projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: Implement VK_EXT_discard_rectangles.
[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 static void radv_handle_image_transition(struct radv_cmd_buffer *cmd_buffer,
41 struct radv_image *image,
42 VkImageLayout src_layout,
43 VkImageLayout dst_layout,
44 uint32_t src_family,
45 uint32_t dst_family,
46 const VkImageSubresourceRange *range,
47 VkImageAspectFlags pending_clears);
48
49 const struct radv_dynamic_state default_dynamic_state = {
50 .viewport = {
51 .count = 0,
52 },
53 .scissor = {
54 .count = 0,
55 },
56 .line_width = 1.0f,
57 .depth_bias = {
58 .bias = 0.0f,
59 .clamp = 0.0f,
60 .slope = 0.0f,
61 },
62 .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
63 .depth_bounds = {
64 .min = 0.0f,
65 .max = 1.0f,
66 },
67 .stencil_compare_mask = {
68 .front = ~0u,
69 .back = ~0u,
70 },
71 .stencil_write_mask = {
72 .front = ~0u,
73 .back = ~0u,
74 },
75 .stencil_reference = {
76 .front = 0u,
77 .back = 0u,
78 },
79 };
80
81 static void
82 radv_bind_dynamic_state(struct radv_cmd_buffer *cmd_buffer,
83 const struct radv_dynamic_state *src)
84 {
85 struct radv_dynamic_state *dest = &cmd_buffer->state.dynamic;
86 uint32_t copy_mask = src->mask;
87 uint32_t dest_mask = 0;
88
89 /* Make sure to copy the number of viewports/scissors because they can
90 * only be specified at pipeline creation time.
91 */
92 dest->viewport.count = src->viewport.count;
93 dest->scissor.count = src->scissor.count;
94 dest->discard_rectangle.count = src->discard_rectangle.count;
95
96 if (copy_mask & RADV_DYNAMIC_VIEWPORT) {
97 if (memcmp(&dest->viewport.viewports, &src->viewport.viewports,
98 src->viewport.count * sizeof(VkViewport))) {
99 typed_memcpy(dest->viewport.viewports,
100 src->viewport.viewports,
101 src->viewport.count);
102 dest_mask |= RADV_DYNAMIC_VIEWPORT;
103 }
104 }
105
106 if (copy_mask & RADV_DYNAMIC_SCISSOR) {
107 if (memcmp(&dest->scissor.scissors, &src->scissor.scissors,
108 src->scissor.count * sizeof(VkRect2D))) {
109 typed_memcpy(dest->scissor.scissors,
110 src->scissor.scissors, src->scissor.count);
111 dest_mask |= RADV_DYNAMIC_SCISSOR;
112 }
113 }
114
115 if (copy_mask & RADV_DYNAMIC_LINE_WIDTH) {
116 if (dest->line_width != src->line_width) {
117 dest->line_width = src->line_width;
118 dest_mask |= RADV_DYNAMIC_LINE_WIDTH;
119 }
120 }
121
122 if (copy_mask & RADV_DYNAMIC_DEPTH_BIAS) {
123 if (memcmp(&dest->depth_bias, &src->depth_bias,
124 sizeof(src->depth_bias))) {
125 dest->depth_bias = src->depth_bias;
126 dest_mask |= RADV_DYNAMIC_DEPTH_BIAS;
127 }
128 }
129
130 if (copy_mask & RADV_DYNAMIC_BLEND_CONSTANTS) {
131 if (memcmp(&dest->blend_constants, &src->blend_constants,
132 sizeof(src->blend_constants))) {
133 typed_memcpy(dest->blend_constants,
134 src->blend_constants, 4);
135 dest_mask |= RADV_DYNAMIC_BLEND_CONSTANTS;
136 }
137 }
138
139 if (copy_mask & RADV_DYNAMIC_DEPTH_BOUNDS) {
140 if (memcmp(&dest->depth_bounds, &src->depth_bounds,
141 sizeof(src->depth_bounds))) {
142 dest->depth_bounds = src->depth_bounds;
143 dest_mask |= RADV_DYNAMIC_DEPTH_BOUNDS;
144 }
145 }
146
147 if (copy_mask & RADV_DYNAMIC_STENCIL_COMPARE_MASK) {
148 if (memcmp(&dest->stencil_compare_mask,
149 &src->stencil_compare_mask,
150 sizeof(src->stencil_compare_mask))) {
151 dest->stencil_compare_mask = src->stencil_compare_mask;
152 dest_mask |= RADV_DYNAMIC_STENCIL_COMPARE_MASK;
153 }
154 }
155
156 if (copy_mask & RADV_DYNAMIC_STENCIL_WRITE_MASK) {
157 if (memcmp(&dest->stencil_write_mask, &src->stencil_write_mask,
158 sizeof(src->stencil_write_mask))) {
159 dest->stencil_write_mask = src->stencil_write_mask;
160 dest_mask |= RADV_DYNAMIC_STENCIL_WRITE_MASK;
161 }
162 }
163
164 if (copy_mask & RADV_DYNAMIC_STENCIL_REFERENCE) {
165 if (memcmp(&dest->stencil_reference, &src->stencil_reference,
166 sizeof(src->stencil_reference))) {
167 dest->stencil_reference = src->stencil_reference;
168 dest_mask |= RADV_DYNAMIC_STENCIL_REFERENCE;
169 }
170 }
171
172 if (copy_mask & RADV_DYNAMIC_DISCARD_RECTANGLE) {
173 if (memcmp(&dest->discard_rectangle.rectangles, &src->discard_rectangle.rectangles,
174 src->discard_rectangle.count * sizeof(VkRect2D))) {
175 typed_memcpy(dest->discard_rectangle.rectangles,
176 src->discard_rectangle.rectangles,
177 src->discard_rectangle.count);
178 dest_mask |= RADV_DYNAMIC_DISCARD_RECTANGLE;
179 }
180 }
181
182 cmd_buffer->state.dirty |= dest_mask;
183 }
184
185 bool radv_cmd_buffer_uses_mec(struct radv_cmd_buffer *cmd_buffer)
186 {
187 return cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE &&
188 cmd_buffer->device->physical_device->rad_info.chip_class >= CIK;
189 }
190
191 enum ring_type radv_queue_family_to_ring(int f) {
192 switch (f) {
193 case RADV_QUEUE_GENERAL:
194 return RING_GFX;
195 case RADV_QUEUE_COMPUTE:
196 return RING_COMPUTE;
197 case RADV_QUEUE_TRANSFER:
198 return RING_DMA;
199 default:
200 unreachable("Unknown queue family");
201 }
202 }
203
204 static VkResult radv_create_cmd_buffer(
205 struct radv_device * device,
206 struct radv_cmd_pool * pool,
207 VkCommandBufferLevel level,
208 VkCommandBuffer* pCommandBuffer)
209 {
210 struct radv_cmd_buffer *cmd_buffer;
211 unsigned ring;
212 cmd_buffer = vk_zalloc(&pool->alloc, sizeof(*cmd_buffer), 8,
213 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
214 if (cmd_buffer == NULL)
215 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
216
217 cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
218 cmd_buffer->device = device;
219 cmd_buffer->pool = pool;
220 cmd_buffer->level = level;
221
222 if (pool) {
223 list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
224 cmd_buffer->queue_family_index = pool->queue_family_index;
225
226 } else {
227 /* Init the pool_link so we can safefly call list_del when we destroy
228 * the command buffer
229 */
230 list_inithead(&cmd_buffer->pool_link);
231 cmd_buffer->queue_family_index = RADV_QUEUE_GENERAL;
232 }
233
234 ring = radv_queue_family_to_ring(cmd_buffer->queue_family_index);
235
236 cmd_buffer->cs = device->ws->cs_create(device->ws, ring);
237 if (!cmd_buffer->cs) {
238 vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
239 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
240 }
241
242 *pCommandBuffer = radv_cmd_buffer_to_handle(cmd_buffer);
243
244 list_inithead(&cmd_buffer->upload.list);
245
246 return VK_SUCCESS;
247 }
248
249 static void
250 radv_cmd_buffer_destroy(struct radv_cmd_buffer *cmd_buffer)
251 {
252 list_del(&cmd_buffer->pool_link);
253
254 list_for_each_entry_safe(struct radv_cmd_buffer_upload, up,
255 &cmd_buffer->upload.list, list) {
256 cmd_buffer->device->ws->buffer_destroy(up->upload_bo);
257 list_del(&up->list);
258 free(up);
259 }
260
261 if (cmd_buffer->upload.upload_bo)
262 cmd_buffer->device->ws->buffer_destroy(cmd_buffer->upload.upload_bo);
263 cmd_buffer->device->ws->cs_destroy(cmd_buffer->cs);
264 free(cmd_buffer->push_descriptors.set.mapped_ptr);
265 vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
266 }
267
268 static VkResult
269 radv_reset_cmd_buffer(struct radv_cmd_buffer *cmd_buffer)
270 {
271
272 cmd_buffer->device->ws->cs_reset(cmd_buffer->cs);
273
274 list_for_each_entry_safe(struct radv_cmd_buffer_upload, up,
275 &cmd_buffer->upload.list, list) {
276 cmd_buffer->device->ws->buffer_destroy(up->upload_bo);
277 list_del(&up->list);
278 free(up);
279 }
280
281 cmd_buffer->push_constant_stages = 0;
282 cmd_buffer->scratch_size_needed = 0;
283 cmd_buffer->compute_scratch_size_needed = 0;
284 cmd_buffer->esgs_ring_size_needed = 0;
285 cmd_buffer->gsvs_ring_size_needed = 0;
286 cmd_buffer->tess_rings_needed = false;
287 cmd_buffer->sample_positions_needed = false;
288
289 if (cmd_buffer->upload.upload_bo)
290 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs,
291 cmd_buffer->upload.upload_bo, 8);
292 cmd_buffer->upload.offset = 0;
293
294 cmd_buffer->record_result = VK_SUCCESS;
295
296 cmd_buffer->ring_offsets_idx = -1;
297
298 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
299 void *fence_ptr;
300 radv_cmd_buffer_upload_alloc(cmd_buffer, 8, 0,
301 &cmd_buffer->gfx9_fence_offset,
302 &fence_ptr);
303 cmd_buffer->gfx9_fence_bo = cmd_buffer->upload.upload_bo;
304 }
305
306 cmd_buffer->status = RADV_CMD_BUFFER_STATUS_INITIAL;
307
308 return cmd_buffer->record_result;
309 }
310
311 static bool
312 radv_cmd_buffer_resize_upload_buf(struct radv_cmd_buffer *cmd_buffer,
313 uint64_t min_needed)
314 {
315 uint64_t new_size;
316 struct radeon_winsys_bo *bo;
317 struct radv_cmd_buffer_upload *upload;
318 struct radv_device *device = cmd_buffer->device;
319
320 new_size = MAX2(min_needed, 16 * 1024);
321 new_size = MAX2(new_size, 2 * cmd_buffer->upload.size);
322
323 bo = device->ws->buffer_create(device->ws,
324 new_size, 4096,
325 RADEON_DOMAIN_GTT,
326 RADEON_FLAG_CPU_ACCESS|
327 RADEON_FLAG_NO_INTERPROCESS_SHARING);
328
329 if (!bo) {
330 cmd_buffer->record_result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
331 return false;
332 }
333
334 radv_cs_add_buffer(device->ws, cmd_buffer->cs, bo, 8);
335 if (cmd_buffer->upload.upload_bo) {
336 upload = malloc(sizeof(*upload));
337
338 if (!upload) {
339 cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
340 device->ws->buffer_destroy(bo);
341 return false;
342 }
343
344 memcpy(upload, &cmd_buffer->upload, sizeof(*upload));
345 list_add(&upload->list, &cmd_buffer->upload.list);
346 }
347
348 cmd_buffer->upload.upload_bo = bo;
349 cmd_buffer->upload.size = new_size;
350 cmd_buffer->upload.offset = 0;
351 cmd_buffer->upload.map = device->ws->buffer_map(cmd_buffer->upload.upload_bo);
352
353 if (!cmd_buffer->upload.map) {
354 cmd_buffer->record_result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
355 return false;
356 }
357
358 return true;
359 }
360
361 bool
362 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
363 unsigned size,
364 unsigned alignment,
365 unsigned *out_offset,
366 void **ptr)
367 {
368 uint64_t offset = align(cmd_buffer->upload.offset, alignment);
369 if (offset + size > cmd_buffer->upload.size) {
370 if (!radv_cmd_buffer_resize_upload_buf(cmd_buffer, size))
371 return false;
372 offset = 0;
373 }
374
375 *out_offset = offset;
376 *ptr = cmd_buffer->upload.map + offset;
377
378 cmd_buffer->upload.offset = offset + size;
379 return true;
380 }
381
382 bool
383 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
384 unsigned size, unsigned alignment,
385 const void *data, unsigned *out_offset)
386 {
387 uint8_t *ptr;
388
389 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size, alignment,
390 out_offset, (void **)&ptr))
391 return false;
392
393 if (ptr)
394 memcpy(ptr, data, size);
395
396 return true;
397 }
398
399 static void
400 radv_emit_write_data_packet(struct radeon_winsys_cs *cs, uint64_t va,
401 unsigned count, const uint32_t *data)
402 {
403 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 2 + count, 0));
404 radeon_emit(cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
405 S_370_WR_CONFIRM(1) |
406 S_370_ENGINE_SEL(V_370_ME));
407 radeon_emit(cs, va);
408 radeon_emit(cs, va >> 32);
409 radeon_emit_array(cs, data, count);
410 }
411
412 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer *cmd_buffer)
413 {
414 struct radv_device *device = cmd_buffer->device;
415 struct radeon_winsys_cs *cs = cmd_buffer->cs;
416 uint64_t va;
417
418 va = radv_buffer_get_va(device->trace_bo);
419 if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY)
420 va += 4;
421
422 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 7);
423
424 ++cmd_buffer->state.trace_id;
425 radv_cs_add_buffer(device->ws, cs, device->trace_bo, 8);
