SSE2 SIMD implementation of Huffman encoding

Full-color compression speedups relative to libjpeg-turbo 1.4.2: 2.8 GHz Intel Xeon W3530, Linux, 64-bit: 2.2-18% (avg. 9.5%) 2.8 GHz Intel Xeon W3530, Linux, 32-bit: 10-25% (avg. 17%) 2.3 GHz AMD A10-4600M APU, Linux, 64-bit: 4.9-17% (avg. 11%) 2.3 GHz AMD A10-4600M APU, Linux, 32-bit: 8.8-19% (avg. 15%) 3.0 GHz Intel Core i7, OS X, 64-bit: 3.5-16% (avg. 10%) 3.0 GHz Intel Core i7, OS X, 32-bit: 4.8-14% (avg. 11%) 2.6 GHz AMD Athlon 64 X2 5050e: Performance-neutral (give or take a few percent) Full-color compression speedups relative to IPP: 2.8 GHz Intel Xeon W3530, Linux, 64-bit: 4.8-34% (avg. 19%) 2.8 GHz Intel Xeon W3530, Linux, 32-bit: -19%-7.0% (avg. -7.0%) Refer to libjpeg-turbo#42 for discussion. Numerous other approaches were attempted, but this one proved to be the most performant across all platforms. This commit also fixes libjpeg-turbo#3 (works around, really-- the clang-compiled version of jchuff.c still performs 20% worse than its GCC-compiled counterpart, but that code is now bypassed by the new SSE2 Huffman algorithm.) Based on: mayeut@2cb4d41 mayeut@36c94e0
lilith · Jan 12, 2016 · f3a8684 · f3a8684
1 parent eb59b6e
commit f3a8684
Show file tree

Hide file tree

Showing 18 changed files with 5,157 additions and 84 deletions.
diff --git a/BUILDING.md b/BUILDING.md
@@ -38,19 +38,7 @@ Build Requirements
 
   NOTE: the NASM build will fail if texinfo is not installed.
 
-- GCC v4.1 or later recommended for best performance
-  * Beginning with Xcode 4, Apple stopped distributing GCC and switched to
-    the LLVM compiler.  Xcode v4.0 through v4.6 provides a GCC front end
-    called LLVM-GCC.  Unfortunately, as of this writing, neither LLVM-GCC nor
-    the LLVM (clang) compiler produces optimal performance with libjpeg-turbo.
-    Building libjpeg-turbo with LLVM-GCC v4.2 results in a 10% performance
-    degradation when compressing using 64-bit code, relative to building
-    libjpeg-turbo with GCC v4.2.  Building libjpeg-turbo with LLVM (clang)
-    results in a 20% performance degradation when compressing using 64-bit
-    code, relative to building libjpeg-turbo with GCC v4.2.  If you are
-    running Snow Leopard or earlier, it is suggested that you continue to use
-    Xcode v3.2.6, which provides GCC v4.2.  If you are using Lion or later, it
-    is suggested that you install Apple GCC v4.2 or GCC v5 through MacPorts.
+- GCC v4.1 (or later) or clang recommended for best performance
 
 - If building the TurboJPEG Java wrapper, JDK or OpenJDK 1.5 or later is
   required.  Some systems, such as Solaris 10 and later and Red Hat Enterprise
@@ -89,38 +77,38 @@ for 64-bit build instructions.)
 
 This will generate the following files under .libs/:
 
-**libjpeg.a**  
+**libjpeg.a**
 Static link library for the libjpeg API
 
-**libjpeg.so.{version}** (Linux, Unix)  
-**libjpeg.{version}.dylib** (OS X)  
-**cygjpeg-{version}.dll** (Cygwin)  
+**libjpeg.so.{version}** (Linux, Unix)
+**libjpeg.{version}.dylib** (OS X)
+**cygjpeg-{version}.dll** (Cygwin)
 Shared library for the libjpeg API
 
 By default, *{version}* is 62.1.0, 7.1.0, or 8.0.2, depending on whether
 libjpeg v6b (default), v7, or v8 emulation is enabled.  If using Cygwin,
 *{version}* is 62, 7, or 8.
 
