1 /*
2 * Copyright © 2020 Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 #include "brw_kernel.h"
25 #include "brw_nir.h"
26 #include "elk/elk_nir_options.h"
27 #include "intel_nir.h"
28
29 #include "intel_nir.h"
30 #include "nir_clc_helpers.h"
31 #include "compiler/nir/nir_builder.h"
32 #include "compiler/spirv/nir_spirv.h"
33 #include "compiler/spirv/spirv_info.h"
34 #include "dev/intel_debug.h"
35 #include "util/u_atomic.h"
36 #include "util/u_dynarray.h"
37
38 static const nir_shader *
load_clc_shader(struct brw_compiler * compiler,struct disk_cache * disk_cache,const nir_shader_compiler_options * nir_options,const struct spirv_to_nir_options * spirv_options)39 load_clc_shader(struct brw_compiler *compiler, struct disk_cache *disk_cache,
40 const nir_shader_compiler_options *nir_options,
41 const struct spirv_to_nir_options *spirv_options)
42 {
43 if (compiler->clc_shader)
44 return compiler->clc_shader;
45
46 nir_shader *nir = nir_load_libclc_shader(64, disk_cache,
47 spirv_options, nir_options,
48 disk_cache != NULL);
49 if (nir == NULL)
50 return NULL;
51
52 const nir_shader *old_nir =
53 p_atomic_cmpxchg(&compiler->clc_shader, NULL, nir);
54 if (old_nir == NULL) {
55 /* We won the race */
56 ralloc_steal(compiler, nir);
57 return nir;
58 } else {
59 /* Someone else built the shader first */
60 ralloc_free(nir);
61 return old_nir;
62 }
63 }
64
65 static nir_builder
builder_init_new_impl(nir_function * func)66 builder_init_new_impl(nir_function *func)
67 {
68 nir_function_impl *impl = nir_function_impl_create(func);
69 return nir_builder_at(nir_before_impl(impl));
70 }
71
72 static void
implement_atomic_builtin(nir_function * func,nir_atomic_op atomic_op,enum glsl_base_type data_base_type,nir_variable_mode mode)73 implement_atomic_builtin(nir_function *func, nir_atomic_op atomic_op,
74 enum glsl_base_type data_base_type,
75 nir_variable_mode mode)
76 {
77 nir_builder b = builder_init_new_impl(func);
78 const struct glsl_type *data_type = glsl_scalar_type(data_base_type);
79
80 unsigned p = 0;
81
82 nir_deref_instr *ret = NULL;
83 ret = nir_build_deref_cast(&b, nir_load_param(&b, p++),
84 nir_var_function_temp, data_type, 0);
85
86 nir_intrinsic_op op = nir_intrinsic_deref_atomic;
87 nir_intrinsic_instr *atomic = nir_intrinsic_instr_create(b.shader, op);
88 nir_intrinsic_set_atomic_op(atomic, atomic_op);
89
90 for (unsigned i = 0; i < nir_intrinsic_infos[op].num_srcs; i++) {
91 nir_def *src = nir_load_param(&b, p++);
92 if (i == 0) {
93 /* The first source is our deref */
94 assert(nir_intrinsic_infos[op].src_components[i] == -1);
95 src = &nir_build_deref_cast(&b, src, mode, data_type, 0)->def;
96 }
97 atomic->src[i] = nir_src_for_ssa(src);
98 }
99
100 nir_def_init_for_type(&atomic->instr, &atomic->def, data_type);
101
102 nir_builder_instr_insert(&b, &atomic->instr);
103 nir_store_deref(&b, ret, &atomic->def, ~0);
104 }
105
106 static void
implement_sub_group_ballot_builtin(nir_function * func)107 implement_sub_group_ballot_builtin(nir_function *func)
108 {
109 nir_builder b = builder_init_new_impl(func);
110 nir_deref_instr *ret =
111 nir_build_deref_cast(&b, nir_load_param(&b, 0),
112 nir_var_function_temp, glsl_uint_type(), 0);
113 nir_def *cond = nir_load_param(&b, 1);
114
115 nir_intrinsic_instr *ballot =
116 nir_intrinsic_instr_create(b.shader, nir_intrinsic_ballot);
117 ballot->src[0] = nir_src_for_ssa(cond);
118 ballot->num_components = 1;
119 nir_def_init(&ballot->instr, &ballot->def, 1, 32);
120 nir_builder_instr_insert(&b, &ballot->instr);
121
122 nir_store_deref(&b, ret, &ballot->def, ~0);
123 }
124
125 static bool
implement_intel_builtins(nir_shader * nir)126 implement_intel_builtins(nir_shader *nir)
127 {
128 bool progress = false;
129
130 nir_foreach_function(func, nir) {
131 if (strcmp(func->name, "_Z10atomic_minPU3AS1Vff") == 0) {
132 /* float atom_min(__global float volatile *p, float val) */
133 implement_atomic_builtin(func, nir_atomic_op_fmin,
134 GLSL_TYPE_FLOAT, nir_var_mem_global);
135 progress = true;
136 } else if (strcmp(func->name, "_Z10atomic_maxPU3AS1Vff") == 0) {
137 /* float atom_max(__global float volatile *p, float val) */
138 implement_atomic_builtin(func, nir_atomic_op_fmax,
139 GLSL_TYPE_FLOAT, nir_var_mem_global);
140 progress = true;
141 } else if (strcmp(func->name, "_Z10atomic_minPU3AS3Vff") == 0) {
142 /* float atomic_min(__shared float volatile *, float) */
143 implement_atomic_builtin(func, nir_atomic_op_fmin,
144 GLSL_TYPE_FLOAT, nir_var_mem_shared);
145 progress = true;
146 } else if (strcmp(func->name, "_Z10atomic_maxPU3AS3Vff") == 0) {
147 /* float atomic_max(__shared float volatile *, float) */
148 implement_atomic_builtin(func, nir_atomic_op_fmax,
149 GLSL_TYPE_FLOAT, nir_var_mem_shared);
150 progress = true;
151 } else if (strcmp(func->name, "intel_sub_group_ballot") == 0) {
152 implement_sub_group_ballot_builtin(func);
153 progress = true;
154 }
155 }
156
157 nir_shader_preserve_all_metadata(nir);
158
159 return progress;
160 }
161
162 static bool
lower_kernel_intrinsics(nir_shader * nir)163 lower_kernel_intrinsics(nir_shader *nir)
164 {
165 nir_function_impl *impl = nir_shader_get_entrypoint(nir);
166
167 bool progress = false;
168
169 unsigned kernel_sysvals_start = 0;
170 unsigned kernel_arg_start = sizeof(struct brw_kernel_sysvals);
171 nir->num_uniforms += kernel_arg_start;
172
173 nir_builder b = nir_builder_create(impl);
174
175 nir_foreach_block(block, impl) {
176 nir_foreach_instr_safe(instr, block) {
177 if (instr->type != nir_instr_type_intrinsic)
178 continue;
179
180 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
181 switch (intrin->intrinsic) {
182 case nir_intrinsic_load_kernel_input: {
183 b.cursor = nir_instr_remove(&intrin->instr);
184
185 nir_intrinsic_instr *load =
186 nir_intrinsic_instr_create(nir, nir_intrinsic_load_uniform);
187 load->num_components = intrin->num_components;
188 load->src[0] = nir_src_for_ssa(nir_u2u32(&b, intrin->src[0].ssa));
189 nir_intrinsic_set_base(load, kernel_arg_start);
190 nir_intrinsic_set_range(load, nir->num_uniforms);
191 nir_def_init(&load->instr, &load->def,
192 intrin->def.num_components,
193 intrin->def.bit_size);
194 nir_builder_instr_insert(&b, &load->instr);
195
196 nir_def_rewrite_uses(&intrin->def, &load->def);
197 progress = true;
198 break;
199 }
200
201 case nir_intrinsic_load_constant_base_ptr: {
202 b.cursor = nir_instr_remove(&intrin->instr);
203 nir_def *const_data_base_addr = nir_pack_64_2x32_split(&b,
204 nir_load_reloc_const_intel(&b, BRW_SHADER_RELOC_CONST_DATA_ADDR_LOW),
205 nir_load_reloc_const_intel(&b, BRW_SHADER_RELOC_CONST_DATA_ADDR_HIGH));
206 nir_def_rewrite_uses(&intrin->def, const_data_base_addr);
207 progress = true;
208 break;
209 }
210
211 case nir_intrinsic_load_num_workgroups: {
212 b.cursor = nir_instr_remove(&intrin->instr);
213
214 nir_intrinsic_instr *load =
215 nir_intrinsic_instr_create(nir, nir_intrinsic_load_uniform);
216 load->num_components = 3;
217 load->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
218 nir_intrinsic_set_base(load, kernel_sysvals_start +
219 offsetof(struct brw_kernel_sysvals, num_work_groups));
220 nir_intrinsic_set_range(load, 3 * 4);
221 nir_def_init(&load->instr, &load->def, 3, 32);
222 nir_builder_instr_insert(&b, &load->instr);
223 nir_def_rewrite_uses(&intrin->def, &load->def);
224 progress = true;
225 break;
226 }
227
228 default:
229 break;
230 }
231 }
232 }
233
234 if (progress) {
235 nir_metadata_preserve(impl, nir_metadata_control_flow);
236 } else {
237 nir_metadata_preserve(impl, nir_metadata_all);
238 }
239
240 return progress;
241 }
242
243 static const struct spirv_capabilities spirv_caps = {
244 .Addresses = true,
245 .Float16 = true,
