[Mesa-dev] [PATCH 1/2] mesa: Add a streaming load memcpy implementation.

Matt Turner mattst88 at gmail.com
Sat Nov 9 09:05:56 PST 2013


On Sat, Nov 9, 2013 at 3:54 AM, Steven Newbury <steve at snewbury.org.uk> wrote:
> On Wed, 2013-11-06 at 17:06 -0800, Chad Versace wrote:
>> On 11/06/2013 02:44 PM, Eric Anholt wrote:
>> > Matt Turner <mattst88 at gmail.com> writes:
>> >
>> >> Uses SSE 4.1's MOVNTDQA instruction (streaming load) to read from
>> >> uncached memory without polluting the cache.
>> >> ---
>> >> We should add runtime detection support later.
>> >
>> > I'd really like to see runtime detection with this.  Effectively not
>> > supporting this on 32-bit sucks, and doing it this way will encourage
>> > people to use compiler flags to get this path, even when those compiler
>> > flags hurt performance in general (in my testing, at least).
>>
>> This patch, as-is without runtime detection, benefits platform-specific
>> Linux builds such as ChromiumOS. Seeing that the patch benefits
>> immediately for some distros, I see no reason to block the patch for lack
>> of runtime detection.
>>
>> I would also like to see runtime detection as a follow-up.
>>
>> Patch 1 is
>> Reviewed-by: Chad Versace <chad.versace at linux.intel.com>
>>
>> By the way, I tested this series by playing around on ChromiumOS,
>> but I don't believe that warrants a Tested-by.
>>
> Building with --disable-asm results in:
> dlopen /usr/lib64/dri/i965_dri.so failed (/usr/lib64/dri/i965_dri.so:
> undefined symbol: _mesa_streaming_load_memcpy
>
> The i965 driver either needs to only use _mesa_streaming_load_memcpy()
> when asm is enabled, or there needs to be a non-asm implementation in
> streaming_load_memcpy.c.
>

Indeed, you're correct. Would you like to send a patch?


More information about the mesa-dev mailing list