[Mesa-dev] [PATCH] i965/vec4/nir: Emit single MOV to generate a scalar constant.

[Eduardo Lima Mitev]

On 08/12/2015 10:58 PM, Matt Turner wrote:
> If an immediate is written to multiple channels, we can load it in a
> single writemasked MOV.
> 
> total instructions in shared programs: 6285144 -> 6261991 (-0.37%)
> instructions in affected programs:     718991 -> 695838 (-3.22%)
> helped:                                5762
> ---
> The shader-db numbers are from a tree with the scalar_vs fixup.
> 

Ohh, I had a similar patch but fixes the problem just partially:

https://github.com/Igalia/mesa/commit/56f07bdd70ef41e837dd9801e48fcc1802a8c8ea

Yours is more generic, so great!

>  src/mesa/drivers/dri/i965/brw_vec4_nir.cpp | 21 ++++++++++++++++++---
>  1 file changed, 18 insertions(+), 3 deletions(-)
> 
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp b/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
> index 923e2d3..632e409 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
> @@ -458,13 +458,28 @@ vec4_visitor::nir_emit_load_const(nir_load_const_instr *instr)
>     dst_reg reg = dst_reg(GRF, alloc.allocate(1));
>     reg.type =  BRW_REGISTER_TYPE_F;
>  
> +   unsigned remaining = brw_writemask_for_size(instr->def.num_components);
> +
>     /* @FIXME: consider emitting vector operations to save some MOVs in
>      * cases where the components are representable in 8 bits.
> -    * By now, we emit a MOV for each component.
> +    * For now, we emit a MOV for each distinct value.
>      */
> -   for (unsigned i = 0; i < instr->def.num_components; ++i) {
> -      reg.writemask = 1 << i;
> +   for (unsigned i = 0; i < instr->def.num_components; i++) {
> +      unsigned writemask = 1 << i;
> +
> +      if ((remaining & writemask) == 0)
> +         continue;
> +
> +      for (unsigned j = i; j < instr->def.num_components; j++) {
> +         if (instr->value.u[i] == instr->value.u[j]) {
> +            writemask |= 1 << j;
> +         }
> +      }
> +
> +      reg.writemask = writemask;
>        emit(MOV(reg, src_reg(instr->value.f[i])));
> +
> +      remaining &= ~writemask;
>     }
>  
>     /* Set final writemask */
>