[Mesa-stable] [PATCH 1/7] radeon/compute: Fix reported values for MAX_GLOBAL_SIZE and MAX_MEM_ALLOC_SIZE

Emil Velikov emil.l.velikov at gmail.com
Wed Aug 27 05:21:09 PDT 2014


On 08/08/14 15:16, Tom Stellard wrote:
> There is a hard limit in older kernels of 256 MB for buffer allocations,
> so report this value as MAX_MEM_ALLOC_SIZE and adjust MAX_GLOBAL_SIZE
> to statisfy requirements of OpenCL.
> 
Hi Tom,

Following Dave's report I've gone through the series and pulled out (rejected)
patches 2, 5 and 6. The rest of the series is still queued for 10.2 yet I
would prefer if someone can give them a spin.

Thanks
Emil

> CC: "10.2" <mesa-stable at lists.freedesktop.org>
> ---
>  src/gallium/drivers/radeon/r600_pipe_common.c | 32 ++++++++++++++++-----------
>  1 file changed, 19 insertions(+), 13 deletions(-)
> 
> diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
> index 3476021..0886b02 100644
> --- a/src/gallium/drivers/radeon/r600_pipe_common.c
> +++ b/src/gallium/drivers/radeon/r600_pipe_common.c
> @@ -474,13 +474,21 @@ static int r600_get_compute_param(struct pipe_screen *screen,
>  	case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
>  		if (ret) {
>  			uint64_t *max_global_size = ret;
> -			/* XXX: This is what the proprietary driver reports, we
> -			 * may want to use a different value. */
> -			/* XXX: Not sure what to put here for SI. */
> -			if (rscreen->chip_class >= SI)
> -				*max_global_size = 2000000000;
> -			else
> -				*max_global_size = 201326592;
> +			uint64_t max_mem_alloc_size;
> +
> +			r600_get_compute_param(screen,
> +				PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE,
> +				&max_mem_alloc_size);
> +
> +			/* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
> +			 * 1/4 of the MAX_GLOBAL_SIZE.  Since the
> +			 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
> +			 * make sure we never report more than
> +			 * 4 * MAX_MEM_ALLOC_SIZE.
> +			 */
> +			*max_global_size = MIN2(4 * max_mem_alloc_size,
> +				rscreen->info.gart_size +
> +				rscreen->info.vram_size);
>  		}
>  		return sizeof(uint64_t);
>  
> @@ -504,13 +512,11 @@ static int r600_get_compute_param(struct pipe_screen *screen,
>  		if (ret) {
>  			uint64_t max_global_size;
>  			uint64_t *max_mem_alloc_size = ret;
> -			r600_get_compute_param(screen, PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE, &max_global_size);
> -			/* OpenCL requres this value be at least
> -			 * max(MAX_GLOBAL_SIZE / 4, 128 * 1024 *1024)
> -			 * I'm really not sure what value to report here, but
> -			 * MAX_GLOBAL_SIZE / 4 seems resonable.
> +
> +			/* XXX: The limit in older kernels is 256 MB.  We
> +			 * should add a query here for newer kernels.
>  			 */
> -			*max_mem_alloc_size = max_global_size / 4;
> +			*max_mem_alloc_size = 256 * 1024 * 1024;
>  		}
>  		return sizeof(uint64_t);
>  
> 



More information about the mesa-stable mailing list