[Mesa-dev] [PATCH 3/3] i965/fs/nir: Optimize integer multiply by a 16-bit constant.

[Ian Romanick]

With just a quick look, this patch is

Reviewed-by: Ian Romanick <ian.d.romanick at intel.com>

On 02/17/2015 11:46 AM, Matt Turner wrote:
> Gen8+ support was just broken, since MUL now consumes 32-bits from both
> sources. Fixes 986 piglit tests on my BDW.
> 
> total instructions in shared programs: 8010869 -> 8010307 (-0.01%)
> instructions in affected programs:     37244 -> 36682 (-1.51%)
> helped:                                113
> GAINED:                                97
> ---
>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 24 +++++++++++++++++++++++-
>  1 file changed, 23 insertions(+), 1 deletion(-)
> 
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> index 266249f..05ac804 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> @@ -825,7 +825,29 @@ fs_visitor::nir_emit_alu(nir_alu_instr *instr)
>        break;
>  
>     case nir_op_imul: {
> -      /* TODO put in the 16-bit constant optimization once we have SSA */
> +      if (brw->gen >= 8) {
> +         emit(MUL(result, op[0], op[1]));
> +         break;
> +      } else {
> +         nir_const_value *value0 = nir_src_as_const_value(instr->src[0].src);
> +         nir_const_value *value1 = nir_src_as_const_value(instr->src[1].src);
> +
> +         if (value0 && value0->u[0] < (1 << 16)) {
> +            if (brw->gen < 7) {
> +               emit(MUL(result, op[0], op[1]));
> +            } else {
> +               emit(MUL(result, op[1], op[0]));
> +            }
> +            break;
> +         } else if (value1 && value1->u[0] < (1 << 16)) {
> +            if (brw->gen < 7) {
> +               emit(MUL(result, op[1], op[0]));
> +            } else {
> +               emit(MUL(result, op[0], op[1]));
> +            }
> +            break;
> +         }
> +      }
>  
>        if (brw->gen >= 7)
>           no16("SIMD16 explicit accumulator operands unsupported\n");
>