#include "aco_interface.h"
#include "aco_ir.h"
#include "vulkan/radv_shader.h"
-#include "c11/threads.h"
-#include "util/debug.h"
+#include "vulkan/radv_shader_args.h"
#include <iostream>
#include <sstream>
-namespace aco {
-uint64_t debug_flags = 0;
-
-static const struct debug_control aco_debug_options[] = {
- {"validateir", DEBUG_VALIDATE},
- {"validatera", DEBUG_VALIDATE_RA},
- {"perfwarn", DEBUG_PERFWARN},
- {NULL, 0}
+static radv_compiler_statistic_info statistic_infos[] = {
+ [aco::statistic_hash] = {"Hash", "CRC32 hash of code and constant data"},
+ [aco::statistic_instructions] = {"Instructions", "Instruction count"},
+ [aco::statistic_copies] = {"Copies", "Copy instructions created for pseudo-instructions"},
+ [aco::statistic_branches] = {"Branches", "Branch instructions"},
+ [aco::statistic_cycles] = {"Busy Cycles", "Estimate of busy cycles"},
+ [aco::statistic_vmem_clauses] = {"VMEM Clause", "Number of VMEM clauses (includes 1-sized clauses)"},
+ [aco::statistic_smem_clauses] = {"SMEM Clause", "Number of SMEM clauses (includes 1-sized clauses)"},
+ [aco::statistic_vmem_score] = {"VMEM Score", "Average VMEM def-use distances"},
+ [aco::statistic_smem_score] = {"SMEM Score", "Average SMEM def-use distances"},
+ [aco::statistic_sgpr_presched] = {"Pre-Sched SGPRs", "SGPR usage before scheduling"},
+ [aco::statistic_vgpr_presched] = {"Pre-Sched VGPRs", "VGPR usage before scheduling"},
};
-static once_flag init_once_flag = ONCE_FLAG_INIT;
-
-static void init()
+static void validate(aco::Program *program)
{
- debug_flags = parse_debug_string(getenv("ACO_DEBUG"), aco_debug_options);
+ if (!(aco::debug_flags & aco::DEBUG_VALIDATE_IR))
+ return;
- #ifndef NDEBUG
- /* enable some flags by default on debug builds */
- debug_flags |= aco::DEBUG_VALIDATE;
- #endif
-}
+ ASSERTED bool is_valid = aco::validate_ir(program);
+ assert(is_valid);
}
void aco_compile_shader(unsigned shader_count,
struct nir_shader *const *shaders,
struct radv_shader_binary **binary,
- struct radv_shader_info *info,
- struct radv_nir_compiler_options *options)
+ struct radv_shader_args *args)
{
- call_once(&aco::init_once_flag, aco::init);
+ aco::init();
ac_shader_config config = {0};
std::unique_ptr<aco::Program> program{new aco::Program};
+ program->collect_statistics = args->options->record_stats;
+ if (program->collect_statistics)
+ memset(program->statistics, 0, sizeof(program->statistics));
+
+ program->debug.func = args->options->debug.func;
+ program->debug.private_data = args->options->debug.private_data;
+
/* Instruction Selection */
- aco::select_program(program.get(), shader_count, shaders, &config, info, options);
- if (options->dump_preoptir) {
+ if (args->is_gs_copy_shader)
+ aco::select_gs_copy_shader(program.get(), shaders[0], &config, args);
+ else if (args->is_trap_handler_shader)
+ aco::select_trap_handler_shader(program.get(), shaders[0], &config, args);
+ else
+ aco::select_program(program.get(), shader_count, shaders, &config, args);
+ if (args->options->dump_preoptir) {
std::cerr << "After Instruction Selection:\n";
aco_print_program(program.get(), stderr);
}
- aco::validate(program.get(), stderr);
- /* Boolean phi lowering */
- aco::lower_bool_phis(program.get());
- //std::cerr << "After Boolean Phi Lowering:\n";
- //aco_print_program(program.get(), stderr);
+ aco::live live_vars;
+ if (!args->is_trap_handler_shader) {
+ /* Phi lowering */
+ aco::lower_phis(program.get());
+ aco::dominator_tree(program.get());
+ validate(program.get());
+
+ /* Optimization */
+ if (!(aco::debug_flags & aco::DEBUG_NO_VN))
+ aco::value_numbering(program.get());
+ if (!(aco::debug_flags & aco::DEBUG_NO_OPT))
+ aco::optimize(program.get());
+
+ /* cleanup and exec mask handling */
+ aco::setup_reduce_temp(program.get());
+ aco::insert_exec_mask(program.get());
+ validate(program.get());
+
+ /* spilling and scheduling */
+ live_vars = aco::live_var_analysis(program.get(), args->options);
+ aco::spill(program.get(), live_vars, args->options);
+ }
- aco::dominator_tree(program.get());
+ std::string llvm_ir;
+ if (args->options->record_ir) {
+ char *data = NULL;
+ size_t size = 0;
+ FILE *f = open_memstream(&data, &size);
+ if (f) {
+ aco_print_program(program.get(), f);
+ fputc(0, f);
+ fclose(f);
+ }
+
+ llvm_ir = std::string(data, data + size);
+ free(data);
+ }
- /* Optimization */
- aco::value_numbering(program.get());
- aco::optimize(program.get());
- aco::validate(program.get(), stderr);
+ if (program->collect_statistics)
+ aco::collect_presched_stats(program.get());
- aco::setup_reduce_temp(program.get());
- aco::insert_exec_mask(program.get());
- aco::validate(program.get(), stderr);
