1 /*
2 * Copyright 2013 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 */
24
25 #include "tgsi/tgsi_parse.h"
26 #include "util/u_async_debug.h"
27 #include "util/u_memory.h"
28 #include "util/u_upload_mgr.h"
29
30 #include "amd_kernel_code_t.h"
31 #include "radeon/r600_cs.h"
32 #include "si_pipe.h"
33 #include "si_compute.h"
34 #include "sid.h"
35
36 #define COMPUTE_DBG(rscreen, fmt, args...) \
37 do { \
38 if ((rscreen->debug_flags & DBG(COMPUTE))) fprintf(stderr, fmt, ##args); \
39 } while (0);
40
41 struct dispatch_packet {
42 uint16_t header;
43 uint16_t setup;
44 uint16_t workgroup_size_x;
45 uint16_t workgroup_size_y;
46 uint16_t workgroup_size_z;
47 uint16_t reserved0;
48 uint32_t grid_size_x;
49 uint32_t grid_size_y;
50 uint32_t grid_size_z;
51 uint32_t private_segment_size;
52 uint32_t group_segment_size;
53 uint64_t kernel_object;
54 uint64_t kernarg_address;
55 uint64_t reserved2;
56 };
57
si_compute_get_code_object(const struct si_compute * program,uint64_t symbol_offset)58 static const amd_kernel_code_t *si_compute_get_code_object(
59 const struct si_compute *program,
60 uint64_t symbol_offset)
61 {
62 if (!program->use_code_object_v2) {
63 return NULL;
64 }
65 return (const amd_kernel_code_t*)
66 (program->shader.binary.code + symbol_offset);
67 }
68
code_object_to_config(const amd_kernel_code_t * code_object,struct si_shader_config * out_config)69 static void code_object_to_config(const amd_kernel_code_t *code_object,
70 struct si_shader_config *out_config) {
71
72 uint32_t rsrc1 = code_object->compute_pgm_resource_registers;
73 uint32_t rsrc2 = code_object->compute_pgm_resource_registers >> 32;
74 out_config->num_sgprs = code_object->wavefront_sgpr_count;
75 out_config->num_vgprs = code_object->workitem_vgpr_count;
76 out_config->float_mode = G_00B028_FLOAT_MODE(rsrc1);
77 out_config->rsrc1 = rsrc1;
78 out_config->lds_size = MAX2(out_config->lds_size, G_00B84C_LDS_SIZE(rsrc2));
79 out_config->rsrc2 = rsrc2;
80 out_config->scratch_bytes_per_wave =
81 align(code_object->workitem_private_segment_byte_size * 64, 1024);
82 }
83
84 /* Asynchronous compute shader compilation. */
si_create_compute_state_async(void * job,int thread_index)85 static void si_create_compute_state_async(void *job, int thread_index)
86 {
87 struct si_compute *program = (struct si_compute *)job;
88 struct si_shader *shader = &program->shader;
89 struct si_shader_selector sel;
90 LLVMTargetMachineRef tm;
91 struct pipe_debug_callback *debug = &program->compiler_ctx_state.debug;
92
93 assert(!debug->debug_message || debug->async);
94 assert(thread_index >= 0);
95 assert(thread_index < ARRAY_SIZE(program->screen->tm));
96 tm = program->screen->tm[thread_index];
97
98 memset(&sel, 0, sizeof(sel));
99
100 sel.screen = program->screen;
101 tgsi_scan_shader(program->tokens, &sel.info);
102 sel.tokens = program->tokens;
103 sel.type = PIPE_SHADER_COMPUTE;
104 sel.local_size = program->local_size;
105 si_get_active_slot_masks(&sel.info,
106 &program->active_const_and_shader_buffers,
107 &program->active_samplers_and_images);
108
109 program->shader.selector = &sel;
110 program->shader.is_monolithic = true;
111 program->uses_grid_size = sel.info.uses_grid_size;
112 program->uses_block_size = sel.info.uses_block_size;
113 program->uses_bindless_samplers = sel.info.uses_bindless_samplers;
114 program->uses_bindless_images = sel.info.uses_bindless_images;
115
116 if (si_shader_create(program->screen, tm, &program->shader, debug)) {
117 program->shader.compilation_failed = true;
118 } else {
119 bool scratch_enabled = shader->config.scratch_bytes_per_wave > 0;
120 unsigned user_sgprs = SI_NUM_RESOURCE_SGPRS +
121 (sel.info.uses_grid_size ? 3 : 0) +
122 (sel.info.uses_block_size ? 3 : 0);
123
124 shader->config.rsrc1 =
125 S_00B848_VGPRS((shader->config.num_vgprs - 1) / 4) |
126 S_00B848_SGPRS((shader->config.num_sgprs - 1) / 8) |
127 S_00B848_DX10_CLAMP(1) |
128 S_00B848_FLOAT_MODE(shader->config.float_mode);
129
130 shader->config.rsrc2 =
131 S_00B84C_USER_SGPR(user_sgprs) |
132 S_00B84C_SCRATCH_EN(scratch_enabled) |
133 S_00B84C_TGID_X_EN(sel.info.uses_block_id[0]) |
134 S_00B84C_TGID_Y_EN(sel.info.uses_block_id[1]) |
135 S_00B84C_TGID_Z_EN(sel.info.uses_block_id[2]) |
