libjpeg-turbo SIMD Extensions

This directory contains the libjpeg-turbo SIMD extensions, hand-coded assembly or compiler intrinsics modules that plug into the modular interfaces of the libjpeg API library and use various SIMD instruction sets to accelerate most of the 8-bit-per-sample lossy JPEG compression and decompression algorithms.

(Note that, since the TurboJPEG API library wraps the libjpeg API library, it uses the libjpeg-turbo SIMD extensions implicitly.)

Algorithm Coverage

The following 8-bit-per-sample lossy JPEG compression algorithms are currently implemented as SIMD modules:

• Color Conversion (see jccolor.c)
  • RGB-to-YCbCr Color Conversion
  • RGB-to-Grayscale Color Conversion
  • NULL Color Conversion (infrequently used)
• Downsampling (see jcsample.c)
  • H2V1 (4:2:2) Downsampling
  • H2V2 (4:2:0) Downsampling
  • H2V2 (4:2:0) Smooth Downsampling (infrequently used)
• Sample Conversion (see jcdctmgr.c)
  • Integer Sample Conversion
  • Floating Point Sample Conversion (legacy feature)
• Forward DCT (see jcdctmgr.c)
  • Accurate Integer Forward DCT
  • Fast Integer Forward DCT (legacy feature)
  • Floating Point Forward DCT (legacy feature)
• Quantization (see jcdctmgr.c)
  • Integer Quantization
  • Floating Point Quantization (legacy feature)
• Entropy Encoding
  • Huffman Encoding (see jchuff.c)
  • Progressive Huffman Encoding (see jcphuff.c)

The following 8-bit-per-sample lossy JPEG decompression algorithms are currently implemented as SIMD modules:

• Inverse DCT (see jddctmgr.c)
  • Accurate Integer Inverse DCT
  • Fast Integer Inverse DCT (legacy feature)
  • Floating Point Inverse DCT (legacy feature)
  • 2x2 (1/4 Scaling) Integer Inverse DCT (infrequently used)
  • 4x4 (1/2 Scaling) Integer Inverse DCT (infrequently used)
  • 6x6 (3/4 Scaling) Integer Inverse DCT (infrequently used)
  • 12x12 (3/2 Scaling) Integer Inverse DCT (infrequently used)
• Upsampling (see jdsample.c)
  • H2V1 (4:2:2) Fancy (Smooth) Upsampling
  • H2V2 (4:2:0) Fancy (Smooth) Upsampling
  • H1V2 (4:4:0) Fancy (Smooth) Upsampling
  • H2V1 (4:2:2) Plain Upsampling (infrequently used)
  • H2V2 (4:2:0) Plain Upsampling (infrequently used)
  • Integer Upsampling (infrequently used)
• Merged Upsampling (see jdmerge.c)
  • H2V1 (4:2:2) Merged Upsampling/Color Conversion
  • H2V2 (4:2:0) Merged Upsampling/Color Conversion
• Color Deconversion (see jdcolor.c)
  • YCbCr-to-RGB Color Conversion
  • YCbCr-to-RGB565 Color Conversion (infrequently used)

Refer to https://libjpeg-turbo.org/About/SIMDCoverage for a list of SIMD modules that are implemented for the algorithms above using specific SIMD instruction sets.

Legacy features are features that were designed to work around hardware performance limitations that no longer exist. They generally have little or no utility on modern hardware and are retained only for backward compatibility with libjpeg.

Infrequently used features may be useful for specific applications but are not the “common case” for JPEG compression and decompression.

SIMD Dispatcher Operation

When initializing a particular compression or decompression module (which occurs during jpeg_start_compress(), jpeg_start_decompress(), tj3Compress*(), or tj3Decompress*()), the libjpeg API library calls a SIMD dispatcher function (jsimd_set_*()) for each algorithm listed above. The SIMD dispatcher functions are defined in jsimd.h, jsimddct.h, and jsimd.c. Each function

• determines which SIMD instruction sets are available for the current architecture (if such has not already been determined for the current API instance),
• determines whether a compatible SIMD module exists for the specified algorithm, and
• (if so) plugs the best available SIMD module for that algorithm into the appropriate compression or decompression module, thus causing a SIMD implementation of the algorithm to be used instead of the scalar/C implementation.

