Mesa (main): aco/optimizer: fuse v_mul_f64 + v_add_f64 -> v_fma_f64

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: be16ebc5cafb2864228bf4ad29935cf0187e7f4d
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=be16ebc5cafb2864228bf4ad29935cf0187e7f4d

Author: Daniel Schürmann <daniel at schuermann.dev>
Date:   Wed Jun 30 19:20:49 2021 +0200

aco/optimizer: fuse v_mul_f64 + v_add_f64 -> v_fma_f64

No fossil-db changes.

Reviewed-by: Rhys Perry <pendingchaos02 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11678>

---

 src/amd/compiler/aco_optimizer.cpp | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index f739257afaf..8874673f823 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -1422,6 +1422,7 @@ label_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
       if (!ctx.program->needs_wqm)
          ctx.info[instr->definitions[0].tempId()].set_constant(ctx.program->chip_class, 0u);
       break;
+   case aco_opcode::v_mul_f64: ctx.info[instr->definitions[0].tempId()].set_mul(instr.get()); break;
    case aco_opcode::v_mul_f16:
    case aco_opcode::v_mul_f32: { /* omod */
       ctx.info[instr->definitions[0].tempId()].set_mul(instr.get());
@@ -3238,9 +3239,11 @@ combine_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
                 instr->opcode == aco_opcode::v_subrev_f32;
    bool mad16 = instr->opcode == aco_opcode::v_add_f16 || instr->opcode == aco_opcode::v_sub_f16 ||
                 instr->opcode == aco_opcode::v_subrev_f16;
-   if (mad16 || mad32) {
-      bool need_fma = mad32 ? (ctx.fp_mode.denorm32 != 0 || ctx.program->chip_class >= GFX10_3)
-                            : (ctx.fp_mode.denorm16_64 != 0 || ctx.program->chip_class >= GFX10);
+   bool mad64 = instr->opcode == aco_opcode::v_add_f64;
+   if (mad16 || mad32 || mad64) {
+      bool need_fma =
+         mad32 ? (ctx.fp_mode.denorm32 != 0 || ctx.program->chip_class >= GFX10_3)
+               : (ctx.fp_mode.denorm16_64 != 0 || ctx.program->chip_class >= GFX10 || mad64);
       if (need_fma && instr->definitions[0].isPrecise())
          return;
       if (need_fma && mad32 && !ctx.program->dev.has_fast_fma32)
@@ -3325,6 +3328,8 @@ combine_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
                                                                  : aco_opcode::v_fma_f16)
                               : (ctx.program->chip_class == GFX8 ? aco_opcode::v_mad_legacy_f16
                                                                  : aco_opcode::v_mad_f16);
+         if (mad64)
+            mad_op = aco_opcode::v_fma_f64;
 
          aco_ptr<VOP3_instruction> mad{
             create_instruction<VOP3_instruction>(mad_op, Format::VOP3, 3, 1)};
@@ -3591,7 +3596,7 @@ select_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
          mad_info = NULL;
       }
       /* check literals */
-      else if (!instr->usesModifiers()) {
+      else if (!instr->usesModifiers() && instr->opcode != aco_opcode::v_fma_f64) {
          /* FMA can only take literals on GFX10+ */
          if ((instr->opcode == aco_opcode::v_fma_f32 || instr->opcode == aco_opcode::v_fma_f16) &&
              ctx.program->chip_class < GFX10)