projects / mesa.git / blob
? search:


cbf3320a05422b13fa58b0d874a89ca3a45e75e2
[mesa.git] / src / amd / vulkan / radv_cmd_buffer.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 #include "radv_private.h"
29 #include "radv_radeon_winsys.h"
30 #include "radv_shader.h"
31 #include "radv_cs.h"
32 #include "sid.h"
33 #include "gfx9d.h"
34 #include "vk_format.h"
35 #include "radv_debug.h"
36 #include "radv_meta.h"
37
38 #include "ac_debug.h"
39
40 enum {
41 RADV_PREFETCH_VBO_DESCRIPTORS = (1 << 0),
42 RADV_PREFETCH_VS = (1 << 1),
43 RADV_PREFETCH_TCS = (1 << 2),
44 RADV_PREFETCH_TES = (1 << 3),
45 RADV_PREFETCH_GS = (1 << 4),
46 RADV_PREFETCH_PS = (1 << 5),
47 RADV_PREFETCH_SHADERS = (RADV_PREFETCH_VS |
48 RADV_PREFETCH_TCS |
49 RADV_PREFETCH_TES |
50 RADV_PREFETCH_GS |
51 RADV_PREFETCH_PS)
52 };
53
54 static void radv_handle_image_transition(struct radv_cmd_buffer *cmd_buffer,
55 struct radv_image *image,
56 VkImageLayout src_layout,
57 VkImageLayout dst_layout,
58 uint32_t src_family,
59 uint32_t dst_family,
60 const VkImageSubresourceRange *range,
61 VkImageAspectFlags pending_clears);
62
63 const struct radv_dynamic_state default_dynamic_state = {
64 .viewport = {
65 .count = 0,
66 },
67 .scissor = {
68 .count = 0,
69 },
70 .line_width = 1.0f,
71 .depth_bias = {
72 .bias = 0.0f,
73 .clamp = 0.0f,
74 .slope = 0.0f,
75 },
76 .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
77 .depth_bounds = {
78 .min = 0.0f,
79 .max = 1.0f,
80 },
81 .stencil_compare_mask = {
82 .front = ~0u,
83 .back = ~0u,
84 },
85 .stencil_write_mask = {
86 .front = ~0u,
87 .back = ~0u,
88 },
89 .stencil_reference = {
90 .front = 0u,
91 .back = 0u,
92 },
93 };
94
95 static void
96 radv_bind_dynamic_state(struct radv_cmd_buffer *cmd_buffer,
97 const struct radv_dynamic_state *src)
98 {
99 struct radv_dynamic_state *dest = &cmd_buffer->state.dynamic;
100 uint32_t copy_mask = src->mask;
101 uint32_t dest_mask = 0;
102
103 /* Make sure to copy the number of viewports/scissors because they can
104 * only be specified at pipeline creation time.
105 */
106 dest->viewport.count = src->viewport.count;
107 dest->scissor.count = src->scissor.count;
108 dest->discard_rectangle.count = src->discard_rectangle.count;
109
110 if (copy_mask & RADV_DYNAMIC_VIEWPORT) {
111 if (memcmp(&dest->viewport.viewports, &src->viewport.viewports,
112 src->viewport.count * sizeof(VkViewport))) {
113 typed_memcpy(dest->viewport.viewports,
114 src->viewport.viewports,
115 src->viewport.count);
116 dest_mask |= RADV_DYNAMIC_VIEWPORT;
117 }
118 }
119
120 if (copy_mask & RADV_DYNAMIC_SCISSOR) {
121 if (memcmp(&dest->scissor.scissors, &src->scissor.scissors,
122 src->scissor.count * sizeof(VkRect2D))) {
123 typed_memcpy(dest->scissor.scissors,
124 src->scissor.scissors, src->scissor.count);
125 dest_mask |= RADV_DYNAMIC_SCISSOR;
126 }
127 }
128
129 if (copy_mask & RADV_DYNAMIC_LINE_WIDTH) {
130 if (dest->line_width != src->line_width) {
131 dest->line_width = src->line_width;
132 dest_mask |= RADV_DYNAMIC_LINE_WIDTH;
133 }
134 }
135
136 if (copy_mask & RADV_DYNAMIC_DEPTH_BIAS) {
137 if (memcmp(&dest->depth_bias, &src->depth_bias,
138 sizeof(src->depth_bias))) {
139 dest->depth_bias = src->depth_bias;
140 dest_mask |= RADV_DYNAMIC_DEPTH_BIAS;
141 }
142 }
143
144 if (copy_mask & RADV_DYNAMIC_BLEND_CONSTANTS) {
145 if (memcmp(&dest->blend_constants, &src->blend_constants,
146 sizeof(src->blend_constants))) {
147 typed_memcpy(dest->blend_constants,
148 src->blend_constants, 4);
149 dest_mask |= RADV_DYNAMIC_BLEND_CONSTANTS;
150 }
151 }
152
153 if (copy_mask & RADV_DYNAMIC_DEPTH_BOUNDS) {
154 if (memcmp(&dest->depth_bounds, &src->depth_bounds,
155 sizeof(src->depth_bounds))) {
156 dest->depth_bounds = src->depth_bounds;
157 dest_mask |= RADV_DYNAMIC_DEPTH_BOUNDS;
158 }
159 }
160
161 if (copy_mask & RADV_DYNAMIC_STENCIL_COMPARE_MASK) {
162 if (memcmp(&dest->stencil_compare_mask,
163 &src->stencil_compare_mask,
164 sizeof(src->stencil_compare_mask))) {
165 dest->stencil_compare_mask = src->stencil_compare_mask;
166 dest_mask |= RADV_DYNAMIC_STENCIL_COMPARE_MASK;
167 }
168 }
169
170 if (copy_mask & RADV_DYNAMIC_STENCIL_WRITE_MASK) {
171 if (memcmp(&dest->stencil_write_mask, &src->stencil_write_mask,
172 sizeof(src->stencil_write_mask))) {
173 dest->stencil_write_mask = src->stencil_write_mask;
174 dest_mask |= RADV_DYNAMIC_STENCIL_WRITE_MASK;
175 }
176 }
177
178 if (copy_mask & RADV_DYNAMIC_STENCIL_REFERENCE) {
179 if (memcmp(&dest->stencil_reference, &src->stencil_reference,
180 sizeof(src->stencil_reference))) {
181 dest->stencil_reference = src->stencil_reference;
182 dest_mask |= RADV_DYNAMIC_STENCIL_REFERENCE;
183 }
184 }
185
186 if (copy_mask & RADV_DYNAMIC_DISCARD_RECTANGLE) {
187 if (memcmp(&dest->discard_rectangle.rectangles, &src->discard_rectangle.rectangles,
188 src->discard_rectangle.count * sizeof(VkRect2D))) {
189 typed_memcpy(dest->discard_rectangle.rectangles,
190 src->discard_rectangle.rectangles,
191 src->discard_rectangle.count);
192 dest_mask |= RADV_DYNAMIC_DISCARD_RECTANGLE;
193 }
194 }
195
196 cmd_buffer->state.dirty |= dest_mask;
197 }
198
199 bool radv_cmd_buffer_uses_mec(struct radv_cmd_buffer *cmd_buffer)
200 {
201 return cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE &&
202 cmd_buffer->device->physical_device->rad_info.chip_class >= CIK;
203 }
204
205 enum ring_type radv_queue_family_to_ring(int f) {
206 switch (f) {
207 case RADV_QUEUE_GENERAL:
208 return RING_GFX;
209 case RADV_QUEUE_COMPUTE:
210 return RING_COMPUTE;
211 case RADV_QUEUE_TRANSFER:
212 return RING_DMA;
213 default:
214 unreachable("Unknown queue family");
215 }
216 }
217
218 static VkResult radv_create_cmd_buffer(
219 struct radv_device * device,
220 struct radv_cmd_pool * pool,
221 VkCommandBufferLevel level,
222 VkCommandBuffer* pCommandBuffer)
223 {
224 struct radv_cmd_buffer *cmd_buffer;
225 unsigned ring;
226 cmd_buffer = vk_zalloc(&pool->alloc, sizeof(*cmd_buffer), 8,
227 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
228 if (cmd_buffer == NULL)
229 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
230
231 cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
232 cmd_buffer->device = device;
233 cmd_buffer->pool = pool;
234 cmd_buffer->level = level;
235
236 if (pool) {
237 list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
238 cmd_buffer->queue_family_index = pool->queue_family_index;
239
240 } else {
241 /* Init the pool_link so we can safefly call list_del when we destroy
242 * the command buffer
243 */
244 list_inithead(&cmd_buffer->pool_link);
245 cmd_buffer->queue_family_index = RADV_QUEUE_GENERAL;
246 }
247
248 ring = radv_queue_family_to_ring(cmd_buffer->queue_family_index);
249
250 cmd_buffer->cs = device->ws->cs_create(device->ws, ring);
251 if (!cmd_buffer->cs) {
252 vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
253 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
254 }
255
256 *pCommandBuffer = radv_cmd_buffer_to_handle(cmd_buffer);
257
258 list_inithead(&cmd_buffer->upload.list);
259
260 return VK_SUCCESS;
261 }
262
263 static void
264 radv_cmd_buffer_destroy(struct radv_cmd_buffer *cmd_buffer)
265 {
266 list_del(&cmd_buffer->pool_link);
267
268 list_for_each_entry_safe(struct radv_cmd_buffer_upload, up,
269 &cmd_buffer->upload.list, list) {
270 cmd_buffer->device->ws->buffer_destroy(up->upload_bo);
271 list_del(&up->list);
272 free(up);
273 }
274
275 if (cmd_buffer->upload.upload_bo)
276 cmd_buffer->device->ws->buffer_destroy(cmd_buffer->upload.upload_bo);
277 cmd_buffer->device->ws->cs_destroy(cmd_buffer->cs);
278
279 for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++)
280 free(cmd_buffer->descriptors[i].push_set.set.mapped_ptr);
281
282 vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
283 }
284
285 static VkResult
286 radv_reset_cmd_buffer(struct radv_cmd_buffer *cmd_buffer)
287 {
288
289 cmd_buffer->device->ws->cs_reset(cmd_buffer->cs);
290
291 list_for_each_entry_safe(struct radv_cmd_buffer_upload, up,
292 &cmd_buffer->upload.list, list) {
293 cmd_buffer->device->ws->buffer_destroy(up->upload_bo);
294 list_del(&up->list);
295 free(up);
296 }
297
298 cmd_buffer->push_constant_stages = 0;
299 cmd_buffer->scratch_size_needed = 0;
300 cmd_buffer->compute_scratch_size_needed = 0;
301 cmd_buffer->esgs_ring_size_needed = 0;
302 cmd_buffer->gsvs_ring_size_needed = 0;
303 cmd_buffer->tess_rings_needed = false;
304 cmd_buffer->sample_positions_needed = false;
305
306 if (cmd_buffer->upload.upload_bo)
307 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs,
308 cmd_buffer->upload.upload_bo, 8);
309 cmd_buffer->upload.offset = 0;
310
311 cmd_buffer->record_result = VK_SUCCESS;
312
313 cmd_buffer->ring_offsets_idx = -1;
314
315 for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++) {
316 cmd_buffer->descriptors[i].dirty = 0;
317 cmd_buffer->descriptors[i].valid = 0;
318 cmd_buffer->descriptors[i].push_dirty = false;
319 }
320
321 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
322 void *fence_ptr;
323 radv_cmd_buffer_upload_alloc(cmd_buffer, 8, 0,
324 &cmd_buffer->gfx9_fence_offset,
325 &fence_ptr);
326 cmd_buffer->gfx9_fence_bo = cmd_buffer->upload.upload_bo;
327 }
328
329 cmd_buffer->status = RADV_CMD_BUFFER_STATUS_INITIAL;
330
331 return cmd_buffer->record_result;
332 }
333
334 static bool
335 radv_cmd_buffer_resize_upload_buf(struct radv_cmd_buffer *cmd_buffer,
336 uint64_t min_needed)
337 {
338 uint64_t new_size;
339 struct radeon_winsys_bo *bo;
340 struct radv_cmd_buffer_upload *upload;
341 struct radv_device *device = cmd_buffer->device;
342
343 new_size = MAX2(min_needed, 16 * 1024);
344 new_size = MAX2(new_size, 2 * cmd_buffer->upload.size);
345
346 bo = device->ws->buffer_create(device->ws,
347 new_size, 4096,
348 RADEON_DOMAIN_GTT,
349 RADEON_FLAG_CPU_ACCESS|
350 RADEON_FLAG_NO_INTERPROCESS_SHARING);
351
352 if (!bo) {
353 cmd_buffer->record_result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
354 return false;
355 }
356
357 radv_cs_add_buffer(device->ws, cmd_buffer->cs, bo, 8);
358 if (cmd_buffer->upload.upload_bo) {
359 upload = malloc(sizeof(*upload));
360
361 if (!upload) {
362 cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
363 device->ws->buffer_destroy(bo);
364 return false;
365 }
366
367 memcpy(upload, &cmd_buffer->upload, sizeof(*upload));
368 list_add(&upload->list, &cmd_buffer->upload.list);
369 }
370
371 cmd_buffer->upload.upload_bo = bo;
372 cmd_buffer->upload.size = new_size;
373 cmd_buffer->upload.offset = 0;
374 cmd_buffer->upload.map = device->ws->buffer_map(cmd_buffer->upload.upload_bo);
375
376 if (!cmd_buffer->upload.map) {
377 cmd_buffer->record_result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
378 return false;
379 }
380
381 return true;
382 }
383
384 bool
385 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
386 unsigned size,
387 unsigned alignment,
388 unsigned *out_offset,
389 void **ptr)
390 {
391 uint64_t offset = align(cmd_buffer->upload.offset, alignment);
392 if (offset + size > cmd_buffer->upload.size) {
393 if (!radv_cmd_buffer_resize_upload_buf(cmd_buffer, size))
394 return false;
395 offset = 0;
396 }
397
398 *out_offset = offset;
399 *ptr = cmd_buffer->upload.map + offset;
400
401 cmd_buffer->upload.offset = offset + size;
402 return true;
403 }
404
405 bool
406 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
407 unsigned size, unsigned alignment,
408 const void *data, unsigned *out_offset)
409 {
410 uint8_t *ptr;
411
412 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size, alignment,
413 out_offset, (void **)&ptr))
414 return false;
415
416 if (ptr)
417 memcpy(ptr, data, size);
418
419 return true;
420 }
421
422 static void
423 radv_emit_write_data_packet(struct radeon_winsys_cs *cs, uint64_t va,
424 unsigned count, const uint32_t *data)
425 {
426 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 2 + count, 0));
427 radeon_emit(cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
428 S_370_WR_CONFIRM(1) |
429 S_370_ENGINE_SEL(V_370_ME));
430 radeon_emit(cs, va);
431 radeon_emit(cs, va >> 32);
432 radeon_emit_array(cs, data, count);
433 }
434
435 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer *cmd_buffer)
436 {
437 struct radv_device *device = cmd_buffer->device;
438 struct radeon_winsys_cs *cs = cmd_buffer->cs;
439 uint64_t va;
440
441 va = radv_buffer_get_va(device->trace_bo);
442 if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY)
443 va += 4;
444
445 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 7);
446
447 ++cmd_buffer->state.trace_id;
448 radv_cs_add_buffer(device->ws, cs, device->trace_bo, 8);
449 radv_emit_write_data_packet(cs, va, 1, &cmd_buffer->state.trace_id);
450 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
451 radeon_emit(cs, AC_ENCODE_TRACE_POINT(cmd_buffer->state.trace_id));
452 }
453
454 static void
455 radv_cmd_buffer_after_draw(struct radv_cmd_buffer *cmd_buffer,
456 enum radv_cmd_flush_bits flags)
457 {
458 if (cmd_buffer->device->instance->debug_flags & RADV_DEBUG_SYNC_SHADERS) {
459 uint32_t *ptr = NULL;
460 uint64_t va = 0;
461
462 assert(flags & (RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
463 RADV_CMD_FLAG_CS_PARTIAL_FLUSH));
464
465 if (cmd_buffer->device->physical_device->rad_info.chip_class == GFX9) {
466 va = radv_buffer_get_va(cmd_buffer->gfx9_fence_bo) +
467 cmd_buffer->gfx9_fence_offset;
468 ptr = &cmd_buffer->gfx9_fence_idx;
469 }
470
471 /* Force wait for graphics or compute engines to be idle. */
472 si_cs_emit_cache_flush(cmd_buffer->cs,
473 cmd_buffer->device->physical_device->rad_info.chip_class,
474 ptr, va,
475 radv_cmd_buffer_uses_mec(cmd_buffer),
476 flags);
477 }
478
479 if (unlikely(cmd_buffer->device->trace_bo))
480 radv_cmd_buffer_trace_emit(cmd_buffer);
481 }
482
483 static void
484 radv_save_pipeline(struct radv_cmd_buffer *cmd_buffer,
485 struct radv_pipeline *pipeline, enum ring_type ring)
486 {
487 struct radv_device *device = cmd_buffer->device;
488 struct radeon_winsys_cs *cs = cmd_buffer->cs;
489 uint32_t data[2];
490 uint64_t va;
491
492 va = radv_buffer_get_va(device->trace_bo);
493
494 switch (ring) {
495 case RING_GFX:
496 va += 8;
497 break;
498 case RING_COMPUTE:
499 va += 16;
500 break;
501 default:
502 assert(!"invalid ring type");
503 }
504
505 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(device->ws,
506 cmd_buffer->cs, 6);
507
508 data[0] = (uintptr_t)pipeline;
509 data[1] = (uintptr_t)pipeline >> 32;
510
511 radv_cs_add_buffer(device->ws, cs, device->trace_bo, 8);
512 radv_emit_write_data_packet(cs, va, 2, data);
513 }
514
515 void radv_set_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
516 VkPipelineBindPoint bind_point,
517 struct radv_descriptor_set *set,
518 unsigned idx)
519 {
520 struct radv_descriptor_state *descriptors_state =
521 radv_get_descriptors_state(cmd_buffer, bind_point);
522
523 descriptors_state->sets[idx] = set;
524 if (set)
525 descriptors_state->valid |= (1u << idx);
526 else
527 descriptors_state->valid &= ~(1u << idx);
528 descriptors_state->dirty |= (1u << idx);
529 }
530
531 static void
532 radv_save_descriptors(struct radv_cmd_buffer *cmd_buffer,
533 VkPipelineBindPoint bind_point)
534 {
535 struct radv_descriptor_state *descriptors_state =
536 radv_get_descriptors_state(cmd_buffer, bind_point);
537 struct radv_device *device = cmd_buffer->device;
538 struct radeon_winsys_cs *cs = cmd_buffer->cs;
539 uint32_t data[MAX_SETS * 2] = {};
540 uint64_t va;
541 unsigned i;
542 va = radv_buffer_get_va(device->trace_bo) + 24;
543
544 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(device->ws,
545 cmd_buffer->cs, 4 + MAX_SETS * 2);
546
547 for_each_bit(i, descriptors_state->valid) {
548 struct radv_descriptor_set *set = descriptors_state->sets[i];
549 data[i * 2] = (uintptr_t)set;
550 data[i * 2 + 1] = (uintptr_t)set >> 32;
551 }
552
553 radv_cs_add_buffer(device->ws, cs, device->trace_bo, 8);
554 radv_emit_write_data_packet(cs, va, MAX_SETS * 2, data);
555 }
556
557 struct radv_userdata_info *
558 radv_lookup_user_sgpr(struct radv_pipeline *pipeline,
559 gl_shader_stage stage,
560 int idx)
561 {
562 if (stage == MESA_SHADER_VERTEX) {
563 if (pipeline->shaders[MESA_SHADER_VERTEX])
564 return &pipeline->shaders[MESA_SHADER_VERTEX]->info.user_sgprs_locs.shader_data[idx];
565 if (pipeline->shaders[MESA_SHADER_TESS_CTRL])
566 return &pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.user_sgprs_locs.shader_data[idx];
567 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
568 return &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.user_sgprs_locs.shader_data[idx];
569 } else if (stage == MESA_SHADER_TESS_EVAL) {
570 if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
571 return &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.user_sgprs_locs.shader_data[idx];
572 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
573 return &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.user_sgprs_locs.shader_data[idx];
574 }
575 return &pipeline->shaders[stage]->info.user_sgprs_locs.shader_data[idx];
576 }
577
578 static void
579 radv_emit_userdata_address(struct radv_cmd_buffer *cmd_buffer,
580 struct radv_pipeline *pipeline,
581 gl_shader_stage stage,
582 int idx, uint64_t va)
583 {
584 struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, stage, idx);
585 uint32_t base_reg = pipeline->user_data_0[stage];
586 if (loc->sgpr_idx == -1)
587 return;
588 assert(loc->num_sgprs == 2);
589 assert(!loc->indirect);
590 radeon_set_sh_reg_seq(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, 2);
591 radeon_emit(cmd_buffer->cs, va);
592 radeon_emit(cmd_buffer->cs, va >> 32);
593 }
594
595 static void
596 radv_update_multisample_state(struct radv_cmd_buffer *cmd_buffer,
597 struct radv_pipeline *pipeline)
598 {
599 int num_samples = pipeline->graphics.ms.num_samples;
600 struct radv_multisample_state *ms = &pipeline->graphics.ms;
601 struct radv_pipeline *old_pipeline = cmd_buffer->state.emitted_pipeline;
602
603 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.needs_sample_positions)
604 cmd_buffer->sample_positions_needed = true;
605
606 if (old_pipeline && num_samples == old_pipeline->graphics.ms.num_samples)
607 return;
608
609 radeon_set_context_reg_seq(cmd_buffer->cs, R_028BDC_PA_SC_LINE_CNTL, 2);
610 radeon_emit(cmd_buffer->cs, ms->pa_sc_line_cntl);
611 radeon_emit(cmd_buffer->cs, ms->pa_sc_aa_config);
612
613 radeon_set_context_reg(cmd_buffer->cs, R_028A48_PA_SC_MODE_CNTL_0, ms->pa_sc_mode_cntl_0);
614
615 radv_cayman_emit_msaa_sample_locs(cmd_buffer->cs, num_samples);
616
617 /* GFX9: Flush DFSM when the AA mode changes. */
618 if (cmd_buffer->device->dfsm_allowed) {
619 radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
620 radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
621 }
622 }
623
624 static void
625 radv_emit_shader_prefetch(struct radv_cmd_buffer *cmd_buffer,
626 struct radv_shader_variant *shader)
627 {
628 struct radeon_winsys *ws = cmd_buffer->device->ws;
629 struct radeon_winsys_cs *cs = cmd_buffer->cs;
630 uint64_t va;
631
632 if (!shader)
633 return;
634
635 va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
636
637 radv_cs_add_buffer(ws, cs, shader->bo, 8);
638 si_cp_dma_prefetch(cmd_buffer, va, shader->code_size);
639 }
640
641 static void
642 radv_emit_prefetch_L2(struct radv_cmd_buffer *cmd_buffer,
643 struct radv_pipeline *pipeline,
644 bool vertex_stage_only)
645 {
646 struct radv_cmd_state *state = &cmd_buffer->state;
647 uint32_t mask = state->prefetch_L2_mask;
648
649 if (vertex_stage_only) {
650 /* Fast prefetch path for starting draws as soon as possible.
