The core has one FIFO between the link and the video output, sized with Output FIFO depth (words) in the IP GUI: 128 to 16,384 words, default 2,048. One word is one output beat of 2 x LANE_NUM pixels. The embedded data output, when built, has a second FIFO of its own.
2.9.1NATIVE: the longest line
Native video replays each line in one burst once it has arrived, so the buffer must hold a whole line. 32 words are reserved for flow control.
| Depth | 1 lane | 2 lanes | 4 lanes | 6 lanes | 8 lanes |
|---|---|---|---|---|---|
| 128 | 192 | 384 | 768 | 1,152 | 1,536 |
| 256 | 448 | 896 | 1,792 | 2,688 | 3,584 |
| 512 | 960 | 1,920 | 3,840 | 5,760 | 7,680 |
| 1,024 | 1,984 | 3,968 | 7,936 | 11,904 | 15,872 |
| 2,048 (default) | 4,032 | 8,064 | 16,128 | 24,192 | 32,256 |
| 4,096 | 8,128 | 16,256 | 32,512 | 48,768 | 65,024 |
| 8,192 | 16,320 | 32,640 | 65,280 | 97,920 | 130,560 |
| 16,384 | 32,704 | 65,408 | 130,816 | 196,224 | 261,632 |
A full-width IMX530 line (5328 pixels) needs 1,332 beats on 2 lanes, so 2,048 words; on 8 lanes it needs 333 beats and 512 words are enough. 1920-pixel lines on 2 lanes fit exactly in 512 words; 1,024 leaves room for other modes.
2.9.2AXI4S: how long the sink may stall
The AXI4-Stream FIFO never has to hold a line. It sets how long m_axis_tready may stay low
before a frame is cut. The lane count drops out: more lanes deliver more pixels per clock, but each word
holds more pixels too.
| Depth | RAW8 | RAW10 | RAW12 | RAW14 | RAW16 |
|---|---|---|---|---|---|
| 128 | 0.51 | 0.63 | 0.76 | 0.88 | 1.0 |
| 256 | 1.0 | 1.3 | 1.6 | 1.8 | 2.1 |
| 512 | 2.1 | 2.7 | 3.2 | 3.7 | 4.2 |
| 1,024 | 4.3 | 5.3 | 6.4 | 7.5 | 8.6 |
| 2,048 (default) | 8.6 | 10.7 | 12.9 | 15.0 | 17.2 |
| 4,096 | 17.2 | 21.5 | 25.8 | 30.1 | 34.4 |
| 8,192 | 34.4 | 43.1 | 51.7 | 60.3 | 68.9 |
| 16,384 | 68.9 | 86.2 | 103 | 121 | 138 |
A VDMA or a pipeline that is always ready needs little buffering: 512 to 2,048 words is comfortable.
A sink that stalls for a known time, such as a frame buffer writer behind a busy memory controller, gets
that time plus margin. A deeper FIFO costs block RAM only; it is first-word fall-through and adds no
latency while tready is high. The stream clock only has to carry the average beat rate of
the video; within a line the FIFO smooths the difference.
2.9.3Embedded data FIFO
Sized with EBD FIFO depth (words): 32 to 16,384 words, default 1,024. It holds the embedded data lines of one frame while the sink is busy. A sink that reads them only in the frame blanking needs
Two embedded data lines of 3,840 pixels on 4 lanes are 960 words, so 1,024. A sink that keeps up while the lines arrive needs only the minimum of 32.