246 .Float64 = true,
247 .Groups = true,
248 .StorageImageWriteWithoutFormat = true,
249 .Int8 = true,
250 .Int16 = true,
251 .Int64 = true,
252 .Int64Atomics = true,
253 .Kernel = true,
254 .Linkage = true, /* We receive linked kernel from clc */
255 .DenormFlushToZero = true,
256 .DenormPreserve = true,
257 .SignedZeroInfNanPreserve = true,
258 .RoundingModeRTE = true,
259 .RoundingModeRTZ = true,
260 .GenericPointer = true,
261 .GroupNonUniform = true,
262 .GroupNonUniformArithmetic = true,
263 .GroupNonUniformClustered = true,
264 .GroupNonUniformBallot = true,
265 .GroupNonUniformQuad = true,
266 .GroupNonUniformShuffle = true,
267 .GroupNonUniformVote = true,
268 .SubgroupDispatch = true,
269
270 .SubgroupShuffleINTEL = true,
271 .SubgroupBufferBlockIOINTEL = true,
272 };
273
274 bool
brw_kernel_from_spirv(struct brw_compiler * compiler,struct disk_cache * disk_cache,struct brw_kernel * kernel,void * log_data,void * mem_ctx,const uint32_t * spirv,size_t spirv_size,const char * entrypoint_name,char ** error_str)275 brw_kernel_from_spirv(struct brw_compiler *compiler,
276 struct disk_cache *disk_cache,
277 struct brw_kernel *kernel,
278 void *log_data, void *mem_ctx,
279 const uint32_t *spirv, size_t spirv_size,
280 const char *entrypoint_name,
281 char **error_str)
282 {
283 const struct intel_device_info *devinfo = compiler->devinfo;
284 const nir_shader_compiler_options *nir_options =
285 compiler->nir_options[MESA_SHADER_KERNEL];
286
287 struct spirv_to_nir_options spirv_options = {
288 .environment = NIR_SPIRV_OPENCL,
289 .capabilities = &spirv_caps,
290 .printf = true,
291 .shared_addr_format = nir_address_format_62bit_generic,
292 .global_addr_format = nir_address_format_62bit_generic,
293 .temp_addr_format = nir_address_format_62bit_generic,
294 .constant_addr_format = nir_address_format_64bit_global,
295 };
296
297 spirv_options.clc_shader = load_clc_shader(compiler, disk_cache,
298 nir_options, &spirv_options);
299 if (spirv_options.clc_shader == NULL) {
300 fprintf(stderr, "ERROR: libclc shader missing."
301 " Consider installing the libclc package\n");
302 abort();
303 }
304
305 assert(spirv_size % 4 == 0);
306 nir_shader *nir =
307 spirv_to_nir(spirv, spirv_size / 4, NULL, 0, MESA_SHADER_KERNEL,
308 entrypoint_name, &spirv_options, nir_options);
309 nir_validate_shader(nir, "after spirv_to_nir");
310 ralloc_steal(mem_ctx, nir);
311 nir->info.name = ralloc_strdup(nir, entrypoint_name);
312
313 if (INTEL_DEBUG(DEBUG_CS)) {
314 /* Re-index SSA defs so we print more sensible numbers. */
315 nir_foreach_function_impl(impl, nir) {
316 nir_index_ssa_defs(impl);
317 }
318
319 fprintf(stderr, "NIR (from SPIR-V) for kernel\n");
320 nir_print_shader(nir, stderr);
321 }
322
323 nir_lower_printf_options printf_opts = {
324 .ptr_bit_size = 64,
325 .max_buffer_size = 1024 * 1024,
326 .use_printf_base_identifier = true,
327 };
328 NIR_PASS_V(nir, nir_lower_printf, &printf_opts);
329
330 NIR_PASS_V(nir, implement_intel_builtins);
331 NIR_PASS_V(nir, nir_link_shader_functions, spirv_options.clc_shader);
332
333 /* We have to lower away local constant initializers right before we
334 * inline functions. That way they get properly initialized at the top
335 * of the function and not at the top of its caller.
336 */
337 NIR_PASS_V(nir, nir_lower_variable_initializers, nir_var_function_temp);
338 NIR_PASS_V(nir, nir_lower_returns);
339 NIR_PASS_V(nir, nir_inline_functions);
340 NIR_PASS_V(nir, nir_copy_prop);
341 NIR_PASS_V(nir, nir_opt_deref);
342
343 /* Pick off the single entrypoint that we want */
344 nir_remove_non_entrypoints(nir);
345
346 /* Now that we've deleted all but the main function, we can go ahead and
347 * lower the rest of the constant initializers. We do this here so that
348 * nir_remove_dead_variables and split_per_member_structs below see the
349 * corresponding stores.