651 */
652 mask = state->prefetch_L2_mask & (RADV_PREFETCH_VS |
653 RADV_PREFETCH_VBO_DESCRIPTORS);
654 }
655
656 if (mask & RADV_PREFETCH_VS)
657 radv_emit_shader_prefetch(cmd_buffer,
658 pipeline->shaders[MESA_SHADER_VERTEX]);
659
660 if (mask & RADV_PREFETCH_VBO_DESCRIPTORS)
661 si_cp_dma_prefetch(cmd_buffer, state->vb_va, state->vb_size);
662
663 if (mask & RADV_PREFETCH_TCS)
664 radv_emit_shader_prefetch(cmd_buffer,
665 pipeline->shaders[MESA_SHADER_TESS_CTRL]);
666
667 if (mask & RADV_PREFETCH_TES)
668 radv_emit_shader_prefetch(cmd_buffer,
669 pipeline->shaders[MESA_SHADER_TESS_EVAL]);
670
671 if (mask & RADV_PREFETCH_GS) {
672 radv_emit_shader_prefetch(cmd_buffer,
673 pipeline->shaders[MESA_SHADER_GEOMETRY]);
674 radv_emit_shader_prefetch(cmd_buffer, pipeline->gs_copy_shader);
675 }
676
677 if (mask & RADV_PREFETCH_PS)
678 radv_emit_shader_prefetch(cmd_buffer,
679 pipeline->shaders[MESA_SHADER_FRAGMENT]);
680
681 state->prefetch_L2_mask &= ~mask;
682 }
683
684 static void
685 radv_emit_graphics_pipeline(struct radv_cmd_buffer *cmd_buffer)
686 {
687 struct radv_pipeline *pipeline = cmd_buffer->state.pipeline;
688
689 if (!pipeline || cmd_buffer->state.emitted_pipeline == pipeline)
690 return;
691
692 radv_update_multisample_state(cmd_buffer, pipeline);
693
694 cmd_buffer->scratch_size_needed =
695 MAX2(cmd_buffer->scratch_size_needed,
696 pipeline->max_waves * pipeline->scratch_bytes_per_wave);
697
698 if (!cmd_buffer->state.emitted_pipeline ||
699 cmd_buffer->state.emitted_pipeline->graphics.can_use_guardband !=
700 pipeline->graphics.can_use_guardband)
701 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_SCISSOR;
702
703 radeon_emit_array(cmd_buffer->cs, pipeline->cs.buf, pipeline->cs.cdw);
704
705 if (unlikely(cmd_buffer->device->trace_bo))
706 radv_save_pipeline(cmd_buffer, pipeline, RING_GFX);
707
708 cmd_buffer->state.emitted_pipeline = pipeline;
709
710 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_PIPELINE;
711 }
712
713 static void
714 radv_emit_viewport(struct radv_cmd_buffer *cmd_buffer)
715 {
716 si_write_viewport(cmd_buffer->cs, 0, cmd_buffer->state.dynamic.viewport.count,
717 cmd_buffer->state.dynamic.viewport.viewports);
718 }
719
720 static void
721 radv_emit_scissor(struct radv_cmd_buffer *cmd_buffer)
722 {
723 uint32_t count = cmd_buffer->state.dynamic.scissor.count;
724
725 /* Vega10/Raven scissor bug workaround. This must be done before VPORT
726 * scissor registers are changed. There is also a more efficient but
727 * more involved alternative workaround.
728 */
729 if (cmd_buffer->device->physical_device->has_scissor_bug) {
730 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
731 si_emit_cache_flush(cmd_buffer);
732 }
733 si_write_scissors(cmd_buffer->cs, 0, count,
734 cmd_buffer->state.dynamic.scissor.scissors,
735 cmd_buffer->state.dynamic.viewport.viewports,
736 cmd_buffer->state.emitted_pipeline->graphics.can_use_guardband);
737 }
738
739 static void
740 radv_emit_discard_rectangle(struct radv_cmd_buffer *cmd_buffer)
741 {
742 if (!cmd_buffer->state.dynamic.discard_rectangle.count)
743 return;
744
745 radeon_set_context_reg_seq(cmd_buffer->cs, R_028210_PA_SC_CLIPRECT_0_TL,
746 cmd_buffer->state.dynamic.discard_rectangle.count * 2);
747 for (unsigned i = 0; i < cmd_buffer->state.dynamic.discard_rectangle.count; ++i) {
748 VkRect2D rect = cmd_buffer->state.dynamic.discard_rectangle.rectangles[i];
749 radeon_emit(cmd_buffer->cs, S_028210_TL_X(rect.offset.x) | S_028210_TL_Y(rect.offset.y));
750 radeon_emit(cmd_buffer->cs, S_028214_BR_X(rect.offset.x + rect.extent.width) |
751 S_028214_BR_Y(rect.offset.y + rect.extent.height));
752 }
753 }
754
755 static void
756 radv_emit_line_width(struct radv_cmd_buffer *cmd_buffer)
757 {
758 unsigned width = cmd_buffer->state.dynamic.line_width * 8;
759
760 radeon_set_context_reg(cmd_buffer->cs, R_028A08_PA_SU_LINE_CNTL,
761 S_028A08_WIDTH(CLAMP(width, 0, 0xFFF)));
762 }
763
764 static void
765 radv_emit_blend_constants(struct radv_cmd_buffer *cmd_buffer)
766 {
767 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
768
769 radeon_set_context_reg_seq(cmd_buffer->cs, R_028414_CB_BLEND_RED, 4);
770 radeon_emit_array(cmd_buffer->cs, (uint32_t *)d->blend_constants, 4);
771 }
772
773 static void
774 radv_emit_stencil(struct radv_cmd_buffer *cmd_buffer)
775 {
776 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
777
778 radeon_set_context_reg_seq(cmd_buffer->cs,
779 R_028430_DB_STENCILREFMASK, 2);
780 radeon_emit(cmd_buffer->cs,
781 S_028430_STENCILTESTVAL(d->stencil_reference.front) |
782 S_028430_STENCILMASK(d->stencil_compare_mask.front) |
783 S_028430_STENCILWRITEMASK(d->stencil_write_mask.front) |
784 S_028430_STENCILOPVAL(1));
785 radeon_emit(cmd_buffer->cs,
786 S_028434_STENCILTESTVAL_BF(d->stencil_reference.back) |
787 S_028434_STENCILMASK_BF(d->stencil_compare_mask.back) |
788 S_028434_STENCILWRITEMASK_BF(d->stencil_write_mask.back) |
789 S_028434_STENCILOPVAL_BF(1));
790 }
791
792 static void
793 radv_emit_depth_bounds(struct radv_cmd_buffer *cmd_buffer)
794 {
795 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
796
797 radeon_set_context_reg(cmd_buffer->cs, R_028020_DB_DEPTH_BOUNDS_MIN,
798 fui(d->depth_bounds.min));
799 radeon_set_context_reg(cmd_buffer->cs, R_028024_DB_DEPTH_BOUNDS_MAX,
800 fui(d->depth_bounds.max));
801 }
802
803 static void
804 radv_emit_depth_bias(struct radv_cmd_buffer *cmd_buffer)
805 {
806 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
807 unsigned slope = fui(d->depth_bias.slope * 16.0f);
808 unsigned bias = fui(d->depth_bias.bias * cmd_buffer->state.offset_scale);
809
810
811 radeon_set_context_reg_seq(cmd_buffer->cs,
812 R_028B7C_PA_SU_POLY_OFFSET_CLAMP, 5);
813 radeon_emit(cmd_buffer->cs, fui(d->depth_bias.clamp)); /* CLAMP */
814 radeon_emit(cmd_buffer->cs, slope); /* FRONT SCALE */
815 radeon_emit(cmd_buffer->cs, bias); /* FRONT OFFSET */
816 radeon_emit(cmd_buffer->cs, slope); /* BACK SCALE */
817 radeon_emit(cmd_buffer->cs, bias); /* BACK OFFSET */
818 }
819
820 static void
821 radv_emit_fb_color_state(struct radv_cmd_buffer *cmd_buffer,
822 int index,
823 struct radv_attachment_info *att,
824 struct radv_image *image,
825 VkImageLayout layout)
826 {
827 bool is_vi = cmd_buffer->device->physical_device->rad_info.chip_class >= VI;
828 struct radv_color_buffer_info *cb = &att->cb;
829 uint32_t cb_color_info = cb->cb_color_info;
830
831 if (!radv_layout_dcc_compressed(image, layout,
832 radv_image_queue_family_mask(image,
833 cmd_buffer->queue_family_index,
834 cmd_buffer->queue_family_index))) {
835 cb_color_info &= C_028C70_DCC_ENABLE;
836 }
837
838 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
839 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C60_CB_COLOR0_BASE + index * 0x3c, 11);
840 radeon_emit(cmd_buffer->cs, cb->cb_color_base);
841 radeon_emit(cmd_buffer->cs, S_028C64_BASE_256B(cb->cb_color_base >> 32));
842 radeon_emit(cmd_buffer->cs, cb->cb_color_attrib2);
843 radeon_emit(cmd_buffer->cs, cb->cb_color_view);
844 radeon_emit(cmd_buffer->cs, cb_color_info);
845 radeon_emit(cmd_buffer->cs, cb->cb_color_attrib);
846 radeon_emit(cmd_buffer->cs, cb->cb_dcc_control);
847 radeon_emit(cmd_buffer->cs, cb->cb_color_cmask);
848 radeon_emit(cmd_buffer->cs, S_028C80_BASE_256B(cb->cb_color_cmask >> 32));
849 radeon_emit(cmd_buffer->cs, cb->cb_color_fmask);
850 radeon_emit(cmd_buffer->cs, S_028C88_BASE_256B(cb->cb_color_fmask >> 32));
851
852 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C94_CB_COLOR0_DCC_BASE + index * 0x3c, 2);
853 radeon_emit(cmd_buffer->cs, cb->cb_dcc_base);
854 radeon_emit(cmd_buffer->cs, S_028C98_BASE_256B(cb->cb_dcc_base >> 32));
855
856 radeon_set_context_reg(cmd_buffer->cs, R_0287A0_CB_MRT0_EPITCH + index * 4,
857 S_0287A0_EPITCH(att->attachment->image->surface.u.gfx9.surf.epitch));
858 } else {
859 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C60_CB_COLOR0_BASE + index * 0x3c, 11);
860 radeon_emit(cmd_buffer->cs, cb->cb_color_base);
861 radeon_emit(cmd_buffer->cs, cb->cb_color_pitch);
862 radeon_emit(cmd_buffer->cs, cb->cb_color_slice);
863 radeon_emit(cmd_buffer->cs, cb->cb_color_view);
864 radeon_emit(cmd_buffer->cs, cb_color_info);
865 radeon_emit(cmd_buffer->cs, cb->cb_color_attrib);
866 radeon_emit(cmd_buffer->cs, cb->cb_dcc_control);
867 radeon_emit(cmd_buffer->cs, cb->cb_color_cmask);
868 radeon_emit(cmd_buffer->cs, cb->cb_color_cmask_slice);
869 radeon_emit(cmd_buffer->cs, cb->cb_color_fmask);
870 radeon_emit(cmd_buffer->cs, cb->cb_color_fmask_slice);
871
872 if (is_vi) { /* DCC BASE */
873 radeon_set_context_reg(cmd_buffer->cs, R_028C94_CB_COLOR0_DCC_BASE + index * 0x3c, cb->cb_dcc_base);
874 }
875 }
876 }
877
878 static void
879 radv_emit_fb_ds_state(struct radv_cmd_buffer *cmd_buffer,
880 struct radv_ds_buffer_info *ds,
881 struct radv_image *image,
882 VkImageLayout layout)
883 {
884 uint32_t db_z_info = ds->db_z_info;
885 uint32_t db_stencil_info = ds->db_stencil_info;
886
887 if (!radv_layout_has_htile(image, layout,
888 radv_image_queue_family_mask(image,
889 cmd_buffer->queue_family_index,
890 cmd_buffer->queue_family_index))) {
891 db_z_info &= C_028040_TILE_SURFACE_ENABLE;
892 db_stencil_info |= S_028044_TILE_STENCIL_DISABLE(1);
893 }
894
895 radeon_set_context_reg(cmd_buffer->cs, R_028008_DB_DEPTH_VIEW, ds->db_depth_view);
896 radeon_set_context_reg(cmd_buffer->cs, R_028ABC_DB_HTILE_SURFACE, ds->db_htile_surface);
897
898
899 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
900 radeon_set_context_reg_seq(cmd_buffer->cs, R_028014_DB_HTILE_DATA_BASE, 3);
901 radeon_emit(cmd_buffer->cs, ds->db_htile_data_base);
902 radeon_emit(cmd_buffer->cs, S_028018_BASE_HI(ds->db_htile_data_base >> 32));
903 radeon_emit(cmd_buffer->cs, ds->db_depth_size);
904
905 radeon_set_context_reg_seq(cmd_buffer->cs, R_028038_DB_Z_INFO, 10);
906 radeon_emit(cmd_buffer->cs, db_z_info); /* DB_Z_INFO */
907 radeon_emit(cmd_buffer->cs, db_stencil_info); /* DB_STENCIL_INFO */
908 radeon_emit(cmd_buffer->cs, ds->db_z_read_base); /* DB_Z_READ_BASE */
909 radeon_emit(cmd_buffer->cs, S_028044_BASE_HI(ds->db_z_read_base >> 32)); /* DB_Z_READ_BASE_HI */
910 radeon_emit(cmd_buffer->cs, ds->db_stencil_read_base); /* DB_STENCIL_READ_BASE */
911 radeon_emit(cmd_buffer->cs, S_02804C_BASE_HI(ds->db_stencil_read_base >> 32)); /* DB_STENCIL_READ_BASE_HI */
912 radeon_emit(cmd_buffer->cs, ds->db_z_write_base); /* DB_Z_WRITE_BASE */
913 radeon_emit(cmd_buffer->cs, S_028054_BASE_HI(ds->db_z_write_base >> 32)); /* DB_Z_WRITE_BASE_HI */
914 radeon_emit(cmd_buffer->cs, ds->db_stencil_write_base); /* DB_STENCIL_WRITE_BASE */
915 radeon_emit(cmd_buffer->cs, S_02805C_BASE_HI(ds->db_stencil_write_base >> 32)); /* DB_STENCIL_WRITE_BASE_HI */
916
917 radeon_set_context_reg_seq(cmd_buffer->cs, R_028068_DB_Z_INFO2, 2);
918 radeon_emit(cmd_buffer->cs, ds->db_z_info2);
919 radeon_emit(cmd_buffer->cs, ds->db_stencil_info2);
920 } else {
921 radeon_set_context_reg(cmd_buffer->cs, R_028014_DB_HTILE_DATA_BASE, ds->db_htile_data_base);
922
923 radeon_set_context_reg_seq(cmd_buffer->cs, R_02803C_DB_DEPTH_INFO, 9);
924 radeon_emit(cmd_buffer->cs, ds->db_depth_info); /* R_02803C_DB_DEPTH_INFO */
925 radeon_emit(cmd_buffer->cs, db_z_info); /* R_028040_DB_Z_INFO */
926 radeon_emit(cmd_buffer->cs, db_stencil_info); /* R_028044_DB_STENCIL_INFO */
927 radeon_emit(cmd_buffer->cs, ds->db_z_read_base); /* R_028048_DB_Z_READ_BASE */
928 radeon_emit(cmd_buffer->cs, ds->db_stencil_read_base); /* R_02804C_DB_STENCIL_READ_BASE */
929 radeon_emit(cmd_buffer->cs, ds->db_z_write_base); /* R_028050_DB_Z_WRITE_BASE */
930 radeon_emit(cmd_buffer->cs, ds->db_stencil_write_base); /* R_028054_DB_STENCIL_WRITE_BASE */
931 radeon_emit(cmd_buffer->cs, ds->db_depth_size); /* R_028058_DB_DEPTH_SIZE */
932 radeon_emit(cmd_buffer->cs, ds->db_depth_slice); /* R_02805C_DB_DEPTH_SLICE */
933
934 }
935
936 radeon_set_context_reg(cmd_buffer->cs, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
937 ds->pa_su_poly_offset_db_fmt_cntl);
938 }
939
940 void
941 radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
942 struct radv_image *image,
943 VkClearDepthStencilValue ds_clear_value,
944 VkImageAspectFlags aspects)
945 {
946 uint64_t va = radv_buffer_get_va(image->bo);
947 va += image->offset + image->clear_value_offset;
948 unsigned reg_offset = 0, reg_count = 0;
949
950 assert(image->surface.htile_size);
951
952 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
953 ++reg_count;
954 } else {
955 ++reg_offset;
956 va += 4;
957 }
958 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
959 ++reg_count;
960
961 radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 2 + reg_count, 0));
962 radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
963 S_370_WR_CONFIRM(1) |
964 S_370_ENGINE_SEL(V_370_PFP));
965 radeon_emit(cmd_buffer->cs, va);
966 radeon_emit(cmd_buffer->cs, va >> 32);
967 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT)
968 radeon_emit(cmd_buffer->cs, ds_clear_value.stencil);
969 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
970 radeon_emit(cmd_buffer->cs, fui(ds_clear_value.depth));
971
972 radeon_set_context_reg_seq(cmd_buffer->cs, R_028028_DB_STENCIL_CLEAR + 4 * reg_offset, reg_count);
973 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT)
974 radeon_emit(cmd_buffer->cs, ds_clear_value.stencil); /* R_028028_DB_STENCIL_CLEAR */
975 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
976 radeon_emit(cmd_buffer->cs, fui(ds_clear_value.depth)); /* R_02802C_DB_DEPTH_CLEAR */
977 }
978
979 static void
980 radv_load_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
981 struct radv_image *image)
982 {
983 VkImageAspectFlags aspects = vk_format_aspects(image->vk_format);
984 uint64_t va = radv_buffer_get_va(image->bo);
985 va += image->offset + image->clear_value_offset;
986 unsigned reg_offset = 0, reg_count = 0;
987
988 if (!image->surface.htile_size)
989 return;
990
991 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
992 ++reg_count;
993 } else {
994 ++reg_offset;
995 va += 4;
996 }
997 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
998 ++reg_count;
999
1000 radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, 0));
1001 radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
1002 COPY_DATA_DST_SEL(COPY_DATA_REG) |
1003 (reg_count == 2 ? COPY_DATA_COUNT_SEL : 0));
1004 radeon_emit(cmd_buffer->cs, va);
1005 radeon_emit(cmd_buffer->cs, va >> 32);
1006 radeon_emit(cmd_buffer->cs, (R_028028_DB_STENCIL_CLEAR + 4 * reg_offset) >> 2);
1007 radeon_emit(cmd_buffer->cs, 0);
1008
1009 radeon_emit(cmd_buffer->cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
1010 radeon_emit(cmd_buffer->cs, 0);
1011 }
1012
1013 /*
1014 *with DCC some colors don't require CMASK elimiation before being
1015 * used as a texture. This sets a predicate value to determine if the
1016 * cmask eliminate is required.