-**libjpeg.so** (Linux, Unix)  
-**libjpeg.dylib** (OS X)  
+**libjpeg.so** (Linux, Unix)
+**libjpeg.dylib** (OS X)
 Development symlink for the libjpeg API
 
-**libjpeg.dll.a** (Cygwin)  
+**libjpeg.dll.a** (Cygwin)
 Import library for the libjpeg API
 
-**libturbojpeg.a**  
+**libturbojpeg.a**
 Static link library for the TurboJPEG API
 
-**libturbojpeg.so.0.1.0** (Linux, Unix)  
-**libturbojpeg.0.1.0.dylib** (OS X)  
-**cygturbojpeg-0.dll** (Cygwin)  
+**libturbojpeg.so.0.1.0** (Linux, Unix)
+**libturbojpeg.0.1.0.dylib** (OS X)
+**cygturbojpeg-0.dll** (Cygwin)
 Shared library for the TurboJPEG API
 
-**libturbojpeg.so** (Linux, Unix)  
-**libturbojpeg.dylib** (OS X)  
+**libturbojpeg.so** (Linux, Unix)
+**libturbojpeg.dylib** (OS X)
 Development symlink for the TurboJPEG API
 
-**libturbojpeg.dll.a** (Cygwin)  
+**libturbojpeg.dll.a** (Cygwin)
 Import library for the TurboJPEG API
 
 
@@ -333,16 +321,16 @@ Set the following shell variables for simplicity:
     IOS_SYSROOT=$IOS_PLATFORMDIR/Developer/SDKs/iPhoneOS*.sdk
     IOS_GCC=$IOS_PLATFORMDIR/Developer/usr/bin/arm-apple-darwin10-llvm-gcc-4.2
 
-  *ARMv6 (code will run on all iOS devices, not SIMD-accelerated)*  
+  *ARMv6 (code will run on all iOS devices, not SIMD-accelerated)*
   [NOTE: Requires Xcode 4.4.x or earlier]
 
     IOS_CFLAGS="-march=armv6 -mcpu=arm1176jzf-s -mfpu=vfp"
 
-  *ARMv7 (code will run on iPhone 3GS-4S/iPad 1st-3rd Generation and newer)*  
+  *ARMv7 (code will run on iPhone 3GS-4S/iPad 1st-3rd Generation and newer)*
 
     IOS_CFLAGS="-march=armv7 -mcpu=cortex-a8 -mtune=cortex-a8 -mfpu=neon"
 
-  *ARMv7s (code will run on iPhone 5/iPad 4th Generation and newer)*  
+  *ARMv7s (code will run on iPhone 5/iPad 4th Generation and newer)*
   [NOTE: Requires Xcode 4.5 or later]
 
     IOS_CFLAGS="-march=armv7s -mcpu=swift -mtune=swift -mfpu=neon"
@@ -365,11 +353,11 @@ Set the following shell variables for simplicity:
     IOS_SYSROOT=$IOS_PLATFORMDIR/Developer/SDKs/iPhoneOS*.sdk
     IOS_GCC=/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang
 
-  *ARMv7 (code will run on iPhone 3GS-4S/iPad 1st-3rd Generation and newer)*  
+  *ARMv7 (code will run on iPhone 3GS-4S/iPad 1st-3rd Generation and newer)*
 
     IOS_CFLAGS="-arch armv7"
 
-  *ARMv7s (code will run on iPhone 5/iPad 4th Generation and newer)*  
+  *ARMv7s (code will run on iPhone 5/iPad 4th Generation and newer)*
 
     IOS_CFLAGS="-arch armv7s"
 
@@ -527,22 +515,22 @@ on which version of cl.exe is in the `PATH`.
 
 The following files will be generated under *{build_directory}*:
 
-**jpeg-static.lib**  
+**jpeg-static.lib**
 Static link library for the libjpeg API
 
-**sharedlib/jpeg{version}.dll**  
+**sharedlib/jpeg{version}.dll**
 DLL for the libjpeg API
 
-**sharedlib/jpeg.lib**  
+**sharedlib/jpeg.lib**
 Import library for the libjpeg API
- 
-**turbojpeg-static.lib**  
+
+**turbojpeg-static.lib**
 Static link library for the TurboJPEG API
 
-**turbojpeg.dll**  
+**turbojpeg.dll**
 DLL for the TurboJPEG API
 
-**turbojpeg.lib**  
+**turbojpeg.lib**
 Import library for the TurboJPEG API
 
 *{version}* is 62, 7, or 8, depending on whether libjpeg v6b (default), v7, or
@@ -569,22 +557,22 @@ build of libjpeg-turbo.
 
