Using CMSIS-DSP to Accelerate JPEG Decoding: Limitations and Alternatives

[JiapengLi]

I am maintaining a lightweight JPEG decoding library and would like to leverage the CMSIS-DSP library to improve performance.

However, based on the CMSIS-DSP documentation (https://arm-software.github.io/CMSIS-DSP/latest/group__DCT4__IDCT4.html
), it appears that only DCT type-IV is supported.

Is it possible to use CMSIS-DSP to accelerate the JPEG decoding process despite this limitation?
Are there alternative methods to implement DCT type-II or III using CMSIS-DSP or other ARM-optimized routines?

[christophe0606]

@JiapengLi Unfortunately, the IDCT in CMSIS-DSP had been marked as deprecated for a long time and have finally been removed. They were old legacy code and no more maintained.

The only functions that may help you are the real FFTs. But the function is generic (for any power of 2 size) so may be less efficient than a function made for a specific size.

Also warning when using RFTT:

• Use only initialization function specific to a size (not the generic one). it is for code size
• For legacy reason, the RFFT F32 is packing the real Nyquist frequency value in the imaginary part of the first bin value. It is to have an output buffer of N samples rather than N+1. It is not very clean. So you'll have to unpack this value. (It is explained in documentation).

Functions are provided to get the lengths of different buffer involved.
(The CMSIS-Pack is a bit old and does not yet contain those new functions so use the github version. CMSIS-Pack will be updated)