Hi Tom,
I see that you're reviewing the libgomp changes. Please disregard the
following hunk:
On 07/11/2018 12:13 PM, Cesar Philippidis wrote:
> @@ -1199,12 +1202,59 @@ nvptx_exec (void (*fn), size_t mapnum, void
> **hostaddrs, void **devaddrs,
> default_dims[GOMP_DIM_VECTOR]);
> }
> pthread_mutex_unlock (&ptx_dev_lock);
> + int vectors = default_dims[GOMP_DIM_VECTOR];
> + int workers = default_dims[GOMP_DIM_WORKER];
> + int gangs = default_dims[GOMP_DIM_GANG];
> +
> + if (nvptx_thread()->ptx_dev->driver_version > 6050)
> + {
> + int grids, blocks;
> + CUDA_CALL_ASSERT (cuOccupancyMaxPotentialBlockSize, &grids,
> + &blocks, function, NULL, 0,
> + dims[GOMP_DIM_WORKER] * dims[GOMP_DIM_VECTOR]);
> + GOMP_PLUGIN_debug (0, "cuOccupancyMaxPotentialBlockSize: "
> + "grid = %d, block = %d\n", grids, blocks);
> +
> + gangs = grids * dev_size;
> + workers = blocks / vectors;
> + }
I revisited this change yesterday and I noticed it was setting gangs
incorrectly. Basically, gangs should be set as follows
gangs = grids * (blocks / warp_size);
or to be more closer to og8 as
gangs = 2 * grids * (blocks / warp_size);
The use of that magic constant 2 is to prevent thread starvation. That's
a similar concept behind make -j<2*#threads>.
Anyway, I'm still experimenting with that change. There are still some
discrepancies between the way that I select num_workers and how the
driver does. The driver appears to be a little bit more conservative,
but according to the thread occupancy calculator, that should yield
greater performance on GPUs.
I just wanted to give you a heads up because you seem to be working on this.
Thanks for all of your reviews!
By the way, are you now maintainer of the libgomp nvptx plugin?
Cesar
