*, uint8_t*, int32_t);
+template void UnpackRowRGB24_AVX2<true>(const uint8_t*, uint8_t*, int32_t);
+
+}  // namespace gfx
+}  // namespace mozilla

gfx/2d/SwizzleSSSE3.cpp

@@ -0,0 +1,67 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "Swizzle.h"
+
+#include <emmintrin.h>
+#include <tmmintrin.h>
+
+namespace mozilla {
+namespace gfx {
+
+template <bool aSwapRB>
+void UnpackRowRGB24(const uint8_t* aSrc, uint8_t* aDst, int32_t aLength);
+
+template <bool aSwapRB>
+void UnpackRowRGB24_SSSE3(const uint8_t* aSrc, uint8_t* aDst, int32_t aLength) {
+  // Because this implementation will read an additional 4 bytes of data that
+  // is ignored and masked over, we cannot use the accelerated version for the
+  // last 1-5 pixels (3-15 bytes remaining) to guarantee we don't access memory
+  // outside the buffer (we read in 16 byte chunks).
+  if (aLength < 6) {
+    UnpackRowRGB24<aSwapRB>(aSrc, aDst, aLength);
+    return;
+  }
+
+  // Because we are expanding, we can only process the data back to front in
+  // case we are performing this in place.
+  int32_t alignedRow = (aLength - 2) & ~3;
+  int32_t remainder = aLength - alignedRow;
+
+  const uint8_t* src = aSrc + alignedRow * 3;
+  uint8_t* dst = aDst + alignedRow * 4;
+
+  // Handle 2-5 remaining pixels.
+  UnpackRowRGB24<aSwapRB>(src, dst, remainder);
+
+  __m128i mask;
+  if (aSwapRB) {
+    mask = _mm_set_epi8(15, 9, 10, 11, 14, 6, 7, 8, 13, 3, 4, 5, 12, 0, 1, 2);
+  } else {
+    mask = _mm_set_epi8(15, 11, 10, 9, 14, 8, 7, 6, 13, 5, 4, 3, 12, 2, 1, 0);
+  }
+
+  __m128i alpha = _mm_set1_epi32(0xFF000000);
+
+  // Process all 4-pixel chunks as one vector.
+  src -= 4 * 3;
+  dst -= 4 * 4;
+  while (src >= aSrc) {
+    __m128i px = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src));
+    px = _mm_shuffle_epi8(px, mask);
+    px = _mm_or_si128(px, alpha);
+    _mm_storeu_si128(reinterpret_cast<__m128i*>(dst), px);
+    src -= 4 * 3;
+    dst -= 4 * 4;
+  }
+}
+
+// Force instantiation of swizzle variants here.
+template void UnpackRowRGB24_SSSE3<false>(const uint8_t*, uint8_t*, int32_t);
+template void UnpackRowRGB24_SSSE3<true>(const uint8_t*, uint8_t*, int32_t);
+
+}  // namespace gfx
+}  // namespace mozilla

gfx/2d/moz.build

@@ -145,7 +145,9 @@ if CONFIG['INTEL_ARCHITECTURE']:
         'FilterProcessingSSE2.cpp',
         'ImageScalingSSE2.cpp',
         'ssse3-scaler.c',
+        'SwizzleAVX2.cpp',
         'SwizzleSSE2.cpp',
+        'SwizzleSSSE3.cpp',
     ]
     DEFINES['USE_SSE2'] = True
     # The file uses SSE2 intrinsics, so it needs special compile flags on some
@@ -153,7 +155,9 @@ if CONFIG['INTEL_ARCHITECTURE']:
     SOURCES['BlurSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
     SOURCES['FilterProcessingSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
     SOURCES['ImageScalingSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
+    SOURCES['SwizzleAVX2.cpp'].flags += ['-mavx2']
     SOURCES['SwizzleSSE2.cpp'].flags += CONFIG['SSE2_FLAGS']
+    SOURCES['SwizzleSSSE3.cpp'].flags += CONFIG['SSSE3_FLAGS']
     SOURCES['ssse3-scaler.c'].flags += CONFIG['SSSE3_FLAGS']
 elif CONFIG['CPU_ARCH'].startswith('mips'):
     SOURCES += [