1017 */
1018 void
1019 radv_set_dcc_need_cmask_elim_pred(struct radv_cmd_buffer *cmd_buffer,
1020 struct radv_image *image,
1021 bool value)
1022 {
1023 uint64_t pred_val = value;
1024 uint64_t va = radv_buffer_get_va(image->bo);
1025 va += image->offset + image->dcc_pred_offset;
1026
1027 assert(image->surface.dcc_size);
1028
1029 radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 4, 0));
1030 radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
1031 S_370_WR_CONFIRM(1) |
1032 S_370_ENGINE_SEL(V_370_PFP));
1033 radeon_emit(cmd_buffer->cs, va);
1034 radeon_emit(cmd_buffer->cs, va >> 32);
1035 radeon_emit(cmd_buffer->cs, pred_val);
1036 radeon_emit(cmd_buffer->cs, pred_val >> 32);
1037 }
1038
1039 void
1040 radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1041 struct radv_image *image,
1042 int idx,
1043 uint32_t color_values[2])
1044 {
1045 uint64_t va = radv_buffer_get_va(image->bo);
1046 va += image->offset + image->clear_value_offset;
1047
1048 assert(image->cmask.size || image->surface.dcc_size);
1049
1050 radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 4, 0));
1051 radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
1052 S_370_WR_CONFIRM(1) |
1053 S_370_ENGINE_SEL(V_370_PFP));
1054 radeon_emit(cmd_buffer->cs, va);
1055 radeon_emit(cmd_buffer->cs, va >> 32);
1056 radeon_emit(cmd_buffer->cs, color_values[0]);
1057 radeon_emit(cmd_buffer->cs, color_values[1]);
1058
1059 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C8C_CB_COLOR0_CLEAR_WORD0 + idx * 0x3c, 2);
1060 radeon_emit(cmd_buffer->cs, color_values[0]);
1061 radeon_emit(cmd_buffer->cs, color_values[1]);
1062 }
1063
1064 static void
1065 radv_load_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1066 struct radv_image *image,
1067 int idx)
1068 {
1069 uint64_t va = radv_buffer_get_va(image->bo);
1070 va += image->offset + image->clear_value_offset;
1071
1072 if (!image->cmask.size && !image->surface.dcc_size)
1073 return;
1074
1075 uint32_t reg = R_028C8C_CB_COLOR0_CLEAR_WORD0 + idx * 0x3c;
1076
1077 radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, cmd_buffer->state.predicating));
1078 radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
1079 COPY_DATA_DST_SEL(COPY_DATA_REG) |
1080 COPY_DATA_COUNT_SEL);
1081 radeon_emit(cmd_buffer->cs, va);
1082 radeon_emit(cmd_buffer->cs, va >> 32);
1083 radeon_emit(cmd_buffer->cs, reg >> 2);
1084 radeon_emit(cmd_buffer->cs, 0);
1085
1086 radeon_emit(cmd_buffer->cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
1087 radeon_emit(cmd_buffer->cs, 0);
1088 }
1089
1090 static void
1091 radv_emit_framebuffer_state(struct radv_cmd_buffer *cmd_buffer)
1092 {
1093 int i;
1094 struct radv_framebuffer *framebuffer = cmd_buffer->state.framebuffer;
1095 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
1096
1097 /* this may happen for inherited secondary recording */
1098 if (!framebuffer)
1099 return;
1100
1101 for (i = 0; i < 8; ++i) {
1102 if (i >= subpass->color_count || subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED) {
1103 radeon_set_context_reg(cmd_buffer->cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
1104 S_028C70_FORMAT(V_028C70_COLOR_INVALID));
1105 continue;
1106 }
1107
1108 int idx = subpass->color_attachments[i].attachment;
1109 struct radv_attachment_info *att = &framebuffer->attachments[idx];
1110 struct radv_image *image = att->attachment->image;
1111 VkImageLayout layout = subpass->color_attachments[i].layout;
1112
1113 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, att->attachment->bo, 8);
1114
1115 assert(att->attachment->aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT);
1116 radv_emit_fb_color_state(cmd_buffer, i, att, image, layout);
1117
1118 radv_load_color_clear_regs(cmd_buffer, image, i);
1119 }
1120
1121 if(subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
1122 int idx = subpass->depth_stencil_attachment.attachment;
1123 VkImageLayout layout = subpass->depth_stencil_attachment.layout;
1124 struct radv_attachment_info *att = &framebuffer->attachments[idx];
1125 struct radv_image *image = att->attachment->image;
1126 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, att->attachment->bo, 8);
1127 MAYBE_UNUSED uint32_t queue_mask = radv_image_queue_family_mask(image,
1128 cmd_buffer->queue_family_index,
1129 cmd_buffer->queue_family_index);
1130 /* We currently don't support writing decompressed HTILE */
1131 assert(radv_layout_has_htile(image, layout, queue_mask) ==
1132 radv_layout_is_htile_compressed(image, layout, queue_mask));
1133
1134 radv_emit_fb_ds_state(cmd_buffer, &att->ds, image, layout);
1135
1136 if (att->ds.offset_scale != cmd_buffer->state.offset_scale) {
1137 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
1138 cmd_buffer->state.offset_scale = att->ds.offset_scale;
1139 }
1140 radv_load_depth_clear_regs(cmd_buffer, image);
1141 } else {
1142 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1143 radeon_set_context_reg_seq(cmd_buffer->cs, R_028038_DB_Z_INFO, 2);
1144 else
1145 radeon_set_context_reg_seq(cmd_buffer->cs, R_028040_DB_Z_INFO, 2);
1146
1147 radeon_emit(cmd_buffer->cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* DB_Z_INFO */
1148 radeon_emit(cmd_buffer->cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* DB_STENCIL_INFO */
1149 }
1150 radeon_set_context_reg(cmd_buffer->cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
1151 S_028208_BR_X(framebuffer->width) |
1152 S_028208_BR_Y(framebuffer->height));
1153
1154 if (cmd_buffer->device->dfsm_allowed) {
1155 radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1156 radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
1157 }
1158
1159 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_FRAMEBUFFER;
1160 }
1161
1162 static void
1163 radv_emit_index_buffer(struct radv_cmd_buffer *cmd_buffer)
1164 {
1165 struct radeon_winsys_cs *cs = cmd_buffer->cs;
1166 struct radv_cmd_state *state = &cmd_buffer->state;
1167
1168 if (state->index_type != state->last_index_type) {
1169 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1170 radeon_set_uconfig_reg_idx(cs, R_03090C_VGT_INDEX_TYPE,
1171 2, state->index_type);
1172 } else {
1173 radeon_emit(cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
1174 radeon_emit(cs, state->index_type);
1175 }
1176
1177 state->last_index_type = state->index_type;
1178 }
1179
1180 radeon_emit(cs, PKT3(PKT3_INDEX_BASE, 1, 0));
1181 radeon_emit(cs, state->index_va);
1182 radeon_emit(cs, state->index_va >> 32);
1183
1184 radeon_emit(cs, PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
1185 radeon_emit(cs, state->max_index_count);
1186
1187 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_INDEX_BUFFER;
1188 }
1189
1190 void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer)
1191 {
1192 struct radv_pipeline *pipeline = cmd_buffer->state.pipeline;
1193 uint32_t pa_sc_mode_cntl_1 =
1194 pipeline ? pipeline->graphics.ms.pa_sc_mode_cntl_1 : 0;
1195 uint32_t db_count_control;
1196
1197 if(!cmd_buffer->state.active_occlusion_queries) {
1198 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
1199 if (G_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(pa_sc_mode_cntl_1) &&
1200 pipeline->graphics.disable_out_of_order_rast_for_occlusion) {
1201 /* Re-enable out-of-order rasterization if the
1202 * bound pipeline supports it and if it's has
1203 * been disabled before starting occlusion
1204 * queries.
1205 */
1206 radeon_set_context_reg(cmd_buffer->cs,
1207 R_028A4C_PA_SC_MODE_CNTL_1,
1208 pa_sc_mode_cntl_1);
1209 }
1210 db_count_control = 0;
1211 } else {
1212 db_count_control = S_028004_ZPASS_INCREMENT_DISABLE(1);
1213 }
1214 } else {
1215 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
1216 uint32_t sample_rate = subpass ? util_logbase2(subpass->max_sample_count) : 0;
1217 bool perfect = cmd_buffer->state.perfect_occlusion_queries_enabled;
1218
1219 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
1220 db_count_control =
1221 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1222 S_028004_SAMPLE_RATE(sample_rate) |
1223 S_028004_ZPASS_ENABLE(1) |
1224 S_028004_SLICE_EVEN_ENABLE(1) |
1225 S_028004_SLICE_ODD_ENABLE(1);
1226
1227 if (G_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(pa_sc_mode_cntl_1) &&
1228 pipeline->graphics.disable_out_of_order_rast_for_occlusion) {
1229 /* If the bound pipeline has enabled
1230 * out-of-order rasterization, we should
1231 * disable it before starting occlusion
1232 * queries.
1233 */
1234 pa_sc_mode_cntl_1 &= C_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE;
1235
1236 radeon_set_context_reg(cmd_buffer->cs,
1237 R_028A4C_PA_SC_MODE_CNTL_1,
1238 pa_sc_mode_cntl_1);
1239 }
1240 } else {
1241 db_count_control = S_028004_PERFECT_ZPASS_COUNTS(1) |
1242 S_028004_SAMPLE_RATE(sample_rate);
1243 }
1244 }
1245
1246 radeon_set_context_reg(cmd_buffer->cs, R_028004_DB_COUNT_CONTROL, db_count_control);
1247 }
1248
1249 static void
1250 radv_cmd_buffer_flush_dynamic_state(struct radv_cmd_buffer *cmd_buffer)
1251 {
1252 uint32_t states = cmd_buffer->state.dirty & cmd_buffer->state.emitted_pipeline->graphics.needed_dynamic_state;
1253
1254 if (states & (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT))
1255 radv_emit_viewport(cmd_buffer);
1256
1257 if (states & (RADV_CMD_DIRTY_DYNAMIC_SCISSOR | RADV_CMD_DIRTY_DYNAMIC_VIEWPORT))
1258 radv_emit_scissor(cmd_buffer);
1259
1260 if (states & RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH)
1261 radv_emit_line_width(cmd_buffer);
1262
1263 if (states & RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS)
1264 radv_emit_blend_constants(cmd_buffer);
1265
1266 if (states & (RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE |
1267 RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK |
1268 RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK))
1269 radv_emit_stencil(cmd_buffer);
1270
1271 if (states & RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS)
1272 radv_emit_depth_bounds(cmd_buffer);
1273
1274 if (states & RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS)
1275 radv_emit_depth_bias(cmd_buffer);
1276
1277 if (states & RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE)
1278 radv_emit_discard_rectangle(cmd_buffer);
1279
1280 cmd_buffer->state.dirty &= ~states;
1281 }
1282
1283 static void
1284 emit_stage_descriptor_set_userdata(struct radv_cmd_buffer *cmd_buffer,
1285 struct radv_pipeline *pipeline,
1286 int idx,
1287 uint64_t va,
1288 gl_shader_stage stage)
1289 {
1290 struct radv_userdata_info *desc_set_loc = &pipeline->shaders[stage]->info.user_sgprs_locs.descriptor_sets[idx];
1291 uint32_t base_reg = pipeline->user_data_0[stage];
1292
1293 if (desc_set_loc->sgpr_idx == -1 || desc_set_loc->indirect)
1294 return;
1295
1296 assert(!desc_set_loc->indirect);
1297 assert(desc_set_loc->num_sgprs == 2);
1298 radeon_set_sh_reg_seq(cmd_buffer->cs,
1299 base_reg + desc_set_loc->sgpr_idx * 4, 2);
1300 radeon_emit(cmd_buffer->cs, va);
1301 radeon_emit(cmd_buffer->cs, va >> 32);
1302 }
1303
1304 static void
1305 radv_emit_descriptor_set_userdata(struct radv_cmd_buffer *cmd_buffer,
1306 VkShaderStageFlags stages,
1307 struct radv_descriptor_set *set,
1308 unsigned idx)
1309 {
1310 if (cmd_buffer->state.pipeline) {
1311 radv_foreach_stage(stage, stages) {
1312 if (cmd_buffer->state.pipeline->shaders[stage])
1313 emit_stage_descriptor_set_userdata(cmd_buffer, cmd_buffer->state.pipeline,
1314 idx, set->va,
1315 stage);
1316 }
1317 }
1318
1319 if (cmd_buffer->state.compute_pipeline && (stages & VK_SHADER_STAGE_COMPUTE_BIT))
1320 emit_stage_descriptor_set_userdata(cmd_buffer, cmd_buffer->state.compute_pipeline,
1321 idx, set->va,
1322 MESA_SHADER_COMPUTE);
1323 }
1324
1325 static void
1326 radv_flush_push_descriptors(struct radv_cmd_buffer *cmd_buffer,
1327 VkPipelineBindPoint bind_point)
1328 {
1329 struct radv_descriptor_state *descriptors_state =
1330 radv_get_descriptors_state(cmd_buffer, bind_point);
1331 struct radv_descriptor_set *set = &descriptors_state->push_set.set;
1332 unsigned bo_offset;
1333
1334 if (!radv_cmd_buffer_upload_data(cmd_buffer, set->size, 32,
1335 set->mapped_ptr,
1336 &bo_offset))
1337 return;
1338
1339 set->va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1340 set->va += bo_offset;
1341 }
1342
1343 static void
1344 radv_flush_indirect_descriptor_sets(struct radv_cmd_buffer *cmd_buffer,
1345 VkPipelineBindPoint bind_point)
1346 {
1347 struct radv_descriptor_state *descriptors_state =
1348 radv_get_descriptors_state(cmd_buffer, bind_point);
1349 uint32_t size = MAX_SETS * 2 * 4;
1350 uint32_t offset;
1351 void *ptr;
1352
1353 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size,
1354 256, &offset, &ptr))
1355 return;
1356
1357 for (unsigned i = 0; i < MAX_SETS; i++) {
1358 uint32_t *uptr = ((uint32_t *)ptr) + i * 2;
1359 uint64_t set_va = 0;
1360 struct radv_descriptor_set *set = descriptors_state->sets[i];
1361 if (descriptors_state->valid & (1u << i))
1362 set_va = set->va;
1363 uptr[0] = set_va & 0xffffffff;
1364 uptr[1] = set_va >> 32;
1365 }
1366
1367 uint64_t va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1368 va += offset;
1369
1370 if (cmd_buffer->state.pipeline) {
1371 if (cmd_buffer->state.pipeline->shaders[MESA_SHADER_VERTEX])
1372 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_VERTEX,
1373 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1374
1375 if (cmd_buffer->state.pipeline->shaders[MESA_SHADER_FRAGMENT])
1376 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_FRAGMENT,
1377 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1378
1379 if (radv_pipeline_has_gs(cmd_buffer->state.pipeline))
1380 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_GEOMETRY,
1381 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1382
1383 if (radv_pipeline_has_tess(cmd_buffer->state.pipeline))
1384 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_TESS_CTRL,
1385 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1386
1387 if (radv_pipeline_has_tess(cmd_buffer->state.pipeline))
1388 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_TESS_EVAL,
1389 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1390 }
1391
1392 if (cmd_buffer->state.compute_pipeline)
1393 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.compute_pipeline, MESA_SHADER_COMPUTE,
