Use NNVM memory when the codec is CODE_RAW
panda9095 opened this issue · comments
Hi,
Enabling NNVM memory did not bring performance gain when the codec is CODE_RAW.
Specifically, the pipeline is as follows: v4l2src device=/dev/video0 ! video/x-raw, format=(string)YUY2, width=(int)1920, height=(int)1080 ! appsink name=mysink. With that pipeline, gst_caps_features_contains(gstCapsFeatures, GST_CAPS_FEATURE_MEMORY_NVMM)) returns false, then the CPU copy from gstBuffer is used, rather than NNVM memory.
I am wondering if we can use NNVM memory when the codec is CODE_RAW?
Thx
@panda9095 NVVM isn't used with uncompressed/raw v4l2src because the image was never in NVMM memory to begin with (because there was no decoder used or no hardware MIPI CSI video capture used). v4l2src element is not capable of outputting NVMM memory. Hence there is always going to be one memory copy at some point.
@dusty-nv Understood.
I am using jetson-utils for V4L2 camera capturing (1080p@30fps) on a Jetson Xavier, which works in a MAXN mode. Typically the memcpy cost only 10+ms to copy 6MB data from gstBuffer, but the cost time can go to 45ms in some cases. Is there any way to improve the copy peformance?
Thx.
Hmm, how frequently does that occur? Since you are already in MAX-N, have you tried running jetson_clocks too?
@dusty-nv Yes, the jetson_clocks is on too.
I have two Xaviers whose copy peformance differs a lot. Maybe it is because of the power supply? The xavier in office shows stable copy performance (about 10+ms), but the other xavier which is on a AGV reports the slow-copy issue. Here is some log:
[DEBUG] [1682327771.625429631]: gst2buffer cpy cost 14 [DEBUG] [1682327771.657340061]: gst2buffer cpy cost 13 [DEBUG] [1682327771.690871921]: gst2buffer cpy cost 13 [DEBUG] [1682327771.724910560]: gst2buffer cpy cost 14 [DEBUG] [1682327771.757805779]: gst2buffer cpy cost 14 [DEBUG] [1682327771.792562600]: gst2buffer cpy cost 15 [DEBUG] [1682327771.870543572]: gst2buffer cpy cost 54 [DEBUG] [1682327771.900762552]: gst2buffer cpy cost 18 [DEBUG] [1682327771.919667236]: gst2buffer cpy cost 17 [DEBUG] [1682327771.939682410]: gst2buffer cpy cost 18 [DEBUG] [1682327771.960783371]: gst2buffer cpy cost 16 [DEBUG] [1682327771.994610735]: gst2buffer cpy cost 16 [DEBUG] [1682327772.024462719]: gst2buffer cpy cost 14 [DEBUG] [1682327772.068030729]: gst2buffer cpy cost 21 [DEBUG] [1682327772.092950387]: gst2buffer cpy cost 15 [DEBUG] [1682327772.124845073]: gst2buffer cpy cost 14 [DEBUG] [1682327772.174974454]: gst2buffer cpy cost 30 [DEBUG] [1682327772.193428458]: gst2buffer cpy cost 14 [DEBUG] [1682327772.226540136]: gst2buffer cpy cost 15 [DEBUG] [1682327772.260683036]: gst2buffer cpy cost 16 [DEBUG] [1682327772.292891466]: gst2buffer cpy cost 15 [DEBUG] [1682327772.324445366]: gst2buffer cpy cost 14 [DEBUG] [1682327772.367985181]: gst2buffer cpy cost 24 [DEBUG] [1682327772.391454395]: gst2buffer cpy cost 14 [DEBUG] [1682327772.425777401]: gst2buffer cpy cost 15 [DEBUG] [1682327772.460853952]: gst2buffer cpy cost 16 [DEBUG] [1682327772.495665176]: gst2buffer cpy cost 17 [DEBUG] [1682327772.528703026]: gst2buffer cpy cost 16 [DEBUG] [1682327772.579597664]: gst2buffer cpy cost 35 [DEBUG] [1682327772.604195994]: gst2buffer cpy cost 23 [DEBUG] [1682327772.631831989]: gst2buffer cpy cost 21 [DEBUG] [1682327772.682291244]: gst2buffer cpy cost 19 [DEBUG] [1682327772.709569652]: gst2buffer cpy cost 16 [DEBUG] [1682327772.736462536]: gst2buffer cpy cost 18 [DEBUG] [1682327772.775740909]: gst2buffer cpy cost 31 [DEBUG] [1682327772.802227499]: gst2buffer cpy cost 20 [DEBUG] [1682327772.826399534]: gst2buffer cpy cost 15 [DEBUG] [1682327772.859011890]: gst2buffer cpy cost 14 [DEBUG] [1682327772.893431605]: gst2buffer cpy cost 15 [DEBUG] [1682327772.926534610]: gst2buffer cpy cost 15 [DEBUG] [1682327772.959032208]: gst2buffer cpy cost 15 [DEBUG] [1682327772.991639507]: gst2buffer cpy cost 14 [DEBUG] [1682327773.053437700]: gst2buffer cpy cost 36 [DEBUG] [1682327773.101971540]: gst2buffer cpy cost 45 [DEBUG] [1682327773.143257060]: gst2buffer cpy cost 38 [DEBUG] [1682327773.194106096]: gst2buffer cpy cost 38 [DEBUG] [1682327773.216031708]: gst2buffer cpy cost 17 [DEBUG] [1682327773.238165427]: gst2buffer cpy cost 19 [DEBUG] [1682327773.260354733]: gst2buffer cpy cost 21 [DEBUG] [1682327773.286129829]: gst2buffer cpy cost 17 [DEBUG] [1682327773.306565762]: gst2buffer cpy cost 19 [DEBUG] [1682327773.334529295]: gst2buffer cpy cost 20 [DEBUG] [1682327773.359945620]: gst2buffer cpy cost 16 [DEBUG] [1682327773.394564418]: gst2buffer cpy cost 17 [DEBUG] [1682327773.427492609]: gst2buffer cpy cost 17 [DEBUG] [1682327773.460830334]: gst2buffer cpy cost 16 [DEBUG] [1682327773.494362884]: gst2buffer cpy cost 17 [DEBUG] [1682327773.528675441]: gst2buffer cpy cost 16 [DEBUG] [1682327773.561738431]: gst2buffer cpy cost 17 [DEBUG] [1682327773.591849627]: gst2buffer cpy cost 14 [DEBUG] [1682327773.625788597]: gst2buffer cpy cost 15 [DEBUG] [1682327773.659019724]: gst2buffer cpy cost 15 [DEBUG] [1682327773.693013897]: gst2buffer cpy cost 15 [DEBUG] [1682327773.726197821]: gst2buffer cpy cost 15 [DEBUG] [1682327773.762367303]: gst2buffer cpy cost 16 [DEBUG] [1682327773.792656715]: gst2buffer cpy cost 15 [DEBUG] [1682327773.827684284]: gst2buffer cpy cost 17 [DEBUG] [1682327773.860415066]: gst2buffer cpy cost 16 [DEBUG] [1682327773.892839400]: gst2buffer cpy cost 15 [DEBUG] [1682327773.928279725]: gst2buffer cpy cost 17 [DEBUG] [1682327773.962264074]: gst2buffer cpy cost 17
but the other xavier which is on a AGV reports the slow-copy issue
I'm not sure if the Xavier on your AGV is inside some enclosure or housing, but I wonder how the temperatures are and if some degree of thermal throttling is occurring? You might want to try keeping an eye on the clocks / system temps with tegrastats or jtop.