1 /*
2 * Copyright 2013 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26 #include "si_compute.h"
27
28 #include "ac_rtld.h"
29 #include "amd_kernel_code_t.h"
30 #include "nir/tgsi_to_nir.h"
31 #include "si_build_pm4.h"
32 #include "util/u_async_debug.h"
33 #include "util/u_memory.h"
34 #include "util/u_upload_mgr.h"
35
36 #define COMPUTE_DBG(sscreen, fmt, args...) \
37 do { \
38 if ((sscreen->debug_flags & DBG(COMPUTE))) \
39 fprintf(stderr, fmt, ##args); \
40 } while (0);
41
42 struct dispatch_packet {
43 uint16_t header;
44 uint16_t setup;
45 uint16_t workgroup_size_x;
46 uint16_t workgroup_size_y;
47 uint16_t workgroup_size_z;
48 uint16_t reserved0;
49 uint32_t grid_size_x;
50 uint32_t grid_size_y;
51 uint32_t grid_size_z;
52 uint32_t private_segment_size;
53 uint32_t group_segment_size;
54 uint64_t kernel_object;
55 uint64_t kernarg_address;
56 uint64_t reserved2;
57 };
58
si_compute_get_code_object(const struct si_compute * program,uint64_t symbol_offset)59 static const amd_kernel_code_t *si_compute_get_code_object(const struct si_compute *program,
60 uint64_t symbol_offset)
61 {
62 const struct si_shader_selector *sel = &program->sel;
63
64 if (program->ir_type != PIPE_SHADER_IR_NATIVE)
65 return NULL;
66
67 struct ac_rtld_binary rtld;
68 if (!ac_rtld_open(&rtld,
69 (struct ac_rtld_open_info){.info = &sel->screen->info,
70 .shader_type = MESA_SHADER_COMPUTE,
71 .wave_size = sel->screen->compute_wave_size,
72 .num_parts = 1,
73 .elf_ptrs = &program->shader.binary.elf_buffer,
74 .elf_sizes = &program->shader.binary.elf_size}))
75 return NULL;
76
77 const amd_kernel_code_t *result = NULL;
78 const char *text;
79 size_t size;
80 if (!ac_rtld_get_section_by_name(&rtld, ".text", &text, &size))
81 goto out;
82
83 if (symbol_offset + sizeof(amd_kernel_code_t) > size)
84 goto out;
85
86 result = (const amd_kernel_code_t *)(text + symbol_offset);
87
88 out:
89 ac_rtld_close(&rtld);
90 return result;
91 }
92
code_object_to_config(const amd_kernel_code_t * code_object,struct ac_shader_config * out_config)93 static void code_object_to_config(const amd_kernel_code_t *code_object,
94 struct ac_shader_config *out_config)
95 {
96
97 uint32_t rsrc1 = code_object->compute_pgm_resource_registers;
98 uint32_t rsrc2 = code_object->compute_pgm_resource_registers >> 32;
99 out_config->num_sgprs = code_object->wavefront_sgpr_count;
100 out_config->num_vgprs = code_object->workitem_vgpr_count;
101 out_config->float_mode = G_00B028_FLOAT_MODE(rsrc1);
102 out_config->rsrc1 = rsrc1;
103 out_config->lds_size = MAX2(out_config->lds_size, G_00B84C_LDS_SIZE(rsrc2));
104 out_config->rsrc2 = rsrc2;
105 out_config->scratch_bytes_per_wave =
106 align(code_object->workitem_private_segment_byte_size * 64, 1024);
107 }
108
109 /* Asynchronous compute shader compilation. */
si_create_compute_state_async(void * job,void * gdata,int thread_index)110 static void si_create_compute_state_async(void *job, void *gdata, int thread_index)
111 {
112 struct si_compute *program = (struct si_compute *)job;
113 struct si_shader_selector *sel = &program->sel;
114 struct si_shader *shader = &program->shader;
115 struct ac_llvm_compiler *compiler;
116 struct pipe_debug_callback *debug = &sel->compiler_ctx_state.debug;
117 struct si_screen *sscreen = sel->screen;
118
119 assert(!debug->debug_message || debug->async);
120 assert(thread_index >= 0);
121 assert(thread_index < ARRAY_SIZE(sscreen->compiler));
122 compiler = &sscreen->compiler[thread_index];
123
124 if (!compiler->passes)
125 si_init_compiler(sscreen, compiler);
126
127 assert(program->ir_type == PIPE_SHADER_IR_NIR);
128 si_nir_scan_shader(sel->nir, &sel->info);
129
130 si_get_active_slot_masks(&sel->info, &sel->active_const_and_shader_buffers,
131 &sel->active_samplers_and_images);
132
133 program->shader.is_monolithic = true;
134
135 /* Variable block sizes need 10 bits (1 + log2(SI_MAX_VARIABLE_THREADS_PER_BLOCK)) per dim.
136 * We pack them into a single user SGPR.
137 */
138 unsigned user_sgprs = SI_NUM_RESOURCE_SGPRS + (sel->info.uses_grid_size ? 3 : 0) +
139 (sel->info.uses_variable_block_size ? 1 : 0) +
140 sel->info.base.cs.user_data_components_amd;
141
142 /* Fast path for compute shaders - some descriptors passed via user SGPRs. */
143 /* Shader buffers in user SGPRs. */
144 for (unsigned i = 0; i < MIN2(3, sel->info.base.num_ssbos) && user_sgprs <= 12; i++) {
145 user_sgprs = align(user_sgprs, 4);
146 if (i == 0)
147 sel->cs_shaderbufs_sgpr_index = user_sgprs;
148 user_sgprs += 4;
149 sel->cs_num_shaderbufs_in_user_sgprs++;
150 }
151
152 /* Images in user SGPRs. */
153 unsigned non_msaa_images = u_bit_consecutive(0, sel->info.base.num_images) &
154 ~sel->info.base.msaa_images;
155
156 for (unsigned i = 0; i < 3 && non_msaa_images & (1 << i); i++) {
157 unsigned num_sgprs = sel->info.base.image_buffers & (1 << i) ? 4 : 8;
158
159 if (align(user_sgprs, num_sgprs) + num_sgprs > 16)
160 break;
161
162 user_sgprs = align(user_sgprs, num_sgprs);
163 if (i == 0)
164 sel->cs_images_sgpr_index = user_sgprs;
165 user_sgprs += num_sgprs;
166 sel->cs_num_images_in_user_sgprs++;
167 }
168 sel->cs_images_num_sgprs = user_sgprs - sel->cs_images_sgpr_index;
169 assert(user_sgprs <= 16);
170
171 unsigned char ir_sha1_cache_key[20];
172 si_get_ir_cache_key(sel, false, false, ir_sha1_cache_key);
173
174 /* Try to load the shader from the shader cache. */
175 simple_mtx_lock(&sscreen->shader_cache_mutex);
176
177 if (si_shader_cache_load_shader(sscreen, ir_sha1_cache_key, shader)) {
178 simple_mtx_unlock(&sscreen->shader_cache_mutex);
179
180 si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
181 si_shader_dump(sscreen, shader, debug, stderr, true);
182
183 if (!si_shader_binary_upload(sscreen, shader, 0))
184 program->shader.compilation_failed = true;
185 } else {
186 simple_mtx_unlock(&sscreen->shader_cache_mutex);
187
188 if (!si_create_shader_variant(sscreen, compiler, &program->shader, debug)) {
189 program->shader.compilation_failed = true;
190 return;
191 }
192
193 bool scratch_enabled = shader->config.scratch_bytes_per_wave > 0;
194
195 shader->config.rsrc1 = S_00B848_VGPRS((shader->config.num_vgprs - 1) /
196 ((sscreen->compute_wave_size == 32 ||
197 sscreen->info.wave64_vgpr_alloc_granularity == 8) ? 8 : 4)) |
198 S_00B848_DX10_CLAMP(1) |
199 S_00B848_MEM_ORDERED(si_shader_mem_ordered(shader)) |
200 S_00B848_WGP_MODE(sscreen->info.chip_class >= GFX10) |
201 S_00B848_FLOAT_MODE(shader->config.float_mode);
202
203 if (sscreen->info.chip_class < GFX10) {
204 shader->config.rsrc1 |= S_00B848_SGPRS((shader->config.num_sgprs - 1) / 8);
205 }
206
207 shader->config.rsrc2 = S_00B84C_USER_SGPR(user_sgprs) | S_00B84C_SCRATCH_EN(scratch_enabled) |
208 S_00B84C_TGID_X_EN(sel->info.uses_block_id[0]) |
209 S_00B84C_TGID_Y_EN(sel->info.uses_block_id[1]) |
210 S_00B84C_TGID_Z_EN(sel->info.uses_block_id[2]) |
211 S_00B84C_TG_SIZE_EN(sel->info.uses_subgroup_info) |
212 S_00B84C_TIDIG_COMP_CNT(sel->info.uses_thread_id[2]
213 ? 2
214 : sel->info.uses_thread_id[1] ? 1 : 0) |
215 S_00B84C_LDS_SIZE(shader->config.lds_size);
216
217 simple_mtx_lock(&sscreen->shader_cache_mutex);
218 si_shader_cache_insert_shader(sscreen, ir_sha1_cache_key, shader, true);
219 simple_mtx_unlock(&sscreen->shader_cache_mutex);
220 }
221
222 ralloc_free(sel->nir);
223 sel->nir = NULL;
224 }
225
si_create_compute_state(struct pipe_context * ctx,const struct pipe_compute_state * cso)226 static void *si_create_compute_state(struct pipe_context *ctx, const struct pipe_compute_state *cso)
227 {
228 struct si_context *sctx = (struct si_context *)ctx;
229 struct si_screen *sscreen = (struct si_screen *)ctx->screen;
230 struct si_compute *program = CALLOC_STRUCT(si_compute);
231 struct si_shader_selector *sel = &program->sel;
232
233 pipe_reference_init(&sel->base.reference, 1);
234 sel->info.stage = MESA_SHADER_COMPUTE;
235 sel->screen = sscreen;
236 sel->const_and_shader_buf_descriptors_index =
237 si_const_and_shader_buffer_descriptors_idx(PIPE_SHADER_COMPUTE);
238 sel->sampler_and_images_descriptors_index =
239 si_sampler_and_image_descriptors_idx(PIPE_SHADER_COMPUTE);
240 sel->info.base.shared_size = cso->req_local_mem;
241 program->shader.selector = &program->sel;
242 program->ir_type = cso->ir_type;
243 program->private_size = cso->req_private_mem;
244 program->input_size = cso->req_input_mem;
245
246 if (cso->ir_type != PIPE_SHADER_IR_NATIVE) {
247 if (cso->ir_type == PIPE_SHADER_IR_TGSI) {
248 program->ir_type = PIPE_SHADER_IR_NIR;
249 sel->nir = tgsi_to_nir(cso->prog, ctx->screen, true);
250 } else {
251 assert(cso->ir_type == PIPE_SHADER_IR_NIR);
252 sel->nir = (struct nir_shader *)cso->prog;
253 }
254
255 sel->compiler_ctx_state.debug = sctx->debug;
256 sel->compiler_ctx_state.is_debug_context = sctx->is_debug;
257 p_atomic_inc(&sscreen->num_shaders_created);
258
259 si_schedule_initial_compile(sctx, MESA_SHADER_COMPUTE, &sel->ready, &sel->compiler_ctx_state,
260 program, si_create_compute_state_async);
261 } else {
262 const struct pipe_binary_program_header *header;
263 header = cso->prog;
264
265 program->shader.binary.elf_size = header->num_bytes;
266 program->shader.binary.elf_buffer = malloc(header->num_bytes);
267 if (!program->shader.binary.elf_buffer) {
268 FREE(program);
269 return NULL;
270 }
271 memcpy((void *)program->shader.binary.elf_buffer, header->blob, header->num_bytes);
272
273 const amd_kernel_code_t *code_object = si_compute_get_code_object(program, 0);
274 code_object_to_config(code_object, &program->shader.config);
275
276 si_shader_dump(sctx->screen, &program->shader, &sctx->debug, stderr, true);