1394 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1395 }
1396
1397 static void
1398 radv_flush_descriptors(struct radv_cmd_buffer *cmd_buffer,
1399 VkShaderStageFlags stages)
1400 {
1401 VkPipelineBindPoint bind_point = stages & VK_SHADER_STAGE_COMPUTE_BIT ?
1402 VK_PIPELINE_BIND_POINT_COMPUTE :
1403 VK_PIPELINE_BIND_POINT_GRAPHICS;
1404 struct radv_descriptor_state *descriptors_state =
1405 radv_get_descriptors_state(cmd_buffer, bind_point);
1406 unsigned i;
1407
1408 if (!descriptors_state->dirty)
1409 return;
1410
1411 if (descriptors_state->push_dirty)
1412 radv_flush_push_descriptors(cmd_buffer, bind_point);
1413
1414 if ((cmd_buffer->state.pipeline && cmd_buffer->state.pipeline->need_indirect_descriptor_sets) ||
1415 (cmd_buffer->state.compute_pipeline && cmd_buffer->state.compute_pipeline->need_indirect_descriptor_sets)) {
1416 radv_flush_indirect_descriptor_sets(cmd_buffer, bind_point);
1417 }
1418
1419 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
1420 cmd_buffer->cs,
1421 MAX_SETS * MESA_SHADER_STAGES * 4);
1422
1423 for_each_bit(i, descriptors_state->dirty) {
1424 struct radv_descriptor_set *set = descriptors_state->sets[i];
1425 if (!(descriptors_state->valid & (1u << i)))
1426 continue;
1427
1428 radv_emit_descriptor_set_userdata(cmd_buffer, stages, set, i);
1429 }
1430 descriptors_state->dirty = 0;
1431 descriptors_state->push_dirty = false;
1432
1433 if (unlikely(cmd_buffer->device->trace_bo))
1434 radv_save_descriptors(cmd_buffer, bind_point);
1435
1436 assert(cmd_buffer->cs->cdw <= cdw_max);
1437 }
1438
1439 static void
1440 radv_flush_constants(struct radv_cmd_buffer *cmd_buffer,
1441 struct radv_pipeline *pipeline,
1442 VkShaderStageFlags stages)
1443 {
1444 struct radv_pipeline_layout *layout = pipeline->layout;
1445 unsigned offset;
1446 void *ptr;
1447 uint64_t va;
1448
1449 stages &= cmd_buffer->push_constant_stages;
1450 if (!stages ||
1451 (!layout->push_constant_size && !layout->dynamic_offset_count))
1452 return;
1453
1454 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, layout->push_constant_size +
1455 16 * layout->dynamic_offset_count,
1456 256, &offset, &ptr))
1457 return;
1458
1459 memcpy(ptr, cmd_buffer->push_constants, layout->push_constant_size);
1460 memcpy((char*)ptr + layout->push_constant_size, cmd_buffer->dynamic_buffers,
1461 16 * layout->dynamic_offset_count);
1462
1463 va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1464 va += offset;
1465
1466 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
1467 cmd_buffer->cs, MESA_SHADER_STAGES * 4);
1468
1469 radv_foreach_stage(stage, stages) {
1470 if (pipeline->shaders[stage]) {
1471 radv_emit_userdata_address(cmd_buffer, pipeline, stage,
1472 AC_UD_PUSH_CONSTANTS, va);
1473 }
1474 }
1475
1476 cmd_buffer->push_constant_stages &= ~stages;
1477 assert(cmd_buffer->cs->cdw <= cdw_max);
1478 }
1479
1480 static bool
1481 radv_cmd_buffer_update_vertex_descriptors(struct radv_cmd_buffer *cmd_buffer, bool pipeline_is_dirty)
1482 {
1483 if ((pipeline_is_dirty ||
1484 (cmd_buffer->state.dirty & RADV_CMD_DIRTY_VERTEX_BUFFER)) &&
1485 cmd_buffer->state.pipeline->vertex_elements.count &&
1486 radv_get_vertex_shader(cmd_buffer->state.pipeline)->info.info.vs.has_vertex_buffers) {
1487 struct radv_vertex_elements_info *velems = &cmd_buffer->state.pipeline->vertex_elements;
1488 unsigned vb_offset;
1489 void *vb_ptr;
1490 uint32_t i = 0;
1491 uint32_t count = velems->count;
1492 uint64_t va;
1493
1494 /* allocate some descriptor state for vertex buffers */
1495 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, count * 16, 256,
1496 &vb_offset, &vb_ptr))
1497 return false;
1498
1499 for (i = 0; i < count; i++) {
1500 uint32_t *desc = &((uint32_t *)vb_ptr)[i * 4];
1501 uint32_t offset;
1502 int vb = velems->binding[i];
1503 struct radv_buffer *buffer = cmd_buffer->vertex_bindings[vb].buffer;
1504 uint32_t stride = cmd_buffer->state.pipeline->binding_stride[vb];
1505
1506 va = radv_buffer_get_va(buffer->bo);
1507
1508 offset = cmd_buffer->vertex_bindings[vb].offset + velems->offset[i];
1509 va += offset + buffer->offset;
1510 desc[0] = va;
1511 desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) | S_008F04_STRIDE(stride);
1512 if (cmd_buffer->device->physical_device->rad_info.chip_class <= CIK && stride)
1513 desc[2] = (buffer->size - offset - velems->format_size[i]) / stride + 1;
1514 else
1515 desc[2] = buffer->size - offset;
1516 desc[3] = velems->rsrc_word3[i];
1517 }
1518
1519 va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1520 va += vb_offset;
1521
1522 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_VERTEX,
1523 AC_UD_VS_VERTEX_BUFFERS, va);
1524
1525 cmd_buffer->state.vb_va = va;
1526 cmd_buffer->state.vb_size = count * 16;
1527 cmd_buffer->state.prefetch_L2_mask |= RADV_PREFETCH_VBO_DESCRIPTORS;
1528 }
1529 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_VERTEX_BUFFER;
1530
1531 return true;
1532 }
1533
1534 static bool
1535 radv_upload_graphics_shader_descriptors(struct radv_cmd_buffer *cmd_buffer, bool pipeline_is_dirty)
1536 {
1537 if (!radv_cmd_buffer_update_vertex_descriptors(cmd_buffer, pipeline_is_dirty))
1538 return false;
1539
1540 radv_flush_descriptors(cmd_buffer, VK_SHADER_STAGE_ALL_GRAPHICS);
1541 radv_flush_constants(cmd_buffer, cmd_buffer->state.pipeline,
1542 VK_SHADER_STAGE_ALL_GRAPHICS);
1543
1544 return true;
1545 }
1546
1547 static void
1548 radv_emit_draw_registers(struct radv_cmd_buffer *cmd_buffer, bool indexed_draw,
1549 bool instanced_draw, bool indirect_draw,
1550 uint32_t draw_vertex_count)
1551 {
1552 struct radeon_info *info = &cmd_buffer->device->physical_device->rad_info;
1553 struct radv_cmd_state *state = &cmd_buffer->state;
1554 struct radeon_winsys_cs *cs = cmd_buffer->cs;
1555 uint32_t ia_multi_vgt_param;
1556 int32_t primitive_reset_en;
1557
1558 /* Draw state. */
1559 ia_multi_vgt_param =
1560 si_get_ia_multi_vgt_param(cmd_buffer, instanced_draw,
1561 indirect_draw, draw_vertex_count);
1562
1563 if (state->last_ia_multi_vgt_param != ia_multi_vgt_param) {
1564 if (info->chip_class >= GFX9) {
1565 radeon_set_uconfig_reg_idx(cs,
1566 R_030960_IA_MULTI_VGT_PARAM,
1567 4, ia_multi_vgt_param);
1568 } else if (info->chip_class >= CIK) {
1569 radeon_set_context_reg_idx(cs,
1570 R_028AA8_IA_MULTI_VGT_PARAM,
1571 1, ia_multi_vgt_param);
1572 } else {
1573 radeon_set_context_reg(cs, R_028AA8_IA_MULTI_VGT_PARAM,
1574 ia_multi_vgt_param);
1575 }
1576 state->last_ia_multi_vgt_param = ia_multi_vgt_param;
1577 }
1578
1579 /* Primitive restart. */
1580 primitive_reset_en =
1581 indexed_draw && state->pipeline->graphics.prim_restart_enable;
1582
1583 if (primitive_reset_en != state->last_primitive_reset_en) {
1584 state->last_primitive_reset_en = primitive_reset_en;
1585 if (info->chip_class >= GFX9) {
1586 radeon_set_uconfig_reg(cs,
1587 R_03092C_VGT_MULTI_PRIM_IB_RESET_EN,
1588 primitive_reset_en);
1589 } else {
1590 radeon_set_context_reg(cs,
1591 R_028A94_VGT_MULTI_PRIM_IB_RESET_EN,
1592 primitive_reset_en);
1593 }
1594 }
1595
1596 if (primitive_reset_en) {
1597 uint32_t primitive_reset_index =
1598 state->index_type ? 0xffffffffu : 0xffffu;
1599
1600 if (primitive_reset_index != state->last_primitive_reset_index) {
1601 radeon_set_context_reg(cs,
1602 R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX,
1603 primitive_reset_index);
1604 state->last_primitive_reset_index = primitive_reset_index;
1605 }
1606 }
1607 }
1608
1609 static void radv_stage_flush(struct radv_cmd_buffer *cmd_buffer,
1610 VkPipelineStageFlags src_stage_mask)
1611 {
1612 if (src_stage_mask & (VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
1613 VK_PIPELINE_STAGE_TRANSFER_BIT |
1614 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT |
1615 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)) {
1616 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH;
1617 }
1618
1619 if (src_stage_mask & (VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT |
1620 VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT |
1621 VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT |
1622 VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
1623 VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
1624 VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
1625 VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT |
1626 VK_PIPELINE_STAGE_TRANSFER_BIT |
1627 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT |
1628 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT |
1629 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)) {
1630 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
1631 } else if (src_stage_mask & (VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT |
1632 VK_PIPELINE_STAGE_VERTEX_INPUT_BIT |
1633 VK_PIPELINE_STAGE_VERTEX_SHADER_BIT)) {
1634 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VS_PARTIAL_FLUSH;
1635 }
1636 }
1637
1638 static enum radv_cmd_flush_bits
1639 radv_src_access_flush(struct radv_cmd_buffer *cmd_buffer,
1640 VkAccessFlags src_flags)
1641 {
1642 enum radv_cmd_flush_bits flush_bits = 0;
1643 uint32_t b;
1644 for_each_bit(b, src_flags) {
1645 switch ((VkAccessFlagBits)(1 << b)) {
1646 case VK_ACCESS_SHADER_WRITE_BIT:
1647 flush_bits |= RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2;
1648 break;
1649 case VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT:
1650 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1651 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
1652 break;
1653 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT:
1654 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
1655 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
1656 break;
1657 case VK_ACCESS_TRANSFER_WRITE_BIT:
1658 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1659 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
1660 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
1661 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META |
1662 RADV_CMD_FLAG_INV_GLOBAL_L2;
1663 break;
1664 default:
1665 break;
1666 }
1667 }
1668 return flush_bits;
1669 }
1670
1671 static enum radv_cmd_flush_bits
1672 radv_dst_access_flush(struct radv_cmd_buffer *cmd_buffer,
1673 VkAccessFlags dst_flags,
1674 struct radv_image *image)
1675 {
1676 enum radv_cmd_flush_bits flush_bits = 0;
1677 uint32_t b;
1678 for_each_bit(b, dst_flags) {
1679 switch ((VkAccessFlagBits)(1 << b)) {
1680 case VK_ACCESS_INDIRECT_COMMAND_READ_BIT:
1681 case VK_ACCESS_INDEX_READ_BIT:
1682 break;
1683 case VK_ACCESS_UNIFORM_READ_BIT:
1684 flush_bits |= RADV_CMD_FLAG_INV_VMEM_L1 | RADV_CMD_FLAG_INV_SMEM_L1;
1685 break;
1686 case VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT:
1687 case VK_ACCESS_SHADER_READ_BIT:
1688 case VK_ACCESS_TRANSFER_READ_BIT:
1689 case VK_ACCESS_INPUT_ATTACHMENT_READ_BIT:
1690 flush_bits |= RADV_CMD_FLAG_INV_VMEM_L1 |
1691 RADV_CMD_FLAG_INV_GLOBAL_L2;
1692 break;
1693 case VK_ACCESS_COLOR_ATTACHMENT_READ_BIT:
1694 /* TODO: change to image && when the image gets passed
1695 * through from the subpass. */
1696 if (!image || (image->usage & VK_IMAGE_USAGE_STORAGE_BIT))
1697 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1698 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
1699 break;
1700 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT:
1701 if (!image || (image->usage & VK_IMAGE_USAGE_STORAGE_BIT))
1702 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
1703 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
1704 break;
1705 default:
1706 break;
1707 }
1708 }
1709 return flush_bits;
1710 }
1711
1712 static void radv_subpass_barrier(struct radv_cmd_buffer *cmd_buffer, const struct radv_subpass_barrier *barrier)
1713 {
1714 cmd_buffer->state.flush_bits |= radv_src_access_flush(cmd_buffer, barrier->src_access_mask);
1715 radv_stage_flush(cmd_buffer, barrier->src_stage_mask);
1716 cmd_buffer->state.flush_bits |= radv_dst_access_flush(cmd_buffer, barrier->dst_access_mask,
1717 NULL);
1718 }
1719
1720 static void radv_handle_subpass_image_transition(struct radv_cmd_buffer *cmd_buffer,
1721 VkAttachmentReference att)
1722 {
1723 unsigned idx = att.attachment;
1724 struct radv_image_view *view = cmd_buffer->state.framebuffer->attachments[idx].attachment;
1725 VkImageSubresourceRange range;
1726 range.aspectMask = 0;
1727 range.baseMipLevel = view->base_mip;
1728 range.levelCount = 1;
1729 range.baseArrayLayer = view->base_layer;
1730 range.layerCount = cmd_buffer->state.framebuffer->layers;
1731
1732 radv_handle_image_transition(cmd_buffer,
1733 view->image,
1734 cmd_buffer->state.attachments[idx].current_layout,
1735 att.layout, 0, 0, &range,
1736 cmd_buffer->state.attachments[idx].pending_clear_aspects);
1737
1738 cmd_buffer->state.attachments[idx].current_layout = att.layout;
1739
1740
1741 }
1742
1743 void
1744 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
1745 const struct radv_subpass *subpass, bool transitions)
1746 {
1747 if (transitions) {
1748 radv_subpass_barrier(cmd_buffer, &subpass->start_barrier);
1749
1750 for (unsigned i = 0; i < subpass->color_count; ++i) {
1751 if (subpass->color_attachments[i].attachment != VK_ATTACHMENT_UNUSED)
1752 radv_handle_subpass_image_transition(cmd_buffer,
1753 subpass->color_attachments[i]);
1754 }
1755
1756 for (unsigned i = 0; i < subpass->input_count; ++i) {
1757 radv_handle_subpass_image_transition(cmd_buffer,
1758 subpass->input_attachments[i]);
1759 }
1760
1761 if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
1762 radv_handle_subpass_image_transition(cmd_buffer,
1763 subpass->depth_stencil_attachment);
1764 }
1765 }
1766
1767 cmd_buffer->state.subpass = subpass;
1768
1769 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_FRAMEBUFFER;
1770 }
1771
1772 static VkResult
1773 radv_cmd_state_setup_attachments(struct radv_cmd_buffer *cmd_buffer,
1774 struct radv_render_pass *pass,
1775 const VkRenderPassBeginInfo *info)
1776 {
1777 struct radv_cmd_state *state = &cmd_buffer->state;
1778
1779 if (pass->attachment_count == 0) {
1780 state->attachments = NULL;
1781 return VK_SUCCESS;
1782 }
1783
1784 state->attachments = vk_alloc(&cmd_buffer->pool->alloc,
1785 pass->attachment_count *
1786 sizeof(state->attachments[0]),
1787 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1788 if (state->attachments == NULL) {
1789 cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
1790 return cmd_buffer->record_result;
1791 }
1792
1793 for (uint32_t i = 0; i < pass->attachment_count; ++i) {
1794 struct radv_render_pass_attachment *att = &pass->attachments[i];
1795 VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
1796 VkImageAspectFlags clear_aspects = 0;
1797
1798 if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
1799 /* color attachment */
1800 if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
1801 clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
1802 }
1803 } else {
1804 /* depthstencil attachment */
1805 if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
1806 att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
1807 clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
1808 if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
1809 att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_DONT_CARE)
1810 clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
1811 }
1812 if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
1813 att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
1814 clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
1815 }
1816 }
1817
1818 state->attachments[i].pending_clear_aspects = clear_aspects;
1819 state->attachments[i].cleared_views = 0;
1820 if (clear_aspects && info) {
1821 assert(info->clearValueCount > i);
1822 state->attachments[i].clear_value = info->pClearValues[i];
1823 }
1824
1825 state->attachments[i].current_layout = att->initial_layout;
1826 }
1827
1828 return VK_SUCCESS;
1829 }
1830
1831 VkResult radv_AllocateCommandBuffers(
1832 VkDevice _device,
1833 const VkCommandBufferAllocateInfo *pAllocateInfo,
1834 VkCommandBuffer *pCommandBuffers)
1835 {
1836 RADV_FROM_HANDLE(radv_device, device, _device);
1837 RADV_FROM_HANDLE(radv_cmd_pool, pool, pAllocateInfo->commandPool);
1838
1839 VkResult result = VK_SUCCESS;
1840 uint32_t i;
1841
1842 for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
1843
1844 if (!list_empty(&pool->free_cmd_buffers)) {
1845 struct radv_cmd_buffer *cmd_buffer = list_first_entry(&pool->free_cmd_buffers, struct radv_cmd_buffer, pool_link);
1846
1847 list_del(&cmd_buffer->pool_link);
1848 list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
1849
1850 result = radv_reset_cmd_buffer(cmd_buffer);
1851 cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1852 cmd_buffer->level = pAllocateInfo->level;
1853
1854 pCommandBuffers[i] = radv_cmd_buffer_to_handle(cmd_buffer);
1855 } else {
1856 result = radv_create_cmd_buffer(device, pool, pAllocateInfo->level,
1857 &pCommandBuffers[i]);
1858 }
1859 if (result != VK_SUCCESS)
1860 break;
1861 }
1862
1863 if (result != VK_SUCCESS) {
1864 radv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
1865 i, pCommandBuffers);
1866
1867 /* From the Vulkan 1.0.66 spec:
1868 *
1869 * "vkAllocateCommandBuffers can be used to create multiple
1870 * command buffers. If the creation of any of those command
1871 * buffers fails, the implementation must destroy all
1872 * successfully created command buffer objects from this
1873 * command, set all entries of the pCommandBuffers array to
1874 * NULL and return the error."
1875 */
1876 memset(pCommandBuffers, 0,
1877 sizeof(*pCommandBuffers) * pAllocateInfo->commandBufferCount);
1878 }
1879
1880 return result;
1881 }
1882
1883 void radv_FreeCommandBuffers(
1884 VkDevice device,
1885 VkCommandPool commandPool,
1886 uint32_t commandBufferCount,
1887 const VkCommandBuffer *pCommandBuffers)
1888 {
1889 for (uint32_t i = 0; i < commandBufferCount; i++) {
1890 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
1891
1892 if (cmd_buffer) {
1893 if (cmd_buffer->pool) {
1894 list_del(&cmd_buffer->pool_link);
1895 list_addtail(&cmd_buffer->pool_link, &cmd_buffer->pool->free_cmd_buffers);
1896 } else
1897 radv_cmd_buffer_destroy(cmd_buffer);
1898
1899 }
1900 }
1901 }
1902
1903 VkResult radv_ResetCommandBuffer(
1904 VkCommandBuffer commandBuffer,
1905 VkCommandBufferResetFlags flags)
1906 {
1907 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1908 return radv_reset_cmd_buffer(cmd_buffer);
1909 }
1910
1911 static void emit_gfx_buffer_state(struct radv_cmd_buffer *cmd_buffer)
1912 {
1913 struct radv_device *device = cmd_buffer->device;
1914 if (device->gfx_init) {
1915 uint64_t va = radv_buffer_get_va(device->gfx_init);
1916 radv_cs_add_buffer(device->ws, cmd_buffer->cs, device->gfx_init, 8);
1917 radeon_emit(cmd_buffer->cs, PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0));
1918 radeon_emit(cmd_buffer->cs, va);
1919 radeon_emit(cmd_buffer->cs, va >> 32);
1920 radeon_emit(cmd_buffer->cs, device->gfx_init_size_dw & 0xffff);
1921 } else
1922 si_init_config(cmd_buffer);
1923 }
1924
1925 VkResult radv_BeginCommandBuffer(
1926 VkCommandBuffer commandBuffer,
1927 const VkCommandBufferBeginInfo *pBeginInfo)
1928 {
1929 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1930 VkResult result = VK_SUCCESS;
1931
1932 if (cmd_buffer->status != RADV_CMD_BUFFER_STATUS_INITIAL) {
1933 /* If the command buffer has already been resetted with
1934 * vkResetCommandBuffer, no need to do it again.
