Re: [gomp4] Lack of OpenACC NVPTX devices is not an error during scanning

[Thomas Schwinge]

Hi Julian!

On Tue, 19 May 2015 11:36:58 +0100, Julian Brown <jul...@codesourcery.com> 
wrote:
> This patch fixes an oversight whereby if the CUDA libraries are
> available for some reason on a system that doesn't actually contain an
> nVidia card, an OpenACC program will raise an error if the NVPTX
> backend is picked as a default instead of falling back to some other
> device instead.

Thanks for fixing this!  (Has already been committed to trunk in r223352,
and to gomp-4_0-branch in r223351.)

Your patch:

> --- a/libgomp/plugin/plugin-nvptx.c
> +++ b/libgomp/plugin/plugin-nvptx.c
> @@ -781,7 +781,13 @@ nvptx_get_num_devices (void)
>       until cuInit has been called.  Just call it now (but don't yet do any
>       further initialization).  */
>    if (instantiated_devices == 0)
> -    cuInit (0);
> +    {
> +      r = cuInit (0);
> +      /* This is not an error: e.g. we may have CUDA libraries installed but
> +         no devices available.  */
> +      if (r != CUDA_SUCCESS)
> +        return 0;
> +    }
>  
>    r = cuDeviceGetCount (&n);
>    if (r!= CUDA_SUCCESS)

In early March, I had noticed the same problem, and came up with the
following patch -- but :-( unfortunately never got around to pushing it
upstream.  I'm now posting my patch just for completeness; I think yours
is sufficient/better: no safe-guard should be needed to the cuInit call
in nvptx_init, because when that is called, we're rightfully expecting to
be able to initialize a PTX device, and in nvptx_get_num_devices, yours
is "more conservative" in doing the right thing ("no PTX offloading
device available") for all kinds of cuInit errors.

commit 6032dde185d0d45d779a1bbf0a5baee7131c0b8c
Author: Thomas Schwinge <tho...@codesourcery.com>
Date:   Sun Mar 1 14:36:02 2015 +0100

    libgomp nvptx plugin: Gracefully handle CUDA_ERROR_NO_DEVICE.
---
 libgomp/plugin/plugin-nvptx.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git libgomp/plugin/plugin-nvptx.c libgomp/plugin/plugin-nvptx.c
index 78e705f..0c1e826 100644
--- libgomp/plugin/plugin-nvptx.c
+++ libgomp/plugin/plugin-nvptx.c
@@ -592,6 +592,8 @@ nvptx_init (void)
     return -1;
 
   r = cuInit (0);
+  if (r == CUDA_ERROR_NO_DEVICE)
+    r = CUDA_SUCCESS;
   if (r != CUDA_SUCCESS)
     GOMP_PLUGIN_fatal ("cuInit error: %s", cuda_error (r));
 
@@ -715,7 +717,13 @@ nvptx_get_num_devices (void)
      until cuInit has been called.  Just call it now (but don't yet do any
      further initialization).  */
   if (!ptx_inited)
-    cuInit (0);
+    {
+      r = cuInit (0);
+      if (r == CUDA_ERROR_NO_DEVICE)
+       return 0;
+      if (r != CUDA_SUCCESS)
+       GOMP_PLUGIN_fatal ("cuInit error: %s", cuda_error (r));
+    }
 
   r = cuDeviceGetCount (&n);
   if (r!= CUDA_SUCCESS)


Grüße,
 Thomas

signature.asc
Description: PGP signature