[Mesa-dev] [PATCH] i965/vec4: Calculate register allocation q values manually.

Ian Romanick idr at freedesktop.org
Wed Feb 18 11:32:33 PST 2015


On 02/18/2015 10:57 AM, Mark Janes wrote:
> This patch fixes a performance regression that causes our piglit CI
> system to slow by 50%.  If anyone has time to take a look at this and
> provide a RB, it will improve the rate at which you get your piglit
> results.

When you send a message like this, you should also say

Tested-by: Mark Janes <mark.a.janes at intel.com>

 :)

> -Mark
> 
> Francisco Jerez <currojerez at riseup.net> writes:
> 
>> This fixes a regression in the running time of Piglit introduced by
>> commit 78e9043475d4bed8b50f7e413963c960fa0935bb, which increased the
>> number of register allocation classes set up by the VEC4 back-end
>> from 2 to 16.  The algorithm used by ra_set_finalize() to calculate
>> them is unnecessarily expensive, do it manually like the FS back-end
>> does.
>>
>> Reported-by: Mark Janes <mark.a.janes at intel.com>
>> ---
>>  src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp | 18 +++++++++++++++++-
>>  1 file changed, 17 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp b/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp
>> index 46f0bfd..a286f8a 100644
>> --- a/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp
>> +++ b/src/mesa/drivers/dri/i965/brw_vec4_reg_allocate.cpp
>> @@ -129,10 +129,13 @@ brw_vec4_alloc_reg_set(struct intel_screen *screen)
>>      * between them and the base GRF registers (and also each other).
>>      */
>>     int reg = 0;
>> +   unsigned *q_values[MAX_VGRF_SIZE];
>>     for (int i = 0; i < class_count; i++) {
>>        int class_reg_count = base_reg_count - (class_sizes[i] - 1);
>>        screen->vec4_reg_set.classes[i] = ra_alloc_reg_class(screen->vec4_reg_set.regs);
>>  
>> +      q_values[i] = new unsigned[MAX_VGRF_SIZE];
>> +
>>        for (int j = 0; j < class_reg_count; j++) {
>>  	 ra_class_add_reg(screen->vec4_reg_set.regs, screen->vec4_reg_set.classes[i], reg);
>>  
>> @@ -146,10 +149,23 @@ brw_vec4_alloc_reg_set(struct intel_screen *screen)
>>  
>>  	 reg++;
>>        }
>> +
>> +      for (int j = 0; j < class_count; j++) {
>> +         /* Calculate the q values manually because the algorithm used by
>> +          * ra_set_finalize() to do it has higher complexity affecting the
>> +          * start-up time of some applications.  q(i, j) is just the maximum
>> +          * number of registers from class i a register from class j can
>> +          * conflict with.
>> +          */
>> +         q_values[i][j] = class_sizes[i] + class_sizes[j] - 1;
>> +      }
>>     }
>>     assert(reg == ra_reg_count);
>>  
>> -   ra_set_finalize(screen->vec4_reg_set.regs, NULL);
>> +   ra_set_finalize(screen->vec4_reg_set.regs, q_values);
>> +
>> +   for (int i = 0; i < MAX_VGRF_SIZE; i++)
>> +      delete[] q_values[i];
>>  }
>>  
>>  void
>> -- 
>> 2.1.3
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
> 



More information about the mesa-dev mailing list