cudaError_t cudaMalloc(void **devPtr, size_t size)
{
static cudaError_t (*nv_cudaMalloc)(void **, size_t) = NULL;
cudaError_t ret;
struct timeval t;
if(!nv_cudaMalloc) {
nv_cudaMalloc = dlsym(RTLD_NEXT, "cudaMalloc");
//nv_cudaMalloc = dlsym(RTLD_NEXT, "cudaMalloc_v2");
if(!nv_cudaMalloc) {
fprintf(stderr, "failed to find symbol cudaMalloc: %s\n", dlerror());
show_stackframe();
return cudaErrorSharedObjectSymbolNotFound;
}
}
gettimeofday(&t, NULL);
printf("[gvm] %lf intercepting cudaMalloc at %lx\n", t.tv_sec + t.tv_usec / 1000000.0, (unsigned long)devPtr);
do {
ret = nv_cudaMalloc(devPtr, size);
} while (ret != cudaSuccess);
gettimeofday(&t, NULL);
printf("[gvm] %lf intercepted cudaMalloc( %lx %ld ) = %d\n", t.tv_sec + t.tv_usec / 1000000.0, (unsigned long)(*devPtr), size, (int)ret);
return ret;
}
// GMM-specific: allowing passing dptr array hints.
GMM_EXPORT
cudaError_t cudaMallocEx(void **devPtr, size_t size, int flags)
{
if (initialized)
return gmm_cudaMalloc(devPtr, size, flags);
else {
gprint(WARN, "cudaMallocEx called outside GMM\n");
return nv_cudaMalloc(devPtr, size);
}
}
GMM_EXPORT
cudaError_t cudaMalloc(void **devPtr, size_t size)
{
cudaError_t ret;
if (initialized)
ret = gmm_cudaMalloc(devPtr, size, 0);
else {
// TODO: We may need to remember those device memory allocated
// before GMM was initialized, so that later when they are
// used in cudaMemcpy or other functions we can treat them
// specially.
gprint(WARN, "cudaMalloc called outside GMM\n");
ret = nv_cudaMalloc(devPtr, size);
}
return ret;
}
