Weird malloc problem

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi guys,

So I am having an issue with the order in which I malloc my variables and being able to access them. Specifically for a code like:

[codebox] cudaMalloc((void**)&kxihlG,nElkxihl*sizeof(float));

    cudaMalloc((void**)&kxihrG,nElkxihr*sizeof(float));

    cudaMemcpy(kxihlG, kxihl, nElkxihl*sizeof(float), cudaMemcpyHostToDevice);

    cudaMemcpy(kxihrG, kxihr, nElkxihr*sizeof(float), cudaMemcpyHostToDevice);

.

.

. lots of variables

cudaMalloc((void**)&TARGETVAR,nElTARGETVAR*sizeof(float));

    cudaMemcpy(TARGETVARG, TARGETVAR, nElTARGETVAR*sizeof(float), cudaMemcpyHostToDevice);

      test<<<1,1>>>(TARGETVARG);

      Error = cudaThreadSynchronize();

      fprintf(stderr,"@TEST1 Error = %d \n",Error);[/codebox]

if I run it like this cudaThreadSynchronize returns a failure of ‘4’, but if move the cudaMalloc for TARGETVAR to the top of the list then the test<<<>>> kernel runs succesfuly:

THIS WORKS

[codebox]

cudaMalloc((void**)&TARGETVAR,nElTARGETVAR*sizeof(float));

cudaMalloc((void**)&kxihlG,nElkxihl*sizeof(float));

    cudaMalloc((void**)&kxihrG,nElkxihr*sizeof(float));

    cudaMemcpy(kxihlG, kxihl, nElkxihl*sizeof(float), cudaMemcpyHostToDevice);

    cudaMemcpy(kxihrG, kxihr, nElkxihr*sizeof(float), cudaMemcpyHostToDevice);

.

.

. lots of variables

cudaMemcpy(TARGETVARG, TARGETVAR, nElTARGETVAR*sizeof(float), cudaMemcpyHostToDevice);

      test<<<1,1>>>(TARGETVARG);

      Error = cudaThreadSynchronize();

      fprintf(stderr,"@TEST1 Error = %d \n",Error);[/codebox]

Can anyone explain this? I dont think the machine is full? Do I have to put some delay in? why does the order matter as long as malloc for a given variable is before the memcpy for it?

How can this be fixed because this problem is occuring elsewhere with other variable? Thank you for your time!

#2

Did you think something happens with TARGETVAR?

why dont you debug the error of that statement?
Error = cudaMalloc((void**)&TARGETVAR,nElTARGETVAR*sizeof(float));
printf(“CUDA Error: %s\n”, cudaGetErrorString(Error));

#3

If nElTARGETVAR is especially large, then it could be a classic problem of memory address space packing. It’s harder to allocate a contiguous large chunk than smaller chunks (which can fit into “cracks” better). This isn’t a GPU problem, it happens all the time on the CPU too.

Classic answer: always allocate from large to small. Do this as a reflex in all your coding on GPU and CPU.
When this fails, the next strategy is usually to start being fancier with your memory allocator, reusing old mallocs, etc, and/or changing your algorithm to need less large contiguous blocks.