Mesa (main): aco: don't use v_mad_mix on GFX9 if 16-bit denormals must be preserved

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 5b4e41e4db744e837998b482e3f53e0006946009
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=5b4e41e4db744e837998b482e3f53e0006946009

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Fri Apr  1 19:51:55 2022 +0100

aco: don't use v_mad_mix on GFX9 if 16-bit denormals must be preserved

This probably effectively disables the v_mad_mix optimization on GFX9.

fossil-db (Vega):
Totals from 11545 (7.15% of 161366) affected shaders:
MaxWaves: 43025 -> 42780 (-0.57%); split: +0.06%, -0.63%
Instrs: 18571635 -> 18734201 (+0.88%); split: -0.00%, +0.88%
CodeSize: 96483568 -> 96611012 (+0.13%); split: -0.11%, +0.24%
SGPRs: 1079056 -> 1077616 (-0.13%); split: -0.14%, +0.01%
VGPRs: 819248 -> 821868 (+0.32%); split: -0.04%, +0.36%
SpillSGPRs: 13313 -> 12464 (-6.38%)
Latency: 293804093 -> 295046122 (+0.42%); split: -0.09%, +0.51%
InvThroughput: 110002239 -> 110994978 (+0.90%); split: -0.03%, +0.93%
VClause: 342458 -> 342596 (+0.04%); split: -0.12%, +0.16%
SClause: 648566 -> 648046 (-0.08%); split: -0.12%, +0.04%
Copies: 1728225 -> 1726679 (-0.09%); split: -0.66%, +0.57%
Branches: 552973 -> 552963 (-0.00%); split: -0.02%, +0.02%
PreSGPRs: 862360 -> 856820 (-0.64%); split: -0.69%, +0.05%
PreVGPRs: 773689 -> 776818 (+0.40%); split: -0.02%, +0.42%

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof at gmail.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6178
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15718>

---

 src/amd/compiler/aco_optimizer.cpp        |  4 ++++
 src/amd/compiler/tests/test_optimizer.cpp | 20 ++++++++++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 5ac2ebc7335..b0d08774379 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -3473,6 +3473,10 @@ can_use_mad_mix(opt_ctx& ctx, aco_ptr<Instruction>& instr)
    if (ctx.program->chip_class < GFX9)
       return false;
 
+   /* v_mad_mix* on GFX9 always flushes denormals for 16-bit inputs/outputs */
+   if (ctx.program->chip_class == GFX9 && ctx.fp_mode.denorm16_64)
+      return false;
+
    switch (instr->opcode) {
    case aco_opcode::v_add_f32:
    case aco_opcode::v_sub_f32:
diff --git a/src/amd/compiler/tests/test_optimizer.cpp b/src/amd/compiler/tests/test_optimizer.cpp
index 3f1da2dbd9a..f4ff665143f 100644
--- a/src/amd/compiler/tests/test_optimizer.cpp
+++ b/src/amd/compiler/tests/test_optimizer.cpp
@@ -1164,6 +1164,8 @@ BEGIN_TEST(optimize.mad_mix.input_conv.basic)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1197,6 +1199,8 @@ BEGIN_TEST(optimize.mad_mix.input_conv.precision)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1248,6 +1252,8 @@ BEGIN_TEST(optimize.mad_mix.input_conv.modifiers)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1341,6 +1347,8 @@ BEGIN_TEST(optimize.mad_mix.output_conv.basic)
       if (!setup_cs("v1 v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp c = inputs[2];
@@ -1381,6 +1389,8 @@ BEGIN_TEST(optimize.mad_mix.output_conv.precision)
       if (!setup_cs("v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a16 = inputs[0];
 
       //! v2b: %res0_tmp = v_mul_f16 %a16, %a16
@@ -1403,6 +1413,8 @@ BEGIN_TEST(optimize.mad_mix.output_conv.modifiers)
       if (!setup_cs("v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp a16 = inputs[2];
@@ -1450,6 +1462,8 @@ BEGIN_TEST(optimize.mad_mix.fma.basic)
       if (!setup_cs("v1 v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp c = inputs[2];
@@ -1502,6 +1516,8 @@ BEGIN_TEST(optimize.mad_mix.fma.precision)
       if (!setup_cs("v1 v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp c = inputs[2];
@@ -1562,6 +1578,8 @@ BEGIN_TEST(optimize.mad_mix.clamp)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1587,6 +1605,8 @@ BEGIN_TEST(optimize.mad_mix.cast)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];