[libgomp, nvptx] Fix too-many-resources fatal error condition and message
2018-04-30 Tom de Vries <tom@codesourcery.com>
* plugin/plugin-nvptx.c (nvptx_exec): Fix
insufficient-resources-to-launch fatal error condition and message.
---
libgomp/plugin/plugin-nvptx.c | 19 +++++++++----------
1 file changed, 9 insertions(+), 10 deletions(-)
@@ -834,16 +834,15 @@ nvptx_exec (void (*fn), size_t mapnum, void **hostaddrs, void **devaddrs,
/* Check if the accelerator has sufficient hardware resources to
launch the offloaded kernel. */
- if (dims[GOMP_DIM_WORKER] > 1)
- {
- if (reg_granularity > 0 && dims[GOMP_DIM_WORKER] > threads_per_block)
- GOMP_PLUGIN_fatal ("The Nvidia accelerator has insufficient resources "
- "to launch '%s'; recompile the program with "
- "'num_workers = %d' on that offloaded region or "
- "'-fopenacc-dim=-:%d'.\n",
- targ_fn->launch->fn, threads_per_block,
- threads_per_block);
- }
+ if (dims[GOMP_DIM_WORKER] * dims[GOMP_DIM_VECTOR]
+ > targ_fn->max_threads_per_block)
+ GOMP_PLUGIN_fatal ("The Nvidia accelerator has insufficient resources to"
+ " launch '%s' with num_workers = %d and vector_length ="
+ " %d; recompile the program with 'num_workers = x and"
+ " vector_length = y' on that offloaded region or "
+ "'-fopenacc-dim=-:x:y' where x * y <= %d.\n",
+ targ_fn->launch->fn, dims[GOMP_DIM_WORKER],
+ dims[GOMP_DIM_VECTOR], targ_fn->max_threads_per_block);
GOMP_PLUGIN_debug (0, " %s: kernel %s: launch"
" gangs=%u, workers=%u, vectors=%u\n",
Hi, atm parallel-dims.c fails on Titan-V, with a cuda launch failure: ... libgomp: cuLaunchKernel error: too many resources requested for launch ... We've got a check in the libgomp nvptx plugin to prevent the cuda launch failure and give a more informative error message: ... /* Check if the accelerator has sufficient hardware resources to launch the offloaded kernel. */ if (dims[GOMP_DIM_WORKER] > 1) { if (reg_granularity > 0 && dims[GOMP_DIM_WORKER] > threads_per_block) GOMP_PLUGIN_fatal ("The Nvidia accelerator has insufficient resources " "to launch '%s'; recompile the program with " "'num_workers = %d' on that offloaded region or " "'-fopenacc-dim=-:%d'.\n", targ_fn->launch->fn, threads_per_block, threads_per_block); } ... The message doesn't trigger, because reg_granularity == -1. This value comes from dev->register_allocation_granularity which defaults to -1 because libgomp does not have a hardcoded constant for sm_70. The hardcoded constants that are present match 'Warp Allocation Granularity' in the GPU Data table in CUDA_Occupancy_calculator.xls, but AFAICT there's no column published yet for sm_70. Furthermore, the comparison to threads_per_block is not correct. What we want here is the maximum amount of threads per block, while the threads_per_block variable contains an approximation of that, and the exact amount required is already available from the CUDA runtime and stored at targ_fn->max_threads_per_block. Then, the comparison to dims[GOMP_DIM_WORKER] is incorrect. It used to be correct before "[nvptx] Handle large vectors in libgomp" when we used to do "threads_per_block /= warp_size", but now we need to compare against dims[GOMP_DIM_WORKER] * dims[GOMP_DIM_VECTOR]. Finally, the message has not been updated to reflect that vector length can be larger than 32. The patch addresses these issues. Committed to og7. Thanks, - Tom