136 S_00B84C_TIDIG_COMP_CNT(sel.info.uses_thread_id[2] ? 2 :
137 sel.info.uses_thread_id[1] ? 1 : 0) |
138 S_00B84C_LDS_SIZE(shader->config.lds_size);
139
140 program->variable_group_size =
141 sel.info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0;
142 }
143
144 FREE(program->tokens);
145 program->shader.selector = NULL;
146 }
147
si_create_compute_state(struct pipe_context * ctx,const struct pipe_compute_state * cso)148 static void *si_create_compute_state(
149 struct pipe_context *ctx,
150 const struct pipe_compute_state *cso)
151 {
152 struct si_context *sctx = (struct si_context *)ctx;
153 struct si_screen *sscreen = (struct si_screen *)ctx->screen;
154 struct si_compute *program = CALLOC_STRUCT(si_compute);
155
156 pipe_reference_init(&program->reference, 1);
157 program->screen = (struct si_screen *)ctx->screen;
158 program->ir_type = cso->ir_type;
159 program->local_size = cso->req_local_mem;
160 program->private_size = cso->req_private_mem;
161 program->input_size = cso->req_input_mem;
162 program->use_code_object_v2 = HAVE_LLVM >= 0x0400 &&
163 cso->ir_type == PIPE_SHADER_IR_NATIVE;
164
165 if (cso->ir_type == PIPE_SHADER_IR_TGSI) {
166 program->tokens = tgsi_dup_tokens(cso->prog);
167 if (!program->tokens) {
168 FREE(program);
169 return NULL;
170 }
171
172 program->compiler_ctx_state.debug = sctx->debug;
173 program->compiler_ctx_state.is_debug_context = sctx->is_debug;
174 p_atomic_inc(&sscreen->num_shaders_created);
175 util_queue_fence_init(&program->ready);
176
177 struct util_async_debug_callback async_debug;
178 bool wait =
179 (sctx->debug.debug_message && !sctx->debug.async) ||
180 sctx->is_debug ||
181 si_can_dump_shader(sscreen, PIPE_SHADER_COMPUTE);
182
183 if (wait) {
184 u_async_debug_init(&async_debug);
185 program->compiler_ctx_state.debug = async_debug.base;
186 }
187
188 util_queue_add_job(&sscreen->shader_compiler_queue,
189 program, &program->ready,
190 si_create_compute_state_async, NULL);
191
192 if (wait) {
193 util_queue_fence_wait(&program->ready);
194 u_async_debug_drain(&async_debug, &sctx->debug);
195 u_async_debug_cleanup(&async_debug);
196 }
197 } else {
198 const struct pipe_llvm_program_header *header;
199 const char *code;
200 header = cso->prog;
201 code = cso->prog + sizeof(struct pipe_llvm_program_header);
202
203 ac_elf_read(code, header->num_bytes, &program->shader.binary);
204 if (program->use_code_object_v2) {
205 const amd_kernel_code_t *code_object =
206 si_compute_get_code_object(program, 0);
207 code_object_to_config(code_object, &program->shader.config);
208 } else {
209 si_shader_binary_read_config(&program->shader.binary,
210 &program->shader.config, 0);
211 }
212 si_shader_dump(sctx->screen, &program->shader, &sctx->debug,
213 PIPE_SHADER_COMPUTE, stderr, true);
214 if (si_shader_binary_upload(sctx->screen, &program->shader) < 0) {
215 fprintf(stderr, "LLVM failed to upload shader\n");
216 FREE(program);
217 return NULL;
218 }
219 }
220
221 return program;
222 }
223
si_bind_compute_state(struct pipe_context * ctx,void * state)224 static void si_bind_compute_state(struct pipe_context *ctx, void *state)
225 {
226 struct si_context *sctx = (struct si_context*)ctx;
227 struct si_compute *program = (struct si_compute*)state;
228
229 sctx->cs_shader_state.program = program;
230 if (!program)
231 return;
232
233 /* Wait because we need active slot usage masks. */
234 if (program->ir_type == PIPE_SHADER_IR_TGSI)
235 util_queue_fence_wait(&program->ready);
236
237 si_set_active_descriptors(sctx,
238 SI_DESCS_FIRST_COMPUTE +
239 SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS,
240 program->active_const_and_shader_buffers);
241 si_set_active_descriptors(sctx,
242 SI_DESCS_FIRST_COMPUTE +
243 SI_SHADER_DESCS_SAMPLERS_AND_IMAGES,
244 program->active_samplers_and_images);
245 }
246
si_set_global_binding(struct pipe_context * ctx,unsigned first,unsigned n,struct pipe_resource ** resources,uint32_t ** handles)247 static void si_set_global_binding(
248 struct pipe_context *ctx, unsigned first, unsigned n,
249 struct pipe_resource **resources,
250 uint32_t **handles)
251 {
252 unsigned i;
253 struct si_context *sctx = (struct si_context*)ctx;
254 struct si_compute *program = sctx->cs_shader_state.program;
255
