[Mesa-dev] [PATCH v2 7/7] i965: make use of nir linking

[Timothy Arceri]

On 26/09/17 17:50, Kenneth Graunke wrote:
> On Monday, September 25, 2017 6:23:13 PM PDT Timothy Arceri wrote:
>> For now linking is just removing unused varyings between stages.
>>
>> shader-db results BDW:
>>
>> total instructions in shared programs: 13198288 -> 13191693 (-0.05%)
>> instructions in affected programs: 48325 -> 41730 (-13.65%)
>> helped: 473
>> HURT: 0
>>
>> total cycles in shared programs: 541184926 -> 541159260 (-0.00%)
>> cycles in affected programs: 213238 -> 187572 (-12.04%)
>> helped: 435
>> HURT: 8
>>
>> V2:
>> - lower indirects on demoted inputs as well as outputs.
>> ---
>>   src/mesa/drivers/dri/i965/brw_link.cpp | 56 ++++++++++++++++++++++++++++++++++
>>   1 file changed, 56 insertions(+)
>>
>> diff --git a/src/mesa/drivers/dri/i965/brw_link.cpp b/src/mesa/drivers/dri/i965/brw_link.cpp
>> index b7fab8d7a25..9ddf0230183 100644
>> --- a/src/mesa/drivers/dri/i965/brw_link.cpp
>> +++ b/src/mesa/drivers/dri/i965/brw_link.cpp
>> @@ -253,6 +253,62 @@ brw_link_shader(struct gl_context *ctx, struct gl_shader_program *shProg)
>>                                    compiler->scalar_stage[stage]);
>>      }
>>   
>> +   /* Determine first and last stage. */
>> +   unsigned first = MESA_SHADER_STAGES;
>> +   unsigned last = 0;
>> +   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
>> +      if (!shProg->_LinkedShaders[i])
>> +         continue;
>> +      if (first == MESA_SHADER_STAGES)
>> +         first = i;
>> +      last = i;
>> +   }
>> +
>> +   /* Linking the stages in the opposite order (from fragment to vertex)
>> +    * ensures that inter-shader outputs written to in an earlier stage
>> +    * are eliminated if they are (transitively) not used in a later
>> +    * stage.
>> +    */
>> +    if (first != last) {
>> +       int next = last;
>> +       for (int i = next - 1; i >= 0; i--) {
>> +          if (shProg->_LinkedShaders[i] == NULL)
>> +             continue;
>> +
>> +            nir_shader *producer = shProg->_LinkedShaders[i]->Program->nir;
>> +            nir_shader *consumer = shProg->_LinkedShaders[next]->Program->nir;
>> +
>> +            nir_remove_dead_variables(producer, nir_var_shader_out);
>> +            nir_remove_dead_variables(consumer, nir_var_shader_in);
>> +
>> +            if (nir_remove_unused_varyings(producer, consumer)) {
>> +               nir_lower_global_vars_to_local(producer);
>> +               nir_lower_global_vars_to_local(consumer);
>> +
>> +               nir_variable_mode indirect_mask = (nir_variable_mode) 0;
>> +               if (compiler->glsl_compiler_options[i].EmitNoIndirectTemp)
>> +                  indirect_mask = (nir_variable_mode) nir_var_local;
>> +
>> +               /* The backend might not be able to handle indirects on
>> +                * temporaries so we need to lower indirects on any of the
>> +                * varyings we have demoted here.
>> +                */
>> +               nir_lower_indirect_derefs(producer, indirect_mask);
>> +               nir_lower_indirect_derefs(consumer, indirect_mask);
>> +
>> +               const bool p_is_scalar = compiler->scalar_stage[producer->stage];
>> +               shProg->_LinkedShaders[i]->Program->nir =
>> +                 brw_nir_optimize(producer, compiler, p_is_scalar);
>> +
>> +               const bool c_is_scalar = compiler->scalar_stage[producer->stage];
>> +               shProg->_LinkedShaders[next]->Program->nir =
>> +                 brw_nir_optimize(consumer, compiler, c_is_scalar);
>> +            }
>> +
>> +            next = i;
>> +       }
>> +    }
>> +
>>      for (stage = 0; stage < ARRAY_SIZE(shProg->_LinkedShaders); stage++) {
>>         struct gl_linked_shader *shader = shProg->_LinkedShaders[stage];
>>         if (!shader)
>>
> 
> A couple more thoughts:
> 
> 1. We're calling brw_nir_optimize even more now...it might make sense to
>     drop it from brw_preprocess_nir, drop it here, and put it once in the
>     final (third) loop, just before brw_shader_gather_info.  At least,
>     it'd be interesting how that affects compile times / quality...

Actually we can't drop the producer brw_nir_optimize call here anyway as 
we want to remove dead code before the next iteration of the loop.

> 
> 2. It would be great to use this in anv as well, especially given that
>     it has no GLSL IR linking to do this sort of stuff.  Plus, the SPIR-V
>     might be compiled independently...and when you build the pipeline,
>     you have all the stages and can do cross-stage linking optimizations
>     like this.
> 
> Series looks great though, and is:
> Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
> 
> Thanks Tim!
>