Extremely slow memcpy to mmap shared region

[paul.doucet]

Please provide complete information as applicable to your setup.

**• Jetson Xavier **
• Deepstream version 6.3
• Jetpack version 4.6.2
**• TensorRT Version 8.5.1.7 **
• Issue Type bugs

Hi,
I’m trying to transfer a frame extracted from the deepstream pipeline to shared memory (mmap) via memcpy, for this I do the following:

// Initialize the shared memory with mmap
char name[25];
sprintf(name, "/shared_mem%zu", 0);
total_size = 1500000

int fd = shm_open(name, O_RDWR | O_CREAT, 0644);

ftruncate(fd, total_size);

void* shared_mem = mmap(NULL, total_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);


// Extract the raw frame from the pipeline
NvBufSurfaceMap (surface, -1, -1, NVBUF_MAP_READ);
NvBufSurfaceSyncForCpu (surface, 0, 0);
NvBufSurfaceParams params = surface->surfaceList[i];
  
guint height = surface->surfaceList[0].height;
guint width = surface->surfaceList[0].width;
guint pitch = surface->surfaceList[0].pitch;
  
// Put the extracted image into an OpenCV datastructure
void* addr = surface->surfaceList[0].mappedAddr.addr[0];
uchar * arr = BGR_mat.isContinuous() ? BGR_mat.data : BGR_mat.clone().data;
uint length = BGR_mat.total() * BGR_mat.channels();
  
cv::Mat BGR_mat = cv::Mat(height, width, CV_8UC1, addr, pitch);

// This memcpy is extremly slow (2s for 1Mb so around 0.5Mb/s)
memcpy(shared_mem + offset, (const void*) arr, length);

I also already put the Xavier into performance mode, so it shouldn’t be causing the problem,

Any ideas on what could be causing this?

[Fiona.Chen]

What do you mean by " Extremely slow memcpy to mmap shared region" ? How slow?

[paul.doucet]

I takes 2000ms to copy 2Mb

[paul.doucet]

It doesn’t seem like the problem comes from mmap, using memcpy to copy 1Mb between any two regions seems to also be that slow.

[Fiona.Chen]

Is the content correct after memcpy?

[paul.doucet]

Yes it is

[paul.doucet]

Executing memcpy like this also takes 1.5 seconds for 1Mb

void* addr1 = malloc(1000000);
void* addr2 = malloc(1000000);

// this line takes 1.5 seconds
memcpy(addr2, addr1, 1000000);

[Fiona.Chen]

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

It does not take so much time in our board with your code.

time ./a.out
real 0m0.007s
user 0m0.000s
sys 0m0.004s

#include <stdlib.h>
#include <string.h>int main(int argc, char ** argv)
{
void* addr1 = malloc(1000000);
void* addr2 = malloc(1000000); // this line takes 1.5 seconds
memcpy(addr2, addr1, 1000000);
}