256 assert(first + n <= MAX_GLOBAL_BUFFERS);
257
258 if (!resources) {
259 for (i = 0; i < n; i++) {
260 pipe_resource_reference(&program->global_buffers[first + i], NULL);
261 }
262 return;
263 }
264
265 for (i = 0; i < n; i++) {
266 uint64_t va;
267 uint32_t offset;
268 pipe_resource_reference(&program->global_buffers[first + i], resources[i]);
269 va = r600_resource(resources[i])->gpu_address;
270 offset = util_le32_to_cpu(*handles[i]);
271 va += offset;
272 va = util_cpu_to_le64(va);
273 memcpy(handles[i], &va, sizeof(va));
274 }
275 }
276
si_initialize_compute(struct si_context * sctx)277 static void si_initialize_compute(struct si_context *sctx)
278 {
279 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
280 uint64_t bc_va;
281
282 radeon_set_sh_reg_seq(cs, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0, 2);
283 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
284 radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
285 radeon_emit(cs, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
286
287 if (sctx->b.chip_class >= CIK) {
288 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
289 radeon_set_sh_reg_seq(cs,
290 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
291 radeon_emit(cs, S_00B864_SH0_CU_EN(0xffff) |
292 S_00B864_SH1_CU_EN(0xffff));
293 radeon_emit(cs, S_00B868_SH0_CU_EN(0xffff) |
294 S_00B868_SH1_CU_EN(0xffff));
295 }
296
297 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
298 * and is now per pipe, so it should be handled in the
299 * kernel if we want to use something other than the default value,
300 * which is now 0x22f.
301 */
302 if (sctx->b.chip_class <= SI) {
303 /* XXX: This should be:
304 * (number of compute units) * 4 * (waves per simd) - 1 */
305
306 radeon_set_sh_reg(cs, R_00B82C_COMPUTE_MAX_WAVE_ID,
307 0x190 /* Default value */);
308 }
309
310 /* Set the pointer to border colors. */
311 bc_va = sctx->border_color_buffer->gpu_address;
312
313 if (sctx->b.chip_class >= CIK) {
314 radeon_set_uconfig_reg_seq(cs, R_030E00_TA_CS_BC_BASE_ADDR, 2);
315 radeon_emit(cs, bc_va >> 8); /* R_030E00_TA_CS_BC_BASE_ADDR */
316 radeon_emit(cs, bc_va >> 40); /* R_030E04_TA_CS_BC_BASE_ADDR_HI */
317 } else {
318 if (sctx->screen->info.drm_major == 3 ||
319 (sctx->screen->info.drm_major == 2 &&
320 sctx->screen->info.drm_minor >= 48)) {
321 radeon_set_config_reg(cs, R_00950C_TA_CS_BC_BASE_ADDR,
322 bc_va >> 8);
323 }
324 }
325
326 sctx->cs_shader_state.emitted_program = NULL;
327 sctx->cs_shader_state.initialized = true;
328 }
329
si_setup_compute_scratch_buffer(struct si_context * sctx,struct si_shader * shader,struct si_shader_config * config)330 static bool si_setup_compute_scratch_buffer(struct si_context *sctx,
331 struct si_shader *shader,
332 struct si_shader_config *config)
333 {
334 uint64_t scratch_bo_size, scratch_needed;
335 scratch_bo_size = 0;
336 scratch_needed = config->scratch_bytes_per_wave * sctx->scratch_waves;
337 if (sctx->compute_scratch_buffer)
338 scratch_bo_size = sctx->compute_scratch_buffer->b.b.width0;
339
340 if (scratch_bo_size < scratch_needed) {
341 r600_resource_reference(&sctx->compute_scratch_buffer, NULL);
342
343 sctx->compute_scratch_buffer = (struct r600_resource*)
344 si_aligned_buffer_create(&sctx->screen->b,
345 R600_RESOURCE_FLAG_UNMAPPABLE,
346 PIPE_USAGE_DEFAULT,
347 scratch_needed, 256);
348
349 if (!sctx->compute_scratch_buffer)
350 return false;
351 }
352
353 if (sctx->compute_scratch_buffer != shader->scratch_bo && scratch_needed) {
354 uint64_t scratch_va = sctx->compute_scratch_buffer->gpu_address;
355
356 si_shader_apply_scratch_relocs(shader, scratch_va);
357
358 if (si_shader_binary_upload(sctx->screen, shader))
359 return false;
360
361 r600_resource_reference(&shader->scratch_bo,
362 sctx->compute_scratch_buffer);
363 }
364
365 return true;
366 }
367
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)368 static bool si_switch_compute_shader(struct si_context *sctx,
369 struct si_compute *program,
370 struct si_shader *shader,
371 const amd_kernel_code_t *code_object,
372 unsigned offset)
373 {
374 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
375 struct si_shader_config inline_config = {0};
376 struct si_shader_config *config;
377 uint64_t shader_va;
378
