int main(int argc, char ** argv){
int i;
if( (argc>=2) && (atoi(argv[1])!=RANK)) error("rank %d mandatory",RANK);
printf("CUDA RANK=%d\n",RANK);
kernel.print();
// build busylist
busylist = (uint32_t*)malloc_file(CNK*sizeof(uint32_t),FMODE_RO,BLIST_FORMAT,RANK);
// put busylist
SafeCall(cuMemHostRegister(busylist,CNK*sizeof*busylist,CU_MEMHOSTREGISTER_DEVICEMAP));
SafeCall(cuMemHostGetDevicePointer(&host_busylist,busylist,0));
SafeCall(cuModuleGetGlobal(&dev_busylist,&bytes,kernel.module[0].module,"busylist"));
if(bytes!=sizeof(host_busylist)) error("busylist!");
SafeCall(cuMemcpyHtoD(dev_busylist,&host_busylist,bytes));
// put array
#ifdef IN_mk_data
mkdir(DATADIR,0755); errno=0;
array = (unsigned char *)malloc_file(abytes(RANK,CNK),1,DATADIR"%d",RANK);
#else
array = (unsigned char *)malloc_file(abytes(RANK,CNK),0,DATADIR"%d",RANK);
#endif
SafeCall(cuMemHostRegister(array,abytes(RANK,CNK),CU_MEMHOSTREGISTER_DEVICEMAP));
SafeCall(cuMemHostGetDevicePointer(&host_array,array,0));
SafeCall(cuModuleGetGlobal(&dev_array,&bytes,kernel.module[0].module,"array"));
if(bytes!=sizeof(host_array)) error("array!");
SafeCall(cuMemcpyHtoD(dev_array,&host_array,bytes));
#define THREADS 512
#define MAXG 65535
uint64_t nado = (cnk[RANK] +(THREADS-1))/THREADS;
uint32_t gridx = nado>MAXG?MAXG:nado;
uint32_t gridy = (nado+(MAXG-1))/MAXG;
printf("gridy=%d gridx=%d THREAD=%d\n",gridy, gridx, THREADS);
kernel.launch(params,THREADS,gridx,gridy);
kernel.wait();
SafeCall(cuMemHostUnregister(busylist));
SafeCall(cuMemHostUnregister(array));
SafeCall(cuModuleGetGlobal(&dev_changed,&bytes,kernel.module[0].module,"changed"));
if(bytes!=sizeof(changed)) error("changed!");
SafeCall(cuMemcpyDtoH(changed,dev_changed,bytes));
for(i=0;i<CACHESIZE;i++)
total += changed[i];
printf("changed=%ju\n",total);
return 0;
}
SEXP R_auto_cuMemHostRegister(SEXP r_p, SEXP r_bytesize, SEXP r_Flags)
{
SEXP r_ans = R_NilValue;
void * p = GET_REF(r_p, void );
size_t bytesize = REAL(r_bytesize)[0];
unsigned int Flags = REAL(r_Flags)[0];
CUresult ans;
ans = cuMemHostRegister(p, bytesize, Flags);
r_ans = Renum_convert_CUresult(ans) ;
return(r_ans);
}
int * transpose(int *matrix, int height, int width) {
if (cuda.get() == NULL) {
cuda.reset(new Cuda());
}
int *result;
result = new int[height*width];
if (width >= BLOCK_DIM_X && height >= BLOCK_DIM_X) {
// gpu_transpose_naive(result, matrix, height, width);
cuMemHostRegister(result, sizeof(int)*height*width, 0); // TODO check result value
gpu_transpose_with_shared_mem(result, matrix, height, width);
cuMemHostUnregister(result); // TODO check result value
} else {
cpu_transpose(result, matrix, height, width);
}
return result;
}
/**
* Call the CUDA register function so we pin the memory in the CUDA
* space.
*/
void mca_common_cuda_register(void *ptr, size_t amount, char *msg) {
int res;
if (!initialized) {
mca_common_cuda_init();
}
if (mca_common_cuda_enabled && mca_common_cuda_register_memory) {
res = cuMemHostRegister(ptr, amount, 0);
if (res != CUDA_SUCCESS) {
/* If registering the memory fails, print a message and continue.
* This is not a fatal error. */
orte_show_help("help-mpi-common-cuda.txt", "cuMemHostRegister failed",
true, ptr, amount, res, msg);
} else {
opal_output_verbose(20, mca_common_cuda_output,
"CUDA: cuMemHostRegister OK on mpool %s: "
"address=%p, bufsize=%d",
msg, ptr, (int)amount);
}
}
}
cl_int
pocl_cuda_alloc_mem_obj (cl_device_id device, cl_mem mem_obj, void *host_ptr)
{
cuCtxSetCurrent (((pocl_cuda_device_data_t *)device->data)->context);
CUresult result;
void *b = NULL;
/* if memory for this global memory is not yet allocated -> do it */
if (mem_obj->device_ptrs[device->global_mem_id].mem_ptr == NULL)
{
cl_mem_flags flags = mem_obj->flags;
if (flags & CL_MEM_USE_HOST_PTR)
{
#if defined __arm__
// cuMemHostRegister is not supported on ARN
// Allocate device memory and perform explicit copies
// before and after running a kernel
result = cuMemAlloc ((CUdeviceptr *)&b, mem_obj->size);
CUDA_CHECK (result, "cuMemAlloc");
#else
result = cuMemHostRegister (host_ptr, mem_obj->size,
CU_MEMHOSTREGISTER_DEVICEMAP);
if (result != CUDA_SUCCESS
&& result != CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED)
CUDA_CHECK (result, "cuMemHostRegister");
result = cuMemHostGetDevicePointer ((CUdeviceptr *)&b, host_ptr, 0);
CUDA_CHECK (result, "cuMemHostGetDevicePointer");
#endif
}
else if (flags & CL_MEM_ALLOC_HOST_PTR)
{
result = cuMemHostAlloc (&mem_obj->mem_host_ptr, mem_obj->size,
CU_MEMHOSTREGISTER_DEVICEMAP);
CUDA_CHECK (result, "cuMemHostAlloc");
result = cuMemHostGetDevicePointer ((CUdeviceptr *)&b,
mem_obj->mem_host_ptr, 0);
CUDA_CHECK (result, "cuMemHostGetDevicePointer");
}
else
{
result = cuMemAlloc ((CUdeviceptr *)&b, mem_obj->size);
if (result != CUDA_SUCCESS)
{
const char *err;
cuGetErrorName (result, &err);
POCL_MSG_PRINT2 (__FUNCTION__, __LINE__,
"-> Failed to allocate memory: %s\n", err);
return CL_MEM_OBJECT_ALLOCATION_FAILURE;
}
}
if (flags & CL_MEM_COPY_HOST_PTR)
{
result = cuMemcpyHtoD ((CUdeviceptr)b, host_ptr, mem_obj->size);
CUDA_CHECK (result, "cuMemcpyHtoD");
}
mem_obj->device_ptrs[device->global_mem_id].mem_ptr = b;
mem_obj->device_ptrs[device->global_mem_id].global_mem_id
= device->global_mem_id;
}
/* copy already allocated global mem info to devices own slot */
mem_obj->device_ptrs[device->dev_id]
= mem_obj->device_ptrs[device->global_mem_id];
return CL_SUCCESS;
}
