Posted on October 07, 2017 at 14:54You have no caching in the block storage subsystem or file system.
Each command to the SD card has very significant overhead, you want to always read/write multiple sectors and have them aligned with the mega-blocks within the geometry of the array, ie not span across two of them. These are likely to be of the order of 128KB
The file data is highly structured, yet you throw out that knowledge and let the computer do random seeks to byte offsets, and do small unaligned reads. This exposes the worst possible characteristics in the system.
When cards say they are Class 4 or 10, or whatever, they are describing the linear streaming of data in the most efficient manner, not jumping around randomly, or skipping forward 1920 bytes for every other read.
Advancing a memory pointer in a buffer by 1920 bytes has no cost, reading a 4K or 8K block is significantly faster than doing multiple 512 byte sector reads and extracting 1920 bytes that falls somewhere inside 4 or 5 them. Repeat over 1080 lines and the inefficiency scales nicely.
DMA helps if you do something with the time it frees up on the processor rather than spinning in a while() waiting for it to complete. It would help significantly to have one thread reading the data, and a second one consuming/processing it, and have multiple buffers.