You can use environment variables to override the dispatchers' choice of SIMD modules:

• JSIMD_FORCENONE=1 disables all SIMD modules.
• JSIMD_NOHUFFENC=1 disables only the Huffman encoding SIMD modules.
• JSIMD_FORCESSE2=1 (x86) enables only the SSE2 and SSE SIMD modules, even if the CPU supports newer instruction sets.
• JSIMD_FORCESSE=1 (i386) enables only the SSE and MMX SIMD modules, even if the CPU supports newer instruction sets.
• JSIMD_FORCEMMX=1 (i386) enables only the MMX SIMD modules, even if the CPU supports newer instruction sets.
• JSIMD_FORCENEON=1 (AArch32) force-enables the Neon SIMD modules, bypassing /proc/cpuinfo feature detection (which may be unreliable in QEMU and other emulation/virtualization environments.)
• JSIMD_FORCEDSPR2=1 (MIPS) force-enables the DSPr2 SIMD modules, bypassing /proc/cpuinfo feature detection (which may be unreliable in QEMU and other emulation/virtualization environments.)
• JSIMD_FORCEMMI=1 (Loongson) force-enables the MMI SIMD modules, bypassing /proc/cpuinfo feature detection (which may be unreliable in QEMU and other emulation/virtualization environments.)

The simdcoverage program reports which SIMD modules will be used, taking into account the current architecture, detected CPU features, and aforementioned overrides.

SIMD Module Performance Profiling

When built with the WITH_PROFILE CMake variable enabled, the libjpeg API library measures the average throughput of each lossy JPEG algorithm as an image is compressed or decompressed, and it prints the results to the command line when jpeg_destroy_compress(), jpeg_destroy_decompress(), or tj3Destroy() is called. This allows developers to easily study the performance of each SIMD module in isolation and compare it to the performance of the corresponding scalar/C module (or a previous implementation of the same SIMD module.)

The most effective way to use the profiling feature is with the TJBench application and a suitably large image, such as one of the 8-bit RGB images from imagecompression.info. Use one of the following command lines to obtain the performance of specific algorithms. (Adjust the warmup and benchmark times to suit your needs.)

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422

  • “Color conversion” reports the performance of the RGB-to-YCbCr Color Conversion algorithm.
  • “Downsampling” reports the performance of the H2V1 (4:2:2) Downsampling algorithm.
  • “Sample conversion” reports the performance of the Integer Sample Conversion algorithm.
  • “Forward DCT” reports the performance of the Accurate Integer Forward DCT algorithm.
  • “Quantization” reports the performance of the Integer Quantization algorithm.
  • “Entropy encoding” reports the performance of the Huffman Encoding algorithm.
  • “Inverse DCT” reports the performance of the Accurate Integer Inverse DCT algorithm.
  • “Upsampling” reports the performance of the H2V1 (4:2:2) Fancy (Smooth) Upsampling algorithm.
  • “Color deconversion” reports the performance of the YCbCr-to-RGB Color Conversion algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp gray

  • “Color conversion” reports the performance of the RGB-to-Grayscale Color Conversion algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 420

  • “Downsampling” reports the performance of the H2V2 (4:2:0) Downsampling algorithm.
  • “Upsampling” reports the performance of the H2V2 (4:2:0) Fancy (Smooth) Upsampling algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 440

  • “Upsampling” reports the performance of the H1V2 (4:4:0) Fancy (Smooth) Upsampling algorithm.

• tjbench {image}.ppm 95 -cmyk -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -nosmooth

  • “Upsampling” reports the performance of the H2V1 (4:2:2) Plain Upsampling algorithm.

• tjbench {image}.ppm 95 -cmyk -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 420 -nosmooth

  • “Upsampling” reports the performance of the H2V2 (4:2:0) Plain Upsampling algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 440 -nosmooth

  • “Upsampling” reports the performance of the Integer Upsampling algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -nosmooth

  • “Merged upsampling” reports the performance of the H2V1 (4:2:2) Merged Upsampling/Color Conversion algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 420 -nosmooth

  • “Merged upsampling” reports the performance of the H2V2 (4:2:0) Merged Upsampling/Color Conversion algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -progressive

  • “Entropy encoding” reports the performance of the Progressive Huffman Encoding algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -scale 1/4

  • “Inverse DCT” reports the performance of the 2x2 (1/4 Scaling) Integer Inverse DCT algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -scale 1/2

  • “Inverse DCT” reports the performance of the 4x4 (1/2 Scaling) Integer Inverse DCT algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -scale 3/4

  • “Inverse DCT” reports the performance of the 6x6 (3/4 Scaling) Integer Inverse DCT algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -scale 3/2

  • “Inverse DCT” reports the performance of the 12x12 (3/2 Scaling) Integer Inverse DCT algorithm.

• tjbench {image}.ppm 95 -rgb -quiet -nowrite -benchtime 10 -warmup 10 -subsamp 422 -dct fast

  • “Forward DCT” reports the performance of the Fast Integer Forward DCT algorithm.
  • “Inverse DCT” reports the performance of the Fast Integer Inverse DCT algorithm.