Mesa (main): radv,aco: scalarize all phis via nir_lower_phis_to_scalar()

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: dc807dff3e69d18619133b773d9a26707f55b5a2
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=dc807dff3e69d18619133b773d9a26707f55b5a2

Author: Daniel Schürmann <daniel at schuermann.dev>
Date:   Fri Jun  4 11:11:52 2021 +0200

radv,aco: scalarize all phis via nir_lower_phis_to_scalar()

This allows to remove some ACO code which did so previously.

Totals from 93 (0.06% of 149839) affected shaders (Navi2):
CodeSize: 582424 -> 582348 (-0.01%); split: -0.10%, +0.08%
Instrs: 107083 -> 107011 (-0.07%); split: -0.08%, +0.01%
Latency: 483338 -> 484881 (+0.32%); split: -0.09%, +0.40%
InvThroughput: 101129 -> 101532 (+0.40%); split: -0.03%, +0.42%
Copies: 9893 -> 9774 (-1.20%); split: -1.28%, +0.08%
Branches: 2862 -> 2858 (-0.14%)
PreSGPRs: 3342 -> 3339 (-0.09%)
PreVGPRs: 4567 -> 4565 (-0.04%)

Reviewed-by: Rhys Perry <pendingchaos02 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11181>

---

 src/amd/compiler/aco_instruction_selection.cpp     | 35 ----------------------
 .../compiler/aco_instruction_selection_setup.cpp   |  2 +-
 src/amd/vulkan/radv_shader.c                       |  2 +-
 3 files changed, 2 insertions(+), 37 deletions(-)

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 3428bd1518f..f4671a45c1f 100644
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -9498,41 +9498,6 @@ void visit_phi(isel_context *ctx, nir_phi_instr *instr)
       return;
    }
 
-   /* try to scalarize vector phis */
-   if (instr->dest.ssa.bit_size != 1 && dst.size() > 1) {
-      // TODO: scalarize linear phis on divergent ifs
-      bool can_scalarize = (opcode == aco_opcode::p_phi || !(ctx->block->kind & block_kind_merge));
-      std::array<Temp, NIR_MAX_VEC_COMPONENTS> new_vec;
-      for (unsigned i = 0; can_scalarize && (i < num_operands); i++) {
-         Operand src = operands[i];
-         if (src.isTemp() && ctx->allocated_vec.find(src.tempId()) == ctx->allocated_vec.end())
-            can_scalarize = false;
-      }
-      if (can_scalarize) {
-         unsigned num_components = instr->dest.ssa.num_components;
-         assert(dst.size() % num_components == 0);
-         RegClass rc = RegClass(dst.type(), dst.size() / num_components);
-
-         aco_ptr<Pseudo_instruction> vec{create_instruction<Pseudo_instruction>(aco_opcode::p_create_vector, Format::PSEUDO, num_components, 1)};
-         for (unsigned k = 0; k < num_components; k++) {
-            phi.reset(create_instruction<Pseudo_instruction>(opcode, Format::PSEUDO, num_operands, 1));
-            for (unsigned i = 0; i < num_operands; i++) {
-               Operand src = operands[i];
-               phi->operands[i] = src.isTemp() ? Operand(ctx->allocated_vec[src.tempId()][k]) : Operand(rc);
-            }
-            Temp phi_dst = ctx->program->allocateTmp(rc);
-            phi->definitions[0] = Definition(phi_dst);
-            ctx->block->instructions.emplace(ctx->block->instructions.begin(), std::move(phi));
-            new_vec[k] = phi_dst;
-            vec->operands[k] = Operand(phi_dst);
-         }
-         vec->definitions[0] = Definition(dst);
-         ctx->block->instructions.emplace_back(std::move(vec));
-         ctx->allocated_vec.emplace(dst.id(), new_vec);
-         return;
-      }
-   }
-
    phi.reset(create_instruction<Pseudo_instruction>(opcode, Format::PSEUDO, num_operands, 1));
    for (unsigned i = 0; i < num_operands; i++)
       phi->operands[i] = operands[i];
diff --git a/src/amd/compiler/aco_instruction_selection_setup.cpp b/src/amd/compiler/aco_instruction_selection_setup.cpp
index ba64a736a57..00fc21d8b76 100644
--- a/src/amd/compiler/aco_instruction_selection_setup.cpp
+++ b/src/amd/compiler/aco_instruction_selection_setup.cpp
@@ -509,7 +509,7 @@ setup_nir(isel_context *ctx, nir_shader *nir)
    setup_variables(ctx, nir);
 
    nir_convert_to_lcssa(nir, true, false);
-   nir_lower_phis_to_scalar(nir, false);
+   nir_lower_phis_to_scalar(nir, true);
 
    nir_function_impl *func = nir_shader_get_entrypoint(nir);
    nir_index_ssa_defs(func);
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 1eb9c281add..3b7adc38a6b 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -146,7 +146,7 @@ radv_optimize_nir(const struct radv_device *device, struct nir_shader *shader,
                nir_var_function_temp | nir_var_shader_in | nir_var_shader_out, NULL);
 
       NIR_PASS_V(shader, nir_lower_alu_to_scalar, NULL, NULL);
-      NIR_PASS_V(shader, nir_lower_phis_to_scalar, false);
+      NIR_PASS_V(shader, nir_lower_phis_to_scalar, true);
 
       NIR_PASS(progress, shader, nir_copy_prop);
       NIR_PASS(progress, shader, nir_opt_remove_phis);