426 radv_emit_write_data_packet(cs, va, 1, &cmd_buffer->state.trace_id);
427 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
428 radeon_emit(cs, AC_ENCODE_TRACE_POINT(cmd_buffer->state.trace_id));
429 }
430
431 static void
432 radv_cmd_buffer_after_draw(struct radv_cmd_buffer *cmd_buffer)
433 {
434 if (cmd_buffer->device->instance->debug_flags & RADV_DEBUG_SYNC_SHADERS) {
435 enum radv_cmd_flush_bits flags;
436
437 /* Force wait for graphics/compute engines to be idle. */
438 flags = RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
439 RADV_CMD_FLAG_CS_PARTIAL_FLUSH;
440
441 si_cs_emit_cache_flush(cmd_buffer->cs, false,
442 cmd_buffer->device->physical_device->rad_info.chip_class,
443 NULL, 0,
444 radv_cmd_buffer_uses_mec(cmd_buffer),
445 flags);
446 }
447
448 if (unlikely(cmd_buffer->device->trace_bo))
449 radv_cmd_buffer_trace_emit(cmd_buffer);
450 }
451
452 static void
453 radv_save_pipeline(struct radv_cmd_buffer *cmd_buffer,
454 struct radv_pipeline *pipeline, enum ring_type ring)
455 {
456 struct radv_device *device = cmd_buffer->device;
457 struct radeon_winsys_cs *cs = cmd_buffer->cs;
458 uint32_t data[2];
459 uint64_t va;
460
461 va = radv_buffer_get_va(device->trace_bo);
462
463 switch (ring) {
464 case RING_GFX:
465 va += 8;
466 break;
467 case RING_COMPUTE:
468 va += 16;
469 break;
470 default:
471 assert(!"invalid ring type");
472 }
473
474 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(device->ws,
475 cmd_buffer->cs, 6);
476
477 data[0] = (uintptr_t)pipeline;
478 data[1] = (uintptr_t)pipeline >> 32;
479
480 radv_cs_add_buffer(device->ws, cs, device->trace_bo, 8);
481 radv_emit_write_data_packet(cs, va, 2, data);
482 }
483
484 void radv_set_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
485 struct radv_descriptor_set *set,
486 unsigned idx)
487 {
488 cmd_buffer->descriptors[idx] = set;
489 if (set)
490 cmd_buffer->state.valid_descriptors |= (1u << idx);
491 else
492 cmd_buffer->state.valid_descriptors &= ~(1u << idx);
493 cmd_buffer->state.descriptors_dirty |= (1u << idx);
494
495 }
496
497 static void
498 radv_save_descriptors(struct radv_cmd_buffer *cmd_buffer)
499 {
500 struct radv_device *device = cmd_buffer->device;
501 struct radeon_winsys_cs *cs = cmd_buffer->cs;
502 uint32_t data[MAX_SETS * 2] = {};
503 uint64_t va;
504 unsigned i;
505 va = radv_buffer_get_va(device->trace_bo) + 24;
506
507 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(device->ws,
508 cmd_buffer->cs, 4 + MAX_SETS * 2);
509
510 for_each_bit(i, cmd_buffer->state.valid_descriptors) {
511 struct radv_descriptor_set *set = cmd_buffer->descriptors[i];
512 data[i * 2] = (uintptr_t)set;
513 data[i * 2 + 1] = (uintptr_t)set >> 32;
514 }
515
516 radv_cs_add_buffer(device->ws, cs, device->trace_bo, 8);
517 radv_emit_write_data_packet(cs, va, MAX_SETS * 2, data);
518 }
519
520 static void
521 radv_emit_graphics_blend_state(struct radv_cmd_buffer *cmd_buffer,
522 struct radv_pipeline *pipeline)
523 {
524 radeon_set_context_reg_seq(cmd_buffer->cs, R_028780_CB_BLEND0_CONTROL, 8);
525 radeon_emit_array(cmd_buffer->cs, pipeline->graphics.blend.cb_blend_control,
526 8);
527 radeon_set_context_reg(cmd_buffer->cs, R_028808_CB_COLOR_CONTROL, pipeline->graphics.blend.cb_color_control);
528 radeon_set_context_reg(cmd_buffer->cs, R_028B70_DB_ALPHA_TO_MASK, pipeline->graphics.blend.db_alpha_to_mask);
529
530 if (cmd_buffer->device->physical_device->has_rbplus) {
531
532 radeon_set_context_reg_seq(cmd_buffer->cs, R_028760_SX_MRT0_BLEND_OPT, 8);
533 radeon_emit_array(cmd_buffer->cs, pipeline->graphics.blend.sx_mrt_blend_opt, 8);
534
535 radeon_set_context_reg_seq(cmd_buffer->cs, R_028754_SX_PS_DOWNCONVERT, 3);
536 radeon_emit(cmd_buffer->cs, 0); /* R_028754_SX_PS_DOWNCONVERT */
537 radeon_emit(cmd_buffer->cs, 0); /* R_028758_SX_BLEND_OPT_EPSILON */
538 radeon_emit(cmd_buffer->cs, 0); /* R_02875C_SX_BLEND_OPT_CONTROL */
539 }
540 }
541
542 static void
543 radv_emit_graphics_depth_stencil_state(struct radv_cmd_buffer *cmd_buffer,
544 struct radv_pipeline *pipeline)
545 {
546 struct radv_depth_stencil_state *ds = &pipeline->graphics.ds;
547 radeon_set_context_reg(cmd_buffer->cs, R_028800_DB_DEPTH_CONTROL, ds->db_depth_control);
548 radeon_set_context_reg(cmd_buffer->cs, R_02842C_DB_STENCIL_CONTROL, ds->db_stencil_control);
549
550 radeon_set_context_reg(cmd_buffer->cs, R_028000_DB_RENDER_CONTROL, ds->db_render_control);
551 radeon_set_context_reg(cmd_buffer->cs, R_028010_DB_RENDER_OVERRIDE2, ds->db_render_override2);
552 }
553
554 struct ac_userdata_info *
555 radv_lookup_user_sgpr(struct radv_pipeline *pipeline,
556 gl_shader_stage stage,
557 int idx)
558 {
559 if (stage == MESA_SHADER_VERTEX) {
560 if (pipeline->shaders[MESA_SHADER_VERTEX])
561 return &pipeline->shaders[MESA_SHADER_VERTEX]->info.user_sgprs_locs.shader_data[idx];
562 if (pipeline->shaders[MESA_SHADER_TESS_CTRL])
563 return &pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.user_sgprs_locs.shader_data[idx];
564 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
565 return &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.user_sgprs_locs.shader_data[idx];
566 } else if (stage == MESA_SHADER_TESS_EVAL) {
567 if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
568 return &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.user_sgprs_locs.shader_data[idx];
569 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
570 return &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.user_sgprs_locs.shader_data[idx];
571 }
572 return &pipeline->shaders[stage]->info.user_sgprs_locs.shader_data[idx];
573 }
574
575 static void
576 radv_emit_userdata_address(struct radv_cmd_buffer *cmd_buffer,
577 struct radv_pipeline *pipeline,
578 gl_shader_stage stage,
579 int idx, uint64_t va)
580 {
581 struct ac_userdata_info *loc = radv_lookup_user_sgpr(pipeline, stage, idx);
582 uint32_t base_reg = pipeline->user_data_0[stage];
583 if (loc->sgpr_idx == -1)
584 return;
585 assert(loc->num_sgprs == 2);
586 assert(!loc->indirect);
587 radeon_set_sh_reg_seq(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, 2);
588 radeon_emit(cmd_buffer->cs, va);
589 radeon_emit(cmd_buffer->cs, va >> 32);
590 }
591
592 static void
593 radv_update_multisample_state(struct radv_cmd_buffer *cmd_buffer,
594 struct radv_pipeline *pipeline)
595 {
596 int num_samples = pipeline->graphics.ms.num_samples;
597 struct radv_multisample_state *ms = &pipeline->graphics.ms;
598 struct radv_pipeline *old_pipeline = cmd_buffer->state.emitted_pipeline;
599
600 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
601 radeon_emit(cmd_buffer->cs, ms->pa_sc_aa_mask[0]);
602 radeon_emit(cmd_buffer->cs, ms->pa_sc_aa_mask[1]);
603
604 radeon_set_context_reg(cmd_buffer->cs, R_028804_DB_EQAA, ms->db_eqaa);
605 radeon_set_context_reg(cmd_buffer->cs, R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
606
607 if (old_pipeline && num_samples == old_pipeline->graphics.ms.num_samples &&
608 old_pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.needs_sample_positions == pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.needs_sample_positions)
609 return;
610
611 radeon_set_context_reg_seq(cmd_buffer->cs, R_028BDC_PA_SC_LINE_CNTL, 2);
612 radeon_emit(cmd_buffer->cs, ms->pa_sc_line_cntl);
613 radeon_emit(cmd_buffer->cs, ms->pa_sc_aa_config);
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 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.needs_sample_positions) {
623 uint32_t offset;
624 struct ac_userdata_info *loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_FRAGMENT, AC_UD_PS_SAMPLE_POS_OFFSET);
625 uint32_t base_reg = pipeline->user_data_0[MESA_SHADER_FRAGMENT];
626 if (loc->sgpr_idx == -1)
627 return;
628 assert(loc->num_sgprs == 1);
629 assert(!loc->indirect);
630 switch (num_samples) {
631 default:
632 offset = 0;
633 break;
634 case 2:
635 offset = 1;
636 break;
637 case 4:
638 offset = 3;
639 break;
640 case 8:
641 offset = 7;
642 break;
643 case 16:
644 offset = 15;
645 break;
646 }
647
648 radeon_set_sh_reg(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, offset);
649 cmd_buffer->sample_positions_needed = true;
650 }
651 }
652
653 static void
654 radv_emit_graphics_raster_state(struct radv_cmd_buffer *cmd_buffer,
655 struct radv_pipeline *pipeline)
656 {
657 struct radv_raster_state *raster = &pipeline->graphics.raster;
658
659 radeon_set_context_reg(cmd_buffer->cs, R_028810_PA_CL_CLIP_CNTL,
660 raster->pa_cl_clip_cntl);
661 radeon_set_context_reg(cmd_buffer->cs, R_0286D4_SPI_INTERP_CONTROL_0,
662 raster->spi_interp_control);
663 radeon_set_context_reg(cmd_buffer->cs, R_028BE4_PA_SU_VTX_CNTL,
664 raster->pa_su_vtx_cntl);
665 radeon_set_context_reg(cmd_buffer->cs, R_028814_PA_SU_SC_MODE_CNTL,
666 raster->pa_su_sc_mode_cntl);
667 }
668
669 static inline void
670 radv_emit_prefetch_TC_L2_async(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
671 unsigned size)
672 {
673 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK)
674 si_cp_dma_prefetch(cmd_buffer, va, size);
675 }
676
677 static void
678 radv_emit_VBO_descriptors_prefetch(struct radv_cmd_buffer *cmd_buffer)
679 {
680 if (cmd_buffer->state.vb_prefetch_dirty) {
681 radv_emit_prefetch_TC_L2_async(cmd_buffer,
682 cmd_buffer->state.vb_va,
683 cmd_buffer->state.vb_size);
684 cmd_buffer->state.vb_prefetch_dirty = false;
685 }
686 }
687
688 static void
689 radv_emit_shader_prefetch(struct radv_cmd_buffer *cmd_buffer,
690 struct radv_shader_variant *shader)
691 {
692 struct radeon_winsys *ws = cmd_buffer->device->ws;
693 struct radeon_winsys_cs *cs = cmd_buffer->cs;
694 uint64_t va;
695
696 if (!shader)
697 return;
698
699 va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
700
701 radv_cs_add_buffer(ws, cs, shader->bo, 8);
702 radv_emit_prefetch_TC_L2_async(cmd_buffer, va, shader->code_size);
703 }
704
705 static void
706 radv_emit_prefetch(struct radv_cmd_buffer *cmd_buffer,
707 struct radv_pipeline *pipeline)
708 {
709 radv_emit_shader_prefetch(cmd_buffer,
710 pipeline->shaders[MESA_SHADER_VERTEX]);
711 radv_emit_VBO_descriptors_prefetch(cmd_buffer);
712 radv_emit_shader_prefetch(cmd_buffer,
713 pipeline->shaders[MESA_SHADER_TESS_CTRL]);
714 radv_emit_shader_prefetch(cmd_buffer,
715 pipeline->shaders[MESA_SHADER_TESS_EVAL]);
716 radv_emit_shader_prefetch(cmd_buffer,
717 pipeline->shaders[MESA_SHADER_GEOMETRY]);
718 radv_emit_shader_prefetch(cmd_buffer, pipeline->gs_copy_shader);
719 radv_emit_shader_prefetch(cmd_buffer,
720 pipeline->shaders[MESA_SHADER_FRAGMENT]);
721 }
722
723 static void
724 radv_emit_hw_vs(struct radv_cmd_buffer *cmd_buffer,
725 struct radv_pipeline *pipeline,
726 struct radv_shader_variant *shader)
727 {
728 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
729
730 radeon_set_context_reg(cmd_buffer->cs, R_0286C4_SPI_VS_OUT_CONFIG,
731 pipeline->graphics.vs.spi_vs_out_config);
732
733 radeon_set_context_reg(cmd_buffer->cs, R_02870C_SPI_SHADER_POS_FORMAT,
734 pipeline->graphics.vs.spi_shader_pos_format);
735
736 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B120_SPI_SHADER_PGM_LO_VS, 4);
737 radeon_emit(cmd_buffer->cs, va >> 8);
738 radeon_emit(cmd_buffer->cs, va >> 40);
739 radeon_emit(cmd_buffer->cs, shader->rsrc1);
740 radeon_emit(cmd_buffer->cs, shader->rsrc2);
741
742 radeon_set_context_reg(cmd_buffer->cs, R_028818_PA_CL_VTE_CNTL,
743 S_028818_VTX_W0_FMT(1) |
744 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
745 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
746 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
747
748
749 radeon_set_context_reg(cmd_buffer->cs, R_02881C_PA_CL_VS_OUT_CNTL,