379 if (sctx->cs_shader_state.emitted_program == program &&
380 sctx->cs_shader_state.offset == offset)
381 return true;
382
383 if (program->ir_type == PIPE_SHADER_IR_TGSI) {
384 config = &shader->config;
385 } else {
386 unsigned lds_blocks;
387
388 config = &inline_config;
389 if (code_object) {
390 code_object_to_config(code_object, config);
391 } else {
392 si_shader_binary_read_config(&shader->binary, config, offset);
393 }
394
395 lds_blocks = config->lds_size;
396 /* XXX: We are over allocating LDS. For SI, the shader reports
397 * LDS in blocks of 256 bytes, so if there are 4 bytes lds
398 * allocated in the shader and 4 bytes allocated by the state
399 * tracker, then we will set LDS_SIZE to 512 bytes rather than 256.
400 */
401 if (sctx->b.chip_class <= SI) {
402 lds_blocks += align(program->local_size, 256) >> 8;
403 } else {
404 lds_blocks += align(program->local_size, 512) >> 9;
405 }
406
407 /* TODO: use si_multiwave_lds_size_workaround */
408 assert(lds_blocks <= 0xFF);
409
410 config->rsrc2 &= C_00B84C_LDS_SIZE;
411 config->rsrc2 |= S_00B84C_LDS_SIZE(lds_blocks);
412 }
413
414 if (!si_setup_compute_scratch_buffer(sctx, shader, config))
415 return false;
416
417 if (shader->scratch_bo) {
418 COMPUTE_DBG(sctx->screen, "Waves: %u; Scratch per wave: %u bytes; "
419 "Total Scratch: %u bytes\n", sctx->scratch_waves,
420 config->scratch_bytes_per_wave,
421 config->scratch_bytes_per_wave *
422 sctx->scratch_waves);
423
424 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
425 shader->scratch_bo, RADEON_USAGE_READWRITE,
426 RADEON_PRIO_SCRATCH_BUFFER);
427 }
428
429 /* Prefetch the compute shader to TC L2.
430 *
431 * We should also prefetch graphics shaders if a compute dispatch was
432 * the last command, and the compute shader if a draw call was the last
433 * command. However, that would add more complexity and we're likely
434 * to get a shader state change in that case anyway.
435 */
436 if (sctx->b.chip_class >= CIK) {
437 cik_prefetch_TC_L2_async(sctx, &program->shader.bo->b.b,
438 0, program->shader.bo->b.b.width0);
439 }
440
441 shader_va = shader->bo->gpu_address + offset;
442 if (program->use_code_object_v2) {
443 /* Shader code is placed after the amd_kernel_code_t
444 * struct. */
445 shader_va += sizeof(amd_kernel_code_t);
446 }
447
448 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, shader->bo,
449 RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
450
451 radeon_set_sh_reg_seq(cs, R_00B830_COMPUTE_PGM_LO, 2);
452 radeon_emit(cs, shader_va >> 8);
453 radeon_emit(cs, shader_va >> 40);
454
455 radeon_set_sh_reg_seq(cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
456 radeon_emit(cs, config->rsrc1);
457 radeon_emit(cs, config->rsrc2);
458
459 COMPUTE_DBG(sctx->screen, "COMPUTE_PGM_RSRC1: 0x%08x "
460 "COMPUTE_PGM_RSRC2: 0x%08x\n", config->rsrc1, config->rsrc2);
461
462 radeon_set_sh_reg(cs, R_00B860_COMPUTE_TMPRING_SIZE,
463 S_00B860_WAVES(sctx->scratch_waves)
464 | S_00B860_WAVESIZE(config->scratch_bytes_per_wave >> 10));
465
466 sctx->cs_shader_state.emitted_program = program;
467 sctx->cs_shader_state.offset = offset;
468 sctx->cs_shader_state.uses_scratch =
469 config->scratch_bytes_per_wave != 0;
470
471 return true;
472 }
473
setup_scratch_rsrc_user_sgprs(struct si_context * sctx,const amd_kernel_code_t * code_object,unsigned user_sgpr)474 static void setup_scratch_rsrc_user_sgprs(struct si_context *sctx,
475 const amd_kernel_code_t *code_object,
476 unsigned user_sgpr)
477 {
478 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
479 uint64_t scratch_va = sctx->compute_scratch_buffer->gpu_address;
480
481 unsigned max_private_element_size = AMD_HSA_BITS_GET(
482 code_object->code_properties,
483 AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE);
484
485 uint32_t scratch_dword0 = scratch_va & 0xffffffff;
486 uint32_t scratch_dword1 =
487 S_008F04_BASE_ADDRESS_HI(scratch_va >> 32) |
488 S_008F04_SWIZZLE_ENABLE(1);
489
490 /* Disable address clamping */
491 uint32_t scratch_dword2 = 0xffffffff;
492 uint32_t scratch_dword3 =
493 S_008F0C_INDEX_STRIDE(3) |
494 S_008F0C_ADD_TID_ENABLE(1);
495
496 if (sctx->b.chip_class >= GFX9) {
497 assert(max_private_element_size == 1); /* always 4 bytes on GFX9 */
498 } else {
499 scratch_dword3 |= S_008F0C_ELEMENT_SIZE(max_private_element_size);
500
501 if (sctx->b.chip_class < VI) {
502 /* BUF_DATA_FORMAT is ignored, but it cannot be
503 * BUF_DATA_FORMAT_INVALID. */
504 scratch_dword3 |=
505 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_8);
506 }
507 }
508
509 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0 +
510 (user_sgpr * 4), 4);
511 radeon_emit(cs, scratch_dword0);
512 radeon_emit(cs, scratch_dword1);
513 radeon_emit(cs, scratch_dword2);
514 radeon_emit(cs, scratch_dword3);
515 }
516
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)517 static void si_setup_user_sgprs_co_v2(struct si_context *sctx,
518 const amd_kernel_code_t *code_object,
519 const struct pipe_grid_info *info,
520 uint64_t kernel_args_va)
521 {
522 struct si_compute *program = sctx->cs_shader_state.program;
523 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
524
525 static const enum amd_code_property_mask_t workgroup_count_masks [] = {
526 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X,
527 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y,
528 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z
529 };
530
531 unsigned i, user_sgpr = 0;
532 if (AMD_HSA_BITS_GET(code_object->code_properties,
533 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER)) {
534 if (code_object->workitem_private_segment_byte_size > 0) {
535 setup_scratch_rsrc_user_sgprs(sctx, code_object,
536 user_sgpr);
537 }
538 user_sgpr += 4;
539 }
540
541 if (AMD_HSA_BITS_GET(code_object->code_properties,
542 AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR)) {
543 struct dispatch_packet dispatch;
544 unsigned dispatch_offset;
545 struct r600_resource *dispatch_buf = NULL;
546 uint64_t dispatch_va;
547
548 /* Upload dispatch ptr */
549 memset(&dispatch, 0, sizeof(dispatch));
550
551 dispatch.workgroup_size_x = info->block[0];
552 dispatch.workgroup_size_y = info->block[1];
553 dispatch.workgroup_size_z = info->block[2];
554
555 dispatch.grid_size_x = info->grid[0] * info->block[0];
556 dispatch.grid_size_y = info->grid[1] * info->block[1];
557 dispatch.grid_size_z = info->grid[2] * info->block[2];
558
559 dispatch.private_segment_size = program->private_size;
560 dispatch.group_segment_size = program->local_size;
561
562 dispatch.kernarg_address = kernel_args_va;
563
564 u_upload_data(sctx->b.b.const_uploader, 0, sizeof(dispatch),
565 256, &dispatch, &dispatch_offset,
566 (struct pipe_resource**)&dispatch_buf);
567
568 if (!dispatch_buf) {
569 fprintf(stderr, "Error: Failed to allocate dispatch "
570 "packet.");
571 }
572 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, dispatch_buf,
573 RADEON_USAGE_READ, RADEON_PRIO_CONST_BUFFER);
574
575 dispatch_va = dispatch_buf->gpu_address + dispatch_offset;
576
577 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0 +
578 (user_sgpr * 4), 2);
579 radeon_emit(cs, dispatch_va);
580 radeon_emit(cs, S_008F04_BASE_ADDRESS_HI(dispatch_va >> 32) |
581 S_008F04_STRIDE(0));
582
583 r600_resource_reference(&dispatch_buf, NULL);
584 user_sgpr += 2;
585 }
586
587 if (AMD_HSA_BITS_GET(code_object->code_properties,
588 AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR)) {
589 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0 +
590 (user_sgpr * 4), 2);
591 radeon_emit(cs, kernel_args_va);
592 radeon_emit(cs, S_008F04_BASE_ADDRESS_HI (kernel_args_va >> 32) |
593 S_008F04_STRIDE(0));
594 user_sgpr += 2;
595 }
596
597 for (i = 0; i < 3 && user_sgpr < 16; i++) {
598 if (code_object->code_properties & workgroup_count_masks[i]) {
599 radeon_set_sh_reg_seq(cs,
600 R_00B900_COMPUTE_USER_DATA_0 +
601 (user_sgpr * 4), 1);
602 radeon_emit(cs, info->grid[i]);
603 user_sgpr += 1;
604 }
605 }
606 }
607
si_upload_compute_input(struct si_context * sctx,const amd_kernel_code_t * code_object,const struct pipe_grid_info * info)608 static bool si_upload_compute_input(struct si_context *sctx,
609 const amd_kernel_code_t *code_object,
610 const struct pipe_grid_info *info)
611 {
612 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
613 struct si_compute *program = sctx->cs_shader_state.program;
614 struct r600_resource *input_buffer = NULL;
615 unsigned kernel_args_size;
616 unsigned num_work_size_bytes = program->use_code_object_v2 ? 0 : 36;
