[Mesa-dev] [PATCH v3 13/14] nvc0: expose ARB_compute_variable_group_size

[Samuel Pitoiset]

On 09/26/2016 07:27 PM, Ilia Mirkin wrote:
> FWIW this limits it to 32 regs on Fermi. IMO that's pretty limiting,
> esp given how shitty our RA is. I think we should do 512 for Fermi and
> 1024 for Kepler+. [A matching adjustment will be needed in codegen.]

Yep, I will improve it, but this can be done just after the mesa/gallium 
bits are upstream. :)

>
> On Mon, Sep 26, 2016 at 1:23 PM, Samuel Pitoiset
> <samuel.pitoiset at gmail.com> wrote:
>> Let's return the same number of threads per block for both fixed and
>> variable sizes.
>>
>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>> ---
>>  src/gallium/drivers/nouveau/nvc0/nvc0_screen.c | 3 +--
>>  1 file changed, 1 insertion(+), 2 deletions(-)
>>
>> diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
>> index df6c6af..6540c31 100644
>> --- a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
>> +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
>> @@ -446,6 +446,7 @@ nvc0_screen_get_compute_param(struct pipe_screen *pscreen,
>>        }
>>     case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
>>        RET(((uint64_t []) { 1024, 1024, 64 }));
>> +   case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
>>     case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
>>        RET((uint64_t []) { 1024 });
>>     case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE: /* g[] */
>> @@ -478,8 +479,6 @@ nvc0_screen_get_compute_param(struct pipe_screen *pscreen,
>>        RET((uint32_t []) { 512 }); /* FIXME: arbitrary limit */
>>     case PIPE_COMPUTE_CAP_ADDRESS_BITS:
>>        RET((uint32_t []) { 64 });
>> -   case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
>> -      RET((uint64_t []) { 0 });
>>     default:
>>        return 0;
>>     }
>> --
>> 2.10.0
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev