Allocate Memory within Device? is it possible?

[glaksmono]

Is it possible?

something like:

CUDA_SAFE_CALL( cudaMalloc((void**) &device_fp, sizeof(FILE)));

in the device

I did this in one of the function of the device

#if DEVICE_EMULATION

if ((support1 == NULL) || (largeItem1 == NULL)) {

printf("out of memory - HERE\n");

exit(1);

}

endif

it gives me the “out of memory” even I’ve allocated memory for the support1 and largeItem1. Hmm…

[glaksmono]

I tried this:

cuInit()

CU_SAFE_CALL(cuMemAlloc(&support1, (sizeof(int) * numItem));

but I got this error:

1>ss.cu(1148) : fatal error C1057: unexpected end of file in macro expansion

what does that mean?

[glaksmono]

I’m not sure, tell me if I’m wrong. It just came up to my mind, since EmuDebug compilation will be run on the CPU, so when I do this on the device

if ((support1 == NULL) || (largeItem1 == NULL)) {

printf("out of memory - HERE\n");

exit(1);

}

even if I’ve allocated memory on the GPU through cudaMemalloc, it prints out “out of memory” because it checks through CPU?

[MisterAnderson42]

cudaMalloc is a host function. It cannot be called from the device.

[glaksmono]

It’s the same with cuMemAlloc, correct?

[Simon_Green]

Yes. You cannot call any CUDA API function from the device.

It is possible to simulate dynamic allocation in CUDA by allocating a big chunk of memory and then allocating sections to threads by using techniques like parallel prefix sum (see the scan example in the SDK).