Get AVX2 into WebP lossless

Change-Id: Ifad3102c9f899a46401985515cd98f3f7a21887f

Author: vrabaud

Makefile.vc

@@ -231,6 +231,7 @@ DSP_DEC_OBJS = \
     $(DIROBJ)\dsp\lossless_neon.obj \
     $(DIROBJ)\dsp\lossless_sse2.obj \
     $(DIROBJ)\dsp\lossless_sse41.obj \
+    $(DIROBJ)\dsp\lossless_avx2.obj \
     $(DIROBJ)\dsp\rescaler.obj \
     $(DIROBJ)\dsp\rescaler_mips32.obj \
     $(DIROBJ)\dsp\rescaler_mips_dsp_r2.obj \
@@ -270,6 +271,7 @@ DSP_ENC_OBJS = \
     $(DIROBJ)\dsp\lossless_enc_neon.obj \
     $(DIROBJ)\dsp\lossless_enc_sse2.obj \
     $(DIROBJ)\dsp\lossless_enc_sse41.obj \
+    $(DIROBJ)\dsp\lossless_enc_avx2.obj \
     $(DIROBJ)\dsp\ssim.obj \
     $(DIROBJ)\dsp\ssim_sse2.obj \
 

cmake/config.h.in

@@ -94,6 +94,9 @@
 /* Set to 1 if SSE4.1 is supported */
 #cmakedefine WEBP_HAVE_SSE41 1
 
+/* Set to 1 if AVX2 is supported */
+#cmakedefine WEBP_HAVE_AVX2 1
+
 /* Set to 1 if TIFF library is installed */
 #cmakedefine WEBP_HAVE_TIFF 1
 

cmake/cpu.cmake

@@ -38,9 +38,9 @@ function(webp_check_compiler_flag WEBP_SIMD_FLAG ENABLE_SIMD)
 endfunction()
 
 # those are included in the names of WEBP_USE_* in c++ code.
-set(WEBP_SIMD_FLAGS "SSE41;SSE2;MIPS32;MIPS_DSP_R2;NEON;MSA")
+set(WEBP_SIMD_FLAGS "AVX2;SSE41;SSE2;MIPS32;MIPS_DSP_R2;NEON;MSA")
 set(WEBP_SIMD_FILE_EXTENSIONS
-    "_sse41.c;_sse2.c;_mips32.c;_mips_dsp_r2.c;_neon.c;_msa.c")
+    "_avx2.c;_sse41.c;_sse2.c;_mips32.c;_mips_dsp_r2.c;_neon.c;_msa.c")
 if(MSVC AND CMAKE_C_COMPILER_ID STREQUAL "MSVC")
   # With at least Visual Studio 12 (2013)+ /arch is not necessary to build SSE2
   # or SSE4 code unless a lesser /arch is forced. MSVC does not have a SSE4
@@ -50,12 +50,12 @@ if(MSVC AND CMAKE_C_COMPILER_ID STREQUAL "MSVC")
   if(MSVC_VERSION GREATER_EQUAL 1800 AND NOT CMAKE_C_FLAGS MATCHES "/arch:")
     set(SIMD_ENABLE_FLAGS)
   else()
-    set(SIMD_ENABLE_FLAGS "/arch:AVX;/arch:SSE2;;;;")
+    set(SIMD_ENABLE_FLAGS "/arch:AVX2;/arch:AVX;/arch:SSE2;;;;")
   endif()
   set(SIMD_DISABLE_FLAGS)
 else()
-  set(SIMD_ENABLE_FLAGS "-msse4.1;-msse2;-mips32;-mdspr2;-mfpu=neon;-mmsa")
-  set(SIMD_DISABLE_FLAGS "-mno-sse4.1;-mno-sse2;;-mno-dspr2;;-mno-msa")
+  set(SIMD_ENABLE_FLAGS "-mavx2;-msse4.1;-msse2;-mips32;-mdspr2;-mfpu=neon;-mmsa")
+  set(SIMD_DISABLE_FLAGS "-mno-avx2;-mno-sse4.1;-mno-sse2;;-mno-dspr2;;-mno-msa")
 endif()
 
