[Mesa-dev] [PATCH] radeonsi/compute: Add Sea Islands support

[Jay Cornwall]

On 2013-10-23 13:05, Tom Stellard wrote:

>> > +		si_pm4_set_reg(pm4, R_00B82C_COMPUTE_MAX_WAVE_ID,
>> > +						0x190 /* Default value */);
>> 
>> Is there any visible effect on performance if this is set very low?
>> 
> 
> I haven't tested with any other values.

OK, I'll give it a test later and send a patch to compute the value if 
it affects performance. I'm not certain what it's used for. (Maybe 
related to COMPUTE_TMPRING_SIZE scratch allocation?)

>> If there is, perhaps we should compute this value for SI. I think all
>> the necessary parameters are provided by DRM (MAX_SE * MAX_SH_PER_SE *
>> MAX_PIPES * 40 wavefronts per CU, off the top of my head).
> 
> The formula from the comment is:
> 
> (number of compute units) * 4 * (waves per simd) - 1
> 
> I think this may be a simplified version of your formula.

Yes, although I omitted the - 1 by mistake. That's why it the value 
struck me as odd, because 0x190 is unlikely to be the last wavefront ID. 
0x18F perhaps, but I calculate 0x31F for Pitcairn, which is 
significantly higher.

Anyhow, it doesn't this patch. Your patch looks good to me.

-- 
Jay Cornwall
http://www.jcornwall.me/