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_vmem_score
] = {"VMEM Score", "Average VMEM def-use distances"},
66 [aco::statistic_smem_score
] = {"SMEM Score", "Average SMEM def-use distances"},
67 [aco::statistic_sgpr_presched
] = {"Pre-Sched SGPRs", "SGPR usage before scheduling"},
68 [aco::statistic_vgpr_presched
] = {"Pre-Sched VGPRs", "VGPR usage before scheduling"},
71 void aco_compile_shader(unsigned shader_count
,
72 struct nir_shader
*const *shaders
,
73 struct radv_shader_binary
**binary
,
74 struct radv_shader_args
*args
)
76 call_once(&aco::init_once_flag
, aco::init
);
78 ac_shader_config config
= {0};
79 std::unique_ptr
<aco::Program
> program
{new aco::Program
};
81 program
->collect_statistics
= args
->options
->record_stats
;
82 if (program
->collect_statistics
)
83 memset(program
->statistics
, 0, sizeof(program
->statistics
));
85 /* Instruction Selection */
86 if (args
->is_gs_copy_shader
)
87 aco::select_gs_copy_shader(program
.get(), shaders
[0], &config
, args
);
89 aco::select_program(program
.get(), shader_count
, shaders
, &config
, args
);
90 if (args
->options
->dump_preoptir
) {
91 std::cerr
<< "After Instruction Selection:\n";
92 aco_print_program(program
.get(), stderr
);
94 aco::validate(program
.get(), stderr
);
96 /* Boolean phi lowering */
97 aco::lower_bool_phis(program
.get());
98 //std::cerr << "After Boolean Phi Lowering:\n";
99 //aco_print_program(program.get(), stderr);
101 aco::dominator_tree(program
.get());
104 aco::value_numbering(program
.get());
105 aco::optimize(program
.get());
106 aco::validate(program
.get(), stderr
);
108 aco::setup_reduce_temp(program
.get());
109 aco::insert_exec_mask(program
.get());
110 aco::validate(program
.get(), stderr
);
112 aco::live live_vars
= aco::live_var_analysis(program
.get(), args
->options
);
113 aco::spill(program
.get(), live_vars
, args
->options
);
115 if (program
->collect_statistics
)
116 aco::collect_presched_stats(program
.get());
118 //std::cerr << "Before Schedule:\n";
119 //aco_print_program(program.get(), stderr);
120 aco::schedule_program(program
.get(), live_vars
);
123 if (args
->options
->record_ir
) {
126 FILE *f
= open_memstream(&data
, &size
);
128 aco_print_program(program
.get(), f
);
133 llvm_ir
= std::string(data
, data
+ size
);
137 /* Register Allocation */
138 aco::register_allocation(program
.get(), live_vars
.live_out
);
139 if (args
->options
->dump_shader
) {
140 std::cerr
<< "After RA:\n";
141 aco_print_program(program
.get(), stderr
);
144 if (aco::validate_ra(program
.get(), args
->options
, stderr
)) {
145 std::cerr
<< "Program after RA validation failure:\n";
146 aco_print_program(program
.get(), stderr
);
150 aco::ssa_elimination(program
.get());
151 /* Lower to HW Instructions */
152 aco::lower_to_hw_instr(program
.get());
153 //std::cerr << "After Eliminate Pseudo Instr:\n";
154 //aco_print_program(program.get(), stderr);
157 aco::insert_wait_states(program
.get());
158 aco::insert_NOPs(program
.get());
160 //std::cerr << "After Insert-Waitcnt:\n";
161 //aco_print_program(program.get(), stderr);
163 if (program
->collect_statistics
)
164 aco::collect_preasm_stats(program
.get());
167 std::vector
<uint32_t> code
;
168 unsigned exec_size
= aco::emit_program(program
.get(), code
);
170 if (program
->collect_statistics
)
171 aco::collect_postasm_stats(program
.get(), code
);
173 bool get_disasm
= args
->options
->dump_shader
|| args
->options
->record_ir
;
175 size_t size
= llvm_ir
.size();
179 std::ostringstream stream
;
180 aco::print_asm(program
.get(), code
, exec_size
/ 4u, stream
);
182 disasm
= stream
.str();
183 size
+= disasm
.size();
186 size_t stats_size
= 0;
187 if (program
->collect_statistics
)
188 stats_size
= sizeof(radv_compiler_statistics
) + aco::num_statistics
* sizeof(uint32_t);
191 size
+= code
.size() * sizeof(uint32_t) + sizeof(radv_shader_binary_legacy
);
192 /* We need to calloc to prevent unintialized data because this will be used
193 * directly for the disk cache. Uninitialized data can appear because of
194 * padding in the struct or because legacy_binary->data can be at an offset
195 * from the start less than sizeof(radv_shader_binary_legacy). */
196 radv_shader_binary_legacy
* legacy_binary
= (radv_shader_binary_legacy
*) calloc(size
, 1);
198 legacy_binary
->base
.type
= RADV_BINARY_TYPE_LEGACY
;
199 legacy_binary
->base
.stage
= shaders
[shader_count
-1]->info
.stage
;
200 legacy_binary
->base
.is_gs_copy_shader
= args
->is_gs_copy_shader
;
201 legacy_binary
->base
.total_size
= size
;
203 if (program
->collect_statistics
) {
204 radv_compiler_statistics
*statistics
= (radv_compiler_statistics
*)legacy_binary
->data
;
205 statistics
->count
= aco::num_statistics
;
206 statistics
->infos
= statistic_infos
;
207 memcpy(statistics
->values
, program
->statistics
, aco::num_statistics
* sizeof(uint32_t));
209 legacy_binary
->stats_size
= stats_size
;
211 memcpy(legacy_binary
->data
+ legacy_binary
->stats_size
, code
.data(), code
.size() * sizeof(uint32_t));
212 legacy_binary
->exec_size
= exec_size
;
213 legacy_binary
->code_size
= code
.size() * sizeof(uint32_t);
215 legacy_binary
->config
= config
;
216 legacy_binary
->disasm_size
= 0;
217 legacy_binary
->ir_size
= llvm_ir
.size();
219 llvm_ir
.copy((char*) legacy_binary
->data
+ legacy_binary
->stats_size
+ legacy_binary
->code_size
, llvm_ir
.size());
222 disasm
.copy((char*) legacy_binary
->data
+ legacy_binary
->stats_size
+ legacy_binary
->code_size
+ llvm_ir
.size(), disasm
.size());
223 legacy_binary
->disasm_size
= disasm
.size();
226 *binary
= (radv_shader_binary
*) legacy_binary
;