commit dc0580d3441152336134d4840a7b70389bd67cd5
author Rhys Perry <pendingchaos02@gmail.com> Fri Nov 13 15:10:58 2020 +0000
committer Dylan Baker <dylan.c.baker@intel.com> Tue Nov 17 10:57:32 2020 -0800
tree 2bad9963b41025233f398115828d1426e9276636
parent ded5cd528a9567fce68909a60c19dbd6b65c1211

aco: disable omod if the sign of zeros should be preserved

The RDNA ISA doc says that omod doesn't preserve -0.0 in 6.2.2. LLVM
appears to always disable omod in this situation, but clamp is unaffected.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Fixes: df645fa369d ("aco: implement VK_KHR_shader_float_controls")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7605>
(cherry picked from commit 558daa73f9adb1275ddcf00515c7f79f726b7ae1)

.pick_status.json

diff --git a/.pick_status.json b/.pick_status.json
index fe67855..6f40c83 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -706,7 +706,7 @@
         "description": "aco: disable omod if the sign of zeros should be preserved",
         "nominated": true,
         "nomination_type": 1,
-        "resolution": 0,
+        "resolution": 1,
         "master_sha": null,
         "because_sha": "df645fa369d12be4d5e0fd9e4f6d4455caf2f4c3"
     },

src/amd/compiler/aco_optimizer.cpp

diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 3f2873a..97eec28 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -2548,9 +2548,14 @@
    if (!instr->isSDWA() && !can_vop3)
       return false;
 
-   /* omod has no effect if denormals are enabled */
-   bool can_use_omod = (instr->definitions[0].bytes() == 4 ? block.fp_mode.denorm32 : block.fp_mode.denorm16_64) == 0;
-   can_use_omod = can_use_omod && (can_vop3 || ctx.program->chip_class >= GFX9); /* SDWA omod is GFX9+ */
+   /* omod flushes -0 to +0 and has no effect if denormals are enabled */
+   bool can_use_omod = (can_vop3 || ctx.program->chip_class >= GFX9); /* SDWA omod is GFX9+ */
+   if (instr->definitions[0].bytes() == 4)
+      can_use_omod = can_use_omod && block.fp_mode.denorm32 == 0 &&
+                     !block.fp_mode.preserve_signed_zero_inf_nan32;
+   else
+      can_use_omod = can_use_omod && block.fp_mode.denorm16_64 == 0 &&
+                     !block.fp_mode.preserve_signed_zero_inf_nan16_64;
 
    ssa_info& def_info = ctx.info[instr->definitions[0].tempId()];