1935 */
1936 result = radv_reset_cmd_buffer(cmd_buffer);
1937 if (result != VK_SUCCESS)
1938 return result;
1939 }
1940
1941 memset(&cmd_buffer->state, 0, sizeof(cmd_buffer->state));
1942 cmd_buffer->state.last_primitive_reset_en = -1;
1943 cmd_buffer->state.last_index_type = -1;
1944 cmd_buffer->state.last_num_instances = -1;
1945 cmd_buffer->state.last_vertex_offset = -1;
1946 cmd_buffer->state.last_first_instance = -1;
1947 cmd_buffer->usage_flags = pBeginInfo->flags;
1948
1949 /* setup initial configuration into command buffer */
1950 if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
1951 switch (cmd_buffer->queue_family_index) {
1952 case RADV_QUEUE_GENERAL:
1953 emit_gfx_buffer_state(cmd_buffer);
1954 break;
1955 case RADV_QUEUE_COMPUTE:
1956 si_init_compute(cmd_buffer);
1957 break;
1958 case RADV_QUEUE_TRANSFER:
1959 default:
1960 break;
1961 }
1962 }
1963
1964 if (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
1965 assert(pBeginInfo->pInheritanceInfo);
1966 cmd_buffer->state.framebuffer = radv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);
1967 cmd_buffer->state.pass = radv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
1968
1969 struct radv_subpass *subpass =
1970 &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
1971
1972 result = radv_cmd_state_setup_attachments(cmd_buffer, cmd_buffer->state.pass, NULL);
1973 if (result != VK_SUCCESS)
1974 return result;
1975
1976 radv_cmd_buffer_set_subpass(cmd_buffer, subpass, false);
1977 }
1978
1979 if (unlikely(cmd_buffer->device->trace_bo))
1980 radv_cmd_buffer_trace_emit(cmd_buffer);
1981
1982 cmd_buffer->status = RADV_CMD_BUFFER_STATUS_RECORDING;
1983
1984 return result;
1985 }
1986
1987 void radv_CmdBindVertexBuffers(
1988 VkCommandBuffer commandBuffer,
1989 uint32_t firstBinding,
1990 uint32_t bindingCount,
1991 const VkBuffer* pBuffers,
1992 const VkDeviceSize* pOffsets)
1993 {
1994 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1995 struct radv_vertex_binding *vb = cmd_buffer->vertex_bindings;
1996 bool changed = false;
1997
1998 /* We have to defer setting up vertex buffer since we need the buffer
1999 * stride from the pipeline. */
2000
2001 assert(firstBinding + bindingCount <= MAX_VBS);
2002 for (uint32_t i = 0; i < bindingCount; i++) {
2003 uint32_t idx = firstBinding + i;
2004
2005 if (!changed &&
2006 (vb[idx].buffer != radv_buffer_from_handle(pBuffers[i]) ||
2007 vb[idx].offset != pOffsets[i])) {
2008 changed = true;
2009 }
2010
2011 vb[idx].buffer = radv_buffer_from_handle(pBuffers[i]);
2012 vb[idx].offset = pOffsets[i];
2013
2014 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs,
2015 vb[idx].buffer->bo, 8);
2016 }
2017
2018 if (!changed) {
2019 /* No state changes. */
2020 return;
2021 }
2022
2023 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_VERTEX_BUFFER;
2024 }
2025
2026 void radv_CmdBindIndexBuffer(
2027 VkCommandBuffer commandBuffer,
2028 VkBuffer buffer,
2029 VkDeviceSize offset,
2030 VkIndexType indexType)
2031 {
2032 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2033 RADV_FROM_HANDLE(radv_buffer, index_buffer, buffer);
2034
2035 if (cmd_buffer->state.index_buffer == index_buffer &&
2036 cmd_buffer->state.index_offset == offset &&
2037 cmd_buffer->state.index_type == indexType) {
2038 /* No state changes. */
2039 return;
2040 }
2041
2042 cmd_buffer->state.index_buffer = index_buffer;
2043 cmd_buffer->state.index_offset = offset;
2044 cmd_buffer->state.index_type = indexType; /* vk matches hw */
2045 cmd_buffer->state.index_va = radv_buffer_get_va(index_buffer->bo);
2046 cmd_buffer->state.index_va += index_buffer->offset + offset;
2047
2048 int index_size_shift = cmd_buffer->state.index_type ? 2 : 1;
2049 cmd_buffer->state.max_index_count = (index_buffer->size - offset) >> index_size_shift;
2050 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_INDEX_BUFFER;
2051 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, index_buffer->bo, 8);
2052 }
2053
2054
2055 static void
2056 radv_bind_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
2057 VkPipelineBindPoint bind_point,
2058 struct radv_descriptor_set *set, unsigned idx)
2059 {
2060 struct radeon_winsys *ws = cmd_buffer->device->ws;
2061
2062 radv_set_descriptor_set(cmd_buffer, bind_point, set, idx);
2063 if (!set)
2064 return;
2065
2066 assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
2067
2068 for (unsigned j = 0; j < set->layout->buffer_count; ++j)
2069 if (set->descriptors[j])
2070 radv_cs_add_buffer(ws, cmd_buffer->cs, set->descriptors[j], 7);
2071
2072 if(set->bo)
2073 radv_cs_add_buffer(ws, cmd_buffer->cs, set->bo, 8);
2074 }
2075
2076 void radv_CmdBindDescriptorSets(
2077 VkCommandBuffer commandBuffer,
2078 VkPipelineBindPoint pipelineBindPoint,
2079 VkPipelineLayout _layout,
2080 uint32_t firstSet,
2081 uint32_t descriptorSetCount,
2082 const VkDescriptorSet* pDescriptorSets,
2083 uint32_t dynamicOffsetCount,
2084 const uint32_t* pDynamicOffsets)
2085 {
2086 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2087 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2088 unsigned dyn_idx = 0;
2089
2090 for (unsigned i = 0; i < descriptorSetCount; ++i) {
2091 unsigned idx = i + firstSet;
2092 RADV_FROM_HANDLE(radv_descriptor_set, set, pDescriptorSets[i]);
2093 radv_bind_descriptor_set(cmd_buffer, pipelineBindPoint, set, idx);
2094
2095 for(unsigned j = 0; j < set->layout->dynamic_offset_count; ++j, ++dyn_idx) {
2096 unsigned idx = j + layout->set[i + firstSet].dynamic_offset_start;
2097 uint32_t *dst = cmd_buffer->dynamic_buffers + idx * 4;
2098 assert(dyn_idx < dynamicOffsetCount);
2099
2100 struct radv_descriptor_range *range = set->dynamic_descriptors + j;
2101 uint64_t va = range->va + pDynamicOffsets[dyn_idx];
2102 dst[0] = va;
2103 dst[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
2104 dst[2] = range->size;
2105 dst[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
2106 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
2107 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
2108 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
2109 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
2110 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
2111 cmd_buffer->push_constant_stages |=
2112 set->layout->dynamic_shader_stages;
2113 }
2114 }
2115 }
2116
2117 static bool radv_init_push_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
2118 struct radv_descriptor_set *set,
2119 struct radv_descriptor_set_layout *layout,
2120 VkPipelineBindPoint bind_point)
2121 {
2122 struct radv_descriptor_state *descriptors_state =
2123 radv_get_descriptors_state(cmd_buffer, bind_point);
2124 set->size = layout->size;
2125 set->layout = layout;
2126
2127 if (descriptors_state->push_set.capacity < set->size) {
2128 size_t new_size = MAX2(set->size, 1024);
2129 new_size = MAX2(new_size, 2 * descriptors_state->push_set.capacity);
2130 new_size = MIN2(new_size, 96 * MAX_PUSH_DESCRIPTORS);
2131
2132 free(set->mapped_ptr);
2133 set->mapped_ptr = malloc(new_size);
2134
2135 if (!set->mapped_ptr) {
2136 descriptors_state->push_set.capacity = 0;
2137 cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
2138 return false;
2139 }
2140
2141 descriptors_state->push_set.capacity = new_size;
2142 }
2143
2144 return true;
2145 }
2146
2147 void radv_meta_push_descriptor_set(
2148 struct radv_cmd_buffer* cmd_buffer,
2149 VkPipelineBindPoint pipelineBindPoint,
2150 VkPipelineLayout _layout,
2151 uint32_t set,
2152 uint32_t descriptorWriteCount,
2153 const VkWriteDescriptorSet* pDescriptorWrites)
2154 {
2155 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2156 struct radv_descriptor_set *push_set = &cmd_buffer->meta_push_descriptors;
2157 unsigned bo_offset;
2158
2159 assert(set == 0);
2160 assert(layout->set[set].layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR);
2161
2162 push_set->size = layout->set[set].layout->size;
2163 push_set->layout = layout->set[set].layout;
2164
2165 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, push_set->size, 32,
2166 &bo_offset,
2167 (void**) &push_set->mapped_ptr))
2168 return;
2169
2170 push_set->va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
2171 push_set->va += bo_offset;
2172
2173 radv_update_descriptor_sets(cmd_buffer->device, cmd_buffer,
2174 radv_descriptor_set_to_handle(push_set),
2175 descriptorWriteCount, pDescriptorWrites, 0, NULL);
2176
2177 radv_set_descriptor_set(cmd_buffer, pipelineBindPoint, push_set, set);
2178 }
2179
2180 void radv_CmdPushDescriptorSetKHR(
2181 VkCommandBuffer commandBuffer,
2182 VkPipelineBindPoint pipelineBindPoint,
2183 VkPipelineLayout _layout,
2184 uint32_t set,
2185 uint32_t descriptorWriteCount,
2186 const VkWriteDescriptorSet* pDescriptorWrites)
2187 {
2188 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2189 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2190 struct radv_descriptor_state *descriptors_state =
2191 radv_get_descriptors_state(cmd_buffer, pipelineBindPoint);
2192 struct radv_descriptor_set *push_set = &descriptors_state->push_set.set;
2193
2194 assert(layout->set[set].layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR);
2195
2196 if (!radv_init_push_descriptor_set(cmd_buffer, push_set,
2197 layout->set[set].layout,
2198 pipelineBindPoint))
2199 return;
2200
2201 radv_update_descriptor_sets(cmd_buffer->device, cmd_buffer,
2202 radv_descriptor_set_to_handle(push_set),
2203 descriptorWriteCount, pDescriptorWrites, 0, NULL);
2204
2205 radv_set_descriptor_set(cmd_buffer, pipelineBindPoint, push_set, set);
2206 descriptors_state->push_dirty = true;
2207 }
2208
2209 void radv_CmdPushDescriptorSetWithTemplateKHR(
2210 VkCommandBuffer commandBuffer,
2211 VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
2212 VkPipelineLayout _layout,
2213 uint32_t set,
2214 const void* pData)
2215 {
2216 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2217 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2218 RADV_FROM_HANDLE(radv_descriptor_update_template, templ, descriptorUpdateTemplate);
2219 struct radv_descriptor_state *descriptors_state =
2220 radv_get_descriptors_state(cmd_buffer, templ->bind_point);
2221 struct radv_descriptor_set *push_set = &descriptors_state->push_set.set;
2222
2223 assert(layout->set[set].layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR);
2224
2225 if (!radv_init_push_descriptor_set(cmd_buffer, push_set,
2226 layout->set[set].layout,
2227 templ->bind_point))
2228 return;
2229
2230 radv_update_descriptor_set_with_template(cmd_buffer->device, cmd_buffer, push_set,
2231 descriptorUpdateTemplate, pData);
2232
2233 radv_set_descriptor_set(cmd_buffer, templ->bind_point, push_set, set);
2234 descriptors_state->push_dirty = true;
2235 }
2236
2237 void radv_CmdPushConstants(VkCommandBuffer commandBuffer,
2238 VkPipelineLayout layout,
2239 VkShaderStageFlags stageFlags,
2240 uint32_t offset,
2241 uint32_t size,
2242 const void* pValues)
2243 {
2244 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2245 memcpy(cmd_buffer->push_constants + offset, pValues, size);
2246 cmd_buffer->push_constant_stages |= stageFlags;
2247 }
2248
2249 VkResult radv_EndCommandBuffer(
2250 VkCommandBuffer commandBuffer)
2251 {
2252 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2253
2254 if (cmd_buffer->queue_family_index != RADV_QUEUE_TRANSFER) {
2255 if (cmd_buffer->device->physical_device->rad_info.chip_class == SI)
2256 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_PS_PARTIAL_FLUSH | RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2;
2257 si_emit_cache_flush(cmd_buffer);
2258 }
2259
2260 vk_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
2261
2262 if (!cmd_buffer->device->ws->cs_finalize(cmd_buffer->cs))
2263 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2264
2265 cmd_buffer->status = RADV_CMD_BUFFER_STATUS_EXECUTABLE;
2266
2267 return cmd_buffer->record_result;
2268 }
2269
2270 static void
2271 radv_emit_compute_pipeline(struct radv_cmd_buffer *cmd_buffer)
2272 {
2273 struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
2274
2275 if (!pipeline || pipeline == cmd_buffer->state.emitted_compute_pipeline)
2276 return;
2277
2278 cmd_buffer->state.emitted_compute_pipeline = pipeline;
2279
2280 radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, pipeline->cs.cdw);
2281 radeon_emit_array(cmd_buffer->cs, pipeline->cs.buf, pipeline->cs.cdw);
2282
2283 cmd_buffer->compute_scratch_size_needed =
2284 MAX2(cmd_buffer->compute_scratch_size_needed,
2285 pipeline->max_waves * pipeline->scratch_bytes_per_wave);
2286
2287 if (unlikely(cmd_buffer->device->trace_bo))
2288 radv_save_pipeline(cmd_buffer, pipeline, RING_COMPUTE);
2289 }
2290
2291 static void radv_mark_descriptor_sets_dirty(struct radv_cmd_buffer *cmd_buffer,
2292 VkPipelineBindPoint bind_point)
2293 {
2294 struct radv_descriptor_state *descriptors_state =
2295 radv_get_descriptors_state(cmd_buffer, bind_point);
2296
2297 descriptors_state->dirty |= descriptors_state->valid;
2298 }
2299
2300 void radv_CmdBindPipeline(
2301 VkCommandBuffer commandBuffer,
2302 VkPipelineBindPoint pipelineBindPoint,
2303 VkPipeline _pipeline)
2304 {
2305 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2306 RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
2307
2308 switch (pipelineBindPoint) {
2309 case VK_PIPELINE_BIND_POINT_COMPUTE:
2310 if (cmd_buffer->state.compute_pipeline == pipeline)
2311 return;
2312 radv_mark_descriptor_sets_dirty(cmd_buffer, pipelineBindPoint);
2313
2314 cmd_buffer->state.compute_pipeline = pipeline;
2315 cmd_buffer->push_constant_stages |= VK_SHADER_STAGE_COMPUTE_BIT;
2316 break;
2317 case VK_PIPELINE_BIND_POINT_GRAPHICS:
2318 if (cmd_buffer->state.pipeline == pipeline)
2319 return;
2320 radv_mark_descriptor_sets_dirty(cmd_buffer, pipelineBindPoint);
2321
2322 cmd_buffer->state.pipeline = pipeline;
2323 if (!pipeline)
2324 break;
2325
2326 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_PIPELINE;
2327 cmd_buffer->push_constant_stages |= pipeline->active_stages;
2328
2329 /* the new vertex shader might not have the same user regs */
2330 cmd_buffer->state.last_first_instance = -1;
2331 cmd_buffer->state.last_vertex_offset = -1;
2332
2333 /* Prefetch all pipeline shaders at first draw time. */
2334 cmd_buffer->state.prefetch_L2_mask |= RADV_PREFETCH_SHADERS;
2335
2336 radv_bind_dynamic_state(cmd_buffer, &pipeline->dynamic_state);
2337
2338 if (pipeline->graphics.esgs_ring_size > cmd_buffer->esgs_ring_size_needed)
2339 cmd_buffer->esgs_ring_size_needed = pipeline->graphics.esgs_ring_size;
2340 if (pipeline->graphics.gsvs_ring_size > cmd_buffer->gsvs_ring_size_needed)
2341 cmd_buffer->gsvs_ring_size_needed = pipeline->graphics.gsvs_ring_size;
2342
2343 if (radv_pipeline_has_tess(pipeline))
2344 cmd_buffer->tess_rings_needed = true;
2345
2346 if (radv_pipeline_has_gs(pipeline)) {
2347 struct radv_userdata_info *loc = radv_lookup_user_sgpr(cmd_buffer->state.pipeline, MESA_SHADER_GEOMETRY,
2348 AC_UD_SCRATCH_RING_OFFSETS);
2349 if (cmd_buffer->ring_offsets_idx == -1)
2350 cmd_buffer->ring_offsets_idx = loc->sgpr_idx;
2351 else if (loc->sgpr_idx != -1)
2352 assert(loc->sgpr_idx == cmd_buffer->ring_offsets_idx);
2353 }
2354 break;
2355 default:
2356 assert(!"invalid bind point");
2357 break;
2358 }
2359 }
2360
2361 void radv_CmdSetViewport(
2362 VkCommandBuffer commandBuffer,
2363 uint32_t firstViewport,
2364 uint32_t viewportCount,
2365 const VkViewport* pViewports)
2366 {
2367 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2368 struct radv_cmd_state *state = &cmd_buffer->state;
2369 MAYBE_UNUSED const uint32_t total_count = firstViewport + viewportCount;
2370
2371 assert(firstViewport < MAX_VIEWPORTS);
2372 assert(total_count >= 1 && total_count <= MAX_VIEWPORTS);
2373
2374 if (cmd_buffer->device->physical_device->has_scissor_bug) {
2375 /* Try to skip unnecessary PS partial flushes when the viewports
2376 * don't change.
