Mesa (master): nir,radv: add and use nir_vectorize_tess_levels()

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: f5adf27fb926a330a13af716f0a03da1a224656d
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=f5adf27fb926a330a13af716f0a03da1a224656d

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Wed Jan  6 14:50:57 2021 +0000

nir,radv: add and use nir_vectorize_tess_levels()

fossil-db (Sienna):
Totals from 1342 (0.97% of 138791) affected shaders:
CodeSize: 3287996 -> 3269572 (-0.56%); split: -0.56%, +0.00%
Instrs: 629896 -> 628191 (-0.27%); split: -0.31%, +0.04%
Cycles: 2619244 -> 2612424 (-0.26%); split: -0.30%, +0.04%
VMEM: 388807 -> 389273 (+0.12%); split: +0.14%, -0.02%
SMEM: 90655 -> 90700 (+0.05%); split: +0.06%, -0.01%
VClause: 21831 -> 21812 (-0.09%)
PreVGPRs: 44155 -> 44058 (-0.22%)

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

---

 src/amd/vulkan/radv_pipeline.c            |  2 +
 src/compiler/nir/nir.h                    |  1 +
 src/compiler/nir/nir_lower_io_to_vector.c | 82 +++++++++++++++++++++++++++++++
 3 files changed, 85 insertions(+)

diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index 678760487d0..63afe44c5df 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -2410,6 +2410,8 @@ radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders,
 		    (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);
+			if (ordered_shaders[i]->info.stage == MESA_SHADER_TESS_CTRL)
+				nir_vectorize_tess_levels(ordered_shaders[i]);
 			nir_opt_combine_stores(ordered_shaders[i], nir_var_shader_out);
 		}
 		if (ordered_shaders[i - 1]->info.stage == MESA_SHADER_GEOMETRY ||
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 04e54d637cf..81450cf62a2 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -4485,6 +4485,7 @@ void nir_lower_io_arrays_to_elements_no_indirects(nir_shader *shader,
 void nir_lower_io_to_scalar(nir_shader *shader, nir_variable_mode mask);
 bool nir_lower_io_to_scalar_early(nir_shader *shader, nir_variable_mode mask);
 bool nir_lower_io_to_vector(nir_shader *shader, nir_variable_mode mask);
+bool nir_vectorize_tess_levels(nir_shader *shader);
 
 bool nir_lower_fragcolor(nir_shader *shader);
 bool nir_lower_fragcoord_wtrans(nir_shader *shader);
diff --git a/src/compiler/nir/nir_lower_io_to_vector.c b/src/compiler/nir/nir_lower_io_to_vector.c
index c201e16a0ec..3d912e1d348 100644
--- a/src/compiler/nir/nir_lower_io_to_vector.c
+++ b/src/compiler/nir/nir_lower_io_to_vector.c
@@ -574,3 +574,85 @@ nir_lower_io_to_vector(nir_shader *shader, nir_variable_mode modes)
 
    return progress;
 }
+
+static bool
+nir_vectorize_tess_levels_impl(nir_function_impl *impl)
+{
+   bool progress = false;
+   nir_builder b;
+   nir_builder_init(&b, impl);
+
+   nir_foreach_block(block, impl) {
+      nir_foreach_instr(instr, block) {
+         if (instr->type != nir_instr_type_intrinsic)
+            continue;
+
+         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+         if (intrin->intrinsic != nir_intrinsic_load_deref &&
+             intrin->intrinsic != nir_intrinsic_store_deref)
+            continue;
+
+         nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+         if (!nir_deref_mode_is(deref, nir_var_shader_out))
+            continue;
+
+         nir_variable *var = nir_deref_instr_get_variable(deref);
+         if (var->data.location != VARYING_SLOT_TESS_LEVEL_OUTER &&
+             var->data.location != VARYING_SLOT_TESS_LEVEL_INNER)
+            continue;
+
+         assert(deref->deref_type == nir_deref_type_array);
+         assert(nir_src_is_const(deref->arr.index));
+         unsigned index = nir_src_as_uint(deref->arr.index);;
+
+         b.cursor = nir_before_instr(instr);
+         nir_ssa_def *new_deref = &nir_build_deref_var(&b, var)->dest.ssa;
+         nir_instr_rewrite_src(instr, &intrin->src[0], nir_src_for_ssa(new_deref));
+
+         nir_deref_instr_remove_if_unused(deref);
+
+         intrin->num_components = glsl_get_vector_elements(var->type);
+
+         if (intrin->intrinsic == nir_intrinsic_store_deref) {
+            nir_intrinsic_set_write_mask(intrin, 1 << index);
+            nir_ssa_def *new_val = nir_ssa_undef(&b, intrin->num_components, 32);
+            new_val = nir_vector_insert_imm(&b, new_val, intrin->src[1].ssa, index);
+            nir_instr_rewrite_src(instr, &intrin->src[1], nir_src_for_ssa(new_val));
+         } else {
+            b.cursor = nir_after_instr(instr);
+            nir_ssa_def *val = &intrin->dest.ssa;
+            nir_ssa_def *comp = nir_channel(&b, val, index);
+            nir_ssa_def_rewrite_uses_after(val, nir_src_for_ssa(comp), comp->parent_instr);
+         }
+
+         progress = true;
+      }
+   }
+
+   return progress;
+}
+
+/* Make the tess factor variables vectors instead of compact arrays, so accesses
+ * can be combined by nir_opt_cse()/nir_opt_combine_stores().
+ */
+bool
+nir_vectorize_tess_levels(nir_shader *shader)
+{
+   bool progress = false;
+
+   nir_foreach_shader_out_variable(var, shader) {
+      if (var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER ||
+          var->data.location == VARYING_SLOT_TESS_LEVEL_INNER) {
+         var->type = glsl_vector_type(GLSL_TYPE_FLOAT, glsl_get_length(var->type));
+         var->data.compact = false;
+         progress = true;
+      }
+   }
+
+   nir_foreach_function(function, shader) {
+      if (function->impl)
+         progress |= nir_vectorize_tess_levels_impl(function->impl);
+   }
+
+   return progress;
+}