[Mesa-dev] [PATCH v2] nir: Try commutative sources in CSE

[Ilia Mirkin]

On Wed, Apr 15, 2015 at 12:37 PM, Ian Romanick <idr at freedesktop.org> wrote:
> From: Ian Romanick <ian.d.romanick at intel.com>
>
> Shader-db results:
>
> GM45 NIR:
> total instructions in shared programs: 4082044 -> 4081919 (-0.00%)
> instructions in affected programs:     27609 -> 27484 (-0.45%)
> helped:                                44
>
> Iron Lake NIR:
> total instructions in shared programs: 5678776 -> 5678646 (-0.00%)
> instructions in affected programs:     27406 -> 27276 (-0.47%)
> helped:                                45
>
> Sandy Bridge NIR:
> total instructions in shared programs: 7329995 -> 7329096 (-0.01%)
> instructions in affected programs:     142035 -> 141136 (-0.63%)
> helped:                                406
> HURT:                                  19
>
> Ivy Bridge NIR:
> total instructions in shared programs: 6769314 -> 6768359 (-0.01%)
> instructions in affected programs:     140820 -> 139865 (-0.68%)
> helped:                                423
> HURT:                                  2
>
> Haswell NIR:
> total instructions in shared programs: 6183693 -> 6183298 (-0.01%)
> instructions in affected programs:     96538 -> 96143 (-0.41%)
> helped:                                303
> HURT:                                  4
>
> Broadwell NIR:
> total instructions in shared programs: 7501711 -> 7498170 (-0.05%)
> instructions in affected programs:     266403 -> 262862 (-1.33%)
> helped:                                705
> HURT:                                  5
> GAINED:                                4
>
> v2: Rebase on top of Connor's fix.
>
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> Reviewed-by: Jordan Justen <jordan.l.justen at intel.com> [v1]
> Cc: Jason Ekstrand <jason.ekstrand at intel.com>
> Cc: Connor Abbott <cwabbott0 at gmail.com>
> ---
>
> The v2 changes were a little more intrusive than before, so it seemed
> worth sending out again.
>
>  src/glsl/nir/nir_opt_cse.c | 27 ++++++++++++++++++---------
>  1 file changed, 18 insertions(+), 9 deletions(-)
>
> diff --git a/src/glsl/nir/nir_opt_cse.c b/src/glsl/nir/nir_opt_cse.c
> index 56d491c..3c5ece7 100644
> --- a/src/glsl/nir/nir_opt_cse.c
> +++ b/src/glsl/nir/nir_opt_cse.c
> @@ -37,18 +37,19 @@ struct cse_state {
>  };
>
>  static bool
> -nir_alu_srcs_equal(nir_alu_instr *alu1, nir_alu_instr *alu2, unsigned src)
> +nir_alu_srcs_equal(nir_alu_instr *alu1, nir_alu_instr *alu2, unsigned src1,
> +                   unsigned src2)
>  {
> -   if (alu1->src[src].abs != alu2->src[src].abs ||
> -       alu1->src[src].negate != alu2->src[src].negate)
> +   if (alu1->src[src1].abs != alu2->src[src2].abs ||
> +       alu1->src[src1].negate != alu2->src[src2].negate)
>        return false;
>
> -   for (unsigned i = 0; i < nir_ssa_alu_instr_src_components(alu1, src); i++) {
> -      if (alu1->src[src].swizzle[i] != alu2->src[src].swizzle[i])
> +   for (unsigned i = 0; i < nir_ssa_alu_instr_src_components(alu1, src1); i++) {
> +      if (alu1->src[src1].swizzle[i] != alu2->src[src2].swizzle[i])
>           return false;
>     }
>
> -   return nir_srcs_equal(alu1->src[src].src, alu2->src[src].src);
> +   return nir_srcs_equal(alu1->src[src1].src, alu2->src[src2].src);
>  }
>
>  static bool
> @@ -71,9 +72,17 @@ nir_instrs_equal(nir_instr *instr1, nir_instr *instr2)
>        if (alu1->dest.dest.ssa.num_components != alu2->dest.dest.ssa.num_components)
>           return false;
>
> -      for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) {
> -         if (!nir_alu_srcs_equal(alu1, alu2, i))
> -            return false;
> +      if (nir_op_infos[alu1->op].num_inputs == 2 &&
> +          (nir_op_infos[alu1->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE)) {

Just a couple of thoughts (could well go into a later patch, or not at all) --

(a) is the NIR_OP_IS_COMMUTATIVE flag ever set when num_inputs != 2?
(b) I see that there is an isub/fsub/etc, which are (hopefully) not
marked as commutative. Would it make sense to first get rid of them
(and introduce negate flags) before CSE is done? I guess it's an
argument that expressions should have a clear canonical form for
maximal CSE, after which various negs can be propagated through.

> +         return (nir_alu_srcs_equal(alu1, alu2, 0, 0) &&
> +                 nir_alu_srcs_equal(alu1, alu2, 1, 1)) ||
> +                (nir_alu_srcs_equal(alu1, alu2, 0, 1) &&
> +                 nir_alu_srcs_equal(alu1, alu2, 1, 0));
> +      } else {
> +         for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) {
> +            if (!nir_alu_srcs_equal(alu1, alu2, i, i))
> +               return false;
> +         }
>        }
>        return true;
>     }
> --
> 2.1.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev