2 * Copyright © 2018 Google
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "aco_interface.h"
26 #include "vulkan/radv_shader.h"
27 #include "vulkan/radv_shader_args.h"
28 #include "c11/threads.h"
29 #include "util/debug.h"
35 uint64_t debug_flags
= 0;
37 static const struct debug_control aco_debug_options
[] = {
38 {"validateir", DEBUG_VALIDATE
},
39 {"validatera", DEBUG_VALIDATE_RA
},
40 {"perfwarn", DEBUG_PERFWARN
},
44 static once_flag init_once_flag
= ONCE_FLAG_INIT
;
48 debug_flags
= parse_debug_string(getenv("ACO_DEBUG"), aco_debug_options
);
51 /* enable some flags by default on debug builds */
52 debug_flags
|= aco::DEBUG_VALIDATE
;
57 static radv_compiler_statistic_info statistic_infos
[] = {
58 [aco::statistic_hash
] = {"Hash", "CRC32 hash of code and constant data"},
59 [aco::statistic_instructions
] = {"Instructions", "Instruction count"},
60 [aco::statistic_copies
] = {"Copies", "Copy instructions created for pseudo-instructions"},
61 [aco::statistic_branches
] = {"Branches", "Branch instructions"},
62 [aco::statistic_cycles
] = {"Busy Cycles", "Estimate of busy cycles"},
63 [aco::statistic_vmem_clauses
] = {"VMEM Clause", "Number of VMEM clauses (includes 1-sized clauses)"},
64 [aco::statistic_smem_clauses
] = {"SMEM Clause", "Number of SMEM clauses (includes 1-sized clauses)"},
65 [aco::statistic_sgpr_presched
] = {"Pre-Sched SGPRs", "SGPR usage before scheduling"},
66 [aco::statistic_vgpr_presched
] = {"Pre-Sched VGPRs", "VGPR usage before scheduling"},
69 void aco_compile_shader(unsigned shader_count
,
70 struct nir_shader
*const *shaders
,
71 struct radv_shader_binary
**binary
,
72 struct radv_shader_args
*args
)
74 call_once(&aco::init_once_flag
, aco::init
);
76 ac_shader_config config
= {0};
77 std::unique_ptr
<aco::Program
> program
{new aco::Program
};
79 program
->collect_statistics
= args
->options
->record_ir
;
80 if (program
->collect_statistics
)
81 memset(program
->statistics
, 0, sizeof(program
->statistics
));
83 /* Instruction Selection */
84 if (args
->is_gs_copy_shader
)
85 aco::select_gs_copy_shader(program
.get(), shaders
[0], &config
, args
);
87 aco::select_program(program
.get(), shader_count
, shaders
, &config
, args
);
88 if (args
->options
->dump_preoptir
) {
89 std::cerr
<< "After Instruction Selection:\n";
90 aco_print_program(program
.get(), stderr
);
92 aco::validate(program
.get(), stderr
);
94 /* Boolean phi lowering */
95 aco::lower_bool_phis(program
.get());
96 //std::cerr << "After Boolean Phi Lowering:\n";
97 //aco_print_program(program.get(), stderr);
99 aco::dominator_tree(program
.get());
102 aco::value_numbering(program
.get());
103 aco::optimize(program
.get());
104 aco::validate(program
.get(), stderr
);
106 aco::setup_reduce_temp(program
.get());
107 aco::insert_exec_mask(program
.get());
108 aco::validate(program
.get(), stderr
);
110 aco::live live_vars
= aco::live_var_analysis(program
.get(), args
->options
);
111 aco::spill(program
.get(), live_vars
, args
->options
);
113 if (program
->collect_statistics
)
114 aco::collect_presched_stats(program
.get());
116 //std::cerr << "Before Schedule:\n";
117 //aco_print_program(program.get(), stderr);
118 aco::schedule_program(program
.get(), live_vars
);
121 if (args
->options
->record_ir
) {
124 FILE *f
= open_memstream(&data
, &size
);
126 aco_print_program(program
.get(), f
);
131 llvm_ir