617 uint32_t kernel_args_offset = 0;
618 uint32_t *kernel_args;
619 void *kernel_args_ptr;
620 uint64_t kernel_args_va;
621 unsigned i;
622
623 /* The extra num_work_size_bytes are for work group / work item size information */
624 kernel_args_size = program->input_size + num_work_size_bytes;
625
626 u_upload_alloc(sctx->b.b.const_uploader, 0, kernel_args_size,
627 sctx->screen->info.tcc_cache_line_size,
628 &kernel_args_offset,
629 (struct pipe_resource**)&input_buffer, &kernel_args_ptr);
630
631 if (unlikely(!kernel_args_ptr))
632 return false;
633
634 kernel_args = (uint32_t*)kernel_args_ptr;
635 kernel_args_va = input_buffer->gpu_address + kernel_args_offset;
636
637 if (!code_object) {
638 for (i = 0; i < 3; i++) {
639 kernel_args[i] = info->grid[i];
640 kernel_args[i + 3] = info->grid[i] * info->block[i];
641 kernel_args[i + 6] = info->block[i];
642 }
643 }
644
645 memcpy(kernel_args + (num_work_size_bytes / 4), info->input,
646 program->input_size);
647
648
649 for (i = 0; i < (kernel_args_size / 4); i++) {
650 COMPUTE_DBG(sctx->screen, "input %u : %u\n", i,
651 kernel_args[i]);
652 }
653
654
655 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, input_buffer,
656 RADEON_USAGE_READ, RADEON_PRIO_CONST_BUFFER);
657
658 if (code_object) {
659 si_setup_user_sgprs_co_v2(sctx, code_object, info, kernel_args_va);
660 } else {
661 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0, 2);
662 radeon_emit(cs, kernel_args_va);
663 radeon_emit(cs, S_008F04_BASE_ADDRESS_HI (kernel_args_va >> 32) |
664 S_008F04_STRIDE(0));
665 }
666
667 r600_resource_reference(&input_buffer, NULL);
668
669 return true;
670 }
671
si_setup_tgsi_grid(struct si_context * sctx,const struct pipe_grid_info * info)672 static void si_setup_tgsi_grid(struct si_context *sctx,
673 const struct pipe_grid_info *info)
674 {
675 struct si_compute *program = sctx->cs_shader_state.program;
676 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
677 unsigned grid_size_reg = R_00B900_COMPUTE_USER_DATA_0 +
678 4 * SI_NUM_RESOURCE_SGPRS;
679 unsigned block_size_reg = grid_size_reg +
680 /* 12 bytes = 3 dwords. */
681 12 * program->uses_grid_size;
682
683 if (info->indirect) {
684 if (program->uses_grid_size) {
685 uint64_t base_va = r600_resource(info->indirect)->gpu_address;
686 uint64_t va = base_va + info->indirect_offset;
687 int i;
688
689 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
690 (struct r600_resource *)info->indirect,
691 RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
692
693 for (i = 0; i < 3; ++i) {
694 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
695 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
696 COPY_DATA_DST_SEL(COPY_DATA_REG));
697 radeon_emit(cs, (va + 4 * i));
698 radeon_emit(cs, (va + 4 * i) >> 32);
699 radeon_emit(cs, (grid_size_reg >> 2) + i);
700 radeon_emit(cs, 0);
701 }
702 }
703 } else {
704 if (program->uses_grid_size) {
705 radeon_set_sh_reg_seq(cs, grid_size_reg, 3);
706 radeon_emit(cs, info->grid[0]);
707 radeon_emit(cs, info->grid[1]);
708 radeon_emit(cs, info->grid[2]);
709 }
710 if (program->variable_group_size && program->uses_block_size) {
711 radeon_set_sh_reg_seq(cs, block_size_reg, 3);
712 radeon_emit(cs, info->block[0]);
713 radeon_emit(cs, info->block[1]);
714 radeon_emit(cs, info->block[2]);
715 }
716 }
717 }
718
si_emit_dispatch_packets(struct si_context * sctx,const struct pipe_grid_info * info)719 static void si_emit_dispatch_packets(struct si_context *sctx,
720 const struct pipe_grid_info *info)
721 {
722 struct si_screen *sscreen = sctx->screen;
723 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
724 bool render_cond_bit = sctx->b.render_cond && !sctx->b.render_cond_force_off;
725 unsigned waves_per_threadgroup =
726 DIV_ROUND_UP(info->block[0] * info->block[1] * info->block[2], 64);
727 unsigned compute_resource_limits =
728 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
729
730 if (sctx->b.chip_class >= CIK) {
731 unsigned num_cu_per_se = sscreen->info.num_good_compute_units /
732 sscreen->info.max_se;
733
734 /* Force even distribution on all SIMDs in CU if the workgroup
735 * size is 64. This has shown some good improvements if # of CUs
736 * per SE is not a multiple of 4.
