cudaMallocPitch error!!! about cudasift HOT 4 CLOSED

I feel you use the wrong '-arch=sm_35' to compile. Check your GPU and modify the Makefile accordingly.

from cudasift.

thanks for the reply.... the '-arch=sm_35' was set appropriately in the settings. I'm trying to implement your cuda kernels on a large dataset... so the kernels have to be called inside a loop. Ideally I'd avoid using cuda kernels inside a loop but I have no choice. I have solved the problem but the solution seems a little unusual. I had to strictly follow the following steps

here is how
img_1_host = cv::Mat::zeros(IMG_1_HOST.rows, IMG_1_HOST.cols,CV_32FC1); // STEP1
SiftData siftDataR; // STEP2
InitSiftData(siftDataR, 65535, true, true); // STEP3
CudaImage IMG_1_DEV;
// declaring cudaimage inside the loop solves the 'OUT OF MEMORY PROBLEM'
cv::cvtColor(IMG_1_HOST,img_1_host,CV_BGR2GRAY); // STEP4
img_1_host.convertTo(img_1_host,CV_32FC1); // STEP5
.......

from cudasift.

Hi @avcs2080 ,
I have got out of memory error when I try to MatchSiftData() of nearly 20000 features from each image. Does it related to memory allocation error.
Thanks

from cudasift.

@rremani
It is hard to say without looking at the code. did you initialize SiftData with 20000 values prior to using it i.e.
something like
SiftData siftDataA;
InitSiftData(siftDataA, 20000, true, true);
In my case, I had nested for loop as discussed above. When I declared CudaImage outside the nested loop, surprisingly I got out of memory error. Declaring it inside the loop somehow got rid of the error. I'm assuming the scope of the variable within the loop might have prevented it. Also, don't forget to free the memory at the bottom.

hope this helps

from cudasift.