/picojpeg

picojpeg: Tiny JPEG decoder for 8/16-bit microcontrollers

Primary LanguageC

picojpeg

Exported from http://code.google.com/p/picojpeg. Original author is Rich Geldreich at richgel99 at gmail.com.

picojpeg is a public domain JPEG decompressor written in plain C in a single source file picojpeg.c and a single header picojpeg.h. It has several fairly unique properties that make it useful on small 8/16-bit embedded devices or in very memory constrained environments:

  • Optimized for minimal memory consumption: Uses only ~2.3KB of work memory.
  • Written assuming ROM code memory is more abundant than available RAM, which is the case on many microcontrollers.
  • All work arrays are limited to less than or equal to 256 bytes in size.
  • Written entirely in C, no reliance at all on the C run-time library (it includes no headers other than picojpeg.h), and does not use any dynamic memory allocation.
  • Simple API: Call pjpeg_decode_init() to initialize the decoder, and pjpeg_decode_mcu() to decode each MCU (Minimum Coded Unit).
  • Majority of expressions use only 8-bit integer operations, most variables are only 8-bits. 16-bit operations are only used when necessary.
  • Uses a Winograd IDCT to minimize the number of multiplies (5 per 1D IDCT, up to 80 total, each signed mul can be implemented as 16 bits*8 bits=24 bits).
  • Whenever possible, multiplies are limited to 8x8 or 16x8 bits, and are done against compile-time constants.
  • As of v1.1 the decoder supports a special "reduce" mode that returns a 1/8th size version of the compressed image. This mode skips several expensive steps: AC dequantization, IDCT, and per-pixel chroma upsampling/colorspace conversion.

picojpeg has several disadantages and known issues compared to other implementations:

  • Quality is traded off for minimal RAM memory consumption and decent performance on small microcontrollers. For example, the chroma upsamplers use only box filtering, 8x8 multiplies, and minimal 16-bit operations so the decoder is not as accurate as it could be.
  • Only supports baseline sequential greyscale, or YCbCr H1V1, H1V2, H2V1, and H2V2 chroma sampling factors. Progressive JPEG's are not supported.
  • The Huffman decoder currently only reads a bit at a time to minimize RAM usage, so it's pretty slow. (However, on microcontroller CPU's with weak integer shift capabilities this method may be reasonable.)
  • All work arrays (approx. 2.3KB) are globals, because this resulted in the best code generation with the embedded compiler I was using during development. I'm assuming either this is not an issue, or the user is using a compiler that allows them to overlap these variables with other unrelated things. picojpeg is not thread safe because of this.

A close variant of picojpeg has been successfully compiled and executed on Microchip's PIC18F4610 microcontroller using SourceBoost Technologies BoostC embedded compiler. (Please email if you would like to see this variant.) picojpeg.c has also been used on the MSP430 and ARM Cortex-M4 CPU's. (For an example usage, see Using picojpeg library on a PIC with ILI9341 320x240 LCD module).

The source distribution includes a sample VS2005 project and precompiled Win32/Win64 command line executables that convert JPG to TGA files using picojpeg for decompression. Sean Barrett's public domain stb_image.c module is used to write TGA files.

picojpeg was originally based off my jpgd decompressor C++ class, which (on modern CPU's) is faster and more capable than picojpeg.c but uses a lot more memory and assumes int's are 32-bits.

Here's picojpeg working on a 6809 CPU (a Tandy Color Computer 3), compiled using gcc6809: picojpeg: Decoding Lena on a Tandy Color Computer 3.

Release History

  • v1.1 - 3/23/2020: Fixed unsigned/signed issue in macro, fixed overflow issue in decoder preventing the decode of very large JPEG's

  • v1.1 - Feb. 19, 2013: Dual licensed as both public domain and (where public domain is not acceptable) the MIT license. Please contact me for the source drop.

  • v1.1 - Feb. 9, 2013: Optimized the IDCT row/col loops to avoid the full inverse transform when only the DC component is non-zero, added "reduce" mode for fast 1/8th res decoding, better error handling, added support for H2V1/H1V2 chroma subsampling factors, ported jpg2tga.cpp to jpg2jpg.c (so all modules are written in plain C now), added code to compare picojpeg's decoded output vs. stb_image'c for testing/verification.

  • v1.0 - Nov. 10, 2010: Initial release. Derived from the original version which was tested on a PIC18F series CPU.

Special Thanks

Thanks to Daniel Swenson swenson@ksu.edu for contributing to picojpeg's development, and Chris Phoenix cphoenix@gmail.com for his MSP430 patches.

For any questions or problems with this module please contact Rich Geldreich at richgel99 at gmail.com. Here's my twitter page.