 This will generate the following files under *{build_directory}*:
 
-**{configuration}/jpeg-static.lib**  
+**{configuration}/jpeg-static.lib**
 Static link library for the libjpeg API
 
-**sharedlib/{configuration}/jpeg{version}.dll**  
+**sharedlib/{configuration}/jpeg{version}.dll**
 DLL for the libjpeg API
 
-**sharedlib/{configuration}/jpeg.lib**  
+**sharedlib/{configuration}/jpeg.lib**
 Import library for the libjpeg API
 
-**{configuration}/turbojpeg-static.lib**  
+**{configuration}/turbojpeg-static.lib**
 Static link library for the TurboJPEG API
 
-**{configuration}/turbojpeg.dll**  
+**{configuration}/turbojpeg.dll**
 DLL for the TurboJPEG API
 
-**{configuration}/turbojpeg.lib**  
+**{configuration}/turbojpeg.lib**
 Import library for the TurboJPEG API
 
 *{configuration}* is Debug, Release, RelWithDebInfo, or MinSizeRel, depending
@@ -603,22 +591,22 @@ cross-compiling on a Linux/Unix machine, then see "Build Recipes" below.
 
 This will generate the following files under *{build_directory}*:
 
-**libjpeg.a**  
+**libjpeg.a**
 Static link library for the libjpeg API
 
-**sharedlib/libjpeg-{version}.dll**  
+**sharedlib/libjpeg-{version}.dll**
 DLL for the libjpeg API
 
-**sharedlib/libjpeg.dll.a**  
+**sharedlib/libjpeg.dll.a**
 Import library for the libjpeg API
 
-**libturbojpeg.a**  
+**libturbojpeg.a**
 Static link library for the TurboJPEG API
 
-**libturbojpeg.dll**  
+**libturbojpeg.dll**
 DLL for the TurboJPEG API
 
-**libturbojpeg.dll.a**  
+**libturbojpeg.dll.a**
 Import library for the TurboJPEG API
 
 *{version}* is 62, 7, or 8, depending on whether libjpeg v6b (default), v7, or

diff --git a/ChangeLog.txt b/ChangeLog.txt
@@ -57,6 +57,16 @@ benchmark from outputting any images.  This removes any potential operating
 system overhead that might be caused by lazy writes to disk and thus improves
 the consistency of the performance measurements.
 
+[12] Added SIMD acceleration for Huffman encoding on SSE2-capable x86 and
+x86-64 platforms.  This speeds up the compression of full-color JPEGs by about
+10-15% on average (relative to libjpeg-turbo 1.4.x) when using modern Intel and
+AMD CPUs.  Additionally, this works around an issue in the clang optimizer that
+prevents it (as of this writing) from achieving the same performance as GCC
+when compiling the C version of the Huffman encoder
+(https://llvm.org/bugs/show_bug.cgi?id=16035). For the purposes of benchmarking
+or regression testing, SIMD-accelerated Huffman encoding can be disabled by
+setting the JSIMD_NOHUFFENC environment variable to 1.
+
 
 1.4.2
 =====

diff --git a/jchuff.c b/jchuff.c
@@ -5,6 +5,7 @@
  * Copyright (C) 1991-1997, Thomas G. Lane.
  * libjpeg-turbo Modifications:
  * Copyright (C) 2009-2011, 2014-2016 D. R. Commander.
+ * Copyright (C) 2015 Matthieu Darbois.
  * For conditions of distribution and use, see the accompanying README.ijg
  * file.
  *
@@ -20,7 +21,7 @@
 #define JPEG_INTERNALS
 #include "jinclude.h"
 #include "jpeglib.h"
-#include "jchuff.h"             /* Declarations shared with jcphuff.c */
+#include "jsimd.h"
 #include "jconfigint.h"
 #include <limits.h>
 