2377 */
2378 if (!(state->dirty & (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT |
2379 RADV_CMD_DIRTY_DYNAMIC_SCISSOR)) &&
2380 !memcmp(state->dynamic.viewport.viewports + firstViewport,
2381 pViewports, viewportCount * sizeof(*pViewports))) {
2382 return;
2383 }
2384 }
2385
2386 memcpy(state->dynamic.viewport.viewports + firstViewport, pViewports,
2387 viewportCount * sizeof(*pViewports));
2388
2389 state->dirty |= RADV_CMD_DIRTY_DYNAMIC_VIEWPORT;
2390 }
2391
2392 void radv_CmdSetScissor(
2393 VkCommandBuffer commandBuffer,
2394 uint32_t firstScissor,
2395 uint32_t scissorCount,
2396 const VkRect2D* pScissors)
2397 {
2398 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2399 struct radv_cmd_state *state = &cmd_buffer->state;
2400 MAYBE_UNUSED const uint32_t total_count = firstScissor + scissorCount;
2401
2402 assert(firstScissor < MAX_SCISSORS);
2403 assert(total_count >= 1 && total_count <= MAX_SCISSORS);
2404
2405 if (cmd_buffer->device->physical_device->has_scissor_bug) {
2406 /* Try to skip unnecessary PS partial flushes when the scissors
2407 * don't change.
2408 */
2409 if (!(state->dirty & (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT |
2410 RADV_CMD_DIRTY_DYNAMIC_SCISSOR)) &&
2411 !memcmp(state->dynamic.scissor.scissors + firstScissor,
2412 pScissors, scissorCount * sizeof(*pScissors))) {
2413 return;
2414 }
2415 }
2416
2417 memcpy(state->dynamic.scissor.scissors + firstScissor, pScissors,
2418 scissorCount * sizeof(*pScissors));
2419
2420 state->dirty |= RADV_CMD_DIRTY_DYNAMIC_SCISSOR;
2421 }
2422
2423 void radv_CmdSetLineWidth(
2424 VkCommandBuffer commandBuffer,
2425 float lineWidth)
2426 {
2427 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2428 cmd_buffer->state.dynamic.line_width = lineWidth;
2429 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
2430 }
2431
2432 void radv_CmdSetDepthBias(
2433 VkCommandBuffer commandBuffer,
2434 float depthBiasConstantFactor,
2435 float depthBiasClamp,
2436 float depthBiasSlopeFactor)
2437 {
2438 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2439
2440 cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
2441 cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
2442 cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
2443
2444 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
2445 }
2446
2447 void radv_CmdSetBlendConstants(
2448 VkCommandBuffer commandBuffer,
2449 const float blendConstants[4])
2450 {
2451 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2452
2453 memcpy(cmd_buffer->state.dynamic.blend_constants,
2454 blendConstants, sizeof(float) * 4);
2455
2456 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
2457 }
2458
2459 void radv_CmdSetDepthBounds(
2460 VkCommandBuffer commandBuffer,
2461 float minDepthBounds,
2462 float maxDepthBounds)
2463 {
2464 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2465
2466 cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
2467 cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
2468
2469 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
2470 }
2471
2472 void radv_CmdSetStencilCompareMask(
2473 VkCommandBuffer commandBuffer,
2474 VkStencilFaceFlags faceMask,
2475 uint32_t compareMask)
2476 {
2477 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2478
2479 if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2480 cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
2481 if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2482 cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
2483
2484 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
2485 }
2486
2487 void radv_CmdSetStencilWriteMask(
2488 VkCommandBuffer commandBuffer,
2489 VkStencilFaceFlags faceMask,
2490 uint32_t writeMask)
2491 {
2492 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2493
2494 if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2495 cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
2496 if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2497 cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
2498
2499 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
2500 }
2501
2502 void radv_CmdSetStencilReference(
2503 VkCommandBuffer commandBuffer,
2504 VkStencilFaceFlags faceMask,
2505 uint32_t reference)
2506 {
2507 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2508
2509 if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2510 cmd_buffer->state.dynamic.stencil_reference.front = reference;
2511 if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2512 cmd_buffer->state.dynamic.stencil_reference.back = reference;
2513
2514 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
2515 }
2516
2517 void radv_CmdSetDiscardRectangleEXT(
2518 VkCommandBuffer commandBuffer,
2519 uint32_t firstDiscardRectangle,
2520 uint32_t discardRectangleCount,
2521 const VkRect2D* pDiscardRectangles)
2522 {
2523 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2524 struct radv_cmd_state *state = &cmd_buffer->state;
2525 MAYBE_UNUSED const uint32_t total_count = firstDiscardRectangle + discardRectangleCount;
2526
2527 assert(firstDiscardRectangle < MAX_DISCARD_RECTANGLES);
2528 assert(total_count >= 1 && total_count <= MAX_DISCARD_RECTANGLES);
2529
2530 typed_memcpy(&state->dynamic.discard_rectangle.rectangles[firstDiscardRectangle],
2531 pDiscardRectangles, discardRectangleCount);
2532
2533 state->dirty |= RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE;
2534 }
2535
2536 void radv_CmdExecuteCommands(
2537 VkCommandBuffer commandBuffer,
2538 uint32_t commandBufferCount,
2539 const VkCommandBuffer* pCmdBuffers)
2540 {
2541 RADV_FROM_HANDLE(radv_cmd_buffer, primary, commandBuffer);
2542
2543 assert(commandBufferCount > 0);
2544
2545 /* Emit pending flushes on primary prior to executing secondary */
2546 si_emit_cache_flush(primary);
2547
2548 for (uint32_t i = 0; i < commandBufferCount; i++) {
2549 RADV_FROM_HANDLE(radv_cmd_buffer, secondary, pCmdBuffers[i]);
2550
2551 primary->scratch_size_needed = MAX2(primary->scratch_size_needed,
2552 secondary->scratch_size_needed);
2553 primary->compute_scratch_size_needed = MAX2(primary->compute_scratch_size_needed,
2554 secondary->compute_scratch_size_needed);
2555
2556 if (secondary->esgs_ring_size_needed > primary->esgs_ring_size_needed)
2557 primary->esgs_ring_size_needed = secondary->esgs_ring_size_needed;
2558 if (secondary->gsvs_ring_size_needed > primary->gsvs_ring_size_needed)
2559 primary->gsvs_ring_size_needed = secondary->gsvs_ring_size_needed;
2560 if (secondary->tess_rings_needed)
2561 primary->tess_rings_needed = true;
2562 if (secondary->sample_positions_needed)
2563 primary->sample_positions_needed = true;
2564
2565 if (secondary->ring_offsets_idx != -1) {
2566 if (primary->ring_offsets_idx == -1)
2567 primary->ring_offsets_idx = secondary->ring_offsets_idx;
2568 else
2569 assert(secondary->ring_offsets_idx == primary->ring_offsets_idx);
2570 }
2571 primary->device->ws->cs_execute_secondary(primary->cs, secondary->cs);
2572
2573
2574 /* When the secondary command buffer is compute only we don't
2575 * need to re-emit the current graphics pipeline.
2576 */
2577 if (secondary->state.emitted_pipeline) {
2578 primary->state.emitted_pipeline =
2579 secondary->state.emitted_pipeline;
2580 }
2581
2582 /* When the secondary command buffer is graphics only we don't
2583 * need to re-emit the current compute pipeline.
2584 */
2585 if (secondary->state.emitted_compute_pipeline) {
2586 primary->state.emitted_compute_pipeline =
2587 secondary->state.emitted_compute_pipeline;
2588 }
2589
2590 /* Only re-emit the draw packets when needed. */
2591 if (secondary->state.last_primitive_reset_en != -1) {
2592 primary->state.last_primitive_reset_en =
2593 secondary->state.last_primitive_reset_en;
2594 }
2595
2596 if (secondary->state.last_primitive_reset_index) {
2597 primary->state.last_primitive_reset_index =
2598 secondary->state.last_primitive_reset_index;
2599 }
2600
2601 if (secondary->state.last_ia_multi_vgt_param) {
2602 primary->state.last_ia_multi_vgt_param =
2603 secondary->state.last_ia_multi_vgt_param;
2604 }
2605
2606 if (secondary->state.last_first_instance != -1) {
2607 primary->state.last_first_instance =
2608 secondary->state.last_first_instance;
2609 }
2610
2611 if (secondary->state.last_num_instances != -1) {
2612 primary->state.last_num_instances =
2613 secondary->state.last_num_instances;
2614 }
2615
2616 if (secondary->state.last_vertex_offset != -1) {
2617 primary->state.last_vertex_offset =
2618 secondary->state.last_vertex_offset;
2619 }
2620
2621 if (secondary->state.last_index_type != -1) {
2622 primary->state.last_index_type =
2623 secondary->state.last_index_type;
2624 }
2625 }
2626
2627 /* After executing commands from secondary buffers we have to dirty
2628 * some states.
2629 */
2630 primary->state.dirty |= RADV_CMD_DIRTY_PIPELINE |
2631 RADV_CMD_DIRTY_INDEX_BUFFER |
2632 RADV_CMD_DIRTY_DYNAMIC_ALL;
2633 radv_mark_descriptor_sets_dirty(primary, VK_PIPELINE_BIND_POINT_GRAPHICS);
2634 radv_mark_descriptor_sets_dirty(primary, VK_PIPELINE_BIND_POINT_COMPUTE);
2635 }
2636
2637 VkResult radv_CreateCommandPool(
2638 VkDevice _device,
2639 const VkCommandPoolCreateInfo* pCreateInfo,
2640 const VkAllocationCallbacks* pAllocator,
2641 VkCommandPool* pCmdPool)
2642 {
2643 RADV_FROM_HANDLE(radv_device, device, _device);
2644 struct radv_cmd_pool *pool;
2645
2646 pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
2647 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2648 if (pool == NULL)
2649 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2650
2651 if (pAllocator)
2652 pool->alloc = *pAllocator;
2653 else
2654 pool->alloc = device->alloc;
2655
2656 list_inithead(&pool->cmd_buffers);
2657 list_inithead(&pool->free_cmd_buffers);
2658
2659 pool->queue_family_index = pCreateInfo->queueFamilyIndex;
2660
2661 *pCmdPool = radv_cmd_pool_to_handle(pool);
2662
2663 return VK_SUCCESS;
2664
2665 }
2666
2667 void radv_DestroyCommandPool(
2668 VkDevice _device,
2669 VkCommandPool commandPool,
2670 const VkAllocationCallbacks* pAllocator)
2671 {
2672 RADV_FROM_HANDLE(radv_device, device, _device);
2673 RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
2674
2675 if (!pool)
2676 return;
2677
2678 list_for_each_entry_safe(struct radv_cmd_buffer, cmd_buffer,
2679 &pool->cmd_buffers, pool_link) {
2680 radv_cmd_buffer_destroy(cmd_buffer);
2681 }
2682
2683 list_for_each_entry_safe(struct radv_cmd_buffer, cmd_buffer,
2684 &pool->free_cmd_buffers, pool_link) {
2685 radv_cmd_buffer_destroy(cmd_buffer);
2686 }
2687
2688 vk_free2(&device->alloc, pAllocator, pool);
2689 }
2690
2691 VkResult radv_ResetCommandPool(
2692 VkDevice device,
2693 VkCommandPool commandPool,
2694 VkCommandPoolResetFlags flags)
2695 {
2696 RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
2697 VkResult result;
2698
2699 list_for_each_entry(struct radv_cmd_buffer, cmd_buffer,
2700 &pool->cmd_buffers, pool_link) {
2701 result = radv_reset_cmd_buffer(cmd_buffer);
2702 if (result != VK_SUCCESS)
2703 return result;
2704 }
2705
2706 return VK_SUCCESS;
2707 }
2708
2709 void radv_TrimCommandPool(
2710 VkDevice device,
2711 VkCommandPool commandPool,
2712 VkCommandPoolTrimFlagsKHR flags)
2713 {
2714 RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
2715
2716 if (!pool)
2717 return;
2718
2719 list_for_each_entry_safe(struct radv_cmd_buffer, cmd_buffer,
2720 &pool->free_cmd_buffers, pool_link) {
2721 radv_cmd_buffer_destroy(cmd_buffer);
2722 }
2723 }
2724
2725 void radv_CmdBeginRenderPass(
2726 VkCommandBuffer commandBuffer,
2727 const VkRenderPassBeginInfo* pRenderPassBegin,
2728 VkSubpassContents contents)
2729 {
2730 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2731 RADV_FROM_HANDLE(radv_render_pass, pass, pRenderPassBegin->renderPass);
2732 RADV_FROM_HANDLE(radv_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
2733
2734 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
2735 cmd_buffer->cs, 2048);
2736 MAYBE_UNUSED VkResult result;
2737
2738 cmd_buffer->state.framebuffer = framebuffer;
2739 cmd_buffer->state.pass = pass;
2740 cmd_buffer->state.render_area = pRenderPassBegin->renderArea;
2741
2742 result = radv_cmd_state_setup_attachments(cmd_buffer, pass, pRenderPassBegin);
2743 if (result != VK_SUCCESS)
2744 return;
2745
2746 radv_cmd_buffer_set_subpass(cmd_buffer, pass->subpasses, true);
2747 assert(cmd_buffer->cs->cdw <= cdw_max);
2748
2749 radv_cmd_buffer_clear_subpass(cmd_buffer);
2750 }
2751
2752 void radv_CmdNextSubpass(
2753 VkCommandBuffer commandBuffer,
2754 VkSubpassContents contents)
2755 {
2756 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2757
2758 radv_cmd_buffer_resolve_subpass(cmd_buffer);
2759
2760 radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs,
2761 2048);
2762
2763 radv_cmd_buffer_set_subpass(cmd_buffer, cmd_buffer->state.subpass + 1, true);
2764 radv_cmd_buffer_clear_subpass(cmd_buffer);
2765 }
2766
2767 static void radv_emit_view_index(struct radv_cmd_buffer *cmd_buffer, unsigned index)
2768 {
2769 struct radv_pipeline *pipeline = cmd_buffer->state.pipeline;
2770 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; ++stage) {
2771 if (!pipeline->shaders[stage])
2772 continue;
2773 struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, stage, AC_UD_VIEW_INDEX);
2774 if (loc->sgpr_idx == -1)
2775 continue;
2776 uint32_t base_reg = pipeline->user_data_0[stage];
2777 radeon_set_sh_reg(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, index);
2778
2779 }
2780 if (pipeline->gs_copy_shader) {
2781 struct radv_userdata_info *loc = &pipeline->gs_copy_shader->info.user_sgprs_locs.shader_data[AC_UD_VIEW_INDEX];
2782 if (loc->sgpr_idx != -1) {
2783 uint32_t base_reg = R_00B130_SPI_SHADER_USER_DATA_VS_0;
2784 radeon_set_sh_reg(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, index);
2785 }
2786 }
2787 }
2788
2789 static void
2790 radv_cs_emit_draw_packet(struct radv_cmd_buffer *cmd_buffer,
2791 uint32_t vertex_count)
2792 {
2793 radeon_emit(cmd_buffer->cs, PKT3(PKT3_DRAW_INDEX_AUTO, 1, cmd_buffer->state.predicating));
2794 radeon_emit(cmd_buffer->cs, vertex_count);
2795 radeon_emit(cmd_buffer->cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX |
2796 S_0287F0_USE_OPAQUE(0));
2797 }
2798
2799 static void
2800 radv_cs_emit_draw_indexed_packet(struct radv_cmd_buffer *cmd_buffer,
2801 uint64_t index_va,
2802 uint32_t index_count)
2803 {
2804 radeon_emit(cmd_buffer->cs, PKT3(PKT3_DRAW_INDEX_2, 4, false));
2805 radeon_emit(cmd_buffer->cs, cmd_buffer->state.max_index_count);
2806 radeon_emit(cmd_buffer->cs, index_va);
2807 radeon_emit(cmd_buffer->cs, index_va >> 32);
2808 radeon_emit(cmd_buffer->cs, index_count);
2809 radeon_emit(cmd_buffer->cs, V_0287F0_DI_SRC_SEL_DMA);
2810 }
2811
2812 static void
2813 radv_cs_emit_indirect_draw_packet(struct radv_cmd_buffer *cmd_buffer,
2814 bool indexed,
2815 uint32_t draw_count,
2816 uint64_t count_va,
2817 uint32_t stride)
2818 {
2819 struct radeon_winsys_cs *cs = cmd_buffer->cs;
2820 unsigned di_src_sel = indexed ? V_0287F0_DI_SRC_SEL_DMA
2821 : V_0287F0_DI_SRC_SEL_AUTO_INDEX;
2822 bool draw_id_enable = radv_get_vertex_shader(cmd_buffer->state.pipeline)->info.info.vs.needs_draw_id;
2823 uint32_t base_reg = cmd_buffer->state.pipeline->graphics.vtx_base_sgpr;
2824 assert(base_reg);
2825
2826 /* just reset draw state for vertex data */
2827 cmd_buffer->state.last_first_instance = -1;
2828 cmd_buffer->state.last_num_instances = -1;
2829 cmd_buffer->state.last_vertex_offset = -1;
2830
2831 if (draw_count == 1 && !count_va && !draw_id_enable) {
2832 radeon_emit(cs, PKT3(indexed ? PKT3_DRAW_INDEX_INDIRECT :
2833 PKT3_DRAW_INDIRECT, 3, false));
2834 radeon_emit(cs, 0);
2835 radeon_emit(cs, (base_reg - SI_SH_REG_OFFSET) >> 2);
2836 radeon_emit(cs, ((base_reg + 4) - SI_SH_REG_OFFSET) >> 2);
2837 radeon_emit(cs, di_src_sel);
2838 } else {
2839 radeon_emit(cs, PKT3(indexed ? PKT3_DRAW_INDEX_INDIRECT_MULTI :
2840 PKT3_DRAW_INDIRECT_MULTI,
2841 8, false));
2842 radeon_emit(cs, 0);
2843 radeon_emit(cs, (base_reg - SI_SH_REG_OFFSET) >> 2);
2844 radeon_emit(cs, ((base_reg + 4) - SI_SH_REG_OFFSET) >> 2);
2845 radeon_emit(cs, (((base_reg + 8) - SI_SH_REG_OFFSET) >> 2) |
2846 S_2C3_DRAW_INDEX_ENABLE(draw_id_enable) |
2847 S_2C3_COUNT_INDIRECT_ENABLE(!!count_va));
2848 radeon_emit(cs, draw_count); /* count */
2849 radeon_emit(cs, count_va); /* count_addr */
2850 radeon_emit(cs, count_va >> 32);
2851 radeon_emit(cs, stride); /* stride */
2852 radeon_emit(cs, di_src_sel);
2853 }
2854 }
2855
2856 struct radv_draw_info {
2857 /**
2858 * Number of vertices.
2859 */
2860 uint32_t count;
2861
2862 /**
2863 * Index of the first vertex.
2864 */
2865 int32_t vertex_offset;
2866
2867 /**
2868 * First instance id.
2869 */
2870 uint32_t first_instance;
2871
2872 /**
2873 * Number of instances.
2874 */
2875 uint32_t instance_count;
2876
2877 /**
2878 * First index (indexed draws only).
2879 */
2880 uint32_t first_index;
2881
2882 /**
2883 * Whether it's an indexed draw.
2884 */
2885 bool indexed;
2886
2887 /**
2888 * Indirect draw parameters resource.
2889 */
2890 struct radv_buffer *indirect;
2891 uint64_t indirect_offset;
2892 uint32_t stride;
2893
2894 /**
2895 * Draw count parameters resource.