750 pipeline->graphics.vs.pa_cl_vs_out_cntl);
751
752 if (cmd_buffer->device->physical_device->rad_info.chip_class <= VI)
753 radeon_set_context_reg(cmd_buffer->cs, R_028AB4_VGT_REUSE_OFF,
754 pipeline->graphics.vs.vgt_reuse_off);
755 }
756
757 static void
758 radv_emit_hw_es(struct radv_cmd_buffer *cmd_buffer,
759 struct radv_pipeline *pipeline,
760 struct radv_shader_variant *shader)
761 {
762 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
763
764 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B320_SPI_SHADER_PGM_LO_ES, 4);
765 radeon_emit(cmd_buffer->cs, va >> 8);
766 radeon_emit(cmd_buffer->cs, va >> 40);
767 radeon_emit(cmd_buffer->cs, shader->rsrc1);
768 radeon_emit(cmd_buffer->cs, shader->rsrc2);
769 }
770
771 static void
772 radv_emit_hw_ls(struct radv_cmd_buffer *cmd_buffer,
773 struct radv_shader_variant *shader)
774 {
775 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
776 uint32_t rsrc2 = shader->rsrc2;
777
778 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
779 radeon_emit(cmd_buffer->cs, va >> 8);
780 radeon_emit(cmd_buffer->cs, va >> 40);
781
782 rsrc2 |= S_00B52C_LDS_SIZE(cmd_buffer->state.pipeline->graphics.tess.lds_size);
783 if (cmd_buffer->device->physical_device->rad_info.chip_class == CIK &&
784 cmd_buffer->device->physical_device->rad_info.family != CHIP_HAWAII)
785 radeon_set_sh_reg(cmd_buffer->cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, rsrc2);
786
787 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
788 radeon_emit(cmd_buffer->cs, shader->rsrc1);
789 radeon_emit(cmd_buffer->cs, rsrc2);
790 }
791
792 static void
793 radv_emit_hw_hs(struct radv_cmd_buffer *cmd_buffer,
794 struct radv_shader_variant *shader)
795 {
796 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
797
798 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
799 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B410_SPI_SHADER_PGM_LO_LS, 2);
800 radeon_emit(cmd_buffer->cs, va >> 8);
801 radeon_emit(cmd_buffer->cs, va >> 40);
802
803 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B428_SPI_SHADER_PGM_RSRC1_HS, 2);
804 radeon_emit(cmd_buffer->cs, shader->rsrc1);
805 radeon_emit(cmd_buffer->cs, shader->rsrc2 |
806 S_00B42C_LDS_SIZE(cmd_buffer->state.pipeline->graphics.tess.lds_size));
807 } else {
808 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B420_SPI_SHADER_PGM_LO_HS, 4);
809 radeon_emit(cmd_buffer->cs, va >> 8);
810 radeon_emit(cmd_buffer->cs, va >> 40);
811 radeon_emit(cmd_buffer->cs, shader->rsrc1);
812 radeon_emit(cmd_buffer->cs, shader->rsrc2);
813 }
814 }
815
816 static void
817 radv_emit_vertex_shader(struct radv_cmd_buffer *cmd_buffer,
818 struct radv_pipeline *pipeline)
819 {
820 struct radv_shader_variant *vs;
821
822 radeon_set_context_reg(cmd_buffer->cs, R_028A84_VGT_PRIMITIVEID_EN, pipeline->graphics.vgt_primitiveid_en);
823
824 /* Skip shaders merged into HS/GS */
825 vs = pipeline->shaders[MESA_SHADER_VERTEX];
826 if (!vs)
827 return;
828
829 if (vs->info.vs.as_ls)
830 radv_emit_hw_ls(cmd_buffer, vs);
831 else if (vs->info.vs.as_es)
832 radv_emit_hw_es(cmd_buffer, pipeline, vs);
833 else
834 radv_emit_hw_vs(cmd_buffer, pipeline, vs);
835 }
836
837
838 static void
839 radv_emit_tess_shaders(struct radv_cmd_buffer *cmd_buffer,
840 struct radv_pipeline *pipeline)
841 {
842 if (!radv_pipeline_has_tess(pipeline))
843 return;
844
845 struct radv_shader_variant *tes, *tcs;
846
847 tcs = pipeline->shaders[MESA_SHADER_TESS_CTRL];
848 tes = pipeline->shaders[MESA_SHADER_TESS_EVAL];
849
850 if (tes) {
851 if (tes->info.tes.as_es)
852 radv_emit_hw_es(cmd_buffer, pipeline, tes);
853 else
854 radv_emit_hw_vs(cmd_buffer, pipeline, tes);
855 }
856
857 radv_emit_hw_hs(cmd_buffer, tcs);
858
859 radeon_set_context_reg(cmd_buffer->cs, R_028B6C_VGT_TF_PARAM,
860 pipeline->graphics.tess.tf_param);
861
862 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK)
863 radeon_set_context_reg_idx(cmd_buffer->cs, R_028B58_VGT_LS_HS_CONFIG, 2,
864 pipeline->graphics.tess.ls_hs_config);
865 else
866 radeon_set_context_reg(cmd_buffer->cs, R_028B58_VGT_LS_HS_CONFIG,
867 pipeline->graphics.tess.ls_hs_config);
868
869 struct ac_userdata_info *loc;
870
871 loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_TESS_CTRL, AC_UD_TCS_OFFCHIP_LAYOUT);
872 if (loc->sgpr_idx != -1) {
873 uint32_t base_reg = pipeline->user_data_0[MESA_SHADER_TESS_CTRL];
874 assert(loc->num_sgprs == 4);
875 assert(!loc->indirect);
876 radeon_set_sh_reg_seq(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, 4);
877 radeon_emit(cmd_buffer->cs, pipeline->graphics.tess.offchip_layout);
878 radeon_emit(cmd_buffer->cs, pipeline->graphics.tess.tcs_out_offsets);
879 radeon_emit(cmd_buffer->cs, pipeline->graphics.tess.tcs_out_layout |
880 pipeline->graphics.tess.num_tcs_input_cp << 26);
881 radeon_emit(cmd_buffer->cs, pipeline->graphics.tess.tcs_in_layout);
882 }
883
884 loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_TESS_EVAL, AC_UD_TES_OFFCHIP_LAYOUT);
885 if (loc->sgpr_idx != -1) {
886 uint32_t base_reg = pipeline->user_data_0[MESA_SHADER_TESS_EVAL];
887 assert(loc->num_sgprs == 1);
888 assert(!loc->indirect);
889
890 radeon_set_sh_reg(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4,
891 pipeline->graphics.tess.offchip_layout);
892 }
893
894 loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_VERTEX, AC_UD_VS_LS_TCS_IN_LAYOUT);
895 if (loc->sgpr_idx != -1) {
896 uint32_t base_reg = pipeline->user_data_0[MESA_SHADER_VERTEX];
897 assert(loc->num_sgprs == 1);
898 assert(!loc->indirect);
899
900 radeon_set_sh_reg(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4,
901 pipeline->graphics.tess.tcs_in_layout);
902 }
903 }
904
905 static void
906 radv_emit_geometry_shader(struct radv_cmd_buffer *cmd_buffer,
907 struct radv_pipeline *pipeline)
908 {
909 struct radv_shader_variant *gs;
910 uint64_t va;
911
912 radeon_set_context_reg(cmd_buffer->cs, R_028A40_VGT_GS_MODE, pipeline->graphics.vgt_gs_mode);
913
914 gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
915 if (!gs)
916 return;
917
918 uint32_t gsvs_itemsize = gs->info.gs.max_gsvs_emit_size >> 2;
919
920 radeon_set_context_reg_seq(cmd_buffer->cs, R_028A60_VGT_GSVS_RING_OFFSET_1, 3);
921 radeon_emit(cmd_buffer->cs, gsvs_itemsize);
922 radeon_emit(cmd_buffer->cs, gsvs_itemsize);
923 radeon_emit(cmd_buffer->cs, gsvs_itemsize);
924
925 radeon_set_context_reg(cmd_buffer->cs, R_028AB0_VGT_GSVS_RING_ITEMSIZE, gsvs_itemsize);
926
927 radeon_set_context_reg(cmd_buffer->cs, R_028B38_VGT_GS_MAX_VERT_OUT, gs->info.gs.vertices_out);
928
929 uint32_t gs_vert_itemsize = gs->info.gs.gsvs_vertex_size;
930 radeon_set_context_reg_seq(cmd_buffer->cs, R_028B5C_VGT_GS_VERT_ITEMSIZE, 4);
931 radeon_emit(cmd_buffer->cs, gs_vert_itemsize >> 2);
932 radeon_emit(cmd_buffer->cs, 0);
933 radeon_emit(cmd_buffer->cs, 0);
934 radeon_emit(cmd_buffer->cs, 0);
935
936 uint32_t gs_num_invocations = gs->info.gs.invocations;
937 radeon_set_context_reg(cmd_buffer->cs, R_028B90_VGT_GS_INSTANCE_CNT,
938 S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
939 S_028B90_ENABLE(gs_num_invocations > 0));
940
941 radeon_set_context_reg(cmd_buffer->cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
942 pipeline->graphics.gs.vgt_esgs_ring_itemsize);
943
944 va = radv_buffer_get_va(gs->bo) + gs->bo_offset;
945
946 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
947 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B210_SPI_SHADER_PGM_LO_ES, 2);
948 radeon_emit(cmd_buffer->cs, va >> 8);
949 radeon_emit(cmd_buffer->cs, va >> 40);
950
951 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
952 radeon_emit(cmd_buffer->cs, gs->rsrc1);
953 radeon_emit(cmd_buffer->cs, gs->rsrc2 |
954 S_00B22C_LDS_SIZE(pipeline->graphics.gs.lds_size));
955
956 radeon_set_context_reg(cmd_buffer->cs, R_028A44_VGT_GS_ONCHIP_CNTL, pipeline->graphics.gs.vgt_gs_onchip_cntl);
957 radeon_set_context_reg(cmd_buffer->cs, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP, pipeline->graphics.gs.vgt_gs_max_prims_per_subgroup);
958 } else {
959 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B220_SPI_SHADER_PGM_LO_GS, 4);
960 radeon_emit(cmd_buffer->cs, va >> 8);
961 radeon_emit(cmd_buffer->cs, va >> 40);
962 radeon_emit(cmd_buffer->cs, gs->rsrc1);
963 radeon_emit(cmd_buffer->cs, gs->rsrc2);
964 }
965
966 radv_emit_hw_vs(cmd_buffer, pipeline, pipeline->gs_copy_shader);
967
968 struct ac_userdata_info *loc = radv_lookup_user_sgpr(cmd_buffer->state.pipeline, MESA_SHADER_GEOMETRY,
969 AC_UD_GS_VS_RING_STRIDE_ENTRIES);
970 if (loc->sgpr_idx != -1) {
971 uint32_t stride = gs->info.gs.max_gsvs_emit_size;
972 uint32_t num_entries = 64;
973 bool is_vi = cmd_buffer->device->physical_device->rad_info.chip_class >= VI;
974
975 if (is_vi)
976 num_entries *= stride;
977
978 stride = S_008F04_STRIDE(stride);
979 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B230_SPI_SHADER_USER_DATA_GS_0 + loc->sgpr_idx * 4, 2);
980 radeon_emit(cmd_buffer->cs, stride);
981 radeon_emit(cmd_buffer->cs, num_entries);
982 }
983 }
984
985 static void
986 radv_emit_fragment_shader(struct radv_cmd_buffer *cmd_buffer,
987 struct radv_pipeline *pipeline)
988 {
989 struct radv_shader_variant *ps;
990 uint64_t va;
991 unsigned spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
992 struct radv_blend_state *blend = &pipeline->graphics.blend;
993 assert (pipeline->shaders[MESA_SHADER_FRAGMENT]);
994
995 ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
996 va = radv_buffer_get_va(ps->bo) + ps->bo_offset;
997
998 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B020_SPI_SHADER_PGM_LO_PS, 4);
999 radeon_emit(cmd_buffer->cs, va >> 8);
1000 radeon_emit(cmd_buffer->cs, va >> 40);
1001 radeon_emit(cmd_buffer->cs, ps->rsrc1);
1002 radeon_emit(cmd_buffer->cs, ps->rsrc2);
1003
1004 radeon_set_context_reg(cmd_buffer->cs, R_02880C_DB_SHADER_CONTROL,
1005 pipeline->graphics.db_shader_control);
1006
1007 radeon_set_context_reg(cmd_buffer->cs, R_0286CC_SPI_PS_INPUT_ENA,
1008 ps->config.spi_ps_input_ena);
1009
1010 radeon_set_context_reg(cmd_buffer->cs, R_0286D0_SPI_PS_INPUT_ADDR,
1011 ps->config.spi_ps_input_addr);
1012
1013 if (ps->info.info.ps.force_persample)
1014 spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
1015
1016 radeon_set_context_reg(cmd_buffer->cs, R_0286D8_SPI_PS_IN_CONTROL,
1017 S_0286D8_NUM_INTERP(ps->info.fs.num_interp));
1018
1019 radeon_set_context_reg(cmd_buffer->cs, R_0286E0_SPI_BARYC_CNTL, spi_baryc_cntl);
1020
1021 radeon_set_context_reg(cmd_buffer->cs, R_028710_SPI_SHADER_Z_FORMAT,
1022 pipeline->graphics.shader_z_format);
1023
1024 radeon_set_context_reg(cmd_buffer->cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
1025
1026 radeon_set_context_reg(cmd_buffer->cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
1027 radeon_set_context_reg(cmd_buffer->cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
1028
1029 if (cmd_buffer->device->dfsm_allowed) {
1030 /* optimise this? */
1031 radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1032 radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
1033 }
1034
1035 if (pipeline->graphics.ps_input_cntl_num) {
1036 radeon_set_context_reg_seq(cmd_buffer->cs, R_028644_SPI_PS_INPUT_CNTL_0, pipeline->graphics.ps_input_cntl_num);
1037 for (unsigned i = 0; i < pipeline->graphics.ps_input_cntl_num; i++) {
1038 radeon_emit(cmd_buffer->cs, pipeline->graphics.ps_input_cntl[i]);
1039 }
1040 }
1041 }
1042
1043 static void
1044 radv_emit_vgt_vertex_reuse(struct radv_cmd_buffer *cmd_buffer,
1045 struct radv_pipeline *pipeline)
1046 {
1047 struct radeon_winsys_cs *cs = cmd_buffer->cs;
1048
1049 if (cmd_buffer->device->physical_device->rad_info.family < CHIP_POLARIS10)
1050 return;
1051
1052 radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
1053 pipeline->graphics.vtx_reuse_depth);
1054 }
1055
1056 static void
1057 radv_emit_binning_state(struct radv_cmd_buffer *cmd_buffer,
1058 struct radv_pipeline *pipeline)
1059 {
1060 struct radeon_winsys_cs *cs = cmd_buffer->cs;
1061
1062 if (cmd_buffer->device->physical_device->rad_info.chip_class < GFX9)
1063 return;
1064
1065 radeon_set_context_reg(cs, R_028C44_PA_SC_BINNER_CNTL_0,
1066 pipeline->graphics.bin.pa_sc_binner_cntl_0);
1067 radeon_set_context_reg(cs, R_028060_DB_DFSM_CONTROL,
1068 pipeline->graphics.bin.db_dfsm_control);
1069 }
1070
1071 static void
1072 radv_emit_graphics_pipeline(struct radv_cmd_buffer *cmd_buffer)
1073 {
1074 struct radv_pipeline *pipeline = cmd_buffer->state.pipeline;
1075
1076 if (!pipeline || cmd_buffer->state.emitted_pipeline == pipeline)
1077 return;
1078
1079 radv_emit_graphics_depth_stencil_state(cmd_buffer, pipeline);
1080 radv_emit_graphics_blend_state(cmd_buffer, pipeline);
1081 radv_emit_graphics_raster_state(cmd_buffer, pipeline);
1082 radv_update_multisample_state(cmd_buffer, pipeline);
1083 radv_emit_vertex_shader(cmd_buffer, pipeline);
1084 radv_emit_tess_shaders(cmd_buffer, pipeline);
1085 radv_emit_geometry_shader(cmd_buffer, pipeline);
1086 radv_emit_fragment_shader(cmd_buffer, pipeline);
1087 radv_emit_vgt_vertex_reuse(cmd_buffer, pipeline);
1088 radv_emit_binning_state(cmd_buffer, pipeline);
1089
1090 cmd_buffer->scratch_size_needed =
1091 MAX2(cmd_buffer->scratch_size_needed,
1092 pipeline->max_waves * pipeline->scratch_bytes_per_wave);
1093
1094 radeon_set_context_reg(cmd_buffer->cs, R_0286E8_SPI_TMPRING_SIZE,
1095 S_0286E8_WAVES(pipeline->max_waves) |
1096 S_0286E8_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
1097
1098 if (!cmd_buffer->state.emitted_pipeline ||
1099 cmd_buffer->state.emitted_pipeline->graphics.can_use_guardband !=
1100 pipeline->graphics.can_use_guardband)
1101 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_SCISSOR;
1102
1103 radeon_set_context_reg(cmd_buffer->cs, R_028B54_VGT_SHADER_STAGES_EN, pipeline->graphics.vgt_shader_stages_en);
1104
1105 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
1106 radeon_set_uconfig_reg_idx(cmd_buffer->cs, R_030908_VGT_PRIMITIVE_TYPE, 1, pipeline->graphics.prim);
1107 } else {
1108 radeon_set_config_reg(cmd_buffer->cs, R_008958_VGT_PRIMITIVE_TYPE, pipeline->graphics.prim);
1109 }
1110 radeon_set_context_reg(cmd_buffer->cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, pipeline->graphics.gs_out);
1111
1112 radeon_set_context_reg(cmd_buffer->cs, R_02820C_PA_SC_CLIPRECT_RULE, pipeline->graphics.pa_sc_cliprect_rule);
1113
1114 if (unlikely(cmd_buffer->device->trace_bo))
1115 radv_save_pipeline(cmd_buffer, pipeline, RING_GFX);
1116
1117 cmd_buffer->state.emitted_pipeline = pipeline;
1118
1119 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_PIPELINE;
1120 }
1121
1122 static void
1123 radv_emit_viewport(struct radv_cmd_buffer *cmd_buffer)
1124 {
1125 si_write_viewport(cmd_buffer->cs, 0, cmd_buffer->state.dynamic.viewport.count,
1126 cmd_buffer->state.dynamic.viewport.viewports);
1127 }
1128
1129 static void
1130 radv_emit_scissor(struct radv_cmd_buffer *cmd_buffer)
1131 {
1132 uint32_t count = cmd_buffer->state.dynamic.scissor.count;
1133
1134 /* Vega10/Raven scissor bug workaround. This must be done before VPORT
1135 * scissor registers are changed. There is also a more efficient but
1136 * more involved alternative workaround.
1137 */
1138 if (cmd_buffer->device->physical_device->has_scissor_bug) {
1139 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
1140 si_emit_cache_flush(cmd_buffer);
1141 }
1142 si_write_scissors(cmd_buffer->cs, 0, count,
1143 cmd_buffer->state.dynamic.scissor.scissors,
1144 cmd_buffer->state.dynamic.viewport.viewports,
1145 cmd_buffer->state.emitted_pipeline->graphics.can_use_guardband);
1146 radeon_set_context_reg(cmd_buffer->cs, R_028A48_PA_SC_MODE_CNTL_0,
1147 cmd_buffer->state.pipeline->graphics.ms.pa_sc_mode_cntl_0 | S_028A48_VPORT_SCISSOR_ENABLE(count ? 1 : 0));
1148 }
1149
1150 static void
1151 radv_emit_discard_rectangle(struct radv_cmd_buffer *cmd_buffer)
1152 {
1153 if (!cmd_buffer->state.dynamic.discard_rectangle.count)
1154 return;
1155
1156 radeon_set_context_reg_seq(cmd_buffer->cs, R_028210_PA_SC_CLIPRECT_0_TL,
1157 cmd_buffer->state.dynamic.discard_rectangle.count * 2);
1158 for (unsigned i = 0; i < cmd_buffer->state.dynamic.discard_rectangle.count; ++i) {
1159 VkRect2D rect = cmd_buffer->state.dynamic.discard_rectangle.rectangles[i];
1160 radeon_emit(cmd_buffer->cs, S_028210_TL_X(rect.offset.x) | S_028210_TL_Y(rect.offset.y));
1161 radeon_emit(cmd_buffer->cs, S_028214_BR_X(rect.offset.x + rect.extent.width) |
1162 S_028214_BR_Y(rect.offset.y + rect.extent.height));
1163 }
1164 }
1165
1166 static void
1167 radv_emit_line_width(struct radv_cmd_buffer *cmd_buffer)
1168 {
1169 unsigned width = cmd_buffer->state.dynamic.line_width * 8;
1170
1171 radeon_set_context_reg(cmd_buffer->cs, R_028A08_PA_SU_LINE_CNTL,
1172 S_028A08_WIDTH(CLAMP(width, 0, 0xFFF)));
1173 }
1174
1175 static void
1176 radv_emit_blend_constants(struct radv_cmd_buffer *cmd_buffer)
1177 {
1178 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
1179
1180 radeon_set_context_reg_seq(cmd_buffer->cs, R_028414_CB_BLEND_RED, 4);
1181 radeon_emit_array(cmd_buffer->cs, (uint32_t *)d->blend_constants, 4);
1182 }
1183
1184 static void
1185 radv_emit_stencil(struct radv_cmd_buffer *cmd_buffer)
1186 {
1187 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
1188
1189 radeon_set_context_reg_seq(cmd_buffer->cs,
1190 R_028430_DB_STENCILREFMASK, 2);
1191 radeon_emit(cmd_buffer->cs,
1192 S_028430_STENCILTESTVAL(d->stencil_reference.front) |
1193 S_028430_STENCILMASK(d->stencil_compare_mask.front) |
1194 S_028430_STENCILWRITEMASK(d->stencil_write_mask.front) |
1195 S_028430_STENCILOPVAL(1));
1196 radeon_emit(cmd_buffer->cs,
1197 S_028434_STENCILTESTVAL_BF(d->stencil_reference.back) |
1198 S_028434_STENCILMASK_BF(d->stencil_compare_mask.back) |
1199 S_028434_STENCILWRITEMASK_BF(d->stencil_write_mask.back) |
1200 S_028434_STENCILOPVAL_BF(1));
1201 }
1202
1203 static void
1204 radv_emit_depth_bounds(struct radv_cmd_buffer *cmd_buffer)
1205 {
1206 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
1207
1208 radeon_set_context_reg(cmd_buffer->cs, R_028020_DB_DEPTH_BOUNDS_MIN,
1209 fui(d->depth_bounds.min));
1210 radeon_set_context_reg(cmd_buffer->cs, R_028024_DB_DEPTH_BOUNDS_MAX,
1211 fui(d->depth_bounds.max));
1212 }
1213
1214 static void
1215 radv_emit_depth_biais(struct radv_cmd_buffer *cmd_buffer)
1216 {
1217 struct radv_raster_state *raster = &cmd_buffer->state.pipeline->graphics.raster;
1218 struct radv_dynamic_state *d = &cmd_buffer->state.dynamic;
1219 unsigned slope = fui(d->depth_bias.slope * 16.0f);
1220 unsigned bias = fui(d->depth_bias.bias * cmd_buffer->state.offset_scale);
1221
1222 if (G_028814_POLY_OFFSET_FRONT_ENABLE(raster->pa_su_sc_mode_cntl)) {
1223 radeon_set_context_reg_seq(cmd_buffer->cs,
1224 R_028B7C_PA_SU_POLY_OFFSET_CLAMP, 5);
1225 radeon_emit(cmd_buffer->cs, fui(d->depth_bias.clamp)); /* CLAMP */
1226 radeon_emit(cmd_buffer->cs, slope); /* FRONT SCALE */
1227 radeon_emit(cmd_buffer->cs, bias); /* FRONT OFFSET */
1228 radeon_emit(cmd_buffer->cs, slope); /* BACK SCALE */
1229 radeon_emit(cmd_buffer->cs, bias); /* BACK OFFSET */
1230 }
1231 }
1232
1233 static void
1234 radv_emit_fb_color_state(struct radv_cmd_buffer *cmd_buffer,
1235 int index,
1236 struct radv_attachment_info *att,
1237 struct radv_image *image,
1238 VkImageLayout layout)
1239 {
1240 bool is_vi = cmd_buffer->device->physical_device->rad_info.chip_class >= VI;
1241 struct radv_color_buffer_info *cb = &att->cb;
1242 uint32_t cb_color_info = cb->cb_color_info;
1243
1244 if (!radv_layout_dcc_compressed(image, layout,
1245 radv_image_queue_family_mask(image,
1246 cmd_buffer->queue_family_index,
1247 cmd_buffer->queue_family_index))) {
1248 cb_color_info &= C_028C70_DCC_ENABLE;
1249 }
1250
1251 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1252 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C60_CB_COLOR0_BASE + index * 0x3c, 11);
1253 radeon_emit(cmd_buffer->cs, cb->cb_color_base);
1254 radeon_emit(cmd_buffer->cs, cb->cb_color_base >> 32);
1255 radeon_emit(cmd_buffer->cs, cb->cb_color_attrib2);
1256 radeon_emit(cmd_buffer->cs, cb->cb_color_view);
1257 radeon_emit(cmd_buffer->cs, cb_color_info);
1258 radeon_emit(cmd_buffer->cs, cb->cb_color_attrib);
1259 radeon_emit(cmd_buffer->cs, cb->cb_dcc_control);
1260 radeon_emit(cmd_buffer->cs, cb->cb_color_cmask);
1261 radeon_emit(cmd_buffer->cs, cb->cb_color_cmask >> 32);
1262 radeon_emit(cmd_buffer->cs, cb->cb_color_fmask);
1263 radeon_emit(cmd_buffer->cs, cb->cb_color_fmask >> 32);
1264
1265 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C94_CB_COLOR0_DCC_BASE + index * 0x3c, 2);
1266 radeon_emit(cmd_buffer->cs, cb->cb_dcc_base);
1267 radeon_emit(cmd_buffer->cs, cb->cb_dcc_base >> 32);
1268
1269 radeon_set_context_reg(cmd_buffer->cs, R_0287A0_CB_MRT0_EPITCH + index * 4,
1270 S_0287A0_EPITCH(att->attachment->image->surface.u.gfx9.surf.epitch));
1271 } else {
1272 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C60_CB_COLOR0_BASE + index * 0x3c, 11);
1273 radeon_emit(cmd_buffer->cs, cb->cb_color_base);
1274 radeon_emit(cmd_buffer->cs, cb->cb_color_pitch);
1275 radeon_emit(cmd_buffer->cs, cb->cb_color_slice);
1276 radeon_emit(cmd_buffer->cs, cb->cb_color_view);
1277 radeon_emit(cmd_buffer->cs, cb_color_info);
1278 radeon_emit(cmd_buffer->cs, cb->cb_color_attrib);
1279 radeon_emit(cmd_buffer->cs, cb->cb_dcc_control);
1280 radeon_emit(cmd_buffer->cs, cb->cb_color_cmask);
1281 radeon_emit(cmd_buffer->cs, cb->cb_color_cmask_slice);
1282 radeon_emit(cmd_buffer->cs, cb->cb_color_fmask);
1283 radeon_emit(cmd_buffer->cs, cb->cb_color_fmask_slice);
1284
1285 if (is_vi) { /* DCC BASE */
1286 radeon_set_context_reg(cmd_buffer->cs, R_028C94_CB_COLOR0_DCC_BASE + index * 0x3c, cb->cb_dcc_base);
1287 }
1288 }
1289 }
1290
1291 static void
1292 radv_emit_fb_ds_state(struct radv_cmd_buffer *cmd_buffer,
1293 struct radv_ds_buffer_info *ds,
1294 struct radv_image *image,
1295 VkImageLayout layout)
1296 {
1297 uint32_t db_z_info = ds->db_z_info;
1298 uint32_t db_stencil_info = ds->db_stencil_info;
1299
1300 if (!radv_layout_has_htile(image, layout,
1301 radv_image_queue_family_mask(image,
1302 cmd_buffer->queue_family_index,
1303 cmd_buffer->queue_family_index))) {
1304 db_z_info &= C_028040_TILE_SURFACE_ENABLE;
1305 db_stencil_info |= S_028044_TILE_STENCIL_DISABLE(1);
1306 }
1307
1308 radeon_set_context_reg(cmd_buffer->cs, R_028008_DB_DEPTH_VIEW, ds->db_depth_view);
1309 radeon_set_context_reg(cmd_buffer->cs, R_028ABC_DB_HTILE_SURFACE, ds->db_htile_surface);
1310
1311
1312 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1313 radeon_set_context_reg_seq(cmd_buffer->cs, R_028014_DB_HTILE_DATA_BASE, 3);
1314 radeon_emit(cmd_buffer->cs, ds->db_htile_data_base);
1315 radeon_emit(cmd_buffer->cs, ds->db_htile_data_base >> 32);
1316 radeon_emit(cmd_buffer->cs, ds->db_depth_size);
1317
1318 radeon_set_context_reg_seq(cmd_buffer->cs, R_028038_DB_Z_INFO, 10);
1319 radeon_emit(cmd_buffer->cs, db_z_info); /* DB_Z_INFO */
1320 radeon_emit(cmd_buffer->cs, db_stencil_info); /* DB_STENCIL_INFO */
1321 radeon_emit(cmd_buffer->cs, ds->db_z_read_base); /* DB_Z_READ_BASE */
1322 radeon_emit(cmd_buffer->cs, ds->db_z_read_base >> 32); /* DB_Z_READ_BASE_HI */
1323 radeon_emit(cmd_buffer->cs, ds->db_stencil_read_base); /* DB_STENCIL_READ_BASE */
1324 radeon_emit(cmd_buffer->cs, ds->db_stencil_read_base >> 32); /* DB_STENCIL_READ_BASE_HI */
1325 radeon_emit(cmd_buffer->cs, ds->db_z_write_base); /* DB_Z_WRITE_BASE */
1326 radeon_emit(cmd_buffer->cs, ds->db_z_write_base >> 32); /* DB_Z_WRITE_BASE_HI */
1327 radeon_emit(cmd_buffer->cs, ds->db_stencil_write_base); /* DB_STENCIL_WRITE_BASE */
1328 radeon_emit(cmd_buffer->cs, ds->db_stencil_write_base >> 32); /* DB_STENCIL_WRITE_BASE_HI */
1329
1330 radeon_set_context_reg_seq(cmd_buffer->cs, R_028068_DB_Z_INFO2, 2);
1331 radeon_emit(cmd_buffer->cs, ds->db_z_info2);
1332 radeon_emit(cmd_buffer->cs, ds->db_stencil_info2);
1333 } else {
1334 radeon_set_context_reg(cmd_buffer->cs, R_028014_DB_HTILE_DATA_BASE, ds->db_htile_data_base);
1335
1336 radeon_set_context_reg_seq(cmd_buffer->cs, R_02803C_DB_DEPTH_INFO, 9);
1337 radeon_emit(cmd_buffer->cs, ds->db_depth_info); /* R_02803C_DB_DEPTH_INFO */
1338 radeon_emit(cmd_buffer->cs, db_z_info); /* R_028040_DB_Z_INFO */
1339 radeon_emit(cmd_buffer->cs, db_stencil_info); /* R_028044_DB_STENCIL_INFO */
1340 radeon_emit(cmd_buffer->cs, ds->db_z_read_base); /* R_028048_DB_Z_READ_BASE */
1341 radeon_emit(cmd_buffer->cs, ds->db_stencil_read_base); /* R_02804C_DB_STENCIL_READ_BASE */
1342 radeon_emit(cmd_buffer->cs, ds->db_z_write_base); /* R_028050_DB_Z_WRITE_BASE */
1343 radeon_emit(cmd_buffer->cs, ds->db_stencil_write_base); /* R_028054_DB_STENCIL_WRITE_BASE */
1344 radeon_emit(cmd_buffer->cs, ds->db_depth_size); /* R_028058_DB_DEPTH_SIZE */
1345 radeon_emit(cmd_buffer->cs, ds->db_depth_slice); /* R_02805C_DB_DEPTH_SLICE */
1346
1347 }
1348
1349 radeon_set_context_reg(cmd_buffer->cs, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1350 ds->pa_su_poly_offset_db_fmt_cntl);
1351 }
1352
1353 void
1354 radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1355 struct radv_image *image,
1356 VkClearDepthStencilValue ds_clear_value,
1357 VkImageAspectFlags aspects)
1358 {
1359 uint64_t va = radv_buffer_get_va(image->bo);
1360 va += image->offset + image->clear_value_offset;
1361 unsigned reg_offset = 0, reg_count = 0;
1362
1363 assert(image->surface.htile_size);
1364
1365 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
1366 ++reg_count;
1367 } else {
1368 ++reg_offset;
1369 va += 4;
1370 }
1371 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
1372 ++reg_count;
1373
1374 radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 2 + reg_count, 0));
1375 radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
1376 S_370_WR_CONFIRM(1) |
1377 S_370_ENGINE_SEL(V_370_PFP));
1378 radeon_emit(cmd_buffer->cs, va);
1379 radeon_emit(cmd_buffer->cs, va >> 32);
1380 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT)
1381 radeon_emit(cmd_buffer->cs, ds_clear_value.stencil);
1382 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
1383 radeon_emit(cmd_buffer->cs, fui(ds_clear_value.depth));
1384
1385 radeon_set_context_reg_seq(cmd_buffer->cs, R_028028_DB_STENCIL_CLEAR + 4 * reg_offset, reg_count);
1386 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT)
1387 radeon_emit(cmd_buffer->cs, ds_clear_value.stencil); /* R_028028_DB_STENCIL_CLEAR */
1388 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
1389 radeon_emit(cmd_buffer->cs, fui(ds_clear_value.depth)); /* R_02802C_DB_DEPTH_CLEAR */
1390 }
1391
1392 static void
1393 radv_load_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1394 struct radv_image *image)
1395 {
1396 VkImageAspectFlags aspects = vk_format_aspects(image->vk_format);
1397 uint64_t va = radv_buffer_get_va(image->bo);
1398 va += image->offset + image->clear_value_offset;
1399 unsigned reg_offset = 0, reg_count = 0;
1400
1401 if (!image->surface.htile_size)
1402 return;
1403
1404 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
1405 ++reg_count;
1406 } else {
1407 ++reg_offset;
1408 va += 4;
1409 }
1410 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
1411 ++reg_count;
1412
1413 radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, 0));
1414 radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
1415 COPY_DATA_DST_SEL(COPY_DATA_REG) |
1416 (reg_count == 2 ? COPY_DATA_COUNT_SEL : 0));
1417 radeon_emit(cmd_buffer->cs, va);
1418 radeon_emit(cmd_buffer->cs, va >> 32);
1419 radeon_emit(cmd_buffer->cs, (R_028028_DB_STENCIL_CLEAR + 4 * reg_offset) >> 2);
1420 radeon_emit(cmd_buffer->cs, 0);
1421
1422 radeon_emit(cmd_buffer->cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
1423 radeon_emit(cmd_buffer->cs, 0);
1424 }
1425
1426 /*
1427 *with DCC some colors don't require CMASK elimiation before being
1428 * used as a texture. This sets a predicate value to determine if the
1429 * cmask eliminate is required.
1430 */
1431 void
1432 radv_set_dcc_need_cmask_elim_pred(struct radv_cmd_buffer *cmd_buffer,
1433 struct radv_image *image,
1434 bool value)
1435 {
1436 uint64_t pred_val = value;
1437 uint64_t va = radv_buffer_get_va(image->bo);
1438 va += image->offset + image->dcc_pred_offset;
1439
1440 assert(image->surface.dcc_size);
1441
1442 radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 4, 0));
1443 radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
1444 S_370_WR_CONFIRM(1) |
1445 S_370_ENGINE_SEL(V_370_PFP));
1446 radeon_emit(cmd_buffer->cs, va);
1447 radeon_emit(cmd_buffer->cs, va >> 32);
1448 radeon_emit(cmd_buffer->cs, pred_val);
1449 radeon_emit(cmd_buffer->cs, pred_val >> 32);
1450 }
1451
1452 void
1453 radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1454 struct radv_image *image,
1455 int idx,
1456 uint32_t color_values[2])
1457 {
1458 uint64_t va = radv_buffer_get_va(image->bo);
1459 va += image->offset + image->clear_value_offset;
1460
1461 assert(image->cmask.size || image->surface.dcc_size);
1462
1463 radeon_emit(cmd_buffer->cs, PKT3(PKT3_WRITE_DATA, 4, 0));
1464 radeon_emit(cmd_buffer->cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
1465 S_370_WR_CONFIRM(1) |
1466 S_370_ENGINE_SEL(V_370_PFP));
1467 radeon_emit(cmd_buffer->cs, va);
1468 radeon_emit(cmd_buffer->cs, va >> 32);
1469 radeon_emit(cmd_buffer->cs, color_values[0]);
1470 radeon_emit(cmd_buffer->cs, color_values[1]);
1471
1472 radeon_set_context_reg_seq(cmd_buffer->cs, R_028C8C_CB_COLOR0_CLEAR_WORD0 + idx * 0x3c, 2);
1473 radeon_emit(cmd_buffer->cs, color_values[0]);
1474 radeon_emit(cmd_buffer->cs, color_values[1]);
1475 }
1476
1477 static void
1478 radv_load_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1479 struct radv_image *image,
1480 int idx)
1481 {
1482 uint64_t va = radv_buffer_get_va(image->bo);
1483 va += image->offset + image->clear_value_offset;
1484
1485 if (!image->cmask.size && !image->surface.dcc_size)
1486 return;
1487
1488 uint32_t reg = R_028C8C_CB_COLOR0_CLEAR_WORD0 + idx * 0x3c;
1489
1490 radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, cmd_buffer->state.predicating));
1491 radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
1492 COPY_DATA_DST_SEL(COPY_DATA_REG) |
1493 COPY_DATA_COUNT_SEL);
1494 radeon_emit(cmd_buffer->cs, va);
1495 radeon_emit(cmd_buffer->cs, va >> 32);
1496 radeon_emit(cmd_buffer->cs, reg >> 2);
1497 radeon_emit(cmd_buffer->cs, 0);
1498
1499 radeon_emit(cmd_buffer->cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
1500 radeon_emit(cmd_buffer->cs, 0);
1501 }
1502
1503 static void
1504 radv_emit_framebuffer_state(struct radv_cmd_buffer *cmd_buffer)
1505 {
1506 int i;
1507 struct radv_framebuffer *framebuffer = cmd_buffer->state.framebuffer;
1508 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
1509
1510 /* this may happen for inherited secondary recording */
1511 if (!framebuffer)
1512 return;
1513
1514 for (i = 0; i < 8; ++i) {
1515 if (i >= subpass->color_count || subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED) {
1516 radeon_set_context_reg(cmd_buffer->cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
1517 S_028C70_FORMAT(V_028C70_COLOR_INVALID));
1518 continue;
1519 }
1520
1521 int idx = subpass->color_attachments[i].attachment;
1522 struct radv_attachment_info *att = &framebuffer->attachments[idx];
1523 struct radv_image *image = att->attachment->image;
1524 VkImageLayout layout = subpass->color_attachments[i].layout;
1525
1526 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, att->attachment->bo, 8);
1527
1528 assert(att->attachment->aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT);
1529 radv_emit_fb_color_state(cmd_buffer, i, att, image, layout);
1530
1531 radv_load_color_clear_regs(cmd_buffer, image, i);
1532 }
1533
1534 if(subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
1535 int idx = subpass->depth_stencil_attachment.attachment;
1536 VkImageLayout layout = subpass->depth_stencil_attachment.layout;
1537 struct radv_attachment_info *att = &framebuffer->attachments[idx];
1538 struct radv_image *image = att->attachment->image;
1539 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, att->attachment->bo, 8);
1540 MAYBE_UNUSED uint32_t queue_mask = radv_image_queue_family_mask(image,
1541 cmd_buffer->queue_family_index,
1542 cmd_buffer->queue_family_index);
1543 /* We currently don't support writing decompressed HTILE */
1544 assert(radv_layout_has_htile(image, layout, queue_mask) ==
1545 radv_layout_is_htile_compressed(image, layout, queue_mask));
1546
1547 radv_emit_fb_ds_state(cmd_buffer, &att->ds, image, layout);
1548
1549 if (att->ds.offset_scale != cmd_buffer->state.offset_scale) {
1550 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
1551 cmd_buffer->state.offset_scale = att->ds.offset_scale;
1552 }
1553 radv_load_depth_clear_regs(cmd_buffer, image);
1554 } else {
1555 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1556 radeon_set_context_reg_seq(cmd_buffer->cs, R_028038_DB_Z_INFO, 2);
1557 else
1558 radeon_set_context_reg_seq(cmd_buffer->cs, R_028040_DB_Z_INFO, 2);
1559
1560 radeon_emit(cmd_buffer->cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* DB_Z_INFO */
1561 radeon_emit(cmd_buffer->cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* DB_STENCIL_INFO */
1562 }
1563 radeon_set_context_reg(cmd_buffer->cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
1564 S_028208_BR_X(framebuffer->width) |
1565 S_028208_BR_Y(framebuffer->height));
1566
1567 if (cmd_buffer->device->dfsm_allowed) {
1568 radeon_emit(cmd_buffer->cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1569 radeon_emit(cmd_buffer->cs, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
1570 }
1571
1572 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_FRAMEBUFFER;
1573 }
1574
1575 static void
1576 radv_emit_index_buffer(struct radv_cmd_buffer *cmd_buffer)
1577 {
1578 struct radeon_winsys_cs *cs = cmd_buffer->cs;
1579 struct radv_cmd_state *state = &cmd_buffer->state;
1580
1581 if (state->index_type != state->last_index_type) {
1582 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1583 radeon_set_uconfig_reg_idx(cs, R_03090C_VGT_INDEX_TYPE,
1584 2, state->index_type);
1585 } else {
1586 radeon_emit(cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
1587 radeon_emit(cs, state->index_type);
1588 }
1589
1590 state->last_index_type = state->index_type;
1591 }
1592
1593 radeon_emit(cs, PKT3(PKT3_INDEX_BASE, 1, 0));
1594 radeon_emit(cs, state->index_va);
1595 radeon_emit(cs, state->index_va >> 32);
1596
1597 radeon_emit(cs, PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
1598 radeon_emit(cs, state->max_index_count);
1599
1600 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_INDEX_BUFFER;
1601 }
1602
1603 void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer)
1604 {
1605 uint32_t db_count_control;
1606
1607 if(!cmd_buffer->state.active_occlusion_queries) {
1608 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
1609 db_count_control = 0;
1610 } else {
1611 db_count_control = S_028004_ZPASS_INCREMENT_DISABLE(1);
1612 }
1613 } else {
1614 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
1615 db_count_control = S_028004_PERFECT_ZPASS_COUNTS(1) |
1616 S_028004_SAMPLE_RATE(0) | /* TODO: set this to the number of samples of the current framebuffer */
1617 S_028004_ZPASS_ENABLE(1) |
1618 S_028004_SLICE_EVEN_ENABLE(1) |
1619 S_028004_SLICE_ODD_ENABLE(1);
1620 } else {
1621 db_count_control = S_028004_PERFECT_ZPASS_COUNTS(1) |
1622 S_028004_SAMPLE_RATE(0); /* TODO: set this to the number of samples of the current framebuffer */
1623 }
1624 }
1625
1626 radeon_set_context_reg(cmd_buffer->cs, R_028004_DB_COUNT_CONTROL, db_count_control);
1627 }
1628
1629 static void
1630 radv_cmd_buffer_flush_dynamic_state(struct radv_cmd_buffer *cmd_buffer)
1631 {
1632 if (G_028810_DX_RASTERIZATION_KILL(cmd_buffer->state.pipeline->graphics.raster.pa_cl_clip_cntl))
1633 return;
1634
1635 if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT))
1636 radv_emit_viewport(cmd_buffer);
1637
1638 if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_DYNAMIC_SCISSOR | RADV_CMD_DIRTY_DYNAMIC_VIEWPORT))
1639 radv_emit_scissor(cmd_buffer);
1640
1641 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH)
1642 radv_emit_line_width(cmd_buffer);
1643
1644 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS)
1645 radv_emit_blend_constants(cmd_buffer);
1646
1647 if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE |
1648 RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK |
1649 RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK))
1650 radv_emit_stencil(cmd_buffer);
1651
1652 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS)
1653 radv_emit_depth_bounds(cmd_buffer);
1654
1655 if (cmd_buffer->state.dirty & (RADV_CMD_DIRTY_PIPELINE |
1656 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS))
1657 radv_emit_depth_biais(cmd_buffer);
1658
1659 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE)
1660 radv_emit_discard_rectangle(cmd_buffer);
1661
1662 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_DYNAMIC_ALL;
1663 }
1664
1665 static void
1666 emit_stage_descriptor_set_userdata(struct radv_cmd_buffer *cmd_buffer,
1667 struct radv_pipeline *pipeline,
1668 int idx,
1669 uint64_t va,
1670 gl_shader_stage stage)
1671 {
1672 struct ac_userdata_info *desc_set_loc = &pipeline->shaders[stage]->info.user_sgprs_locs.descriptor_sets[idx];
1673 uint32_t base_reg = pipeline->user_data_0[stage];
1674
1675 if (desc_set_loc->sgpr_idx == -1 || desc_set_loc->indirect)
1676 return;
1677
1678 assert(!desc_set_loc->indirect);
1679 assert(desc_set_loc->num_sgprs == 2);
1680 radeon_set_sh_reg_seq(cmd_buffer->cs,
1681 base_reg + desc_set_loc->sgpr_idx * 4, 2);
1682 radeon_emit(cmd_buffer->cs, va);
1683 radeon_emit(cmd_buffer->cs, va >> 32);
1684 }
1685
1686 static void
1687 radv_emit_descriptor_set_userdata(struct radv_cmd_buffer *cmd_buffer,
1688 VkShaderStageFlags stages,
1689 struct radv_descriptor_set *set,
1690 unsigned idx)
1691 {
1692 if (cmd_buffer->state.pipeline) {
1693 radv_foreach_stage(stage, stages) {
1694 if (cmd_buffer->state.pipeline->shaders[stage])
1695 emit_stage_descriptor_set_userdata(cmd_buffer, cmd_buffer->state.pipeline,
1696 idx, set->va,
1697 stage);
1698 }
1699 }
1700
1701 if (cmd_buffer->state.compute_pipeline && (stages & VK_SHADER_STAGE_COMPUTE_BIT))
1702 emit_stage_descriptor_set_userdata(cmd_buffer, cmd_buffer->state.compute_pipeline,
1703 idx, set->va,
1704 MESA_SHADER_COMPUTE);
1705 }
1706
1707 static void
1708 radv_flush_push_descriptors(struct radv_cmd_buffer *cmd_buffer)
1709 {
1710 struct radv_descriptor_set *set = &cmd_buffer->push_descriptors.set;
1711 unsigned bo_offset;
1712
1713 if (!radv_cmd_buffer_upload_data(cmd_buffer, set->size, 32,
1714 set->mapped_ptr,
1715 &bo_offset))
1716 return;
1717
1718 set->va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1719 set->va += bo_offset;
1720 }
1721
1722 static void
1723 radv_flush_indirect_descriptor_sets(struct radv_cmd_buffer *cmd_buffer)
1724 {
1725 uint32_t size = MAX_SETS * 2 * 4;
1726 uint32_t offset;
1727 void *ptr;
1728
1729 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, size,
1730 256, &offset, &ptr))
1731 return;
1732
1733 for (unsigned i = 0; i < MAX_SETS; i++) {
1734 uint32_t *uptr = ((uint32_t *)ptr) + i * 2;
1735 uint64_t set_va = 0;
1736 struct radv_descriptor_set *set = cmd_buffer->descriptors[i];
1737 if (cmd_buffer->state.valid_descriptors & (1u << i))
1738 set_va = set->va;
1739 uptr[0] = set_va & 0xffffffff;
1740 uptr[1] = set_va >> 32;
1741 }
1742
1743 uint64_t va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1744 va += offset;
1745
1746 if (cmd_buffer->state.pipeline) {
1747 if (cmd_buffer->state.pipeline->shaders[MESA_SHADER_VERTEX])
1748 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_VERTEX,
1749 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1750
1751 if (cmd_buffer->state.pipeline->shaders[MESA_SHADER_FRAGMENT])
1752 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_FRAGMENT,
1753 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1754
1755 if (radv_pipeline_has_gs(cmd_buffer->state.pipeline))
1756 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_GEOMETRY,
1757 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1758
1759 if (radv_pipeline_has_tess(cmd_buffer->state.pipeline))
1760 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_TESS_CTRL,
1761 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1762
1763 if (radv_pipeline_has_tess(cmd_buffer->state.pipeline))
1764 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_TESS_EVAL,
1765 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1766 }
1767
1768 if (cmd_buffer->state.compute_pipeline)
1769 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.compute_pipeline, MESA_SHADER_COMPUTE,
1770 AC_UD_INDIRECT_DESCRIPTOR_SETS, va);
1771 }
1772
1773 static void
1774 radv_flush_descriptors(struct radv_cmd_buffer *cmd_buffer,
1775 VkShaderStageFlags stages)
1776 {
1777 unsigned i;
1778
1779 if (!cmd_buffer->state.descriptors_dirty)
1780 return;
1781
1782 if (cmd_buffer->state.push_descriptors_dirty)
1783 radv_flush_push_descriptors(cmd_buffer);
1784
1785 if ((cmd_buffer->state.pipeline && cmd_buffer->state.pipeline->need_indirect_descriptor_sets) ||
1786 (cmd_buffer->state.compute_pipeline && cmd_buffer->state.compute_pipeline->need_indirect_descriptor_sets)) {
1787 radv_flush_indirect_descriptor_sets(cmd_buffer);
1788 }
1789
1790 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
1791 cmd_buffer->cs,
1792 MAX_SETS * MESA_SHADER_STAGES * 4);
1793
1794 for_each_bit(i, cmd_buffer->state.descriptors_dirty) {
1795 struct radv_descriptor_set *set = cmd_buffer->descriptors[i];
1796 if (!(cmd_buffer->state.valid_descriptors & (1u << i)))
1797 continue;
1798
1799 radv_emit_descriptor_set_userdata(cmd_buffer, stages, set, i);
1800 }
1801 cmd_buffer->state.descriptors_dirty = 0;
1802 cmd_buffer->state.push_descriptors_dirty = false;
1803
1804 if (unlikely(cmd_buffer->device->trace_bo))
1805 radv_save_descriptors(cmd_buffer);
1806
1807 assert(cmd_buffer->cs->cdw <= cdw_max);
1808 }
1809
1810 static void
1811 radv_flush_constants(struct radv_cmd_buffer *cmd_buffer,
1812 struct radv_pipeline *pipeline,
1813 VkShaderStageFlags stages)
1814 {
1815 struct radv_pipeline_layout *layout = pipeline->layout;
1816 unsigned offset;
1817 void *ptr;
1818 uint64_t va;
1819
1820 stages &= cmd_buffer->push_constant_stages;
1821 if (!stages ||
1822 (!layout->push_constant_size && !layout->dynamic_offset_count))
1823 return;
1824
1825 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, layout->push_constant_size +
1826 16 * layout->dynamic_offset_count,
1827 256, &offset, &ptr))
1828 return;
1829
1830 memcpy(ptr, cmd_buffer->push_constants, layout->push_constant_size);
1831 memcpy((char*)ptr + layout->push_constant_size, cmd_buffer->dynamic_buffers,
1832 16 * layout->dynamic_offset_count);
1833
1834 va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1835 va += offset;
1836
1837 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
1838 cmd_buffer->cs, MESA_SHADER_STAGES * 4);
1839
1840 radv_foreach_stage(stage, stages) {
1841 if (pipeline->shaders[stage]) {
1842 radv_emit_userdata_address(cmd_buffer, pipeline, stage,
1843 AC_UD_PUSH_CONSTANTS, va);
1844 }
1845 }
1846
1847 cmd_buffer->push_constant_stages &= ~stages;
1848 assert(cmd_buffer->cs->cdw <= cdw_max);
1849 }
1850
1851 static bool
1852 radv_cmd_buffer_update_vertex_descriptors(struct radv_cmd_buffer *cmd_buffer, bool pipeline_is_dirty)
1853 {
1854 if ((pipeline_is_dirty ||
1855 (cmd_buffer->state.dirty & RADV_CMD_DIRTY_VERTEX_BUFFER)) &&
1856 cmd_buffer->state.pipeline->vertex_elements.count &&
1857 radv_get_vertex_shader(cmd_buffer->state.pipeline)->info.info.vs.has_vertex_buffers) {
1858 struct radv_vertex_elements_info *velems = &cmd_buffer->state.pipeline->vertex_elements;
1859 unsigned vb_offset;
1860 void *vb_ptr;
1861 uint32_t i = 0;
1862 uint32_t count = velems->count;
1863 uint64_t va;
1864
1865 /* allocate some descriptor state for vertex buffers */
1866 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, count * 16, 256,
1867 &vb_offset, &vb_ptr))
1868 return false;
1869
1870 for (i = 0; i < count; i++) {
1871 uint32_t *desc = &((uint32_t *)vb_ptr)[i * 4];
1872 uint32_t offset;
1873 int vb = velems->binding[i];
1874 struct radv_buffer *buffer = cmd_buffer->vertex_bindings[vb].buffer;
1875 uint32_t stride = cmd_buffer->state.pipeline->binding_stride[vb];
1876
1877 va = radv_buffer_get_va(buffer->bo);
1878
1879 offset = cmd_buffer->vertex_bindings[vb].offset + velems->offset[i];
1880 va += offset + buffer->offset;
1881 desc[0] = va;
1882 desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) | S_008F04_STRIDE(stride);
1883 if (cmd_buffer->device->physical_device->rad_info.chip_class <= CIK && stride)
1884 desc[2] = (buffer->size - offset - velems->format_size[i]) / stride + 1;
1885 else
1886 desc[2] = buffer->size - offset;
1887 desc[3] = velems->rsrc_word3[i];
1888 }
1889
1890 va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1891 va += vb_offset;
1892
1893 radv_emit_userdata_address(cmd_buffer, cmd_buffer->state.pipeline, MESA_SHADER_VERTEX,
1894 AC_UD_VS_VERTEX_BUFFERS, va);
1895
1896 cmd_buffer->state.vb_va = va;
1897 cmd_buffer->state.vb_size = count * 16;
1898 cmd_buffer->state.vb_prefetch_dirty = true;
1899 }
1900 cmd_buffer->state.dirty &= ~RADV_CMD_DIRTY_VERTEX_BUFFER;
1901
1902 return true;
1903 }
1904
1905 static bool
1906 radv_upload_graphics_shader_descriptors(struct radv_cmd_buffer *cmd_buffer, bool pipeline_is_dirty)
1907 {
1908 if (!radv_cmd_buffer_update_vertex_descriptors(cmd_buffer, pipeline_is_dirty))
1909 return false;
1910
1911 radv_flush_descriptors(cmd_buffer, VK_SHADER_STAGE_ALL_GRAPHICS);
1912 radv_flush_constants(cmd_buffer, cmd_buffer->state.pipeline,
1913 VK_SHADER_STAGE_ALL_GRAPHICS);
1914
1915 return true;
1916 }
1917
1918 static void
1919 radv_emit_draw_registers(struct radv_cmd_buffer *cmd_buffer, bool indexed_draw,
1920 bool instanced_draw, bool indirect_draw,
1921 uint32_t draw_vertex_count)
1922 {
1923 struct radeon_info *info = &cmd_buffer->device->physical_device->rad_info;
1924 struct radv_cmd_state *state = &cmd_buffer->state;
1925 struct radeon_winsys_cs *cs = cmd_buffer->cs;
1926 uint32_t ia_multi_vgt_param;
1927 int32_t primitive_reset_en;
1928
1929 /* Draw state. */
1930 ia_multi_vgt_param =
1931 si_get_ia_multi_vgt_param(cmd_buffer, instanced_draw,
1932 indirect_draw, draw_vertex_count);
1933
1934 if (state->last_ia_multi_vgt_param != ia_multi_vgt_param) {
1935 if (info->chip_class >= GFX9) {
1936 radeon_set_uconfig_reg_idx(cs,
1937 R_030960_IA_MULTI_VGT_PARAM,
1938 4, ia_multi_vgt_param);
1939 } else if (info->chip_class >= CIK) {
1940 radeon_set_context_reg_idx(cs,
1941 R_028AA8_IA_MULTI_VGT_PARAM,
1942 1, ia_multi_vgt_param);
1943 } else {
1944 radeon_set_context_reg(cs, R_028AA8_IA_MULTI_VGT_PARAM,
1945 ia_multi_vgt_param);
1946 }
1947 state->last_ia_multi_vgt_param = ia_multi_vgt_param;
1948 }
1949
1950 /* Primitive restart. */
1951 primitive_reset_en =
1952 indexed_draw && state->pipeline->graphics.prim_restart_enable;
1953
1954 if (primitive_reset_en != state->last_primitive_reset_en) {
1955 state->last_primitive_reset_en = primitive_reset_en;
1956 if (info->chip_class >= GFX9) {
1957 radeon_set_uconfig_reg(cs,
1958 R_03092C_VGT_MULTI_PRIM_IB_RESET_EN,
1959 primitive_reset_en);
1960 } else {
1961 radeon_set_context_reg(cs,
1962 R_028A94_VGT_MULTI_PRIM_IB_RESET_EN,
1963 primitive_reset_en);
1964 }
1965 }
1966
1967 if (primitive_reset_en) {
1968 uint32_t primitive_reset_index =
1969 state->index_type ? 0xffffffffu : 0xffffu;
1970
1971 if (primitive_reset_index != state->last_primitive_reset_index) {
1972 radeon_set_context_reg(cs,
1973 R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX,
1974 primitive_reset_index);
1975 state->last_primitive_reset_index = primitive_reset_index;
1976 }
1977 }
1978 }
1979
1980 static void radv_stage_flush(struct radv_cmd_buffer *cmd_buffer,
1981 VkPipelineStageFlags src_stage_mask)
1982 {
1983 if (src_stage_mask & (VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
1984 VK_PIPELINE_STAGE_TRANSFER_BIT |
1985 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT |
1986 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)) {
1987 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH;
1988 }
1989
1990 if (src_stage_mask & (VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT |
1991 VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT |
1992 VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT |
1993 VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
1994 VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
1995 VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
1996 VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT |
1997 VK_PIPELINE_STAGE_TRANSFER_BIT |
1998 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT |
1999 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT |
2000 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)) {
2001 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
2002 } else if (src_stage_mask & (VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT |
2003 VK_PIPELINE_STAGE_VERTEX_INPUT_BIT |
2004 VK_PIPELINE_STAGE_VERTEX_SHADER_BIT)) {
2005 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VS_PARTIAL_FLUSH;
2006 }
2007 }
2008
2009 static enum radv_cmd_flush_bits
2010 radv_src_access_flush(struct radv_cmd_buffer *cmd_buffer,
2011 VkAccessFlags src_flags)
2012 {
2013 enum radv_cmd_flush_bits flush_bits = 0;
2014 uint32_t b;
2015 for_each_bit(b, src_flags) {
2016 switch ((VkAccessFlagBits)(1 << b)) {
2017 case VK_ACCESS_SHADER_WRITE_BIT:
2018 flush_bits |= RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2;
2019 break;
2020 case VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT:
2021 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
2022 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
2023 break;
2024 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT:
2025 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
2026 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
2027 break;
2028 case VK_ACCESS_TRANSFER_WRITE_BIT:
2029 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
2030 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
2031 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
2032 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META |
2033 RADV_CMD_FLAG_INV_GLOBAL_L2;
2034 break;
2035 default:
2036 break;
2037 }
2038 }
2039 return flush_bits;
2040 }
2041
2042 static enum radv_cmd_flush_bits
2043 radv_dst_access_flush(struct radv_cmd_buffer *cmd_buffer,
2044 VkAccessFlags dst_flags,
2045 struct radv_image *image)
2046 {
2047 enum radv_cmd_flush_bits flush_bits = 0;
2048 uint32_t b;
2049 for_each_bit(b, dst_flags) {
2050 switch ((VkAccessFlagBits)(1 << b)) {
2051 case VK_ACCESS_INDIRECT_COMMAND_READ_BIT:
2052 case VK_ACCESS_INDEX_READ_BIT:
2053 break;
2054 case VK_ACCESS_UNIFORM_READ_BIT:
2055 flush_bits |= RADV_CMD_FLAG_INV_VMEM_L1 | RADV_CMD_FLAG_INV_SMEM_L1;
2056 break;
2057 case VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT:
2058 case VK_ACCESS_SHADER_READ_BIT:
2059 case VK_ACCESS_TRANSFER_READ_BIT:
2060 case VK_ACCESS_INPUT_ATTACHMENT_READ_BIT:
2061 flush_bits |= RADV_CMD_FLAG_INV_VMEM_L1 |
2062 RADV_CMD_FLAG_INV_GLOBAL_L2;
2063 break;
2064 case VK_ACCESS_COLOR_ATTACHMENT_READ_BIT:
2065 /* TODO: change to image && when the image gets passed
2066 * through from the subpass. */
2067 if (!image || (image->usage & VK_IMAGE_USAGE_STORAGE_BIT))
2068 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
2069 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
2070 break;
2071 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT:
2072 if (!image || (image->usage & VK_IMAGE_USAGE_STORAGE_BIT))
2073 flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
2074 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
2075 break;
2076 default:
2077 break;
2078 }
2079 }
2080 return flush_bits;
2081 }
2082
2083 static void radv_subpass_barrier(struct radv_cmd_buffer *cmd_buffer, const struct radv_subpass_barrier *barrier)
2084 {
2085 cmd_buffer->state.flush_bits |= radv_src_access_flush(cmd_buffer, barrier->src_access_mask);
2086 radv_stage_flush(cmd_buffer, barrier->src_stage_mask);
2087 cmd_buffer->state.flush_bits |= radv_dst_access_flush(cmd_buffer, barrier->dst_access_mask,
2088 NULL);
2089 }
2090
2091 static void radv_handle_subpass_image_transition(struct radv_cmd_buffer *cmd_buffer,
2092 VkAttachmentReference att)
2093 {
2094 unsigned idx = att.attachment;
2095 struct radv_image_view *view = cmd_buffer->state.framebuffer->attachments[idx].attachment;
2096 VkImageSubresourceRange range;
2097 range.aspectMask = 0;
2098 range.baseMipLevel = view->base_mip;
2099 range.levelCount = 1;
2100 range.baseArrayLayer = view->base_layer;
2101 range.layerCount = cmd_buffer->state.framebuffer->layers;
2102
2103 radv_handle_image_transition(cmd_buffer,
2104 view->image,
2105 cmd_buffer->state.attachments[idx].current_layout,
2106 att.layout, 0, 0, &range,
2107 cmd_buffer->state.attachments[idx].pending_clear_aspects);
2108
2109 cmd_buffer->state.attachments[idx].current_layout = att.layout;
2110
2111
2112 }
2113
2114 void
2115 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
2116 const struct radv_subpass *subpass, bool transitions)
2117 {
2118 if (transitions) {
2119 radv_subpass_barrier(cmd_buffer, &subpass->start_barrier);
2120
2121 for (unsigned i = 0; i < subpass->color_count; ++i) {
2122 if (subpass->color_attachments[i].attachment != VK_ATTACHMENT_UNUSED)
2123 radv_handle_subpass_image_transition(cmd_buffer,
2124 subpass->color_attachments[i]);
2125 }
2126
2127 for (unsigned i = 0; i < subpass->input_count; ++i) {
2128 radv_handle_subpass_image_transition(cmd_buffer,
2129 subpass->input_attachments[i]);
2130 }
2131
2132 if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
2133 radv_handle_subpass_image_transition(cmd_buffer,
2134 subpass->depth_stencil_attachment);
2135 }
2136 }
2137
2138 cmd_buffer->state.subpass = subpass;
2139
2140 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_FRAMEBUFFER;
2141 }
2142
2143 static VkResult
2144 radv_cmd_state_setup_attachments(struct radv_cmd_buffer *cmd_buffer,
2145 struct radv_render_pass *pass,
2146 const VkRenderPassBeginInfo *info)
2147 {
2148 struct radv_cmd_state *state = &cmd_buffer->state;
2149
2150 if (pass->attachment_count == 0) {
2151 state->attachments = NULL;
2152 return VK_SUCCESS;
2153 }
2154
2155 state->attachments = vk_alloc(&cmd_buffer->pool->alloc,
2156 pass->attachment_count *
2157 sizeof(state->attachments[0]),
2158 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2159 if (state->attachments == NULL) {
2160 cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
2161 return cmd_buffer->record_result;
2162 }
2163
2164 for (uint32_t i = 0; i < pass->attachment_count; ++i) {
2165 struct radv_render_pass_attachment *att = &pass->attachments[i];
2166 VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
2167 VkImageAspectFlags clear_aspects = 0;
2168
2169 if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
2170 /* color attachment */
2171 if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
2172 clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
2173 }
2174 } else {
2175 /* depthstencil attachment */
2176 if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
2177 att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
2178 clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
2179 if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
2180 att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_DONT_CARE)
2181 clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
2182 }
2183 if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
2184 att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
2185 clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
2186 }
2187 }
2188
2189 state->attachments[i].pending_clear_aspects = clear_aspects;
2190 state->attachments[i].cleared_views = 0;
2191 if (clear_aspects && info) {
2192 assert(info->clearValueCount > i);
2193 state->attachments[i].clear_value = info->pClearValues[i];
2194 }
2195
2196 state->attachments[i].current_layout = att->initial_layout;
2197 }
2198
2199 return VK_SUCCESS;
2200 }
2201
2202 VkResult radv_AllocateCommandBuffers(
2203 VkDevice _device,
2204 const VkCommandBufferAllocateInfo *pAllocateInfo,
2205 VkCommandBuffer *pCommandBuffers)
2206 {
2207 RADV_FROM_HANDLE(radv_device, device, _device);
2208 RADV_FROM_HANDLE(radv_cmd_pool, pool, pAllocateInfo->commandPool);
2209
2210 VkResult result = VK_SUCCESS;
2211 uint32_t i;
2212
2213 for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
2214
2215 if (!list_empty(&pool->free_cmd_buffers)) {
2216 struct radv_cmd_buffer *cmd_buffer = list_first_entry(&pool->free_cmd_buffers, struct radv_cmd_buffer, pool_link);
2217
2218 list_del(&cmd_buffer->pool_link);
2219 list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
2220
2221 result = radv_reset_cmd_buffer(cmd_buffer);
2222 cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2223 cmd_buffer->level = pAllocateInfo->level;
2224
2225 pCommandBuffers[i] = radv_cmd_buffer_to_handle(cmd_buffer);
2226 } else {
2227 result = radv_create_cmd_buffer(device, pool, pAllocateInfo->level,
2228 &pCommandBuffers[i]);
2229 }
2230 if (result != VK_SUCCESS)
2231 break;
2232 }
2233
2234 if (result != VK_SUCCESS) {
2235 radv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
2236 i, pCommandBuffers);
2237
2238 /* From the Vulkan 1.0.66 spec:
2239 *
2240 * "vkAllocateCommandBuffers can be used to create multiple
2241 * command buffers. If the creation of any of those command
2242 * buffers fails, the implementation must destroy all
2243 * successfully created command buffer objects from this
2244 * command, set all entries of the pCommandBuffers array to
2245 * NULL and return the error."
