Re: [patch] adjust default nvptx launch geometry for OpenACC offloaded regions

[Tom de Vries <tdevries@suse.de>]

[-- Attachment #1: Type: text/plain, Size: 1564 bytes --]

>> Right, in fact there are two separate things you're trying to address
>> here: launch failure and occupancy heuristic, so split the patch.

> That hunk was small, so I included it with this patch. Although if you
> insist, I can remove it.

Please, for future reference, always assume that I insist instead of
asking me, unless you have an argument to present why that is not a good
idea. And just to be clear here: "small" is not such an argument.

Please keep in mind ( https://gcc.gnu.org/contribute.html#patches ):
...
Don't mix together changes made for different reasons. Send them
individually.
...

> +  /* Check if the accelerator has sufficient hardware resources to
> +     launch the offloaded kernel.  */
> +  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);
> +

This is copied from the state on an openacc branch where vector-length
is variable, and the error message text doesn't make sense on current
trunk for that reason. Also, it suggests a syntax for fopenacc-dim
that's not supported on trunk.

Committed as attached.

Thanks,
- Tom

[-- Attachment #2: 0002-libgomp-nvptx-Add-error-with-recompilation-hint-for-launch-failure.patch --]
[-- Type: text/x-patch, Size: 1866 bytes --]

[libgomp, nvptx] Add error with recompilation hint for launch failure

Currently, when a kernel is lauched with too many workers, it results in a cuda
launch failure.  This is triggered f.i. for parallel-loop-1.c at -O0 on a Quadro
M1200.

This patch detects this situation, and errors out with a hint on how to fix it.

Build and reg-tested on x86_64 with nvptx accelerator.

2018-07-26  Cesar Philippidis  <cesar@codesourcery.com>
	    Tom de Vries  <tdevries@suse.de>

	* plugin/plugin-nvptx.c (nvptx_exec): Error if the hardware doesn't have
	sufficient resources to launch a kernel, and give a hint on how to fix
	it.

---
 libgomp/plugin/plugin-nvptx.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/libgomp/plugin/plugin-nvptx.c b/libgomp/plugin/plugin-nvptx.c
index 5d9b5151e95..3a4077a1315 100644
--- a/libgomp/plugin/plugin-nvptx.c
+++ b/libgomp/plugin/plugin-nvptx.c
@@ -1204,6 +1204,21 @@ nvptx_exec (void (*fn), size_t mapnum, void **hostaddrs, void **devaddrs,
 	  dims[i] = default_dims[i];
     }
 
+  /* Check if the accelerator has sufficient hardware resources to
+     launch the offloaded kernel.  */
+  if (dims[GOMP_DIM_WORKER] * dims[GOMP_DIM_VECTOR]
+      > targ_fn->max_threads_per_block)
+    {
+      int suggest_workers
+	= targ_fn->max_threads_per_block / dims[GOMP_DIM_VECTOR];
+      GOMP_PLUGIN_fatal ("The Nvidia accelerator has insufficient resources to"
+			 " launch '%s' with num_workers = %d; recompile the"
+			 " program with 'num_workers = %d' on that offloaded"
+			 " region or '-fopenacc-dim=:%d'",
+			 targ_fn->launch->fn, dims[GOMP_DIM_WORKER],
+			 suggest_workers, suggest_workers);
+    }
+
   /* This reserves a chunk of a pre-allocated page of memory mapped on both
      the host and the device. HP is a host pointer to the new chunk, and DP is
      the corresponding device pointer.  */