radv: vectorize shader I/O

Fixes code size regressions after enabling TCS/TES for ACO.

fossil-db (Sienna):
Totals from 2588 (1.86% of 138791) affected shaders:
SGPRs: 109950 -> 108480 (-1.34%); split: -1.43%, +0.09%
VGPRs: 107764 -> 112060 (+3.99%); split: -0.03%, +4.02%
CodeSize: 5957760 -> 5321656 (-10.68%)
MaxWaves: 31718 -> 30358 (-4.29%); split: +0.03%, -4.32%
Instrs: 1116300 -> 1029000 (-7.82%)
Cycles: 4600344 -> 4251072 (-7.59%)
VMEM: 980024 -> 812978 (-17.05%); split: +1.14%, -18.18%
SMEM: 275458 -> 258227 (-6.26%); split: +2.34%, -8.60%
VClause: 42925 -> 30533 (-28.87%); split: -31.02%, +2.15%
SClause: 31554 -> 31362 (-0.61%); split: -1.79%, +1.18%
Branches: 15689 -> 15697 (+0.05%)
PreVGPRs: 80399 -> 83953 (+4.42%); split: -0.00%, +4.42%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4202>

diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index a496672..6787604 100644 (file)
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -2330,6 +2330,10 @@ radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders,
                ordered_shaders[shader_count++] = shaders[MESA_SHADER_COMPUTE];
        }
 
+       bool has_geom_tess = shaders[MESA_SHADER_GEOMETRY] || shaders[MESA_SHADER_TESS_CTRL];
+       bool merged_gs = shaders[MESA_SHADER_GEOMETRY] &&
+                        pipeline->device->physical_device->rad_info.chip_class >= GFX9;
+
        if (!optimize_conservatively && shader_count > 1) {
                unsigned first = ordered_shaders[shader_count - 1]->info.stage;
                unsigned last = ordered_shaders[0]->info.stage;
@@ -2402,6 +2406,18 @@ radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders,
                nir_compact_varyings(ordered_shaders[i],
                                     ordered_shaders[i - 1], true);
 
+               if (ordered_shaders[i]->info.stage == MESA_SHADER_TESS_CTRL ||
+                   (ordered_shaders[i]->info.stage == MESA_SHADER_VERTEX && has_geom_tess) ||
+                   (ordered_shaders[i]->info.stage == MESA_SHADER_TESS_EVAL && merged_gs)) {
+                       nir_lower_io_to_vector(ordered_shaders[i], nir_var_shader_out);
+                       nir_opt_combine_stores(ordered_shaders[i], nir_var_shader_out);
+               }
+               if (ordered_shaders[i - 1]->info.stage == MESA_SHADER_GEOMETRY ||
+                   ordered_shaders[i - 1]->info.stage == MESA_SHADER_TESS_CTRL ||
+                   ordered_shaders[i - 1]->info.stage == MESA_SHADER_TESS_EVAL) {
+                       nir_lower_io_to_vector(ordered_shaders[i - 1], nir_var_shader_in);
+               }
+
                if (progress) {
                        if (nir_lower_global_vars_to_local(ordered_shaders[i])) {
                                ac_lower_indirect_derefs(ordered_shaders[i],