@@ -108,6 +109,8 @@ typedef struct {
   long * dc_count_ptrs[NUM_HUFF_TBLS];
   long * ac_count_ptrs[NUM_HUFF_TBLS];
 #endif
+
+  int simd;
 } huff_entropy_encoder;
 
 typedef huff_entropy_encoder * huff_entropy_ptr;
@@ -159,6 +162,8 @@ start_pass_huff (j_compress_ptr cinfo, boolean gather_statistics)
     entropy->pub.finish_pass = finish_pass_huff;
   }
 
+  entropy->simd = jsimd_can_huff_encode_one_block();
+
   for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
     compptr = cinfo->cur_comp_info[ci];
     dctbl = compptr->dc_tbl_no;
@@ -480,6 +485,23 @@ flush_bits (working_state * state)
 
 /* Encode a single block's worth of coefficients */
 
+LOCAL(boolean)
+encode_one_block_simd (working_state * state, JCOEFPTR block, int last_dc_val,
+                       c_derived_tbl *dctbl, c_derived_tbl *actbl)
+{
+  JOCTET _buffer[BUFSIZE], *buffer;
+  size_t bytes, bytestocopy;  int localbuf = 0;
+
+  LOAD_BUFFER()
+
+  buffer = jsimd_huff_encode_one_block(state, buffer, block, last_dc_val,
+                                       dctbl, actbl);
+
+  STORE_BUFFER()
+
+  return TRUE;
+}
+
 LOCAL(boolean)
 encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
                   c_derived_tbl *dctbl, c_derived_tbl *actbl)
@@ -640,16 +662,30 @@ encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
   }
 
   /* Encode the MCU data blocks */
-  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
-    ci = cinfo->MCU_membership[blkn];
-    compptr = cinfo->cur_comp_info[ci];
-    if (! encode_one_block(&state,
-                           MCU_data[blkn][0], state.cur.last_dc_val[ci],
-                           entropy->dc_derived_tbls[compptr->dc_tbl_no],
-                           entropy->ac_derived_tbls[compptr->ac_tbl_no]))
-      return FALSE;
-    /* Update last_dc_val */
-    state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+  if (entropy->simd) {
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      if (! encode_one_block_simd(&state,
+                                  MCU_data[blkn][0], state.cur.last_dc_val[ci],
+                                  entropy->dc_derived_tbls[compptr->dc_tbl_no],
+                                  entropy->ac_derived_tbls[compptr->ac_tbl_no]))
+        return FALSE;
+      /* Update last_dc_val */
+      state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+    }
+  } else {
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      if (! encode_one_block(&state,
+                             MCU_data[blkn][0], state.cur.last_dc_val[ci],
+                             entropy->dc_derived_tbls[compptr->dc_tbl_no],
+                             entropy->ac_derived_tbls[compptr->ac_tbl_no]))
+        return FALSE;
+      /* Update last_dc_val */
+      state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+    }
   }
 
   /* Completed MCU, so update state */

diff --git a/jsimd.h b/jsimd.h
@@ -3,13 +3,16 @@
  *
  * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
  * Copyright 2011, 2014 D. R. Commander
+ * Copyright 2015 Matthieu Darbois
  *
  * Based on the x86 SIMD extension for IJG JPEG library,
  * Copyright (C) 1999-2006, MIYASAKA Masaru.
  * For conditions of distribution and use, see copyright notice in jsimdext.inc
  *
  */
 