737 */
738 if (num_cu_per_se % 4 && waves_per_threadgroup == 1)
739 compute_resource_limits |= S_00B854_FORCE_SIMD_DIST(1);
740 }
741
742 radeon_set_sh_reg(cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
743 compute_resource_limits);
744
745 radeon_set_sh_reg_seq(cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
746 radeon_emit(cs, S_00B81C_NUM_THREAD_FULL(info->block[0]));
747 radeon_emit(cs, S_00B820_NUM_THREAD_FULL(info->block[1]));
748 radeon_emit(cs, S_00B824_NUM_THREAD_FULL(info->block[2]));
749
750 unsigned dispatch_initiator =
751 S_00B800_COMPUTE_SHADER_EN(1) |
752 S_00B800_FORCE_START_AT_000(1) |
753 /* If the KMD allows it (there is a KMD hw register for it),
754 * allow launching waves out-of-order. (same as Vulkan) */
755 S_00B800_ORDER_MODE(sctx->b.chip_class >= CIK);
756
757 if (info->indirect) {
758 uint64_t base_va = r600_resource(info->indirect)->gpu_address;
759
760 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
761 (struct r600_resource *)info->indirect,
762 RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
763
764 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0) |
765 PKT3_SHADER_TYPE_S(1));
766 radeon_emit(cs, 1);
767 radeon_emit(cs, base_va);
768 radeon_emit(cs, base_va >> 32);
769
770 radeon_emit(cs, PKT3(PKT3_DISPATCH_INDIRECT, 1, render_cond_bit) |
771 PKT3_SHADER_TYPE_S(1));
772 radeon_emit(cs, info->indirect_offset);
773 radeon_emit(cs, dispatch_initiator);
774 } else {
775 radeon_emit(cs, PKT3(PKT3_DISPATCH_DIRECT, 3, render_cond_bit) |
776 PKT3_SHADER_TYPE_S(1));
777 radeon_emit(cs, info->grid[0]);
778 radeon_emit(cs, info->grid[1]);
779 radeon_emit(cs, info->grid[2]);
780 radeon_emit(cs, dispatch_initiator);
781 }
782 }
783
784
si_launch_grid(struct pipe_context * ctx,const struct pipe_grid_info * info)785 static void si_launch_grid(
786 struct pipe_context *ctx, const struct pipe_grid_info *info)
787 {
788 struct si_context *sctx = (struct si_context*)ctx;
789 struct si_compute *program = sctx->cs_shader_state.program;
790 const amd_kernel_code_t *code_object =
791 si_compute_get_code_object(program, info->pc);
792 int i;
793 /* HW bug workaround when CS threadgroups > 256 threads and async
794 * compute isn't used, i.e. only one compute job can run at a time.
795 * If async compute is possible, the threadgroup size must be limited
796 * to 256 threads on all queues to avoid the bug.
797 * Only SI and certain CIK chips are affected.
798 */
799 bool cs_regalloc_hang =
800 (sctx->b.chip_class == SI ||
801 sctx->b.family == CHIP_BONAIRE ||
802 sctx->b.family == CHIP_KABINI) &&
803 info->block[0] * info->block[1] * info->block[2] > 256;
804
805 if (cs_regalloc_hang)
806 sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
807 SI_CONTEXT_CS_PARTIAL_FLUSH;
808
809 if (program->ir_type == PIPE_SHADER_IR_TGSI &&
810 program->shader.compilation_failed)
811 return;
812
813 if (sctx->b.last_num_draw_calls != sctx->b.num_draw_calls) {
814 si_update_fb_dirtiness_after_rendering(sctx);
815 sctx->b.last_num_draw_calls = sctx->b.num_draw_calls;
816 }
817
818 si_decompress_textures(sctx, 1 << PIPE_SHADER_COMPUTE);
819
820 /* Add buffer sizes for memory checking in need_cs_space. */
821 si_context_add_resource_size(ctx, &program->shader.bo->b.b);
822 /* TODO: add the scratch buffer */
823
824 if (info->indirect) {
825 si_context_add_resource_size(ctx, info->indirect);
826
827 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
828 if (sctx->b.chip_class <= VI &&
829 r600_resource(info->indirect)->TC_L2_dirty) {
830 sctx->b.flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
831 r600_resource(info->indirect)->TC_L2_dirty = false;
832 }
833 }
834
835 si_need_cs_space(sctx);
836
837 if (!sctx->cs_shader_state.initialized)
838 si_initialize_compute(sctx);
839
840 if (sctx->b.flags)
841 si_emit_cache_flush(sctx);
842
843 if (!si_switch_compute_shader(sctx, program, &program->shader,
844 code_object, info->pc))
845 return;
846
847 si_upload_compute_shader_descriptors(sctx);
848 si_emit_compute_shader_pointers(sctx);
849
850 if (si_is_atom_dirty(sctx, sctx->atoms.s.render_cond)) {
851 sctx->atoms.s.render_cond->emit(&sctx->b,
852 sctx->atoms.s.render_cond);
853 si_set_atom_dirty(sctx, sctx->atoms.s.render_cond, false);
854 }
855
856 if ((program->input_size ||
857 program->ir_type == PIPE_SHADER_IR_NATIVE) &&
858 unlikely(!si_upload_compute_input(sctx, code_object, info))) {
859 return;
860 }
861
862 /* Global buffers */
863 for (i = 0; i < MAX_GLOBAL_BUFFERS; i++) {
864 struct r600_resource *buffer =
865 (struct r600_resource*)program->global_buffers[i];
866 if (!buffer) {
867 continue;
868 }
869 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, buffer,
870 RADEON_USAGE_READWRITE,
871 RADEON_PRIO_COMPUTE_GLOBAL);
872 }
873
874 if (program->ir_type == PIPE_SHADER_IR_TGSI)
875 si_setup_tgsi_grid(sctx, info);
876
877 si_emit_dispatch_packets(sctx, info);
878
879 if (unlikely(sctx->current_saved_cs)) {
880 si_trace_emit(sctx);
881 si_log_compute_state(sctx, sctx->b.log);
882 }
883
884 sctx->compute_is_busy = true;
885 sctx->b.num_compute_calls++;
886 if (sctx->cs_shader_state.uses_scratch)
887 sctx->b.num_spill_compute_calls++;
888
889 if (cs_regalloc_hang)
890 sctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
891 }
892
si_destroy_compute(struct si_compute * program)893 void si_destroy_compute(struct si_compute *program)
894 {
895 if (program->ir_type == PIPE_SHADER_IR_TGSI) {
896 util_queue_drop_job(&program->screen->shader_compiler_queue,
897 &program->ready);
898 util_queue_fence_destroy(&program->ready);
899 }
900
901 si_shader_destroy(&program->shader);
902 FREE(program);
903 }
904
si_delete_compute_state(struct pipe_context * ctx,void * state)905 static void si_delete_compute_state(struct pipe_context *ctx, void* state){
906 struct si_compute *program = (struct si_compute *)state;
907 struct si_context *sctx = (struct si_context*)ctx;
908
909 if (!state)
910 return;
911
912 if (program == sctx->cs_shader_state.program)
913 sctx->cs_shader_state.program = NULL;
914
915 if (program == sctx->cs_shader_state.emitted_program)
916 sctx->cs_shader_state.emitted_program = NULL;
917
918 si_compute_reference(&program, NULL);
919 }
920
si_set_compute_resources(struct pipe_context * ctx_,unsigned start,unsigned count,struct pipe_surface ** surfaces)921 static void si_set_compute_resources(struct pipe_context * ctx_,
922 unsigned start, unsigned count,
923 struct pipe_surface ** surfaces) { }
924
si_init_compute_functions(struct si_context * sctx)925 void si_init_compute_functions(struct si_context *sctx)
926 {
927 sctx->b.b.create_compute_state = si_create_compute_state;
928 sctx->b.b.delete_compute_state = si_delete_compute_state;
929 sctx->b.b.bind_compute_state = si_bind_compute_state;
930 /* ctx->context.create_sampler_view = evergreen_compute_create_sampler_view; */
931 sctx->b.b.set_compute_resources = si_set_compute_resources;
932 sctx->b.b.set_global_binding = si_set_global_binding;
933 sctx->b.b.launch_grid = si_launch_grid;
934 }
935