Module: Mesa Branch: main Commit: d200916ca2f656a5f81d6713a9073d7f66a5fa7a URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=d200916ca2f656a5f81d6713a9073d7f66a5fa7a
Author: Rhys Perry <pendingchao...@gmail.com> Date: Wed Oct 4 17:28:38 2023 +0100 aco: add VALU/SALU/VMEM/SMEM statistics This lets us measure optimizations without interference of waitcnt instructions. Signed-off-by: Rhys Perry <pendingchao...@gmail.com> Reviewed-by: Daniel Schürmann <dan...@schuermann.dev> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25559> --- src/amd/compiler/aco_interface.cpp | 4 ++++ src/amd/compiler/aco_shader_info.h | 4 ++++ src/amd/compiler/aco_statistics.cpp | 14 +++++++++++--- 3 files changed, 19 insertions(+), 3 deletions(-) diff --git a/src/amd/compiler/aco_interface.cpp b/src/amd/compiler/aco_interface.cpp index 8b4950823e4..bdf74ca5dd1 100644 --- a/src/amd/compiler/aco_interface.cpp +++ b/src/amd/compiler/aco_interface.cpp @@ -55,6 +55,10 @@ static const std::array<aco_compiler_statistic_info, aco_num_statistics> statist aco_compiler_statistic_info{"Pre-Sched SGPRs", "SGPR usage before scheduling"}; ret[aco_statistic_vgpr_presched] = aco_compiler_statistic_info{"Pre-Sched VGPRs", "VGPR usage before scheduling"}; + ret[aco_statistic_valu] = aco_compiler_statistic_info{"VALU", "Number of VALU instructions"}; + ret[aco_statistic_salu] = aco_compiler_statistic_info{"SALU", "Number of SALU instructions"}; + ret[aco_statistic_vmem] = aco_compiler_statistic_info{"VMEM", "Number of VMEM instructions"}; + ret[aco_statistic_smem] = aco_compiler_statistic_info{"SMEM", "Number of SMEM instructions"}; return ret; }(); diff --git a/src/amd/compiler/aco_shader_info.h b/src/amd/compiler/aco_shader_info.h index a9559f9d783..2664fdae3f8 100644 --- a/src/amd/compiler/aco_shader_info.h +++ b/src/amd/compiler/aco_shader_info.h @@ -226,6 +226,10 @@ enum aco_statistic { aco_statistic_smem_clauses, aco_statistic_sgpr_presched, aco_statistic_vgpr_presched, + aco_statistic_valu, + aco_statistic_salu, + aco_statistic_vmem, + aco_statistic_smem, aco_num_statistics }; diff --git a/src/amd/compiler/aco_statistics.cpp b/src/amd/compiler/aco_statistics.cpp index 89303a43809..e454c2de93b 100644 --- a/src/amd/compiler/aco_statistics.cpp +++ b/src/amd/compiler/aco_statistics.cpp @@ -531,11 +531,15 @@ collect_preasm_stats(Program* program) program->statistics[aco_statistic_instructions] += block.instructions.size(); for (aco_ptr<Instruction>& instr : block.instructions) { - if (instr->isSOPP() && instr->sopp().block != -1) + bool is_branch = instr->isSOPP() && instr->sopp().block != -1; + if (is_branch) program->statistics[aco_statistic_branches]++; - if (instr->opcode == aco_opcode::p_constaddr) - program->statistics[aco_statistic_instructions] += 2; + if (instr->isVALU() || instr->isVINTRP()) + program->statistics[aco_statistic_valu]++; + if (instr->isSALU() && !instr->isSOPP() && + instr_info.classes[(int)instr->opcode] != instr_class::waitcnt) + program->statistics[aco_statistic_salu]++; if ((instr->isVMEM() || instr->isScratch() || instr->isGlobal()) && !instr->operands.empty()) { @@ -544,6 +548,8 @@ collect_preasm_stats(Program* program) { return should_form_clause(instr.get(), other); })) program->statistics[aco_statistic_vmem_clauses]++; vmem_clause.insert(instr.get()); + + program->statistics[aco_statistic_vmem]++; } else { vmem_clause.clear(); } @@ -554,6 +560,8 @@ collect_preasm_stats(Program* program) { return should_form_clause(instr.get(), other); })) program->statistics[aco_statistic_smem_clauses]++; smem_clause.insert(instr.get()); + + program->statistics[aco_statistic_smem]++; } else { smem_clause.clear(); }
