[Mesa-dev] [PATCH 2/2] gallium/radeon: use max_alloc_size

[Vedran Miletić]

On 07/01/2016 01:29 AM, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> also fix max_global_size to take a maximum of {vram_size, gart_size}
> ---
>  src/gallium/drivers/r600/r600_pipe.c          | 2 +-
>  src/gallium/drivers/radeon/r600_pipe_common.c | 9 +++------
>  src/gallium/drivers/radeonsi/si_pipe.c        | 2 +-
>  3 files changed, 5 insertions(+), 8 deletions(-)
>
> diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
> index 119c76b..55bbde1 100644
> --- a/src/gallium/drivers/r600/r600_pipe.c
> +++ b/src/gallium/drivers/r600/r600_pipe.c
> @@ -301,7 +301,7 @@ static int r600_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
>  		return 0;
>
>  	case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
> -		return MIN2(rscreen->b.info.vram_size, 0xFFFFFFFF);
> +		return MIN2(rscreen->b.info.max_alloc_size, 0xFFFFFFFF);
>
>          case PIPE_CAP_MIN_MAP_BUFFER_ALIGNMENT:
>                  return R600_MAP_BUFFER_ALIGNMENT;
> diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
> index d7f1d41..f75fa6c 100644
> --- a/src/gallium/drivers/radeon/r600_pipe_common.c
> +++ b/src/gallium/drivers/radeon/r600_pipe_common.c
> @@ -864,8 +864,8 @@ static int r600_get_compute_param(struct pipe_screen *screen,
>  			 * 4 * MAX_MEM_ALLOC_SIZE.
>  			 */
>  			*max_global_size = MIN2(4 * max_mem_alloc_size,
> -				rscreen->info.gart_size +
> -				rscreen->info.vram_size);
> +						MAX2(rscreen->info.gart_size,
> +						     rscreen->info.vram_size));

Can't you also use info.max_alloc_size here?

>  		}
>  		return sizeof(uint64_t);
>
> @@ -889,10 +889,7 @@ static int r600_get_compute_param(struct pipe_screen *screen,
>  		if (ret) {
>  			uint64_t *max_mem_alloc_size = ret;
>
> -			/* XXX: The limit in older kernels is 256 MB.  We
> -			 * should add a query here for newer kernels.
> -			 */
> -			*max_mem_alloc_size = 256 * 1024 * 1024;
> +			*max_mem_alloc_size = rscreen->info.max_alloc_size;
>  		}
>  		return sizeof(uint64_t);
>
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
> index ad2a86a..30e6253 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.c
> +++ b/src/gallium/drivers/radeonsi/si_pipe.c
> @@ -419,7 +419,7 @@ static int si_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
>  		       HAVE_LLVM >= 0x0307 ? 410 : 330;
>
>  	case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
> -		return MIN2(sscreen->b.info.vram_size, 0xFFFFFFFF);
> +		return MIN2(sscreen->b.info.max_alloc_size, 0xFFFFFFFF);
>
>  	case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
>  		return 0;
>

Had something similar in the works, Bas did as well, but this approach 
is cleaner.

With these changes, in si_pipe.c and r600_pipe.c, you should not return 
max_const_buffer_size anymore, since it can exceed int limits, but 
instead something like

MAX2(0x7ffffff, max_const_buffer_size)

Haven't tried running memory-heavy OpenCL programs, but clinfo other 
than const buffer size looks good.

-- 
Vedran Miletić
vedran.miletic.net