 set(WEBP_SIMD_FILES_TO_INCLUDE)

configure.ac

@@ -161,6 +161,25 @@ AS_IF([test "$GCC" = "yes" ], [
 AC_SUBST([AM_CFLAGS])
 
 dnl === Check for machine specific flags
+AC_ARG_ENABLE([avx2],
+              AS_HELP_STRING([--disable-avx2],
+                             [Disable detection of AVX2 support
+                              @<:@default=auto@:>@]))
+
+AS_IF([test "x$enable_avx2" != "xno" -a "x$enable_sse4_1" != "xno"
+      -a "x$enable_sse2" != "xno"], [
+  AVX2_FLAGS="$INTRINSICS_CFLAGS $AVX2_FLAGS"
+  TEST_AND_ADD_CFLAGS([AVX2_FLAGS], [-mavx2])
+  AS_IF([test -n "$AVX2_FLAGS"], [
+    SAVED_CFLAGS=$CFLAGS
+    CFLAGS="$CFLAGS $AVX2_FLAGS"
+    AC_CHECK_HEADER([immintrin.h],
+                    [AC_DEFINE(WEBP_HAVE_AVX2, [1],
+                     [Set to 1 if AVX2 is supported])],
+                    [AVX2_FLAGS=""])
+    CFLAGS=$SAVED_CFLAGS])
+  AC_SUBST([AVX2_FLAGS])])
+
 AC_ARG_ENABLE([sse4.1],
               AS_HELP_STRING([--disable-sse4.1],
                              [Disable detection of SSE4.1 support

src/dsp/Makefile.am

@@ -5,6 +5,8 @@ noinst_LTLIBRARIES += libwebpdsp_sse2.la
 noinst_LTLIBRARIES += libwebpdspdecode_sse2.la
 noinst_LTLIBRARIES += libwebpdsp_sse41.la
 noinst_LTLIBRARIES += libwebpdspdecode_sse41.la
+noinst_LTLIBRARIES += libwebpdsp_avx2.la
+noinst_LTLIBRARIES += libwebpdspdecode_avx2.la
 noinst_LTLIBRARIES += libwebpdsp_neon.la
 noinst_LTLIBRARIES += libwebpdspdecode_neon.la
 noinst_LTLIBRARIES += libwebpdsp_msa.la
@@ -44,6 +46,11 @@ ENC_SOURCES += lossless_enc.c
 ENC_SOURCES += quant.h
 ENC_SOURCES += ssim.c
 
+libwebpdspdecode_avx2_la_SOURCES =
+libwebpdspdecode_avx2_la_SOURCES += lossless_avx2.c
+libwebpdspdecode_avx2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
+libwebpdspdecode_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
@@ -123,6 +130,12 @@ libwebpdsp_sse41_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
 libwebpdsp_sse41_la_CFLAGS = $(AM_CFLAGS) $(SSE41_FLAGS)
 libwebpdsp_sse41_la_LIBADD = libwebpdspdecode_sse41.la
 
+libwebpdsp_avx2_la_SOURCES =
+libwebpdsp_avx2_la_SOURCES += lossless_enc_avx2.c
+libwebpdsp_avx2_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS)
+libwebpdsp_avx2_la_CFLAGS = $(AM_CFLAGS) $(AVX2_FLAGS)
+libwebpdsp_avx2_la_LIBADD = libwebpdspdecode_avx2.la
+
 libwebpdsp_neon_la_SOURCES =
 libwebpdsp_neon_la_SOURCES += cost_neon.c
 libwebpdsp_neon_la_SOURCES += enc_neon.c
@@ -167,6 +180,7 @@ libwebpdsp_la_LDFLAGS = -lm
 libwebpdsp_la_LIBADD =
 libwebpdsp_la_LIBADD += libwebpdsp_sse2.la
 libwebpdsp_la_LIBADD += libwebpdsp_sse41.la
+libwebpdsp_la_LIBADD += libwebpdsp_avx2.la
 libwebpdsp_la_LIBADD += libwebpdsp_neon.la
 libwebpdsp_la_LIBADD += libwebpdsp_msa.la
 libwebpdsp_la_LIBADD += libwebpdsp_mips32.la
@@ -180,6 +194,7 @@ if BUILD_LIBWEBPDECODER
   libwebpdspdecode_la_LIBADD =
   libwebpdspdecode_la_LIBADD += libwebpdspdecode_sse2.la
   libwebpdspdecode_la_LIBADD += libwebpdspdecode_sse41.la
+  libwebpdspdecode_la_LIBADD += libwebpdspdecode_avx2.la
   libwebpdspdecode_la_LIBADD += libwebpdspdecode_neon.la
   libwebpdspdecode_la_LIBADD += libwebpdspdecode_msa.la
   libwebpdspdecode_la_LIBADD += libwebpdspdecode_mips32.la

