How to pass an already stored data in GPU by GpuMat to a kernel

asked 2017-07-18 06:34:34 -0500

arrfou gravatar image


I have a GpuMat of OpenCV type created as

cv::cuda::GpuMat d_im(h_im.size().height, h_im.size().width, CV_8UC1);

that I am doing some image processing operations using opencv::cuda and then according to the OpenCV documentation I have tried to pass it directly to the kernel function below as :

Kernel_func<<<grid_size, block_size, 0, stream>>>( d_im.ptr<uint8_t>(), output);

but I got wrong results.

However, it was okay if the new d_im is downloaded from GPU to CPU and then copy it again to the GPU by cudaMemcpy as in this code snippet below (with no problems). I know this is not okay to do.

CUDA_CHECK_RETURN(cudaMemcpyAsync(input, h_im_new.ptr<uint8_t>(), sizeof(uint8_t)*size, cudaMemcpyHostToDevice,stream1));

My global function prototype is :

__global__ void Kernel_func(const uint8_t *input, const uint8_t *output);

I am not sure what is wrong in this case, please anyone had similar issue or any suggestions. Thanks for your help

edit retag flag offensive close merge delete


This is my first post here; I made an account just to comment on this! I too am observing this exact behavior, I have tried every possible way of passing the GpuMat to the kernel as a float* pointer, but regardless only the first third or so of the image (roughly 1500 bytes) is valid, the rest is zero.

I tried: Allocating a local GpuMat and using copyTo() to copy the original GpuMat data manually there. *Allocate a float with cudaMalloc(), and using cudaMemcpy(...DeviceToDevice) to store the raw data. *Using both gpumat.ptr<float>(); and functions.

The only thing that works is copying the data to Host memory, then back to a GpuMat using, then using cudaMemcpy(...HostToDevice) to copy it to a float*. I can only conclude that this is a bug with OpenCV!

patrobic gravatar imagepatrobic ( 2017-07-27 16:08:44 -0500 )edit

As the documentation says

In contrast with Mat, in most cases GpuMat::isContinuous() == false . This means that rows are aligned to a size depending on the hardware. Single-row GpuMat is always a continuous matrix.
Tetragramm gravatar imageTetragramm ( 2017-07-27 17:49:21 -0500 )edit