Re: [PATCH] Remove PTX link option

[James Norris <jnorris at codesourcery dot com>]

Jakub,

On 01/08/2016 09:44 AM, Jakub Jelinek wrote:
> On Fri, Jan 08, 2016 at 09:39:11AM -0600, James Norris wrote:
> > On 01/08/2016 09:30 AM, Jakub Jelinek wrote:
> > > On Fri, Jan 08, 2016 at 09:24:14AM -0600, James Norris wrote:
> > > > This patch removes the constraint whereby the PTX
> > > > emitted was only for sm_30 GPU's. With this removal,
> > > > the PTX emitted will be targeted for the current
> > > > context, i.e., attached GPU.
> > > >
> > > > Bootstrapped/regtested on x86_64-linux, ok for trunk?
> > > How does that work?
> > > I see
> > > static void
> > > nvptx_file_start (void)
> > > {
> > >    fputs ("// BEGIN PREAMBLE\n", asm_out_file);
> > >    fputs ("\t.version\t3.1\n", asm_out_file);
> > >    fputs ("\t.target\tsm_30\n", asm_out_file);
> > >    fprintf (asm_out_file, "\t.address_size %d\n", GET_MODE_BITSIZE (Pmode));
> > >    fputs ("// END PREAMBLE\n", asm_out_file);
> > > }
> > > which I'd guess means the PTX code can be executed in sm_30 and newer only
> > > anyway.
> > >
> > > 	Jakub
> >
> > Yes. Per the PTX ISA manual for the .target directive: "In general,
> > generations of SM architectures follow an onion layer model, where
> > each generation adds new features and retains all features of
> > previous generations."
>
> And CU_JIT_TARGET / CU_TARGET_COMPUTE_30 requests JITting only on sm_30 and
> nothing else, or just on sm_30 or later, something else?
> This really should be reviewed by somebody familiar with CUDA more than
> myself.

The former. If one were to execute a program that specified
CU_TARGET_COMPUTE_30 on a Maxwell-class GPU (sm_50), you would
see the runtime error:

libgomp: cuModuleLoadData error: device kernel image is invalid

Jim