350 */
351 NIR_PASS_V(nir, nir_lower_variable_initializers, ~0);
352
353 /* LLVM loves take advantage of the fact that vec3s in OpenCL are 16B
354 * aligned and so it can just read/write them as vec4s. This results in a
355 * LOT of vec4->vec3 casts on loads and stores. One solution to this
356 * problem is to get rid of all vec3 variables.
357 */
358 NIR_PASS_V(nir, nir_lower_vec3_to_vec4,
359 nir_var_shader_temp | nir_var_function_temp |
360 nir_var_mem_shared | nir_var_mem_global|
361 nir_var_mem_constant);
362
363 /* We assign explicit types early so that the optimizer can take advantage
364 * of that information and hopefully get rid of some of our memcpys.
365 */
366 NIR_PASS_V(nir, nir_lower_vars_to_explicit_types,
367 nir_var_uniform |
368 nir_var_shader_temp | nir_var_function_temp |
369 nir_var_mem_shared | nir_var_mem_global,
370 glsl_get_cl_type_size_align);
371
372 struct brw_nir_compiler_opts opts = {};
373 brw_preprocess_nir(compiler, nir, &opts);
374
375 int max_arg_idx = -1;
376 nir_foreach_uniform_variable(var, nir) {
377 assert(var->data.location < 256);
378 max_arg_idx = MAX2(max_arg_idx, var->data.location);
379 }
380
381 kernel->args_size = nir->num_uniforms;
382 kernel->arg_count = max_arg_idx + 1;
383
384 /* No bindings */
385 struct brw_kernel_arg_desc *args =
386 rzalloc_array(mem_ctx, struct brw_kernel_arg_desc, kernel->arg_count);
387 kernel->args = args;
388
389 nir_foreach_uniform_variable(var, nir) {
390 struct brw_kernel_arg_desc arg_desc = {
391 .offset = var->data.driver_location,
392 .size = glsl_get_explicit_size(var->type, false),
393 };
394 assert(arg_desc.offset + arg_desc.size <= nir->num_uniforms);
395
396 assert(var->data.location >= 0);
397 args[var->data.location] = arg_desc;
398 }
399
400 NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_all, NULL);
401
402 /* Lower again, this time after dead-variables to get more compact variable
403 * layouts.
404 */
405 nir->global_mem_size = 0;
406 nir->scratch_size = 0;
407 nir->info.shared_size = 0;
408 NIR_PASS_V(nir, nir_lower_vars_to_explicit_types,
409 nir_var_shader_temp | nir_var_function_temp |
410 nir_var_mem_shared | nir_var_mem_global | nir_var_mem_constant,
411 glsl_get_cl_type_size_align);
412 if (nir->constant_data_size > 0) {
413 assert(nir->constant_data == NULL);
414 nir->constant_data = rzalloc_size(nir, nir->constant_data_size);
415 nir_gather_explicit_io_initializers(nir, nir->constant_data,
416 nir->constant_data_size,
417 nir_var_mem_constant);
418 }
419
420 if (INTEL_DEBUG(DEBUG_CS)) {
421 /* Re-index SSA defs so we print more sensible numbers. */
422 nir_foreach_function_impl(impl, nir) {
423 nir_index_ssa_defs(impl);
424 }
425
426 fprintf(stderr, "NIR (before I/O lowering) for kernel\n");
427 nir_print_shader(nir, stderr);
428 }
429
430 NIR_PASS_V(nir, nir_lower_memcpy);
431
432 NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_constant,
433 nir_address_format_64bit_global);
434
435 NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_uniform,
436 nir_address_format_32bit_offset_as_64bit);
437
438 NIR_PASS_V(nir, nir_lower_explicit_io,
439 nir_var_shader_temp | nir_var_function_temp |
440 nir_var_mem_shared | nir_var_mem_global,
441 nir_address_format_62bit_generic);
442
443 NIR_PASS_V(nir, nir_lower_convert_alu_types, NULL);
444
445 NIR_PASS_V(nir, brw_nir_lower_cs_intrinsics, devinfo, NULL);
446 NIR_PASS_V(nir, lower_kernel_intrinsics);
447
448 struct brw_cs_prog_key key = { };
449
450 memset(&kernel->prog_data, 0, sizeof(kernel->prog_data));
451 kernel->prog_data.base.nr_params = DIV_ROUND_UP(nir->num_uniforms, 4);
452
453 struct brw_compile_cs_params params = {
454 .base = {
455 .nir = nir,
456 .stats = kernel->stats,
457 .log_data = log_data,
458 .mem_ctx = mem_ctx,
459 },
460 .key = &key,
461 .prog_data = &kernel->prog_data,
462 };
463
464 kernel->code = brw_compile_cs(compiler, ¶ms);
465
466 if (error_str)
467 *error_str = params.base.error_str;
468
469 return kernel->code != NULL;
470 }
471