Unlike fma() in GLSL, MAD in ARB programs is 100% splittable. Just emit
the split version and let the optimizer fuse them later.
Shader-db results on Haswell:
total instructions in shared programs: 7560379 -> 7560300 (-0.00%)
instructions in affected programs: 143928 -> 143849 (-0.05%)
helped: 443
HURT: 250
---
src/mesa/program/prog_to_nir.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/src/mesa/program/prog_to_nir.c b/src/mesa/program/prog_to_nir.c
index a6119ae..aa51dec 100644
--- a/src/mesa/program/prog_to_nir.c
+++ b/src/mesa/program/prog_to_nir.c
@@ -706,7 +706,7 @@ static const nir_op op_trans[MAX_OPCODE] = {
[OPCODE_LIT] = 0,
[OPCODE_LOG] = 0,
[OPCODE_LRP] = 0,
- [OPCODE_MAD] = nir_op_ffma,
+ [OPCODE_MAD] = 0,
[OPCODE_MAX] = nir_op_fmax,
[OPCODE_MIN] = nir_op_fmin,
[OPCODE_MOV] = nir_op_fmov,
@@ -801,6 +801,10 @@ ptn_emit_instruction(struct ptn_compile *c, struct
prog_instruction *prog_inst)
ptn_lrp(b, dest, src);
break;
+ case OPCODE_MAD:
+ ptn_move_dest(b, dest, nir_fadd(b, nir_fmul(b, src[0], src[1]), src[2]));
+ break;
+
case OPCODE_DST:
ptn_dst(b, dest, src);
break;
--
2.5.0.400.gff86faf
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