src/dsp/cpu.h

@@ -56,6 +56,11 @@
     (defined(_M_X64) || defined(_M_IX86))
 #define WEBP_MSC_SSE41  // Visual C++ SSE4.1 targets
 #endif
+
+#if defined(_MSC_VER) && _MSC_VER >= 1700 && \
+    (defined(_M_X64) || defined(_M_IX86))
+#define WEBP_MSC_AVX2  // Visual C++ AVX2 targets
+#endif
 #endif
 
 // WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp
@@ -80,6 +85,16 @@
 #define WEBP_HAVE_SSE41
 #endif
 
+#if (defined(__AVX2__) || defined(WEBP_MSC_AVX2)) && \
+    (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_AVX2))
+#define WEBP_USE_AVX2
+#endif
+
+#if defined(WEBP_USE_AVX2) && !defined(WEBP_HAVE_AVX2)
+#define WEBP_HAVE_AVX2
+#endif
+
+#undef WEBP_MSC_AVX2
 #undef WEBP_MSC_SSE41
 #undef WEBP_MSC_SSE2
 

src/dsp/lossless.c

@@ -577,16 +577,21 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
 //------------------------------------------------------------------------------
 
 VP8LProcessDecBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+VP8LProcessDecBlueAndRedFunc VP8LAddGreenToBlueAndRed_SSE;
 VP8LPredictorAddSubFunc VP8LPredictorsAdd[16];
+VP8LPredictorAddSubFunc VP8LPredictorsAdd_SSE[16];
 VP8LPredictorFunc VP8LPredictors[16];
 
 // exposed plain-C implementations
 VP8LPredictorAddSubFunc VP8LPredictorsAdd_C[16];
 
 VP8LTransformColorInverseFunc VP8LTransformColorInverse;
+VP8LTransformColorInverseFunc VP8LTransformColorInverse_SSE;
 
 VP8LConvertFunc VP8LConvertBGRAToRGB;
+VP8LConvertFunc VP8LConvertBGRAToRGB_SSE;
 VP8LConvertFunc VP8LConvertBGRAToRGBA;
+VP8LConvertFunc VP8LConvertBGRAToRGBA_SSE;
 VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
 VP8LConvertFunc VP8LConvertBGRAToRGB565;
 VP8LConvertFunc VP8LConvertBGRAToBGR;
@@ -597,6 +602,7 @@ VP8LMapAlphaFunc VP8LMapColor8b;
 extern VP8CPUInfo VP8GetCPUInfo;
 extern void VP8LDspInitSSE2(void);
 extern void VP8LDspInitSSE41(void);
+extern void VP8LDspInitAVX2(void);
 extern void VP8LDspInitNEON(void);
 extern void VP8LDspInitMIPSdspR2(void);
 extern void VP8LDspInitMSA(void);
@@ -649,6 +655,11 @@ WEBP_DSP_INIT_FUNC(VP8LDspInit) {
 #if defined(WEBP_HAVE_SSE41)
       if (VP8GetCPUInfo(kSSE4_1)) {
         VP8LDspInitSSE41();
+#if defined(WEBP_HAVE_AVX2)
+        if (VP8GetCPUInfo(kAVX2)) {
+          VP8LDspInitAVX2();
+        }
+#endif
       }
 #endif
     }

