On Wed, Nov 02, 2016 at 12:34:47PM -0700, Cesar Philippidis wrote:
> @@ -932,9 +933,84 @@ nvptx_exec (void (*fn), size_t mapnum, void **hostaddrs,
> void **devaddrs,
>
> if (seen_zero)
> {
> + /* See if the user provided GOMP_OPENACC_DIM environment
> + variable to specify runtime defaults. */
> + static int default_dims[GOMP_DIM_MAX];
> +
> + if (!default_dims[0])
> + {
Is this guarded by some lock, or is it just racy if multiple
nvptx_execs are done at the same time?
> + /* We only read the environment variable once. You can't
> + change it in the middle of execution. The sytntax is
syntax
> + the same as for the -fopenacc-dim compilation option. */
> + const char *env_var = getenv ("GOMP_OPENACC_DIM");
> +
> + if (CUDA_SUCCESS == cuDeviceGetAttribute
> + (&block_size, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, dev)
> + && CUDA_SUCCESS == cuDeviceGetAttribute
> + (&warp_size, CU_DEVICE_ATTRIBUTE_WARP_SIZE, dev)
> + && CUDA_SUCCESS == cuDeviceGetAttribute
> + (&dev_size, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, dev)
> + && CUDA_SUCCESS == cuDeviceGetAttribute
> + (&cpu_size, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR,
> dev))
The formatting is wrong. 1) you should use the call should be on lhs of ==,
not rhs 2) ( should be after cuDeviceGetAttribute, not on the next line
3) still the lines are too long.
if (cuDeviceGetAttribute (&block_size,
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK,
dev) == CUDA_SUCCESS
&& cuDeviceGetAttribute (...
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR
is still way too long, perhaps initialize a temporary const var
to that, or use some macro like
DEV_ATTR (MAX_THREADS_PER_MULTIPROCESSOR)
where
#define DEV_ATTR(x) CU_DEVICE_ATTRIBUTE_##x
Otherwise LGTM.
Jakub
