[Mesa-dev] [PATCH 01/11] i965/vs: Add support for splitting virtual GRFs.

[Kenneth Graunke]

On 10/08/2012 09:32 AM, Eric Anholt wrote:
> Kenneth Graunke <kenneth at whitecape.org> writes:
>
>> On 10/04/2012 04:07 PM, Eric Anholt wrote:
>>> This should improve our ability to register allocate without spilling.
>>> Unfortuantely, due to the live variable analysis being ignorant of loops, we
>>> still have register allocation failures on some programs.
>>> ---
>>>    src/mesa/drivers/dri/i965/brw_vec4.cpp      |   54 +++++++++++++++++++++++++++
>>>    src/mesa/drivers/dri/i965/brw_vec4.h        |    1 +
>>>    src/mesa/drivers/dri/i965/brw_vec4_emit.cpp |    1 +
>>>    3 files changed, 56 insertions(+)
>>>
>>> diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp
>>> index 2941729..fbd49ca 100644
>>> --- a/src/mesa/drivers/dri/i965/brw_vec4.cpp
>>> +++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp
>>> @@ -865,4 +865,58 @@ vec4_visitor::opt_compute_to_mrf()
>>>       return progress;
>>>    }
>>>
>>> +/**
>>> + * Splits virtual GRFs requesting more than one contiguous physical register.
>>> + *
>>> + * Unlike in the FS visitor, we have no SEND messages that return more than 1
>>> + * register.  We also don't do any array access in register space, which would
>>> + * have required contiguous physical registers.
>>> + */
>>
>> It would be nice to mention what the VS /does/ have that occupies VGRFs
>> of size > 1.  Structures and matrices?
>>
>> Otherwise, this looks good.
>> Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
>
> New function comment:
>
> /**
>   * Splits virtual GRFs requesting more than one contiguous physical register.
>   *
>   * We initially create large virtual GRFs for temporary structures, arrays,
>   * and matrices, so that the dereference visitor functions can add reg_offsets
>   * to work their way down to the actual member being accessed.
>   *
>   * Unlike in the FS visitor, though, we have no SEND messages that return more
>   * than 1 register.  We also don't do any array access in register space,
>   * which would have required contiguous physical registers.  Thus, all those
>   * large virtual GRFs can be split up into independent single-register virtual
>   * GRFs, making allocation and optimization easier.
>   */

Perfect, thank you!
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>