src/dsp/lossless.h

@@ -64,10 +64,12 @@ typedef void (*VP8LPredictorAddSubFunc)(const uint32_t* in,
                                         uint32_t* WEBP_RESTRICT out);
 extern VP8LPredictorAddSubFunc VP8LPredictorsAdd[16];
 extern VP8LPredictorAddSubFunc VP8LPredictorsAdd_C[16];
+extern VP8LPredictorAddSubFunc VP8LPredictorsAdd_SSE[16];
 
 typedef void (*VP8LProcessDecBlueAndRedFunc)(const uint32_t* src,
                                              int num_pixels, uint32_t* dst);
 extern VP8LProcessDecBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+extern VP8LProcessDecBlueAndRedFunc VP8LAddGreenToBlueAndRed_SSE;
 
 typedef struct {
   // Note: the members are uint8_t, so that any negative values are
@@ -80,6 +82,7 @@ typedef void (*VP8LTransformColorInverseFunc)(const VP8LMultipliers* const m,
                                               const uint32_t* src,
                                               int num_pixels, uint32_t* dst);
 extern VP8LTransformColorInverseFunc VP8LTransformColorInverse;
+extern VP8LTransformColorInverseFunc VP8LTransformColorInverse_SSE;
 
 struct VP8LTransform;  // Defined in dec/vp8li.h.
 
@@ -99,6 +102,8 @@ extern VP8LConvertFunc VP8LConvertBGRAToRGBA;
 extern VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
 extern VP8LConvertFunc VP8LConvertBGRAToRGB565;
 extern VP8LConvertFunc VP8LConvertBGRAToBGR;
+extern VP8LConvertFunc VP8LConvertBGRAToRGB_SSE;
+extern VP8LConvertFunc VP8LConvertBGRAToRGBA_SSE;
 
 // Converts from BGRA to other color spaces.
 void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
@@ -149,21 +154,25 @@ void VP8LDspInit(void);
 
 typedef void (*VP8LProcessEncBlueAndRedFunc)(uint32_t* dst, int num_pixels);
 extern VP8LProcessEncBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
+extern VP8LProcessEncBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed_SSE;
 typedef void (*VP8LTransformColorFunc)(
     const VP8LMultipliers* WEBP_RESTRICT const m, uint32_t* WEBP_RESTRICT dst,
     int num_pixels);
 extern VP8LTransformColorFunc VP8LTransformColor;
+extern VP8LTransformColorFunc VP8LTransformColor_SSE;
 typedef void (*VP8LCollectColorBlueTransformsFunc)(
     const uint32_t* WEBP_RESTRICT argb, int stride,
     int tile_width, int tile_height,
     int green_to_blue, int red_to_blue, uint32_t histo[]);
 extern VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms;
+extern VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms_SSE;
 
 typedef void (*VP8LCollectColorRedTransformsFunc)(
     const uint32_t* WEBP_RESTRICT argb, int stride,
     int tile_width, int tile_height,
     int green_to_red, uint32_t histo[]);
 extern VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms;
+extern VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms_SSE;
 
 // Expose some C-only fallback functions
 void VP8LTransformColor_C(const VP8LMultipliers* WEBP_RESTRICT const m,
@@ -181,6 +190,7 @@ void VP8LCollectColorBlueTransforms_C(const uint32_t* WEBP_RESTRICT argb,
 
 extern VP8LPredictorAddSubFunc VP8LPredictorsSub[16];
 extern VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16];
+extern VP8LPredictorAddSubFunc VP8LPredictorsSub_SSE[16];
 
 // -----------------------------------------------------------------------------
 // Huffman-cost related functions.
@@ -255,6 +265,7 @@ typedef void (*VP8LBundleColorMapFunc)(const uint8_t* WEBP_RESTRICT const row,
                                        int width, int xbits,
                                        uint32_t* WEBP_RESTRICT dst);
 extern VP8LBundleColorMapFunc VP8LBundleColorMap;
+extern VP8LBundleColorMapFunc VP8LBundleColorMap_SSE;
 void VP8LBundleColorMap_C(const uint8_t* WEBP_RESTRICT const row,
                           int width, int xbits, uint32_t* WEBP_RESTRICT dst);