+#include "jchuff.h"             /* Declarations shared with jcphuff.c */
+
 EXTERN(int) jsimd_can_rgb_ycc (void);
 EXTERN(int) jsimd_can_rgb_gray (void);
 EXTERN(int) jsimd_can_ycc_rgb (void);
@@ -82,3 +85,9 @@ EXTERN(void) jsimd_h2v2_merged_upsample
 EXTERN(void) jsimd_h2v1_merged_upsample
         (j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
          JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf);
+
+EXTERN(int) jsimd_can_huff_encode_one_block (void);
+
+EXTERN(JOCTET*) jsimd_huff_encode_one_block
+        (void * state, JOCTET *buffer, JCOEFPTR block, int last_dc_val,
+         c_derived_tbl *dctbl, c_derived_tbl *actbl);
diff --git a/jsimd_none.c b/jsimd_none.c
@@ -3,6 +3,7 @@
  *
  * Copyright 2009 Pierre Ossman <[email protected]> for Cendio AB
  * Copyright 2009-2011, 2014 D. R. Commander
+ * Copyright 2015 Matthieu Darbois
  *
  * Based on the x86 SIMD extension for IJG JPEG library,
  * Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -387,3 +388,16 @@ jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
 {
 }
 
+GLOBAL(int)
+jsimd_can_huff_encode_one_block (void)
+{
+  return 0;
+}
+
+GLOBAL(JOCTET*)
+jsimd_huff_encode_one_block (void * state, JOCTET *buffer, JCOEFPTR block,
+                             int last_dc_val, c_derived_tbl *dctbl,
+                             c_derived_tbl *actbl)
+{
+  return NULL;
+}
diff --git a/jversion.h b/jversion.h
@@ -32,6 +32,7 @@
                         "Copyright (C) 2009-2016 D. R. Commander\n" \
                         "Copyright (C) 2009-2011 Nokia Corporation and/or its subsidiary(-ies)\n" \
                         "Copyright (C) 2013-2014 MIPS Technologies, Inc.\n" \
-                        "Copyright (C) 2013 Linaro Limited"
+                        "Copyright (C) 2013 Linaro Limited\n" \
+                        "Copyright (C) 2015 Matthieu Darbois"
 
 #define JCOPYRIGHT_SHORT "Copyright (C) 1991-2016 The libjpeg-turbo Project and many others"
diff --git a/simd/CMakeLists.txt b/simd/CMakeLists.txt
@@ -22,17 +22,19 @@ endif()
 
 if(SIMD_X86_64)
   set(SIMD_BASENAMES jfdctflt-sse-64 jccolor-sse2-64 jcgray-sse2-64
-    jcsample-sse2-64 jdcolor-sse2-64 jdmerge-sse2-64 jdsample-sse2-64
-    jfdctfst-sse2-64 jfdctint-sse2-64 jidctflt-sse2-64 jidctfst-sse2-64
-    jidctint-sse2-64 jidctred-sse2-64 jquantf-sse2-64 jquanti-sse2-64)
+    jchuff-sse2-64 jcsample-sse2-64 jdcolor-sse2-64 jdmerge-sse2-64
+    jdsample-sse2-64 jfdctfst-sse2-64 jfdctint-sse2-64 jidctflt-sse2-64
+    jidctfst-sse2-64 jidctint-sse2-64 jidctred-sse2-64 jquantf-sse2-64
+    jquanti-sse2-64)
   message(STATUS "Building x86_64 SIMD extensions")
 else()
   set(SIMD_BASENAMES jsimdcpu jfdctflt-3dn jidctflt-3dn jquant-3dn jccolor-mmx
     jcgray-mmx jcsample-mmx jdcolor-mmx jdmerge-mmx jdsample-mmx jfdctfst-mmx
     jfdctint-mmx jidctfst-mmx jidctint-mmx jidctred-mmx jquant-mmx jfdctflt-sse
-    jidctflt-sse jquant-sse jccolor-sse2 jcgray-sse2 jcsample-sse2 jdcolor-sse2
-    jdmerge-sse2 jdsample-sse2 jfdctfst-sse2 jfdctint-sse2 jidctflt-sse2
-    jidctfst-sse2 jidctint-sse2 jidctred-sse2 jquantf-sse2 jquanti-sse2)
+    jidctflt-sse jquant-sse jccolor-sse2 jcgray-sse2 jchuff-sse2 jcsample-sse2
+    jdcolor-sse2 jdmerge-sse2 jdsample-sse2 jfdctfst-sse2 jfdctint-sse2
+    jidctflt-sse2 jidctfst-sse2 jidctint-sse2 jidctred-sse2 jquantf-sse2
+    jquanti-sse2)
   message(STATUS "Building i386 SIMD extensions")
 endif()