2896 */
2897 struct radv_buffer *count_buffer;
2898 uint64_t count_buffer_offset;
2899 };
2900
2901 static void
2902 radv_emit_draw_packets(struct radv_cmd_buffer *cmd_buffer,
2903 const struct radv_draw_info *info)
2904 {
2905 struct radv_cmd_state *state = &cmd_buffer->state;
2906 struct radeon_winsys *ws = cmd_buffer->device->ws;
2907 struct radeon_winsys_cs *cs = cmd_buffer->cs;
2908
2909 if (info->indirect) {
2910 uint64_t va = radv_buffer_get_va(info->indirect->bo);
2911 uint64_t count_va = 0;
2912
2913 va += info->indirect->offset + info->indirect_offset;
2914
2915 radv_cs_add_buffer(ws, cs, info->indirect->bo, 8);
2916
2917 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
2918 radeon_emit(cs, 1);
2919 radeon_emit(cs, va);
2920 radeon_emit(cs, va >> 32);
2921
2922 if (info->count_buffer) {
2923 count_va = radv_buffer_get_va(info->count_buffer->bo);
2924 count_va += info->count_buffer->offset +
2925 info->count_buffer_offset;
2926
2927 radv_cs_add_buffer(ws, cs, info->count_buffer->bo, 8);
2928 }
2929
2930 if (!state->subpass->view_mask) {
2931 radv_cs_emit_indirect_draw_packet(cmd_buffer,
2932 info->indexed,
2933 info->count,
2934 count_va,
2935 info->stride);
2936 } else {
2937 unsigned i;
2938 for_each_bit(i, state->subpass->view_mask) {
2939 radv_emit_view_index(cmd_buffer, i);
2940
2941 radv_cs_emit_indirect_draw_packet(cmd_buffer,
2942 info->indexed,
2943 info->count,
2944 count_va,
2945 info->stride);
2946 }
2947 }
2948 } else {
2949 assert(state->pipeline->graphics.vtx_base_sgpr);
2950
2951 if (info->vertex_offset != state->last_vertex_offset ||
2952 info->first_instance != state->last_first_instance) {
2953 radeon_set_sh_reg_seq(cs, state->pipeline->graphics.vtx_base_sgpr,
2954 state->pipeline->graphics.vtx_emit_num);
2955
2956 radeon_emit(cs, info->vertex_offset);
2957 radeon_emit(cs, info->first_instance);
2958 if (state->pipeline->graphics.vtx_emit_num == 3)
2959 radeon_emit(cs, 0);
2960 state->last_first_instance = info->first_instance;
2961 state->last_vertex_offset = info->vertex_offset;
2962 }
2963
2964 if (state->last_num_instances != info->instance_count) {
2965 radeon_emit(cs, PKT3(PKT3_NUM_INSTANCES, 0, state->predicating));
2966 radeon_emit(cs, info->instance_count);
2967 state->last_num_instances = info->instance_count;
2968 }
2969
2970 if (info->indexed) {
2971 int index_size = state->index_type ? 4 : 2;
2972 uint64_t index_va;
2973
2974 index_va = state->index_va;
2975 index_va += info->first_index * index_size;
2976
2977 if (!state->subpass->view_mask) {
2978 radv_cs_emit_draw_indexed_packet(cmd_buffer,
2979 index_va,
2980 info->count);
2981 } else {
2982 unsigned i;
2983 for_each_bit(i, state->subpass->view_mask) {
2984 radv_emit_view_index(cmd_buffer, i);
2985
2986 radv_cs_emit_draw_indexed_packet(cmd_buffer,
2987 index_va,
2988 info->count);
2989 }
2990 }
2991 } else {
2992 if (!state->subpass->view_mask) {
2993 radv_cs_emit_draw_packet(cmd_buffer, info->count);
2994 } else {
2995 unsigned i;
2996 for_each_bit(i, state->subpass->view_mask) {
2997 radv_emit_view_index(cmd_buffer, i);
2998
2999 radv_cs_emit_draw_packet(cmd_buffer,
3000 info->count);
3001 }
3002 }
3003 }
3004 }
3005 }
3006
3007 static void
3008 radv_emit_all_graphics_states(struct radv_cmd_buffer *cmd_buffer,
3009 const struct radv_draw_info *info)
3010 {
3011 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_PIPELINE)
3012 radv_emit_graphics_pipeline(cmd_buffer);
3013
3014 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_FRAMEBUFFER)
3015 radv_emit_framebuffer_state(cmd_buffer);
3016
3017 if (info->indexed) {
3018 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_INDEX_BUFFER)
3019 radv_emit_index_buffer(cmd_buffer);
3020 } else {
3021 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
3022 * so the state must be re-emitted before the next indexed
3023 * draw.
3024 */
3025 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
3026 cmd_buffer->state.last_index_type = -1;
3027 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_INDEX_BUFFER;
3028 }
3029 }
3030
3031 radv_cmd_buffer_flush_dynamic_state(cmd_buffer);
3032
3033 radv_emit_draw_registers(cmd_buffer, info->indexed,
3034 info->instance_count > 1, info->indirect,
3035 info->indirect ? 0 : info->count);
3036 }
3037
3038 static void
3039 radv_draw(struct radv_cmd_buffer *cmd_buffer,
3040 const struct radv_draw_info *info)
3041 {
3042 bool has_prefetch =
3043 cmd_buffer->device->physical_device->rad_info.chip_class >= CIK;
3044 bool pipeline_is_dirty =
3045 (cmd_buffer->state.dirty & RADV_CMD_DIRTY_PIPELINE) &&
3046 cmd_buffer->state.pipeline &&
3047 cmd_buffer->state.pipeline != cmd_buffer->state.emitted_pipeline;
3048
3049 MAYBE_UNUSED unsigned cdw_max =
3050 radeon_check_space(cmd_buffer->device->ws,
3051 cmd_buffer->cs, 4096);
3052
3053 /* Use optimal packet order based on whether we need to sync the
3054 * pipeline.
3055 */
3056 if (cmd_buffer->state.flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3057 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3058 RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
3059 RADV_CMD_FLAG_CS_PARTIAL_FLUSH)) {
3060 /* If we have to wait for idle, set all states first, so that
3061 * all SET packets are processed in parallel with previous draw
3062 * calls. Then upload descriptors, set shader pointers, and
3063 * draw, and prefetch at the end. This ensures that the time
3064 * the CUs are idle is very short. (there are only SET_SH
3065 * packets between the wait and the draw)
3066 */
3067 radv_emit_all_graphics_states(cmd_buffer, info);
3068 si_emit_cache_flush(cmd_buffer);
3069 /* <-- CUs are idle here --> */
3070
3071 if (!radv_upload_graphics_shader_descriptors(cmd_buffer, pipeline_is_dirty))
3072 return;
3073
3074 radv_emit_draw_packets(cmd_buffer, info);
3075 /* <-- CUs are busy here --> */
3076
3077 /* Start prefetches after the draw has been started. Both will
3078 * run in parallel, but starting the draw first is more
3079 * important.
3080 */
3081 if (has_prefetch && cmd_buffer->state.prefetch_L2_mask) {
3082 radv_emit_prefetch_L2(cmd_buffer,
3083 cmd_buffer->state.pipeline, false);
3084 }
3085 } else {
3086 /* If we don't wait for idle, start prefetches first, then set
3087 * states, and draw at the end.
3088 */
3089 si_emit_cache_flush(cmd_buffer);
3090
3091 if (cmd_buffer->state.prefetch_L2_mask) {
3092 /* Only prefetch the vertex shader and VBO descriptors
3093 * in order to start the draw as soon as possible.
3094 */
3095 radv_emit_prefetch_L2(cmd_buffer,
3096 cmd_buffer->state.pipeline, true);
3097 }
3098
3099 if (!radv_upload_graphics_shader_descriptors(cmd_buffer, pipeline_is_dirty))
3100 return;
3101
3102 radv_emit_all_graphics_states(cmd_buffer, info);
3103 radv_emit_draw_packets(cmd_buffer, info);
3104
3105 /* Prefetch the remaining shaders after the draw has been
3106 * started.
3107 */
3108 if (has_prefetch && cmd_buffer->state.prefetch_L2_mask) {
3109 radv_emit_prefetch_L2(cmd_buffer,
3110 cmd_buffer->state.pipeline, false);
3111 }
3112 }
3113
3114 assert(cmd_buffer->cs->cdw <= cdw_max);
3115 radv_cmd_buffer_after_draw(cmd_buffer, RADV_CMD_FLAG_PS_PARTIAL_FLUSH);
3116 }
3117
3118 void radv_CmdDraw(
3119 VkCommandBuffer commandBuffer,
3120 uint32_t vertexCount,
3121 uint32_t instanceCount,
3122 uint32_t firstVertex,
3123 uint32_t firstInstance)
3124 {
3125 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3126 struct radv_draw_info info = {};
3127
3128 info.count = vertexCount;
3129 info.instance_count = instanceCount;
3130 info.first_instance = firstInstance;
3131 info.vertex_offset = firstVertex;
3132
3133 radv_draw(cmd_buffer, &info);
3134 }
3135
3136 void radv_CmdDrawIndexed(
3137 VkCommandBuffer commandBuffer,
3138 uint32_t indexCount,
3139 uint32_t instanceCount,
3140 uint32_t firstIndex,
3141 int32_t vertexOffset,
3142 uint32_t firstInstance)
3143 {
3144 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3145 struct radv_draw_info info = {};
3146
3147 info.indexed = true;
3148 info.count = indexCount;
3149 info.instance_count = instanceCount;
3150 info.first_index = firstIndex;
3151 info.vertex_offset = vertexOffset;
3152 info.first_instance = firstInstance;
3153
3154 radv_draw(cmd_buffer, &info);
3155 }
3156
3157 void radv_CmdDrawIndirect(
3158 VkCommandBuffer commandBuffer,
3159 VkBuffer _buffer,
3160 VkDeviceSize offset,
3161 uint32_t drawCount,
3162 uint32_t stride)
3163 {
3164 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3165 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3166 struct radv_draw_info info = {};
3167
3168 info.count = drawCount;
3169 info.indirect = buffer;
3170 info.indirect_offset = offset;
3171 info.stride = stride;
3172
3173 radv_draw(cmd_buffer, &info);
3174 }
3175
3176 void radv_CmdDrawIndexedIndirect(
3177 VkCommandBuffer commandBuffer,
3178 VkBuffer _buffer,
3179 VkDeviceSize offset,
3180 uint32_t drawCount,
3181 uint32_t stride)
3182 {
3183 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3184 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3185 struct radv_draw_info info = {};
3186
3187 info.indexed = true;
3188 info.count = drawCount;
3189 info.indirect = buffer;
3190 info.indirect_offset = offset;
3191 info.stride = stride;
3192
3193 radv_draw(cmd_buffer, &info);
3194 }
3195
3196 void radv_CmdDrawIndirectCountAMD(
3197 VkCommandBuffer commandBuffer,
3198 VkBuffer _buffer,
3199 VkDeviceSize offset,
3200 VkBuffer _countBuffer,
3201 VkDeviceSize countBufferOffset,
3202 uint32_t maxDrawCount,
3203 uint32_t stride)
3204 {
3205 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3206 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3207 RADV_FROM_HANDLE(radv_buffer, count_buffer, _countBuffer);
3208 struct radv_draw_info info = {};
3209
3210 info.count = maxDrawCount;
3211 info.indirect = buffer;
3212 info.indirect_offset = offset;
3213 info.count_buffer = count_buffer;
3214 info.count_buffer_offset = countBufferOffset;
3215 info.stride = stride;
3216
3217 radv_draw(cmd_buffer, &info);
3218 }
3219
3220 void radv_CmdDrawIndexedIndirectCountAMD(
3221 VkCommandBuffer commandBuffer,
3222 VkBuffer _buffer,
3223 VkDeviceSize offset,
3224 VkBuffer _countBuffer,
3225 VkDeviceSize countBufferOffset,
3226 uint32_t maxDrawCount,
3227 uint32_t stride)
3228 {
3229 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3230 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3231 RADV_FROM_HANDLE(radv_buffer, count_buffer, _countBuffer);
3232 struct radv_draw_info info = {};
3233
3234 info.indexed = true;
3235 info.count = maxDrawCount;
3236 info.indirect = buffer;
3237 info.indirect_offset = offset;
3238 info.count_buffer = count_buffer;
3239 info.count_buffer_offset = countBufferOffset;
3240 info.stride = stride;
3241
3242 radv_draw(cmd_buffer, &info);
3243 }
3244
3245 struct radv_dispatch_info {
3246 /**
3247 * Determine the layout of the grid (in block units) to be used.
3248 */
3249 uint32_t blocks[3];
3250
3251 /**
3252 * A starting offset for the grid. If unaligned is set, the offset
3253 * must still be aligned.
3254 */
3255 uint32_t offsets[3];
3256 /**
3257 * Whether it's an unaligned compute dispatch.
3258 */
3259 bool unaligned;
3260
3261 /**
3262 * Indirect compute parameters resource.
3263 */
3264 struct radv_buffer *indirect;
3265 uint64_t indirect_offset;
3266 };
3267
3268 static void
3269 radv_emit_dispatch_packets(struct radv_cmd_buffer *cmd_buffer,
3270 const struct radv_dispatch_info *info)
3271 {
3272 struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
3273 struct radv_shader_variant *compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
3274 unsigned dispatch_initiator = cmd_buffer->device->dispatch_initiator;
3275 struct radeon_winsys *ws = cmd_buffer->device->ws;
3276 struct radeon_winsys_cs *cs = cmd_buffer->cs;
3277 struct radv_userdata_info *loc;
3278
3279 loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_COMPUTE,
3280 AC_UD_CS_GRID_SIZE);
3281
3282 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(ws, cs, 25);
3283
3284 if (info->indirect) {
3285 uint64_t va = radv_buffer_get_va(info->indirect->bo);
3286
3287 va += info->indirect->offset + info->indirect_offset;
3288
3289 radv_cs_add_buffer(ws, cs, info->indirect->bo, 8);
3290
3291 if (loc->sgpr_idx != -1) {
3292 for (unsigned i = 0; i < 3; ++i) {
3293 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
3294 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
3295 COPY_DATA_DST_SEL(COPY_DATA_REG));
3296 radeon_emit(cs, (va + 4 * i));
3297 radeon_emit(cs, (va + 4 * i) >> 32);
3298 radeon_emit(cs, ((R_00B900_COMPUTE_USER_DATA_0
3299 + loc->sgpr_idx * 4) >> 2) + i);
3300 radeon_emit(cs, 0);
3301 }
3302 }
3303
3304 if (radv_cmd_buffer_uses_mec(cmd_buffer)) {
3305 radeon_emit(cs, PKT3(PKT3_DISPATCH_INDIRECT, 2, 0) |
3306 PKT3_SHADER_TYPE_S(1));
3307 radeon_emit(cs, va);
3308 radeon_emit(cs, va >> 32);
3309 radeon_emit(cs, dispatch_initiator);
3310 } else {
3311 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0) |
3312 PKT3_SHADER_TYPE_S(1));
3313 radeon_emit(cs, 1);
3314 radeon_emit(cs, va);
3315 radeon_emit(cs, va >> 32);
3316
3317 radeon_emit(cs, PKT3(PKT3_DISPATCH_INDIRECT, 1, 0) |
3318 PKT3_SHADER_TYPE_S(1));
3319 radeon_emit(cs, 0);
3320 radeon_emit(cs, dispatch_initiator);
3321 }
3322 } else {
3323 unsigned blocks[3] = { info->blocks[0], info->blocks[1], info->blocks[2] };
3324 unsigned offsets[3] = { info->offsets[0], info->offsets[1], info->offsets[2] };
3325
3326 if (info->unaligned) {
3327 unsigned *cs_block_size = compute_shader->info.cs.block_size;
3328 unsigned remainder[3];
3329
3330 /* If aligned, these should be an entire block size,
3331 * not 0.
3332 */
3333 remainder[0] = blocks[0] + cs_block_size[0] -
3334 align_u32_npot(blocks[0], cs_block_size[0]);
3335 remainder[1] = blocks[1] + cs_block_size[1] -
3336 align_u32_npot(blocks[1], cs_block_size[1]);
3337 remainder[2] = blocks[2] + cs_block_size[2] -
3338 align_u32_npot(blocks[2], cs_block_size[2]);
3339
3340 blocks[0] = round_up_u32(blocks[0], cs_block_size[0]);
3341 blocks[1] = round_up_u32(blocks[1], cs_block_size[1]);
3342 blocks[2] = round_up_u32(blocks[2], cs_block_size[2]);
3343
3344 for(unsigned i = 0; i < 3; ++i) {
3345 assert(offsets[i] % cs_block_size[i] == 0);
3346 offsets[i] /= cs_block_size[i];
3347 }
3348
3349 radeon_set_sh_reg_seq(cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
3350 radeon_emit(cs,
3351 S_00B81C_NUM_THREAD_FULL(cs_block_size[0]) |
3352 S_00B81C_NUM_THREAD_PARTIAL(remainder[0]));
3353 radeon_emit(cs,
3354 S_00B81C_NUM_THREAD_FULL(cs_block_size[1]) |
3355 S_00B81C_NUM_THREAD_PARTIAL(remainder[1]));
3356 radeon_emit(cs,
3357 S_00B81C_NUM_THREAD_FULL(cs_block_size[2]) |
3358 S_00B81C_NUM_THREAD_PARTIAL(remainder[2]));
3359
3360 dispatch_initiator |= S_00B800_PARTIAL_TG_EN(1);
3361 }
3362
3363 if (loc->sgpr_idx != -1) {
3364 assert(!loc->indirect);
3365 assert(loc->num_sgprs == 3);
3366
3367 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0 +
3368 loc->sgpr_idx * 4, 3);
3369 radeon_emit(cs, blocks[0]);
3370 radeon_emit(cs, blocks[1]);
3371 radeon_emit(cs, blocks[2]);
3372 }
3373
3374 if (offsets[0] || offsets[1] || offsets[2]) {
3375 radeon_set_sh_reg_seq(cs, R_00B810_COMPUTE_START_X, 3);
3376 radeon_emit(cs, offsets[0]);
3377 radeon_emit(cs, offsets[1]);
3378 radeon_emit(cs, offsets[2]);
3379
3380 /* The blocks in the packet are not counts but end values. */
3381 for (unsigned i = 0; i < 3; ++i)
3382 blocks[i] += offsets[i];
3383 } else {
3384 dispatch_initiator |= S_00B800_FORCE_START_AT_000(1);
3385 }
3386
3387 radeon_emit(cs, PKT3(PKT3_DISPATCH_DIRECT, 3, 0) |
3388 PKT3_SHADER_TYPE_S(1));
3389 radeon_emit(cs, blocks[0]);
3390 radeon_emit(cs, blocks[1]);
3391 radeon_emit(cs, blocks[2]);
3392 radeon_emit(cs, dispatch_initiator);
3393 }
3394
3395 assert(cmd_buffer->cs->cdw <= cdw_max);
3396 }
3397
3398 static void
3399 radv_upload_compute_shader_descriptors(struct radv_cmd_buffer *cmd_buffer)
3400 {
3401 radv_flush_descriptors(cmd_buffer, VK_SHADER_STAGE_COMPUTE_BIT);
3402 radv_flush_constants(cmd_buffer, cmd_buffer->state.compute_pipeline,
3403 VK_SHADER_STAGE_COMPUTE_BIT);
3404 }
3405
3406 static void
3407 radv_dispatch(struct radv_cmd_buffer *cmd_buffer,
3408 const struct radv_dispatch_info *info)
3409 {
3410 struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
3411 bool has_prefetch =
3412 cmd_buffer->device->physical_device->rad_info.chip_class >= CIK;
3413 bool pipeline_is_dirty = pipeline &&
3414 pipeline != cmd_buffer->state.emitted_compute_pipeline;
3415
3416 if (cmd_buffer->state.flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3417 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3418 RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
3419 RADV_CMD_FLAG_CS_PARTIAL_FLUSH)) {
3420 /* If we have to wait for idle, set all states first, so that
3421 * all SET packets are processed in parallel with previous draw
3422 * calls. Then upload descriptors, set shader pointers, and
3423 * dispatch, and prefetch at the end. This ensures that the
3424 * time the CUs are idle is very short. (there are only SET_SH
3425 * packets between the wait and the draw)
3426 */
3427 radv_emit_compute_pipeline(cmd_buffer);
3428 si_emit_cache_flush(cmd_buffer);
3429 /* <-- CUs are idle here --> */
3430
3431 radv_upload_compute_shader_descriptors(cmd_buffer);
3432
3433 radv_emit_dispatch_packets(cmd_buffer, info);
3434 /* <-- CUs are busy here --> */
3435
3436 /* Start prefetches after the dispatch has been started. Both
3437 * will run in parallel, but starting the dispatch first is
3438 * more important.