277 if (!si_shader_binary_upload(sctx->screen, &program->shader, 0)) {
278 fprintf(stderr, "LLVM failed to upload shader\n");
279 free((void *)program->shader.binary.elf_buffer);
280 FREE(program);
281 return NULL;
282 }
283 }
284
285 return program;
286 }
287
si_bind_compute_state(struct pipe_context * ctx,void * state)288 static void si_bind_compute_state(struct pipe_context *ctx, void *state)
289 {
290 struct si_context *sctx = (struct si_context *)ctx;
291 struct si_compute *program = (struct si_compute *)state;
292 struct si_shader_selector *sel = &program->sel;
293
294 sctx->cs_shader_state.program = program;
295 if (!program)
296 return;
297
298 /* Wait because we need active slot usage masks. */
299 if (program->ir_type != PIPE_SHADER_IR_NATIVE)
300 util_queue_fence_wait(&sel->ready);
301
302 si_set_active_descriptors(sctx,
303 SI_DESCS_FIRST_COMPUTE + SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS,
304 sel->active_const_and_shader_buffers);
305 si_set_active_descriptors(sctx, SI_DESCS_FIRST_COMPUTE + SI_SHADER_DESCS_SAMPLERS_AND_IMAGES,
306 sel->active_samplers_and_images);
307
308 sctx->compute_shaderbuf_sgprs_dirty = true;
309 sctx->compute_image_sgprs_dirty = true;
310
311 if (unlikely((sctx->screen->debug_flags & DBG(SQTT)) && sctx->thread_trace)) {
312 uint32_t pipeline_code_hash = _mesa_hash_data_with_seed(
313 program->shader.binary.elf_buffer,
314 program->shader.binary.elf_size,
315 0);
316 uint64_t base_address = program->shader.bo->gpu_address;
317
318 struct ac_thread_trace_data *thread_trace_data = sctx->thread_trace;
319 if (!si_sqtt_pipeline_is_registered(thread_trace_data, pipeline_code_hash)) {
320 si_sqtt_register_pipeline(sctx, pipeline_code_hash, base_address, true);
321 }
322
323 si_sqtt_describe_pipeline_bind(sctx, pipeline_code_hash, 1);
324 }
325 }
326
si_set_global_binding(struct pipe_context * ctx,unsigned first,unsigned n,struct pipe_resource ** resources,uint32_t ** handles)327 static void si_set_global_binding(struct pipe_context *ctx, unsigned first, unsigned n,
328 struct pipe_resource **resources, uint32_t **handles)
329 {
330 unsigned i;
331 struct si_context *sctx = (struct si_context *)ctx;
332 struct si_compute *program = sctx->cs_shader_state.program;
333
334 if (first + n > program->max_global_buffers) {
335 unsigned old_max = program->max_global_buffers;
336 program->max_global_buffers = first + n;
337 program->global_buffers = realloc(
338 program->global_buffers, program->max_global_buffers * sizeof(program->global_buffers[0]));
339 if (!program->global_buffers) {
340 fprintf(stderr, "radeonsi: failed to allocate compute global_buffers\n");
341 return;
342 }
343
344 memset(&program->global_buffers[old_max], 0,
345 (program->max_global_buffers - old_max) * sizeof(program->global_buffers[0]));
346 }
347
348 if (!resources) {
349 for (i = 0; i < n; i++) {
350 pipe_resource_reference(&program->global_buffers[first + i], NULL);
351 }
352 return;
353 }
354
355 for (i = 0; i < n; i++) {
356 uint64_t va;
357 uint32_t offset;
358 pipe_resource_reference(&program->global_buffers[first + i], resources[i]);
359 va = si_resource(resources[i])->gpu_address;
360 offset = util_le32_to_cpu(*handles[i]);
361 va += offset;
362 va = util_cpu_to_le64(va);
363 memcpy(handles[i], &va, sizeof(va));
364 }
365 }
366
si_emit_initial_compute_regs(struct si_context * sctx,struct radeon_cmdbuf * cs)367 void si_emit_initial_compute_regs(struct si_context *sctx, struct radeon_cmdbuf *cs)
368 {
369 radeon_begin(cs);
370 radeon_set_sh_reg(R_00B834_COMPUTE_PGM_HI,
371 S_00B834_DATA(sctx->screen->info.address32_hi >> 8));
372
373 radeon_set_sh_reg_seq(R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0, 2);
374 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1,
375 * renamed COMPUTE_DESTINATION_EN_SEn on gfx10. */
376 radeon_emit(S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
377 radeon_emit(S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
378
379 if (sctx->chip_class == GFX6) {
380 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
381 * and is now per pipe, so it should be handled in the
382 * kernel if we want to use something other than the default value.
383 *
384 * TODO: This should be:
385 * (number of compute units) * 4 * (waves per simd) - 1
386 */
387 radeon_set_sh_reg(R_00B82C_COMPUTE_MAX_WAVE_ID, 0x190 /* Default value */);
388
389 if (sctx->screen->info.si_TA_CS_BC_BASE_ADDR_allowed) {
390 uint64_t bc_va = sctx->border_color_buffer->gpu_address;
391
392 radeon_set_config_reg(R_00950C_TA_CS_BC_BASE_ADDR, bc_va >> 8);
393 }
394 }
395
396 if (sctx->chip_class >= GFX7) {
397 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
398 radeon_set_sh_reg_seq(R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
399 radeon_emit(S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
400 radeon_emit(S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
401
402 /* Disable profiling on compute queues. */
403 if (cs != &sctx->gfx_cs || !sctx->screen->info.has_graphics) {
404 radeon_set_sh_reg(R_00B82C_COMPUTE_PERFCOUNT_ENABLE, 0);
405 radeon_set_sh_reg(R_00B878_COMPUTE_THREAD_TRACE_ENABLE, 0);
406 }
407
408 /* Set the pointer to border colors. */
409 /* Aldebaran doesn't support border colors. */
410 if (sctx->border_color_buffer) {
411 uint64_t bc_va = sctx->border_color_buffer->gpu_address;
412
413 radeon_set_uconfig_reg_seq(R_030E00_TA_CS_BC_BASE_ADDR, 2, false);
414 radeon_emit(bc_va >> 8); /* R_030E00_TA_CS_BC_BASE_ADDR */
415 radeon_emit(S_030E04_ADDRESS(bc_va >> 40)); /* R_030E04_TA_CS_BC_BASE_ADDR_HI */
416 }
417 }
418
419 /* cs_preamble_state initializes this for the gfx queue, so only do this
420 * if we are on a compute queue.
421 */
422 if (sctx->chip_class >= GFX9 &&
423 (cs != &sctx->gfx_cs || !sctx->screen->info.has_graphics)) {
424 radeon_set_uconfig_reg(R_0301EC_CP_COHER_START_DELAY,
425 sctx->chip_class >= GFX10 ? 0x20 : 0);
426 }
427
428 if (sctx->chip_class >= GFX10) {
429 radeon_set_sh_reg_seq(R_00B890_COMPUTE_USER_ACCUM_0, 5);
430 radeon_emit(0); /* R_00B890_COMPUTE_USER_ACCUM_0 */
431 radeon_emit(0); /* R_00B894_COMPUTE_USER_ACCUM_1 */
432 radeon_emit(0); /* R_00B898_COMPUTE_USER_ACCUM_2 */
433 radeon_emit(0); /* R_00B89C_COMPUTE_USER_ACCUM_3 */
434 radeon_emit(0); /* R_00B8A0_COMPUTE_PGM_RSRC3 */
435
436 radeon_set_sh_reg(R_00B9F4_COMPUTE_DISPATCH_TUNNEL, 0);
437 }
438 radeon_end();
439 }
440
si_setup_compute_scratch_buffer(struct si_context * sctx,struct si_shader * shader,struct ac_shader_config * config)441 static bool si_setup_compute_scratch_buffer(struct si_context *sctx, struct si_shader *shader,
442 struct ac_shader_config *config)
443 {
444 uint64_t scratch_bo_size, scratch_needed;
445 scratch_bo_size = 0;
446 scratch_needed = config->scratch_bytes_per_wave * sctx->scratch_waves;
447 if (sctx->compute_scratch_buffer)
448 scratch_bo_size = sctx->compute_scratch_buffer->b.b.width0;
449
450 if (scratch_bo_size < scratch_needed) {
451 si_resource_reference(&sctx->compute_scratch_buffer, NULL);
452
453 sctx->compute_scratch_buffer =
454 si_aligned_buffer_create(&sctx->screen->b,
455 SI_RESOURCE_FLAG_UNMAPPABLE | SI_RESOURCE_FLAG_DRIVER_INTERNAL,
456 PIPE_USAGE_DEFAULT,
457 scratch_needed, sctx->screen->info.pte_fragment_size);
458
459 if (!sctx->compute_scratch_buffer)
460 return false;
461 }
462
463 if (sctx->compute_scratch_buffer != shader->scratch_bo && scratch_needed) {
464 uint64_t scratch_va = sctx->compute_scratch_buffer->gpu_address;
465
466 if (!si_shader_binary_upload(sctx->screen, shader, scratch_va))
467 return false;
468
469 si_resource_reference(&shader->scratch_bo, sctx->compute_scratch_buffer);
470 }
471
472 return true;
473 }
474
si_switch_compute_shader(struct si_context * sctx,struct si_compute * program,struct si_shader * shader,const amd_kernel_code_t * code_object,unsigned offset,bool * prefetch)475 static bool si_switch_compute_shader(struct si_context *sctx, struct si_compute *program,
476 struct si_shader *shader, const amd_kernel_code_t *code_object,
477 unsigned offset, bool *prefetch)
478 {
479 struct radeon_cmdbuf *cs = &sctx->gfx_cs;
480 struct ac_shader_config inline_config = {0};
481 struct ac_shader_config *config;
482 uint64_t shader_va;
483
484 *prefetch = false;
485
486 if (sctx->cs_shader_state.emitted_program == program && sctx->cs_shader_state.offset == offset)
487 return true;
488
489 if (program->ir_type != PIPE_SHADER_IR_NATIVE) {
490 config = &shader->config;
491 } else {
492 unsigned lds_blocks;
493
494 config = &inline_config;
495 code_object_to_config(code_object, config);
496
497 lds_blocks = config->lds_size;
498 /* XXX: We are over allocating LDS. For GFX6, the shader reports
499 * LDS in blocks of 256 bytes, so if there are 4 bytes lds
500 * allocated in the shader and 4 bytes allocated by the state
501 * tracker, then we will set LDS_SIZE to 512 bytes rather than 256.
502 */
503 if (sctx->chip_class <= GFX6) {
504 lds_blocks += align(program->sel.info.base.shared_size, 256) >> 8;
505 } else {
506 lds_blocks += align(program->sel.info.base.shared_size, 512) >> 9;
507 }
508
509 /* TODO: use si_multiwave_lds_size_workaround */
510 assert(lds_blocks <= 0xFF);
511
512 config->rsrc2 &= C_00B84C_LDS_SIZE;
513 config->rsrc2 |= S_00B84C_LDS_SIZE(lds_blocks);
514 }
515
516 if (!si_setup_compute_scratch_buffer(sctx, shader, config))
517 return false;
518
519 if (shader->scratch_bo) {
520 COMPUTE_DBG(sctx->screen,
521 "Waves: %u; Scratch per wave: %u bytes; "
522 "Total Scratch: %u bytes\n",
523 sctx->scratch_waves, config->scratch_bytes_per_wave,
524 config->scratch_bytes_per_wave * sctx->scratch_waves);
525
526 radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, shader->scratch_bo, RADEON_USAGE_READWRITE,
527 RADEON_PRIO_SCRATCH_BUFFER);
528 }
529
530 shader_va = shader->bo->gpu_address + offset;
531 if (program->ir_type == PIPE_SHADER_IR_NATIVE) {
532 /* Shader code is placed after the amd_kernel_code_t
533 * struct. */
534 shader_va += sizeof(amd_kernel_code_t);
535 }
536
537 radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, shader->bo, RADEON_USAGE_READ,
538 RADEON_PRIO_SHADER_BINARY);
539
540 radeon_begin(cs);
541 radeon_set_sh_reg(R_00B830_COMPUTE_PGM_LO, shader_va >> 8);
542
543 radeon_set_sh_reg_seq(R_00B848_COMPUTE_PGM_RSRC1, 2);
544 radeon_emit(config->rsrc1);
545 radeon_emit(config->rsrc2);
546
547 COMPUTE_DBG(sctx->screen,
548 "COMPUTE_PGM_RSRC1: 0x%08x "
549 "COMPUTE_PGM_RSRC2: 0x%08x\n",
550 config->rsrc1, config->rsrc2);
551
552 sctx->max_seen_compute_scratch_bytes_per_wave =
553 MAX2(sctx->max_seen_compute_scratch_bytes_per_wave, config->scratch_bytes_per_wave);
554
555 radeon_set_sh_reg(R_00B860_COMPUTE_TMPRING_SIZE,
556 S_00B860_WAVES(sctx->scratch_waves) |
557 S_00B860_WAVESIZE(sctx->max_seen_compute_scratch_bytes_per_wave >> 10));
558 radeon_end();
559
560 sctx->cs_shader_state.emitted_program = program;
561 sctx->cs_shader_state.offset = offset;
562 sctx->cs_shader_state.uses_scratch = config->scratch_bytes_per_wave != 0;
563
564 *prefetch = true;
565 return true;
566 }
567
setup_scratch_rsrc_user_sgprs(struct si_context * sctx,const amd_kernel_code_t * code_object,unsigned user_sgpr)568 static void setup_scratch_rsrc_user_sgprs(struct si_context *sctx,
569 const amd_kernel_code_t *code_object, unsigned user_sgpr)
570 {
571 struct radeon_cmdbuf *cs = &sctx->gfx_cs;
572 uint64_t scratch_va = sctx->compute_scratch_buffer->gpu_address;
573
574 unsigned max_private_element_size =
575 AMD_HSA_BITS_GET(code_object->code_properties, AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE);
576
577 uint32_t scratch_dword0 = scratch_va & 0xffffffff;
578 uint32_t scratch_dword1 =
579 S_008F04_BASE_ADDRESS_HI(scratch_va >> 32) | S_008F04_SWIZZLE_ENABLE(1);
580
581 /* Disable address clamping */
582 uint32_t scratch_dword2 = 0xffffffff;
583 uint32_t scratch_dword3 = S_008F0C_INDEX_STRIDE(3) | S_008F0C_ADD_TID_ENABLE(1);
584
585 if (sctx->chip_class >= GFX9) {
586 assert(max_private_element_size == 1); /* always 4 bytes on GFX9 */
587 } else {
588 scratch_dword3 |= S_008F0C_ELEMENT_SIZE(max_private_element_size);
589
590 if (sctx->chip_class < GFX8) {
591 /* BUF_DATA_FORMAT is ignored, but it cannot be
592 * BUF_DATA_FORMAT_INVALID. */
593 scratch_dword3 |= S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_8);
594 }
595 }
596
597 radeon_begin(cs);
598 radeon_set_sh_reg_seq(R_00B900_COMPUTE_USER_DATA_0 + (user_sgpr * 4), 4);
599 radeon_emit(scratch_dword0);
600 radeon_emit(scratch_dword1);
601 radeon_emit(scratch_dword2);
602 radeon_emit(scratch_dword3);
603 radeon_end();
604 }
605
si_setup_user_sgprs_co_v2(struct si_context * sctx,const amd_kernel_code_t * code_object,const struct pipe_grid_info * info,uint64_t kernel_args_va)606 static void si_setup_user_sgprs_co_v2(struct si_context *sctx, const amd_kernel_code_t *code_object,
607 const struct pipe_grid_info *info, uint64_t kernel_args_va)
608 {
609 struct si_compute *program = sctx->cs_shader_state.program;
610 struct radeon_cmdbuf *cs = &sctx->gfx_cs;
611
612 static const enum amd_code_property_mask_t workgroup_count_masks[] = {
613 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X,
614 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y,
615 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z};
616
617 unsigned i, user_sgpr = 0;
618 if (AMD_HSA_BITS_GET(code_object->code_properties,
619 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER)) {
620 if (code_object->workitem_private_segment_byte_size > 0) {
621 setup_scratch_rsrc_user_sgprs(sctx, code_object, user_sgpr);
622 }
623 user_sgpr += 4;
624 }
625
626 radeon_begin(cs);
627
628 if (AMD_HSA_BITS_GET(code_object->code_properties, AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR)) {
629 struct dispatch_packet dispatch;
630 unsigned dispatch_offset;
631 struct si_resource *dispatch_buf = NULL;
632 uint64_t dispatch_va;
633
634 /* Upload dispatch ptr */
635 memset(&dispatch, 0, sizeof(dispatch));
636
637 dispatch.workgroup_size_x = util_cpu_to_le16(info->block[0]);
638 dispatch.workgroup_size_y = util_cpu_to_le16(info->block[1]);
639 dispatch.workgroup_size_z = util_cpu_to_le16(info->block[2]);
640
641 dispatch.grid_size_x = util_cpu_to_le32(info->grid[0] * info->block[0]);
642 dispatch.grid_size_y = util_cpu_to_le32(info->grid[1] * info->block[1]);
643 dispatch.grid_size_z = util_cpu_to_le32(info->grid[2] * info->block[2]);
644
645 dispatch.private_segment_size = util_cpu_to_le32(program->private_size);
646 dispatch.group_segment_size = util_cpu_to_le32(program->sel.info.base.shared_size);
647
648 dispatch.kernarg_address = util_cpu_to_le64(kernel_args_va);
649
650 u_upload_data(sctx->b.const_uploader, 0, sizeof(dispatch), 256, &dispatch, &dispatch_offset,
651 (struct pipe_resource **)&dispatch_buf);
652
653 if (!dispatch_buf) {
654 fprintf(stderr, "Error: Failed to allocate dispatch "
655 "packet.");
656 }
657 radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, dispatch_buf, RADEON_USAGE_READ,
658 RADEON_PRIO_CONST_BUFFER);
659
660 dispatch_va = dispatch_buf->gpu_address + dispatch_offset;
661
662 radeon_set_sh_reg_seq(R_00B900_COMPUTE_USER_DATA_0 + (user_sgpr * 4), 2);
663 radeon_emit(dispatch_va);
664 radeon_emit(S_008F04_BASE_ADDRESS_HI(dispatch_va >> 32) | S_008F04_STRIDE(0));
665
666 si_resource_reference(&dispatch_buf, NULL);
667 user_sgpr += 2;
668 }
669
670 if (AMD_HSA_BITS_GET(code_object->code_properties,
671 AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR)) {
672 radeon_set_sh_reg_seq(R_00B900_COMPUTE_USER_DATA_0 + (user_sgpr * 4), 2);
673 radeon_emit(kernel_args_va);
674 radeon_emit(S_008F04_BASE_ADDRESS_HI(kernel_args_va >> 32) | S_008F04_STRIDE(0));
675 user_sgpr += 2;
676 }
677
678 for (i = 0; i < 3 && user_sgpr < 16; i++) {
679 if (code_object->code_properties & workgroup_count_masks[i]) {
680 radeon_set_sh_reg_seq(R_00B900_COMPUTE_USER_DATA_0 + (user_sgpr * 4), 1);
681 radeon_emit(info->grid[i]);
682 user_sgpr += 1;
683 }
684 }
685 radeon_end();
686 }
687
si_upload_compute_input(struct si_context * sctx,const amd_kernel_code_t * code_object,const struct pipe_grid_info * info)688 static bool si_upload_compute_input(struct si_context *sctx, const amd_kernel_code_t *code_object,
689 const struct pipe_grid_info *info)
690 {
691 struct si_compute *program = sctx->cs_shader_state.program;
692 struct si_resource *input_buffer = NULL;
693 uint32_t kernel_args_offset = 0;
694 uint32_t *kernel_args;
695 void *kernel_args_ptr;
696 uint64_t kernel_args_va;
697
698 u_upload_alloc(sctx->b.const_uploader, 0, program->input_size,
699 sctx->screen->info.tcc_cache_line_size, &kernel_args_offset,
700 (struct pipe_resource **)&input_buffer, &kernel_args_ptr);
701
702 if (unlikely(!kernel_args_ptr))
703 return false;
704
705 kernel_args = (uint32_t *)kernel_args_ptr;
706 kernel_args_va = input_buffer->gpu_address + kernel_args_offset;
707
708 memcpy(kernel_args, info->input, program->input_size);
709
710 for (unsigned i = 0; i < program->input_size / 4; i++) {
711 COMPUTE_DBG(sctx->screen, "input %u : %u\n", i, kernel_args[i]);
712 }
713
714 radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, input_buffer, RADEON_USAGE_READ,
715 RADEON_PRIO_CONST_BUFFER);
716
717 si_setup_user_sgprs_co_v2(sctx, code_object, info, kernel_args_va);
718 si_resource_reference(&input_buffer, NULL);
719 return true;
720 }
721
si_setup_nir_user_data(struct si_context * sctx,const struct pipe_grid_info * info)722 static void si_setup_nir_user_data(struct si_context *sctx, const struct pipe_grid_info *info)
723 {
724 struct si_compute *program = sctx->cs_shader_state.program;
725 struct si_shader_selector *sel = &program->sel;
726 struct radeon_cmdbuf *cs = &sctx->gfx_cs;
727 unsigned grid_size_reg = R_00B900_COMPUTE_USER_DATA_0 + 4 * SI_NUM_RESOURCE_SGPRS;
728 unsigned block_size_reg = grid_size_reg +
729 /* 12 bytes = 3 dwords. */
730 12 * sel->info.uses_grid_size;
731 unsigned cs_user_data_reg = block_size_reg + 4 * program->sel.info.uses_variable_block_size;
732
733 radeon_begin(cs);
734
735 if (sel->info.uses_grid_size) {
736 if (info->indirect) {
737 radeon_end();
738
739 for (unsigned i = 0; i < 3; ++i) {
740 si_cp_copy_data(sctx, &sctx->gfx_cs, COPY_DATA_REG, NULL, (grid_size_reg >> 2) + i,
741 COPY_DATA_SRC_MEM, si_resource(info->indirect),
742 info->indirect_offset + 4 * i);
743 }
744 radeon_begin_again(cs);
745 } else {
746 radeon_set_sh_reg_seq(grid_size_reg, 3);
747 radeon_emit(info->grid[0]);
748 radeon_emit(info->grid[1]);
749 radeon_emit(info->grid[2]);
750 }
751 }
752
753 if (sel->info.uses_variable_block_size) {
754 radeon_set_sh_reg(block_size_reg,
755 info->block[0] | (info->block[1] << 10) | (info->block[2] << 20));
756 }
757
758 if (sel->info.base.cs.user_data_components_amd) {
759 radeon_set_sh_reg_seq(cs_user_data_reg, sel->info.base.cs.user_data_components_amd);
760 radeon_emit_array(sctx->cs_user_data, sel->info.base.cs.user_data_components_amd);
761 }
762 radeon_end();
763 }
764
si_emit_dispatch_packets(struct si_context * sctx,const struct pipe_grid_info * info)765 static void si_emit_dispatch_packets(struct si_context *sctx, const struct pipe_grid_info *info)
766 {
767 struct si_screen *sscreen = sctx->screen;
768 struct radeon_cmdbuf *cs = &sctx->gfx_cs;
769 bool render_cond_bit = sctx->render_cond_enabled;
770 unsigned threads_per_threadgroup = info->block[0] * info->block[1] * info->block[2];
771 unsigned waves_per_threadgroup =
772 DIV_ROUND_UP(threads_per_threadgroup, sscreen->compute_wave_size);
773 unsigned threadgroups_per_cu = 1;
774
775 if (sctx->chip_class >= GFX10 && waves_per_threadgroup == 1)
776 threadgroups_per_cu = 2;
777
778 if (unlikely(sctx->thread_trace_enabled)) {
779 si_write_event_with_dims_marker(sctx, &sctx->gfx_cs,
780 info->indirect ? EventCmdDispatchIndirect : EventCmdDispatch,
781 info->grid[0], info->grid[1], info->grid[2]);
782 }
783
784 radeon_begin(cs);
785 radeon_set_sh_reg(
786 R_00B854_COMPUTE_RESOURCE_LIMITS,
787 ac_get_compute_resource_limits(&sscreen->info, waves_per_threadgroup,
788 sctx->cs_max_waves_per_sh, threadgroups_per_cu));
789
790 unsigned dispatch_initiator = S_00B800_COMPUTE_SHADER_EN(1) | S_00B800_FORCE_START_AT_000(1) |
791 /* If the KMD allows it (there is a KMD hw register for it),
792 * allow launching waves out-of-order. (same as Vulkan) */
793 S_00B800_ORDER_MODE(sctx->chip_class >= GFX7) |
794 S_00B800_CS_W32_EN(sscreen->compute_wave_size == 32);
795
796 const uint *last_block = info->last_block;
797 bool partial_block_en = last_block[0] || last_block[1] || last_block[2];
798
799 radeon_set_sh_reg_seq(R_00B81C_COMPUTE_NUM_THREAD_X, 3);
800
801 if (partial_block_en) {
802 unsigned partial[3];
803
804 /* If no partial_block, these should be an entire block size, not 0. */
805 partial[0] = last_block[0] ? last_block[0] : info->block[0];
806 partial[1] = last_block[1] ? last_block[1] : info->block[1];
807 partial[2] = last_block[2] ? last_block[2] : info->block[2];
808
809 radeon_emit(S_00B81C_NUM_THREAD_FULL(info->block[0]) |
810 S_00B81C_NUM_THREAD_PARTIAL(partial[0]));
811 radeon_emit(S_00B820_NUM_THREAD_FULL(info->block[1]) |
812 S_00B820_NUM_THREAD_PARTIAL(partial[1]));
813 radeon_emit(S_00B824_NUM_THREAD_FULL(info->block[2]) |
814 S_00B824_NUM_THREAD_PARTIAL(partial[2]));
815
816 dispatch_initiator |= S_00B800_PARTIAL_TG_EN(1);
817 } else {
818 radeon_emit(S_00B81C_NUM_THREAD_FULL(info->block[0]));
819 radeon_emit(S_00B820_NUM_THREAD_FULL(info->block[1]));
820 radeon_emit(S_00B824_NUM_THREAD_FULL(info->block[2]));
821 }
822
823 if (info->indirect) {
824 uint64_t base_va = si_resource(info->indirect)->gpu_address;
825
826 radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, si_resource(info->indirect), RADEON_USAGE_READ,
827 RADEON_PRIO_DRAW_INDIRECT);
828
829 radeon_emit(PKT3(PKT3_SET_BASE, 2, 0) | PKT3_SHADER_TYPE_S(1));
830 radeon_emit(1);
831 radeon_emit(base_va);
832 radeon_emit(base_va >> 32);
833
834 radeon_emit(PKT3(PKT3_DISPATCH_INDIRECT, 1, render_cond_bit) | PKT3_SHADER_TYPE_S(1));
835 radeon_emit(info->indirect_offset);
836 radeon_emit(dispatch_initiator);
837 } else {
838 radeon_emit(PKT3(PKT3_DISPATCH_DIRECT, 3, render_cond_bit) | PKT3_SHADER_TYPE_S(1));
839 radeon_emit(info->grid[0]);
840 radeon_emit(info->grid[1]);
841 radeon_emit(info->grid[2]);
842 radeon_emit(dispatch_initiator);
843 }
844
845 if (unlikely(sctx->thread_trace_enabled && sctx->chip_class >= GFX9)) {
846 radeon_emit(PKT3(PKT3_EVENT_WRITE, 0, 0));
847 radeon_emit(EVENT_TYPE(V_028A90_THREAD_TRACE_MARKER) | EVENT_INDEX(0));
848 }
849 radeon_end();
850 }
851
si_check_needs_implicit_sync(struct si_context * sctx)852 static bool si_check_needs_implicit_sync(struct si_context *sctx)
853 {
854 /* If the compute shader is going to read from a texture/image written by a
855 * previous draw, we must wait for its completion before continuing.
856 * Buffers and image stores (from the draw) are not taken into consideration
857 * because that's the app responsibility.
858 *
859 * The OpenGL 4.6 spec says:
860 *
861 * buffer object and texture stores performed by shaders are not
862 * automatically synchronized
863 *
864 * TODO: Bindless textures are not handled, and thus are not synchronized.
865 */
866 struct si_shader_info *info = &sctx->cs_shader_state.program->sel.info;
867 struct si_samplers *samplers = &sctx->samplers[PIPE_SHADER_COMPUTE];
868 unsigned mask = samplers->enabled_mask & info->base.textures_used[0];
869
870 while (mask) {
871 int i = u_bit_scan(&mask);
872 struct si_sampler_view *sview = (struct si_sampler_view *)samplers->views[i];
873
874 struct si_resource *res = si_resource(sview->base.texture);
875 if (sctx->ws->cs_is_buffer_referenced(&sctx->gfx_cs, res->buf,
876 RADEON_USAGE_NEEDS_IMPLICIT_SYNC))
877 return true;
878 }
879
880 struct si_images *images = &sctx->images[PIPE_SHADER_COMPUTE];
881 mask = u_bit_consecutive(0, info->base.num_images) & images->enabled_mask;
882
883 while (mask) {
884 int i = u_bit_scan(&mask);
885 struct pipe_image_view *sview = &images->views[i];
886
887 struct si_resource *res = si_resource(sview->resource);
888 if (sctx->ws->cs_is_buffer_referenced(&sctx->gfx_cs, res->buf,
889 RADEON_USAGE_NEEDS_IMPLICIT_SYNC))
890 return true;
891 }
892 return false;
893 }
894
si_launch_grid(struct pipe_context * ctx,const struct pipe_grid_info * info)895 static void si_launch_grid(struct pipe_context *ctx, const struct pipe_grid_info *info)
896 {
897 struct si_context *sctx = (struct si_context *)ctx;
898 struct si_screen *sscreen = sctx->screen;
899 struct si_compute *program = sctx->cs_shader_state.program;
900 const amd_kernel_code_t *code_object = si_compute_get_code_object(program, info->pc);
901 int i;
902 bool cs_regalloc_hang = sscreen->info.has_cs_regalloc_hang_bug &&
903 info->block[0] * info->block[1] * info->block[2] > 256;
904
905 if (cs_regalloc_hang)
906 sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH | SI_CONTEXT_CS_PARTIAL_FLUSH;
907
908 if (program->ir_type != PIPE_SHADER_IR_NATIVE && program->shader.compilation_failed)
909 return;
910
911 if (sctx->has_graphics) {
912 if (sctx->last_num_draw_calls != sctx->num_draw_calls) {
913 si_update_fb_dirtiness_after_rendering(sctx);
914 sctx->last_num_draw_calls = sctx->num_draw_calls;
915
916 if (sctx->force_cb_shader_coherent || si_check_needs_implicit_sync(sctx))
917 si_make_CB_shader_coherent(sctx, 0,
918 sctx->framebuffer.CB_has_shader_readable_metadata,
919 sctx->framebuffer.all_DCC_pipe_aligned);
920 }
921
922 si_decompress_textures(sctx, 1 << PIPE_SHADER_COMPUTE);
923 }
924
925 /* Add buffer sizes for memory checking in need_cs_space. */
926 si_context_add_resource_size(sctx, &program->shader.bo->b.b);
927 /* TODO: add the scratch buffer */
928
929 if (info->indirect) {
930 si_context_add_resource_size(sctx, info->indirect);
931
932 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
933 if (sctx->chip_class <= GFX8 && si_resource(info->indirect)->TC_L2_dirty) {
934 sctx->flags |= SI_CONTEXT_WB_L2;
935 si_resource(info->indirect)->TC_L2_dirty = false;
936 }
937 }
938
939 si_need_gfx_cs_space(sctx, 0);
940
941 /* If we're using a secure context, determine if cs must be secure or not */
942 if (unlikely(radeon_uses_secure_bos(sctx->ws))) {
943 bool secure = si_compute_resources_check_encrypted(sctx);
944 if (secure != sctx->ws->cs_is_secure(&sctx->gfx_cs)) {
945 si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW |
946 RADEON_FLUSH_TOGGLE_SECURE_SUBMISSION,
947 NULL);
948 }
949 }
950
951 if (sctx->bo_list_add_all_compute_resources)
952 si_compute_resources_add_all_to_bo_list(sctx);
953
954 if (!sctx->cs_shader_state.initialized) {
955 si_emit_initial_compute_regs(sctx, &sctx->gfx_cs);
956
957 sctx->cs_shader_state.emitted_program = NULL;
958 sctx->cs_shader_state.initialized = true;
959 }
960
961 /* First emit registers. */
962 bool prefetch;
963 if (!si_switch_compute_shader(sctx, program, &program->shader, code_object, info->pc, &prefetch))
964 return;
965
966 si_upload_compute_shader_descriptors(sctx);
967 si_emit_compute_shader_pointers(sctx);
968
969 if (program->ir_type == PIPE_SHADER_IR_NATIVE &&
970 unlikely(!si_upload_compute_input(sctx, code_object, info)))
971 return;
972
973 /* Global buffers */
974 for (i = 0; i < program->max_global_buffers; i++) {
975 struct si_resource *buffer = si_resource(program->global_buffers[i]);
976 if (!buffer) {
977 continue;
978 }
979 radeon_add_to_buffer_list(sctx, &sctx->gfx_cs, buffer, RADEON_USAGE_READWRITE,
980 RADEON_PRIO_COMPUTE_GLOBAL);
981 }
982
983 /* Registers that are not read from memory should be set before this: */
984 if (sctx->flags)
985 sctx->emit_cache_flush(sctx, &sctx->gfx_cs);
986
987 if (sctx->has_graphics && si_is_atom_dirty(sctx, &sctx->atoms.s.render_cond)) {
988 sctx->atoms.s.render_cond.emit(sctx);
989 si_set_atom_dirty(sctx, &sctx->atoms.s.render_cond, false);
990 }
991
992 /* Prefetch the compute shader to L2. */
993 if (sctx->chip_class >= GFX7 && prefetch)
994 si_cp_dma_prefetch(sctx, &program->shader.bo->b.b, 0, program->shader.bo->b.b.width0);
995
996 if (program->ir_type != PIPE_SHADER_IR_NATIVE)
997 si_setup_nir_user_data(sctx, info);
998
999 si_emit_dispatch_packets(sctx, info);
1000
1001 if (unlikely(sctx->current_saved_cs)) {
1002 si_trace_emit(sctx);
1003 si_log_compute_state(sctx, sctx->log);
1004 }
1005
1006 /* Mark displayable DCC as dirty for bound images. */
1007 unsigned display_dcc_store_mask = sctx->images[PIPE_SHADER_COMPUTE].display_dcc_store_mask &
1008 BITFIELD_MASK(program->sel.info.base.num_images);
1009 while (display_dcc_store_mask) {
1010 struct si_texture *tex = (struct si_texture *)
1011 sctx->images[PIPE_SHADER_COMPUTE].views[u_bit_scan(&display_dcc_store_mask)].resource;
1012
1013 si_mark_display_dcc_dirty(sctx, tex);
1014 }
1015
1016 /* TODO: Bindless images don't set displayable_dcc_dirty after image stores. */
1017
1018 sctx->compute_is_busy = true;
1019 sctx->num_compute_calls++;
1020
1021 if (cs_regalloc_hang)
1022 sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
1023 }
1024
si_destroy_compute(struct si_compute * program)1025 void si_destroy_compute(struct si_compute *program)
1026 {
1027 struct si_shader_selector *sel = &program->sel;
1028
1029 if (program->ir_type != PIPE_SHADER_IR_NATIVE) {
1030 util_queue_drop_job(&sel->screen->shader_compiler_queue, &sel->ready);
1031 util_queue_fence_destroy(&sel->ready);
1032 }
1033
1034 for (unsigned i = 0; i < program->max_global_buffers; i++)
1035 pipe_resource_reference(&program->global_buffers[i], NULL);
1036 FREE(program->global_buffers);
1037
1038 si_shader_destroy(&program->shader);
1039 ralloc_free(program->sel.nir);
1040 FREE(program);
1041 }
1042
si_delete_compute_state(struct pipe_context * ctx,void * state)1043 static void si_delete_compute_state(struct pipe_context *ctx, void *state)
1044 {
1045 struct si_compute *program = (struct si_compute *)state;
1046 struct si_context *sctx = (struct si_context *)ctx;
1047
1048 if (!state)
1049 return;
1050
1051 if (program == sctx->cs_shader_state.program)
1052 sctx->cs_shader_state.program = NULL;
1053
1054 if (program == sctx->cs_shader_state.emitted_program)
1055 sctx->cs_shader_state.emitted_program = NULL;
1056
1057 si_compute_reference(&program, NULL);
1058 }
1059
si_set_compute_resources(struct pipe_context * ctx_,unsigned start,unsigned count,struct pipe_surface ** surfaces)1060 static void si_set_compute_resources(struct pipe_context *ctx_, unsigned start, unsigned count,
1061 struct pipe_surface **surfaces)
1062 {
1063 }
1064
si_init_compute_functions(struct si_context * sctx)1065 void si_init_compute_functions(struct si_context *sctx)
1066 {
1067 sctx->b.create_compute_state = si_create_compute_state;
1068 sctx->b.delete_compute_state = si_delete_compute_state;
1069 sctx->b.bind_compute_state = si_bind_compute_state;
1070 sctx->b.set_compute_resources = si_set_compute_resources;
1071 sctx->b.set_global_binding = si_set_global_binding;
1072 sctx->b.launch_grid = si_launch_grid;
1073 }
1074