+ if (!args->is_trap_handler_shader) {
+ if (!(aco::debug_flags & aco::DEBUG_NO_SCHED))
+ aco::schedule_program(program.get(), live_vars);
+ validate(program.get());
- aco::live live_vars = aco::live_var_analysis(program.get(), options);
- aco::spill(program.get(), live_vars, options);
+ /* Register Allocation */
+ aco::register_allocation(program.get(), live_vars.live_out);
+ if (args->options->dump_shader) {
+ std::cerr << "After RA:\n";
+ aco_print_program(program.get(), stderr);
+ }
- //std::cerr << "Before Schedule:\n";
- //aco_print_program(program.get(), stderr);
- aco::schedule_program(program.get(), live_vars);
+ if (aco::validate_ra(program.get(), args->options)) {
+ std::cerr << "Program after RA validation failure:\n";
+ aco_print_program(program.get(), stderr);
+ abort();
+ }
- /* Register Allocation */
- aco::register_allocation(program.get(), live_vars.live_out);
- if (options->dump_shader) {
- std::cerr << "After RA:\n";
- aco_print_program(program.get(), stderr);
- }
+ validate(program.get());
- if (aco::validate_ra(program.get(), options, stderr)) {
- std::cerr << "Program after RA validation failure:\n";
- aco_print_program(program.get(), stderr);
- abort();
+ aco::ssa_elimination(program.get());
}
- aco::ssa_elimination(program.get());
/* Lower to HW Instructions */
aco::lower_to_hw_instr(program.get());
- //std::cerr << "After Eliminate Pseudo Instr:\n";
- //aco_print_program(program.get(), stderr);
/* Insert Waitcnt */
aco::insert_wait_states(program.get());
aco::insert_NOPs(program.get());
- //std::cerr << "After Insert-Waitcnt:\n";
- //aco_print_program(program.get(), stderr);
+ if (program->collect_statistics)
+ aco::collect_preasm_stats(program.get());
/* Assembly */
std::vector<uint32_t> code;
unsigned exec_size = aco::emit_program(program.get(), code);
- bool get_disasm = options->dump_shader || options->record_llvm_ir;
+ if (program->collect_statistics)
+ aco::collect_postasm_stats(program.get(), code);
- size_t size = 0;
+ bool get_disasm = args->options->dump_shader || args->options->record_ir;
+
+ size_t size = llvm_ir.size();
std::string disasm;
if (get_disasm) {
std::ostringstream stream;
- aco::print_asm(program.get(), code, exec_size / 4u, options->family, stream);
+ aco::print_asm(program.get(), code, exec_size / 4u, stream);
stream << '\0';
disasm = stream.str();
size += disasm.size();
}
+ size_t stats_size = 0;
+ if (program->collect_statistics)
+ stats_size = sizeof(radv_compiler_statistics) + aco::num_statistics * sizeof(uint32_t);
+ size += stats_size;
+
size += code.size() * sizeof(uint32_t) + sizeof(radv_shader_binary_legacy);
- radv_shader_binary_legacy* legacy_binary = (radv_shader_binary_legacy*) malloc(size);
+ /* We need to calloc to prevent unintialized data because this will be used
+ * directly for the disk cache. Uninitialized data can appear because of
+ * padding in the struct or because legacy_binary->data can be at an offset
+ * from the start less than sizeof(radv_shader_binary_legacy). */
+ radv_shader_binary_legacy* legacy_binary = (radv_shader_binary_legacy*) calloc(size, 1);
legacy_binary->base.type = RADV_BINARY_TYPE_LEGACY;
legacy_binary->base.stage = shaders[shader_count-1]->info.stage;
- legacy_binary->base.is_gs_copy_shader = false;
+ legacy_binary->base.is_gs_copy_shader = args->is_gs_copy_shader;
legacy_binary->base.total_size = size;
- memcpy(legacy_binary->data, code.data(), code.size() * sizeof(uint32_t));
+ if (program->collect_statistics) {
+ radv_compiler_statistics *statistics = (radv_compiler_statistics *)legacy_binary->data;
+ statistics->count = aco::num_statistics;
+ statistics->infos = statistic_infos;
+ memcpy(statistics->values, program->statistics, aco::num_statistics * sizeof(uint32_t));
+ }
+ legacy_binary->stats_size = stats_size;
+
+ memcpy(legacy_binary->data + legacy_binary->stats_size, code.data(), code.size() * sizeof(uint32_t));
legacy_binary->exec_size = exec_size;
legacy_binary->code_size = code.size() * sizeof(uint32_t);
legacy_binary->config = config;
legacy_binary->disasm_size = 0;
- legacy_binary->llvm_ir_size = 0;
+ legacy_binary->ir_size = llvm_ir.size();
+
+ llvm_ir.copy((char*) legacy_binary->data + legacy_binary->stats_size + legacy_binary->code_size, llvm_ir.size());
if (get_disasm) {
- disasm.copy((char*) legacy_binary->data + legacy_binary->code_size, disasm.size());
- legacy_binary->disasm_size = disasm.size() - 1;
+ disasm.copy((char*) legacy_binary->data + legacy_binary->stats_size + legacy_binary->code_size + llvm_ir.size(), disasm.size());
+ legacy_binary->disasm_size = disasm.size();
}
*binary = (radv_shader_binary*) legacy_binary;
projects / mesa.git / blobdiff