How to improve DMA copy speeds on the STM32MP257FDK

[Cedricl]

I need the CPU to copy camera frames for tasks like screenshot-based recognition and uploading. However, in the data path (sensor → CSI → DCMIPP → DMA → DDR), the DMA mapping layer marks the memory as uncached, resulting in very slow speeds; the CPU read bandwidth is only ~143 MB/s—just one-eleventh of the speed of standard cached memory (1619 MB/s).

Copying a 1280x960 image from DMA takes around 25ms, and the process becomes even slower at higher resolutions, leading to a poor user experience.

Is there a way to make the DMA copy speed match that of standard cached memory (1619 MB/s)?