We have finished a project,called "Vehicle nighttime infrared pedestrian detection system" on DM6437.
the scheme of our project is based on video_preview, just adding viedo process algorithm between video-capture and video_display, like this
/* loop forever performing video capture and display */
while ( status == 0 ) {
/* grab a fresh video input frame */
FVID_exchange(hGioVpfeCcdc, &frameBuffPtr);
/*video process algorithm, frameBuffPtr->frame.frameBufferPtr contain current frame data*/
detect_pedestrian(frameBuffPtr->frame.frameBufferPtr);
/*observe the index that contains current frame data*/ */
printf("%x\n", (void*)(frameBuffPtr->frame.frameBufferPtr));
fflush(stdout);
/* display the video frame */
FVID_exchange(hGioVpbeVid0, &frameBuffPtr);
}
Then my task is to optimize this system from DMA and Cache. I have done some experiments, but still have questions to solve.
1、 the size of one frame image is 720*576*2 bytes, about 829.44Kb, absolutely larger than L1 Memory and L2 Memory.
So we must have Video Frame buffers in DDR2 rather than L1 Memory or L2 Memory, though the speed of DDR2 is much slower .
I had a experiment ,.By enable MAR128-MAR159, as a result ,DDR2 is cacheable. The performance of above loop is proved to be about 8 times improvement .
I wonder if DMA can be used here to improve further? Maybe DMA and Cache should cooperate with each other, can this mechanism make a difference?
2、The VPFE/ VPBE drivers maintain two circular queues called READY queue and a FREE queue for each
channel of display and capture.
/* prime up the video capture channel */
if (status == 0) {
FVID_queue(hGioVpfeCcdc, frameBuffTable[0]);
FVID_queue(hGioVpfeCcdc, frameBuffTable[1]);
FVID_queue(hGioVpfeCcdc, frameBuffTable[2]);
}
/* prime up the video display channel */
if (status == 0) {
FVID_queue(hGioVpbeVid0, frameBuffTable[3]);
FVID_queue(hGioVpbeVid0, frameBuffTable[4]);
FVID_queue(hGioVpbeVid0, frameBuffTable[5]);
}
Above are six frameBuff used in FVID_exchange(hGioVpfeCcdc, &frameBuffPtr); and FVID_exchange(hGioVpbeCcdc, &frameBuffPtr);
I observe the frameBuffPtr->frame.frameBufferPtr that contains current frame data .and the result is in following Stdout window. They are circluar every six loop. The address and order are stable for new loops.
800cac00
80195400
80000380
80195400
8025fc80
8032a480
800cac00
80195400
80000380
80195400
8025fc80
8032a480
I attempted to use DMA to move several lines of the whole frame from DDR2 to L2 SRAM. So I use frameBuffPtr->frame.frameBufferPtr as the souce address of DMA transfer. And a buffer in L2 SRAM as the destination of DMA tranfer
If I want to create another task in bios to accomplish the DMA transfer. What's the priority of two tasks?
DMA transfer should start automaticly as soon as video capture is finished. How should I know the moment?
DMA transfer should also start automaticly as soon as video process is finished. How should I know the moment?
