Linux/PROCESSOR-SDK-AM57X: OpenCL+DSP accelerate

Part Number:PROCESSOR-SDK-AM57X

Tool/software: Linux

Hi Ti，

I am doing OpenCV+DSP accelerate on AM5718 board，SDK version is ti-processor-sdk-linux-rt-am57xx-evm-04.01.00.06.

I found UMat convert to Mat takes too much time。

Here is my test code and result：

code：

int main(int argc, char ** argv)

{

unsigned char *yuv_raw_data = NULL;
int c = -1;
struct timespec tp0, tp1, tp2, tp3, tp4, tp5;

cameraOpenDevice(1);

Mat mat_yuv(CAMERA_HEIGHT , CAMERA_WIDTH, CV_8UC2);
Mat mat_rgb(CAMERA_HEIGHT , CAMERA_WIDTH, CV_8UC3);

UMat umat_yuv(CAMERA_HEIGHT , CAMERA_WIDTH, CV_8UC2);
UMat umat_rgb(CAMERA_HEIGHT , CAMERA_WIDTH, CV_8UC3);

while(1) {
clock_gettime(CLOCK_MONOTONIC, &tp0);
DqBuffer(&yuv_raw_data);//yuv raw data from V4l2 camera
QBuffer();
memcpy(mat_yuv.data, yuv_raw_data, CAMERA_WIDTH*CAMERA_HEIGHT*2);

clock_gettime(CLOCK_MONOTONIC, &tp1);
mat_yuv.copyTo(umat_yuv);

clock_gettime(CLOCK_MONOTONIC, &tp2);

cvtColor(umat_yuv, umat_rgb, CV_YUV2BGR_YUYV);

clock_gettime(CLOCK_MONOTONIC, &tp3);

umat_rgb.copyTo(mat_rgb);
clock_gettime(CLOCK_MONOTONIC, &tp4);

imshow("frame", mat_rgb);
clock_gettime(CLOCK_MONOTONIC, &tp5);

printf("get yuv Mat tdiff=%lf ms \n", tdiff_calc(tp0, tp1));
printf("Mat2UMat tdiff=%lf ms \n", tdiff_calc(tp1, tp2));
printf("cvtColor_YUV2BGR tdiff=%lf ms \n", tdiff_calc(tp2, tp3));
printf("UMat2Mat tdiff=%lf ms \n", tdiff_calc(tp3, tp4));
printf("imshow tdiff=%lf ms \n", tdiff_calc(tp4, tp5));

c = waitKey(1);
if( c == 27 || c == 'q' || c == 'Q' )
break;
}

cameraCloseDevice();
return 0;

}

Then I compile it to an exec bin:  dsp_accelerate_opencv_opencl

result ：

root@ok5718-idk:/home/forlinx/qt# on_dsp_accelerate.sh

get yuv Mat tdiff=5.017779 ms
Mat2UMat tdiff=0.672464 ms
cvtColor_YUV2BGR tdiff=0.165757 ms
UMat2Mat tdiff=151.863810 ms
imshow tdiff=1.787381 ms

root@ok5718-idk:/home/forlinx/qt# off_dsp_accelerate.sh

get yuv Mat tdiff=0.509635 ms
Mat2UMat tdiff=0.479541 ms
cvtColor_YUV2BGR tdiff=11.217494 ms
UMat2Mat tdiff=5.070482 ms
imshow tdiff=2.764032 ms

root@ok5718-idk:/home/forlinx/qt# cat on_dsp_accelerate.sh
export TI_OCL_CACHE_KERNELS=Y
export OPENCV_OPENCL_DEVICE='TI AM57:ACCELERATOR:TI Multicore C66 DSP'
echo "OpenCL on"
./dsp_accelerate_opencv_opencl

root@ok5718-idk:/home/forlinx/qt# cat off_dsp_accelerate.sh
export OPENCV_OPENCL_DEVICE='disabled'
echo "OpenCL off"
./dsp_accelerate_opencv_opencl

The result shows UMat convert to Mat takes too much time when dsp accelerate is working.

I Found TI docs , http://software-dl.ti.com/processor-sdk-linux/esd/docs/latest/linux/Foundational_Components_OpenCV.html#alternative-approach-to-add-new-opencl-kernels-at-opencv-application-level :

It said the UMat cost time includes: waiting time (for DSP to finish) plus actual write and any format conversion (done on CPU). It also depends a lot on data types used, and if floating point operations are involved. This can be accelerated if DSP optimized implementation of remap() is created.

So, my question is：How to optimized remap() with DSP accelerate。

I read the Chapters Creating OpenCL C kernel optimized for C66 core to Alternative approach to add new OpenCL kernels at OpenCV application level， but I found no way to do.

Please give me help, thank you.