3439 */
3440 if (has_prefetch && pipeline_is_dirty) {
3441 radv_emit_shader_prefetch(cmd_buffer,
3442 pipeline->shaders[MESA_SHADER_COMPUTE]);
3443 }
3444 } else {
3445 /* If we don't wait for idle, start prefetches first, then set
3446 * states, and dispatch at the end.
3447 */
3448 si_emit_cache_flush(cmd_buffer);
3449
3450 if (has_prefetch && pipeline_is_dirty) {
3451 radv_emit_shader_prefetch(cmd_buffer,
3452 pipeline->shaders[MESA_SHADER_COMPUTE]);
3453 }
3454
3455 radv_upload_compute_shader_descriptors(cmd_buffer);
3456
3457 radv_emit_compute_pipeline(cmd_buffer);
3458 radv_emit_dispatch_packets(cmd_buffer, info);
3459 }
3460
3461 radv_cmd_buffer_after_draw(cmd_buffer, RADV_CMD_FLAG_CS_PARTIAL_FLUSH);
3462 }
3463
3464 void radv_CmdDispatchBase(
3465 VkCommandBuffer commandBuffer,
3466 uint32_t base_x,
3467 uint32_t base_y,
3468 uint32_t base_z,
3469 uint32_t x,
3470 uint32_t y,
3471 uint32_t z)
3472 {
3473 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3474 struct radv_dispatch_info info = {};
3475
3476 info.blocks[0] = x;
3477 info.blocks[1] = y;
3478 info.blocks[2] = z;
3479
3480 info.offsets[0] = base_x;
3481 info.offsets[1] = base_y;
3482 info.offsets[2] = base_z;
3483 radv_dispatch(cmd_buffer, &info);
3484 }
3485
3486 void radv_CmdDispatch(
3487 VkCommandBuffer commandBuffer,
3488 uint32_t x,
3489 uint32_t y,
3490 uint32_t z)
3491 {
3492 radv_CmdDispatchBase(commandBuffer, 0, 0, 0, x, y, z);
3493 }
3494
3495 void radv_CmdDispatchIndirect(
3496 VkCommandBuffer commandBuffer,
3497 VkBuffer _buffer,
3498 VkDeviceSize offset)
3499 {
3500 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3501 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3502 struct radv_dispatch_info info = {};
3503
3504 info.indirect = buffer;
3505 info.indirect_offset = offset;
3506
3507 radv_dispatch(cmd_buffer, &info);
3508 }
3509
3510 void radv_unaligned_dispatch(
3511 struct radv_cmd_buffer *cmd_buffer,
3512 uint32_t x,
3513 uint32_t y,
3514 uint32_t z)
3515 {
3516 struct radv_dispatch_info info = {};
3517
3518 info.blocks[0] = x;
3519 info.blocks[1] = y;
3520 info.blocks[2] = z;
3521 info.unaligned = 1;
3522
3523 radv_dispatch(cmd_buffer, &info);
3524 }
3525
3526 void radv_CmdEndRenderPass(
3527 VkCommandBuffer commandBuffer)
3528 {
3529 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3530
3531 radv_subpass_barrier(cmd_buffer, &cmd_buffer->state.pass->end_barrier);
3532
3533 radv_cmd_buffer_resolve_subpass(cmd_buffer);
3534
3535 for (unsigned i = 0; i < cmd_buffer->state.framebuffer->attachment_count; ++i) {
3536 VkImageLayout layout = cmd_buffer->state.pass->attachments[i].final_layout;
3537 radv_handle_subpass_image_transition(cmd_buffer,
3538 (VkAttachmentReference){i, layout});
3539 }
3540
3541 vk_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
3542
3543 cmd_buffer->state.pass = NULL;
3544 cmd_buffer->state.subpass = NULL;
3545 cmd_buffer->state.attachments = NULL;
3546 cmd_buffer->state.framebuffer = NULL;
3547 }
3548
3549 /*
3550 * For HTILE we have the following interesting clear words:
3551 * 0xfffff30f: Uncompressed, full depth range, for depth+stencil HTILE
3552 * 0xfffc000f: Uncompressed, full depth range, for depth only HTILE.
3553 * 0xfffffff0: Clear depth to 1.0
3554 * 0x00000000: Clear depth to 0.0
3555 */
3556 static void radv_initialize_htile(struct radv_cmd_buffer *cmd_buffer,
3557 struct radv_image *image,
3558 const VkImageSubresourceRange *range,
3559 uint32_t clear_word)
3560 {
3561 assert(range->baseMipLevel == 0);
3562 assert(range->levelCount == 1 || range->levelCount == VK_REMAINING_ARRAY_LAYERS);
3563 unsigned layer_count = radv_get_layerCount(image, range);
3564 uint64_t size = image->surface.htile_slice_size * layer_count;
3565 uint64_t offset = image->offset + image->htile_offset +
3566 image->surface.htile_slice_size * range->baseArrayLayer;
3567 struct radv_cmd_state *state = &cmd_buffer->state;
3568
3569 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3570 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3571
3572 state->flush_bits |= radv_fill_buffer(cmd_buffer, image->bo, offset,
3573 size, clear_word);
3574
3575 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3576 }
3577
3578 static void radv_handle_depth_image_transition(struct radv_cmd_buffer *cmd_buffer,
3579 struct radv_image *image,
3580 VkImageLayout src_layout,
3581 VkImageLayout dst_layout,
3582 unsigned src_queue_mask,
3583 unsigned dst_queue_mask,
3584 const VkImageSubresourceRange *range,
3585 VkImageAspectFlags pending_clears)
3586 {
3587 if (dst_layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL &&
3588 (pending_clears & vk_format_aspects(image->vk_format)) == vk_format_aspects(image->vk_format) &&
3589 cmd_buffer->state.render_area.offset.x == 0 && cmd_buffer->state.render_area.offset.y == 0 &&
3590 cmd_buffer->state.render_area.extent.width == image->info.width &&
3591 cmd_buffer->state.render_area.extent.height == image->info.height) {
3592 /* The clear will initialize htile. */
3593 return;
3594 } else if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED &&
3595 radv_layout_has_htile(image, dst_layout, dst_queue_mask)) {
3596 /* TODO: merge with the clear if applicable */
3597 radv_initialize_htile(cmd_buffer, image, range, 0);
3598 } else if (!radv_layout_is_htile_compressed(image, src_layout, src_queue_mask) &&
3599 radv_layout_is_htile_compressed(image, dst_layout, dst_queue_mask)) {
3600 uint32_t clear_value = vk_format_is_stencil(image->vk_format) ? 0xfffff30f : 0xfffc000f;
3601 radv_initialize_htile(cmd_buffer, image, range, clear_value);
3602 } else if (radv_layout_is_htile_compressed(image, src_layout, src_queue_mask) &&
3603 !radv_layout_is_htile_compressed(image, dst_layout, dst_queue_mask)) {
3604 VkImageSubresourceRange local_range = *range;
3605 local_range.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
3606 local_range.baseMipLevel = 0;
3607 local_range.levelCount = 1;
3608
3609 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3610 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3611
3612 radv_decompress_depth_image_inplace(cmd_buffer, image, &local_range);
3613
3614 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3615 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3616 }
3617 }
3618
3619 void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
3620 struct radv_image *image, uint32_t value)
3621 {
3622 struct radv_cmd_state *state = &cmd_buffer->state;
3623
3624 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3625 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3626
3627 state->flush_bits |= radv_fill_buffer(cmd_buffer, image->bo,
3628 image->offset + image->cmask.offset,
3629 image->cmask.size, value);
3630
3631 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3632 }
3633
3634 static void radv_handle_cmask_image_transition(struct radv_cmd_buffer *cmd_buffer,
3635 struct radv_image *image,
3636 VkImageLayout src_layout,
3637 VkImageLayout dst_layout,
3638 unsigned src_queue_mask,
3639 unsigned dst_queue_mask,
3640 const VkImageSubresourceRange *range)
3641 {
3642 if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
3643 if (image->fmask.size)
3644 radv_initialise_cmask(cmd_buffer, image, 0xccccccccu);
3645 else
3646 radv_initialise_cmask(cmd_buffer, image, 0xffffffffu);
3647 } else if (radv_layout_can_fast_clear(image, src_layout, src_queue_mask) &&
3648 !radv_layout_can_fast_clear(image, dst_layout, dst_queue_mask)) {
3649 radv_fast_clear_flush_image_inplace(cmd_buffer, image, range);
3650 }
3651 }
3652
3653 void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
3654 struct radv_image *image, uint32_t value)
3655 {
3656 struct radv_cmd_state *state = &cmd_buffer->state;
3657
3658 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3659 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3660
3661 state->flush_bits |= radv_fill_buffer(cmd_buffer, image->bo,
3662 image->offset + image->dcc_offset,
3663 image->surface.dcc_size, value);
3664
3665 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3666 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3667 }
3668
3669 static void radv_handle_dcc_image_transition(struct radv_cmd_buffer *cmd_buffer,
3670 struct radv_image *image,
3671 VkImageLayout src_layout,
3672 VkImageLayout dst_layout,
3673 unsigned src_queue_mask,
3674 unsigned dst_queue_mask,
3675 const VkImageSubresourceRange *range)
3676 {
3677 if (src_layout == VK_IMAGE_LAYOUT_PREINITIALIZED) {
3678 radv_initialize_dcc(cmd_buffer, image, 0xffffffffu);
3679 } else if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
3680 radv_initialize_dcc(cmd_buffer, image,
3681 radv_layout_dcc_compressed(image, dst_layout, dst_queue_mask) ?
3682 0x20202020u : 0xffffffffu);
3683 } else if (radv_layout_dcc_compressed(image, src_layout, src_queue_mask) &&
3684 !radv_layout_dcc_compressed(image, dst_layout, dst_queue_mask)) {
3685 radv_decompress_dcc(cmd_buffer, image, range);
3686 } else if (radv_layout_can_fast_clear(image, src_layout, src_queue_mask) &&
3687 !radv_layout_can_fast_clear(image, dst_layout, dst_queue_mask)) {
3688 radv_fast_clear_flush_image_inplace(cmd_buffer, image, range);
3689 }
3690 }
3691
3692 static void radv_handle_image_transition(struct radv_cmd_buffer *cmd_buffer,
3693 struct radv_image *image,
3694 VkImageLayout src_layout,
3695 VkImageLayout dst_layout,
3696 uint32_t src_family,
3697 uint32_t dst_family,
3698 const VkImageSubresourceRange *range,
3699 VkImageAspectFlags pending_clears)
3700 {
3701 if (image->exclusive && src_family != dst_family) {
3702 /* This is an acquire or a release operation and there will be
3703 * a corresponding release/acquire. Do the transition in the
3704 * most flexible queue. */
3705
3706 assert(src_family == cmd_buffer->queue_family_index ||
3707 dst_family == cmd_buffer->queue_family_index);
3708
3709 if (cmd_buffer->queue_family_index == RADV_QUEUE_TRANSFER)
3710 return;
3711
3712 if (cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE &&
3713 (src_family == RADV_QUEUE_GENERAL ||
3714 dst_family == RADV_QUEUE_GENERAL))
3715 return;
3716 }
3717
3718 unsigned src_queue_mask = radv_image_queue_family_mask(image, src_family, cmd_buffer->queue_family_index);
3719 unsigned dst_queue_mask = radv_image_queue_family_mask(image, dst_family, cmd_buffer->queue_family_index);
3720
3721 if (image->surface.htile_size)
3722 radv_handle_depth_image_transition(cmd_buffer, image, src_layout,
3723 dst_layout, src_queue_mask,
3724 dst_queue_mask, range,
3725 pending_clears);
3726
3727 if (image->cmask.size || image->fmask.size)
3728 radv_handle_cmask_image_transition(cmd_buffer, image, src_layout,
3729 dst_layout, src_queue_mask,
3730 dst_queue_mask, range);
3731
3732 if (image->surface.dcc_size)
3733 radv_handle_dcc_image_transition(cmd_buffer, image, src_layout,
3734 dst_layout, src_queue_mask,
3735 dst_queue_mask, range);
3736 }
3737
3738 void radv_CmdPipelineBarrier(
3739 VkCommandBuffer commandBuffer,
3740 VkPipelineStageFlags srcStageMask,
3741 VkPipelineStageFlags destStageMask,
3742 VkBool32 byRegion,
3743 uint32_t memoryBarrierCount,
3744 const VkMemoryBarrier* pMemoryBarriers,
3745 uint32_t bufferMemoryBarrierCount,
3746 const VkBufferMemoryBarrier* pBufferMemoryBarriers,
3747 uint32_t imageMemoryBarrierCount,
3748 const VkImageMemoryBarrier* pImageMemoryBarriers)
3749 {
3750 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3751 enum radv_cmd_flush_bits src_flush_bits = 0;
3752 enum radv_cmd_flush_bits dst_flush_bits = 0;
3753
3754 for (uint32_t i = 0; i < memoryBarrierCount; i++) {
3755 src_flush_bits |= radv_src_access_flush(cmd_buffer, pMemoryBarriers[i].srcAccessMask);
3756 dst_flush_bits |= radv_dst_access_flush(cmd_buffer, pMemoryBarriers[i].dstAccessMask,
3757 NULL);
3758 }
3759
3760 for (uint32_t i = 0; i < bufferMemoryBarrierCount; i++) {
3761 src_flush_bits |= radv_src_access_flush(cmd_buffer, pBufferMemoryBarriers[i].srcAccessMask);
3762 dst_flush_bits |= radv_dst_access_flush(cmd_buffer, pBufferMemoryBarriers[i].dstAccessMask,
3763 NULL);
3764 }
3765
3766 for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
3767 RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
3768 src_flush_bits |= radv_src_access_flush(cmd_buffer, pImageMemoryBarriers[i].srcAccessMask);
3769 dst_flush_bits |= radv_dst_access_flush(cmd_buffer, pImageMemoryBarriers[i].dstAccessMask,
3770 image);
3771 }
3772
3773 radv_stage_flush(cmd_buffer, srcStageMask);
3774 cmd_buffer->state.flush_bits |= src_flush_bits;
3775
3776 for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
3777 RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
3778 radv_handle_image_transition(cmd_buffer, image,
3779 pImageMemoryBarriers[i].oldLayout,
3780 pImageMemoryBarriers[i].newLayout,
3781 pImageMemoryBarriers[i].srcQueueFamilyIndex,
3782 pImageMemoryBarriers[i].dstQueueFamilyIndex,
3783 &pImageMemoryBarriers[i].subresourceRange,
3784 0);
3785 }
3786
3787 cmd_buffer->state.flush_bits |= dst_flush_bits;
3788 }
3789
3790
3791 static void write_event(struct radv_cmd_buffer *cmd_buffer,
3792 struct radv_event *event,
3793 VkPipelineStageFlags stageMask,
3794 unsigned value)
3795 {
3796 struct radeon_winsys_cs *cs = cmd_buffer->cs;
3797 uint64_t va = radv_buffer_get_va(event->bo);
3798
3799 radv_cs_add_buffer(cmd_buffer->device->ws, cs, event->bo, 8);
3800
3801 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 18);
3802
3803 /* TODO: this is overkill. Probably should figure something out from
3804 * the stage mask. */
3805
3806 si_cs_emit_write_event_eop(cs,
3807 cmd_buffer->state.predicating,
3808 cmd_buffer->device->physical_device->rad_info.chip_class,
3809 radv_cmd_buffer_uses_mec(cmd_buffer),
3810 V_028A90_BOTTOM_OF_PIPE_TS, 0,
3811 1, va, 2, value);
3812
3813 assert(cmd_buffer->cs->cdw <= cdw_max);
3814 }
3815
3816 void radv_CmdSetEvent(VkCommandBuffer commandBuffer,
3817 VkEvent _event,
3818 VkPipelineStageFlags stageMask)
3819 {
3820 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3821 RADV_FROM_HANDLE(radv_event, event, _event);
3822
3823 write_event(cmd_buffer, event, stageMask, 1);
3824 }
3825
3826 void radv_CmdResetEvent(VkCommandBuffer commandBuffer,
3827 VkEvent _event,
3828 VkPipelineStageFlags stageMask)
3829 {
3830 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3831 RADV_FROM_HANDLE(radv_event, event, _event);
3832
3833 write_event(cmd_buffer, event, stageMask, 0);
3834 }
3835
3836 void radv_CmdWaitEvents(VkCommandBuffer commandBuffer,
3837 uint32_t eventCount,
3838 const VkEvent* pEvents,
3839 VkPipelineStageFlags srcStageMask,
3840 VkPipelineStageFlags dstStageMask,
3841 uint32_t memoryBarrierCount,
3842 const VkMemoryBarrier* pMemoryBarriers,
3843 uint32_t bufferMemoryBarrierCount,
3844 const VkBufferMemoryBarrier* pBufferMemoryBarriers,
3845 uint32_t imageMemoryBarrierCount,
3846 const VkImageMemoryBarrier* pImageMemoryBarriers)
3847 {
3848 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3849 struct radeon_winsys_cs *cs = cmd_buffer->cs;
3850
3851 for (unsigned i = 0; i < eventCount; ++i) {
3852 RADV_FROM_HANDLE(radv_event, event, pEvents[i]);
3853 uint64_t va = radv_buffer_get_va(event->bo);
3854
3855 radv_cs_add_buffer(cmd_buffer->device->ws, cs, event->bo, 8);
3856
3857 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 7);
3858
3859 si_emit_wait_fence(cs, false, va, 1, 0xffffffff);
3860 assert(cmd_buffer->cs->cdw <= cdw_max);
3861 }
3862
3863
3864 for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
3865 RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
3866
3867 radv_handle_image_transition(cmd_buffer, image,
3868 pImageMemoryBarriers[i].oldLayout,
3869 pImageMemoryBarriers[i].newLayout,
3870 pImageMemoryBarriers[i].srcQueueFamilyIndex,
3871 pImageMemoryBarriers[i].dstQueueFamilyIndex,
3872 &pImageMemoryBarriers[i].subresourceRange,
3873 0);
3874 }
3875
3876 /* TODO: figure out how to do memory barriers without waiting */
3877 cmd_buffer->state.flush_bits |= RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER |
3878 RADV_CMD_FLAG_INV_GLOBAL_L2 |
3879 RADV_CMD_FLAG_INV_VMEM_L1 |
3880 RADV_CMD_FLAG_INV_SMEM_L1;
3881 }
3882
3883
3884 void radv_CmdSetDeviceMask(VkCommandBuffer commandBuffer,
3885 uint32_t deviceMask)
3886 {
3887 /* No-op */
3888 }