= std::string(data
, data
+ size
);
135 /* Register Allocation */
136 aco::register_allocation(program
.get(), live_vars
.live_out
);
137 if (args
->options
->dump_shader
) {
138 std::cerr
<< "After RA:\n";
139 aco_print_program(program
.get(), stderr
);
142 if (aco::validate_ra(program
.get(), args
->options
, stderr
)) {
143 std::cerr
<< "Program after RA validation failure:\n";
144 aco_print_program(program
.get(), stderr
);
148 aco::ssa_elimination(program
.get());
149 /* Lower to HW Instructions */
150 aco::lower_to_hw_instr(program
.get());
151 //std::cerr << "After Eliminate Pseudo Instr:\n";
152 //aco_print_program(program.get(), stderr);
155 aco::insert_wait_states(program
.get());
156 aco::insert_NOPs(program
.get());
158 //std::cerr << "After Insert-Waitcnt:\n";
159 //aco_print_program(program.get(), stderr);
161 if (program
->collect_statistics
)
162 aco::collect_preasm_stats(program
.get());
165 std::vector
<uint32_t> code
;
166 unsigned exec_size
= aco::emit_program(program
.get(), code
);
168 if (program
->collect_statistics
)
169 aco::collect_postasm_stats(program
.get(), code
);
171 bool get_disasm
= args
->options
->dump_shader
|| args
->options
->record_ir
;
173 size_t size
= llvm_ir
.size();
177 std::ostringstream stream
;
178 aco::print_asm(program
.get(), code
, exec_size
/ 4u, stream
);
180 disasm
= stream
.str();
181 size
+= disasm
.size();
184 size_t stats_size
= 0;
185 if (program
->collect_statistics
)
186 stats_size
= sizeof(radv_compiler_statistics
) + aco::num_statistics
* sizeof(uint32_t);
189 size
+= code
.size() * sizeof(uint32_t) + sizeof(radv_shader_binary_legacy
);
190 /* We need to calloc to prevent unintialized data because this will be used
191 * directly for the disk cache. Uninitialized data can appear because of
192 * padding in the struct or because legacy_binary->data can be at an offset
193 * from the start less than sizeof(radv_shader_binary_legacy). */
194 radv_shader_binary_legacy
* legacy_binary
= (radv_shader_binary_legacy
*) calloc(size
, 1);
196 legacy_binary
->base
.type
= RADV_BINARY_TYPE_LEGACY
;
197 legacy_binary
->base
.stage
= shaders
[shader_count
-1]->info
.stage
;
198 legacy_binary
->base
.is_gs_copy_shader
= args
->is_gs_copy_shader
;
199 legacy_binary
->base
.total_size
= size
;
201 if (program
->collect_statistics
) {
202 radv_compiler_statistics
*statistics
= (radv_compiler_statistics
*)legacy_binary
->data
;
203 statistics
->count
= aco::num_statistics
;
204 statistics
->infos
= statistic_infos
;
205 memcpy(statistics
->values
, program
->statistics
, aco::num_statistics
* sizeof(uint32_t));
207 legacy_binary
->stats_size
= stats_size
;
209 memcpy(legacy_binary
->data
+ legacy_binary
->stats_size
, code
.data(), code
.size() * sizeof(uint32_t));
210 legacy_binary
->exec_size
= exec_size
;
211 legacy_binary
->code_size
= code
.size() * sizeof(uint32_t);
213 legacy_binary
->config
= config
;
214 legacy_binary
->disasm_size
= 0;
215 legacy_binary
->ir_size
= llvm_ir
.size();
217 llvm_ir
.copy((char*) legacy_binary
->data
+ legacy_binary
->stats_size
+ legacy_binary
->code_size
, llvm_ir
.size());
220 disasm
.copy((char*) legacy_binary
->data
+ legacy_binary
->stats_size
+ legacy_binary
->code_size
+ llvm_ir
.size(), disasm
.size());
221 legacy_binary
->disasm_size
= disasm
.size();
224 *binary
= (radv_shader_binary
*) legacy_binary
;