Mesa (main): radv: remove exports without color attachment or writemask

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: fce6ca0f3a674f7531598c1588ed100ec80c89c1
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=fce6ca0f3a674f7531598c1588ed100ec80c89c1

Author: Daniel Schürmann <daniel at schuermann.dev>
Date:   Fri Jan 28 14:29:05 2022 +0100

radv: remove exports without color attachment or writemask

This lets us make use of NIR's more advanced DCE.
This includes removing of CF constructs, PS inputs and VS outputs.

Totals from 1959 (1.45% of 134913) affected shaders: (GFX10.3)
VGPRs: 73464 -> 71944 (-2.07%); split: -3.79%, +1.72%
SpillSGPRs: 6 -> 0 (-inf%)
CodeSize: 4860324 -> 4675248 (-3.81%); split: -4.92%, +1.11%
LDS: 2619904 -> 2781696 (+6.18%); split: -0.37%, +6.55%
MaxWaves: 50614 -> 50852 (+0.47%); split: +1.63%, -1.16%
Instrs: 924233 -> 887836 (-3.94%); split: -5.01%, +1.07%
Latency: 5635532 -> 5418083 (-3.86%); split: -4.53%, +0.67%
InvThroughput: 1107764 -> 1077542 (-2.73%); split: -3.44%, +0.71%
VClause: 17361 -> 16163 (-6.90%); split: -8.38%, +1.47%
SClause: 31886 -> 29323 (-8.04%); split: -8.52%, +0.48%
Copies: 53529 -> 52127 (-2.62%); split: -5.30%, +2.68%
Branches: 22993 -> 22802 (-0.83%); split: -3.44%, +2.61%
PreSGPRs: 53123 -> 51395 (-3.25%); split: -3.60%, +0.35%
PreVGPRs: 59699 -> 57424 (-3.81%); split: -5.13%, +1.32%

Reviewed-by: Timur Kristóf <timur.kristof at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14771>

---

 src/amd/vulkan/radv_pipeline.c | 45 +++++++++++++++++++++++++++++++++++-------
 1 file changed, 38 insertions(+), 7 deletions(-)

diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index b372079c37b..3e76366c2e4 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -2586,15 +2586,46 @@ radv_link_shaders(struct radv_pipeline *pipeline,
       }
    }
 
-   bool uses_xfb = pipeline->graphics.last_vgt_api_stage != -1 &&
-                   radv_nir_stage_uses_xfb(shaders[pipeline->graphics.last_vgt_api_stage]);
-   if (!uses_xfb && !optimize_conservatively) {
-      /* Remove PSIZ from shaders when it's not needed.
-       * This is typically produced by translation layers like Zink or D9VK.
-       */
+   if (!optimize_conservatively) {
+      bool uses_xfb = pipeline->graphics.last_vgt_api_stage != -1 &&
+                      radv_nir_stage_uses_xfb(shaders[pipeline->graphics.last_vgt_api_stage]);
+
       for (unsigned i = 0; i < shader_count; ++i) {
          shader_info *info = &ordered_shaders[i]->info;
-         if (!(info->outputs_written & VARYING_BIT_PSIZ))
+
+         /* Remove exports without color attachment or writemask. */
+         if (info->stage == MESA_SHADER_FRAGMENT) {
+            bool fixup_derefs = false;
+            nir_foreach_variable_with_modes(var, ordered_shaders[i], nir_var_shader_out) {
+               int idx = var->data.location;
+               idx -= FRAG_RESULT_DATA0;
+               if (idx < 0)
+                  continue;
+
+               unsigned col_format = (pipeline_key->ps.col_format >> (4 * idx)) & 0xf;
+               switch (col_format) {
+               case V_028714_SPI_SHADER_ZERO:
+                  info->outputs_written &= ~BITFIELD64_BIT(var->data.location);
+                  var->data.location = 0;
+                  var->data.mode = nir_var_shader_temp;
+                  fixup_derefs = true;
+                  break;
+               default:
+                  break;
+               }
+            }
+            if (fixup_derefs) {
+               nir_fixup_deref_modes(ordered_shaders[i]);
+               nir_remove_dead_variables(ordered_shaders[i], nir_var_shader_temp, NULL);
+               nir_opt_dce(ordered_shaders[i]);
+            }
+            continue;
+         }
+
+         /* Remove PSIZ from shaders when it's not needed.
+          * This is typically produced by translation layers like Zink or D9VK.
+          */
+         if (uses_xfb || !(info->outputs_written & VARYING_BIT_PSIZ))
             continue;
 
          bool next_stage_needs_psiz =