2246 */
2247 memset(pCommandBuffers, 0,
2248 sizeof(*pCommandBuffers) * pAllocateInfo->commandBufferCount);
2249 }
2250
2251 return result;
2252 }
2253
2254 void radv_FreeCommandBuffers(
2255 VkDevice device,
2256 VkCommandPool commandPool,
2257 uint32_t commandBufferCount,
2258 const VkCommandBuffer *pCommandBuffers)
2259 {
2260 for (uint32_t i = 0; i < commandBufferCount; i++) {
2261 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
2262
2263 if (cmd_buffer) {
2264 if (cmd_buffer->pool) {
2265 list_del(&cmd_buffer->pool_link);
2266 list_addtail(&cmd_buffer->pool_link, &cmd_buffer->pool->free_cmd_buffers);
2267 } else
2268 radv_cmd_buffer_destroy(cmd_buffer);
2269
2270 }
2271 }
2272 }
2273
2274 VkResult radv_ResetCommandBuffer(
2275 VkCommandBuffer commandBuffer,
2276 VkCommandBufferResetFlags flags)
2277 {
2278 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2279 return radv_reset_cmd_buffer(cmd_buffer);
2280 }
2281
2282 static void emit_gfx_buffer_state(struct radv_cmd_buffer *cmd_buffer)
2283 {
2284 struct radv_device *device = cmd_buffer->device;
2285 if (device->gfx_init) {
2286 uint64_t va = radv_buffer_get_va(device->gfx_init);
2287 radv_cs_add_buffer(device->ws, cmd_buffer->cs, device->gfx_init, 8);
2288 radeon_emit(cmd_buffer->cs, PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0));
2289 radeon_emit(cmd_buffer->cs, va);
2290 radeon_emit(cmd_buffer->cs, va >> 32);
2291 radeon_emit(cmd_buffer->cs, device->gfx_init_size_dw & 0xffff);
2292 } else
2293 si_init_config(cmd_buffer);
2294 }
2295
2296 VkResult radv_BeginCommandBuffer(
2297 VkCommandBuffer commandBuffer,
2298 const VkCommandBufferBeginInfo *pBeginInfo)
2299 {
2300 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2301 VkResult result = VK_SUCCESS;
2302
2303 if (cmd_buffer->status != RADV_CMD_BUFFER_STATUS_INITIAL) {
2304 /* If the command buffer has already been resetted with
2305 * vkResetCommandBuffer, no need to do it again.
2306 */
2307 result = radv_reset_cmd_buffer(cmd_buffer);
2308 if (result != VK_SUCCESS)
2309 return result;
2310 }
2311
2312 memset(&cmd_buffer->state, 0, sizeof(cmd_buffer->state));
2313 cmd_buffer->state.last_primitive_reset_en = -1;
2314 cmd_buffer->state.last_index_type = -1;
2315 cmd_buffer->usage_flags = pBeginInfo->flags;
2316
2317 /* setup initial configuration into command buffer */
2318 if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
2319 switch (cmd_buffer->queue_family_index) {
2320 case RADV_QUEUE_GENERAL:
2321 emit_gfx_buffer_state(cmd_buffer);
2322 break;
2323 case RADV_QUEUE_COMPUTE:
2324 si_init_compute(cmd_buffer);
2325 break;
2326 case RADV_QUEUE_TRANSFER:
2327 default:
2328 break;
2329 }
2330 }
2331
2332 if (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
2333 assert(pBeginInfo->pInheritanceInfo);
2334 cmd_buffer->state.framebuffer = radv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);
2335 cmd_buffer->state.pass = radv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
2336
2337 struct radv_subpass *subpass =
2338 &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
2339
2340 result = radv_cmd_state_setup_attachments(cmd_buffer, cmd_buffer->state.pass, NULL);
2341 if (result != VK_SUCCESS)
2342 return result;
2343
2344 radv_cmd_buffer_set_subpass(cmd_buffer, subpass, false);
2345 }
2346
2347 if (unlikely(cmd_buffer->device->trace_bo))
2348 radv_cmd_buffer_trace_emit(cmd_buffer);
2349
2350 cmd_buffer->status = RADV_CMD_BUFFER_STATUS_RECORDING;
2351
2352 return result;
2353 }
2354
2355 void radv_CmdBindVertexBuffers(
2356 VkCommandBuffer commandBuffer,
2357 uint32_t firstBinding,
2358 uint32_t bindingCount,
2359 const VkBuffer* pBuffers,
2360 const VkDeviceSize* pOffsets)
2361 {
2362 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2363 struct radv_vertex_binding *vb = cmd_buffer->vertex_bindings;
2364 bool changed = false;
2365
2366 /* We have to defer setting up vertex buffer since we need the buffer
2367 * stride from the pipeline. */
2368
2369 assert(firstBinding + bindingCount <= MAX_VBS);
2370 for (uint32_t i = 0; i < bindingCount; i++) {
2371 uint32_t idx = firstBinding + i;
2372
2373 if (!changed &&
2374 (vb[idx].buffer != radv_buffer_from_handle(pBuffers[i]) ||
2375 vb[idx].offset != pOffsets[i])) {
2376 changed = true;
2377 }
2378
2379 vb[idx].buffer = radv_buffer_from_handle(pBuffers[i]);
2380 vb[idx].offset = pOffsets[i];
2381
2382 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs,
2383 vb[idx].buffer->bo, 8);
2384 }
2385
2386 if (!changed) {
2387 /* No state changes. */
2388 return;
2389 }
2390
2391 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_VERTEX_BUFFER;
2392 }
2393
2394 void radv_CmdBindIndexBuffer(
2395 VkCommandBuffer commandBuffer,
2396 VkBuffer buffer,
2397 VkDeviceSize offset,
2398 VkIndexType indexType)
2399 {
2400 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2401 RADV_FROM_HANDLE(radv_buffer, index_buffer, buffer);
2402
2403 if (cmd_buffer->state.index_buffer == index_buffer &&
2404 cmd_buffer->state.index_offset == offset &&
2405 cmd_buffer->state.index_type == indexType) {
2406 /* No state changes. */
2407 return;
2408 }
2409
2410 cmd_buffer->state.index_buffer = index_buffer;
2411 cmd_buffer->state.index_offset = offset;
2412 cmd_buffer->state.index_type = indexType; /* vk matches hw */
2413 cmd_buffer->state.index_va = radv_buffer_get_va(index_buffer->bo);
2414 cmd_buffer->state.index_va += index_buffer->offset + offset;
2415
2416 int index_size_shift = cmd_buffer->state.index_type ? 2 : 1;
2417 cmd_buffer->state.max_index_count = (index_buffer->size - offset) >> index_size_shift;
2418 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_INDEX_BUFFER;
2419 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, index_buffer->bo, 8);
2420 }
2421
2422
2423 static void
2424 radv_bind_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
2425 struct radv_descriptor_set *set, unsigned idx)
2426 {
2427 struct radeon_winsys *ws = cmd_buffer->device->ws;
2428
2429 radv_set_descriptor_set(cmd_buffer, set, idx);
2430 if (!set)
2431 return;
2432
2433 assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
2434
2435 for (unsigned j = 0; j < set->layout->buffer_count; ++j)
2436 if (set->descriptors[j])
2437 radv_cs_add_buffer(ws, cmd_buffer->cs, set->descriptors[j], 7);
2438
2439 if(set->bo)
2440 radv_cs_add_buffer(ws, cmd_buffer->cs, set->bo, 8);
2441 }
2442
2443 void radv_CmdBindDescriptorSets(
2444 VkCommandBuffer commandBuffer,
2445 VkPipelineBindPoint pipelineBindPoint,
2446 VkPipelineLayout _layout,
2447 uint32_t firstSet,
2448 uint32_t descriptorSetCount,
2449 const VkDescriptorSet* pDescriptorSets,
2450 uint32_t dynamicOffsetCount,
2451 const uint32_t* pDynamicOffsets)
2452 {
2453 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2454 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2455 unsigned dyn_idx = 0;
2456
2457 for (unsigned i = 0; i < descriptorSetCount; ++i) {
2458 unsigned idx = i + firstSet;
2459 RADV_FROM_HANDLE(radv_descriptor_set, set, pDescriptorSets[i]);
2460 radv_bind_descriptor_set(cmd_buffer, set, idx);
2461
2462 for(unsigned j = 0; j < set->layout->dynamic_offset_count; ++j, ++dyn_idx) {
2463 unsigned idx = j + layout->set[i + firstSet].dynamic_offset_start;
2464 uint32_t *dst = cmd_buffer->dynamic_buffers + idx * 4;
2465 assert(dyn_idx < dynamicOffsetCount);
2466
2467 struct radv_descriptor_range *range = set->dynamic_descriptors + j;
2468 uint64_t va = range->va + pDynamicOffsets[dyn_idx];
2469 dst[0] = va;
2470 dst[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
2471 dst[2] = range->size;
2472 dst[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
2473 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
2474 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
2475 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
2476 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
2477 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
2478 cmd_buffer->push_constant_stages |=
2479 set->layout->dynamic_shader_stages;
2480 }
2481 }
2482 }
2483
2484 static bool radv_init_push_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
2485 struct radv_descriptor_set *set,
2486 struct radv_descriptor_set_layout *layout)
2487 {
2488 set->size = layout->size;
2489 set->layout = layout;
2490
2491 if (cmd_buffer->push_descriptors.capacity < set->size) {
2492 size_t new_size = MAX2(set->size, 1024);
2493 new_size = MAX2(new_size, 2 * cmd_buffer->push_descriptors.capacity);
2494 new_size = MIN2(new_size, 96 * MAX_PUSH_DESCRIPTORS);
2495
2496 free(set->mapped_ptr);
2497 set->mapped_ptr = malloc(new_size);
2498
2499 if (!set->mapped_ptr) {
2500 cmd_buffer->push_descriptors.capacity = 0;
2501 cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
2502 return false;
2503 }
2504
2505 cmd_buffer->push_descriptors.capacity = new_size;
2506 }
2507
2508 return true;
2509 }
2510
2511 void radv_meta_push_descriptor_set(
2512 struct radv_cmd_buffer* cmd_buffer,
2513 VkPipelineBindPoint pipelineBindPoint,
2514 VkPipelineLayout _layout,
2515 uint32_t set,
2516 uint32_t descriptorWriteCount,
2517 const VkWriteDescriptorSet* pDescriptorWrites)
2518 {
2519 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2520 struct radv_descriptor_set *push_set = &cmd_buffer->meta_push_descriptors;
2521 unsigned bo_offset;
2522
2523 assert(set == 0);
2524 assert(layout->set[set].layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR);
2525
2526 push_set->size = layout->set[set].layout->size;
2527 push_set->layout = layout->set[set].layout;
2528
2529 if (!radv_cmd_buffer_upload_alloc(cmd_buffer, push_set->size, 32,
2530 &bo_offset,
2531 (void**) &push_set->mapped_ptr))
2532 return;
2533
2534 push_set->va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
2535 push_set->va += bo_offset;
2536
2537 radv_update_descriptor_sets(cmd_buffer->device, cmd_buffer,
2538 radv_descriptor_set_to_handle(push_set),
2539 descriptorWriteCount, pDescriptorWrites, 0, NULL);
2540
2541 radv_set_descriptor_set(cmd_buffer, push_set, set);
2542 }
2543
2544 void radv_CmdPushDescriptorSetKHR(
2545 VkCommandBuffer commandBuffer,
2546 VkPipelineBindPoint pipelineBindPoint,
2547 VkPipelineLayout _layout,
2548 uint32_t set,
2549 uint32_t descriptorWriteCount,
2550 const VkWriteDescriptorSet* pDescriptorWrites)
2551 {
2552 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2553 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2554 struct radv_descriptor_set *push_set = &cmd_buffer->push_descriptors.set;
2555
2556 assert(layout->set[set].layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR);
2557
2558 if (!radv_init_push_descriptor_set(cmd_buffer, push_set, layout->set[set].layout))
2559 return;
2560
2561 radv_update_descriptor_sets(cmd_buffer->device, cmd_buffer,
2562 radv_descriptor_set_to_handle(push_set),
2563 descriptorWriteCount, pDescriptorWrites, 0, NULL);
2564
2565 radv_set_descriptor_set(cmd_buffer, push_set, set);
2566 cmd_buffer->state.push_descriptors_dirty = true;
2567 }
2568
2569 void radv_CmdPushDescriptorSetWithTemplateKHR(
2570 VkCommandBuffer commandBuffer,
2571 VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
2572 VkPipelineLayout _layout,
2573 uint32_t set,
2574 const void* pData)
2575 {
2576 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2577 RADV_FROM_HANDLE(radv_pipeline_layout, layout, _layout);
2578 struct radv_descriptor_set *push_set = &cmd_buffer->push_descriptors.set;
2579
2580 assert(layout->set[set].layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR);
2581
2582 if (!radv_init_push_descriptor_set(cmd_buffer, push_set, layout->set[set].layout))
2583 return;
2584
2585 radv_update_descriptor_set_with_template(cmd_buffer->device, cmd_buffer, push_set,
2586 descriptorUpdateTemplate, pData);
2587
2588 radv_set_descriptor_set(cmd_buffer, push_set, set);
2589 cmd_buffer->state.push_descriptors_dirty = true;
2590 }
2591
2592 void radv_CmdPushConstants(VkCommandBuffer commandBuffer,
2593 VkPipelineLayout layout,
2594 VkShaderStageFlags stageFlags,
2595 uint32_t offset,
2596 uint32_t size,
2597 const void* pValues)
2598 {
2599 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2600 memcpy(cmd_buffer->push_constants + offset, pValues, size);
2601 cmd_buffer->push_constant_stages |= stageFlags;
2602 }
2603
2604 VkResult radv_EndCommandBuffer(
2605 VkCommandBuffer commandBuffer)
2606 {
2607 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2608
2609 if (cmd_buffer->queue_family_index != RADV_QUEUE_TRANSFER) {
2610 if (cmd_buffer->device->physical_device->rad_info.chip_class == SI)
2611 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_PS_PARTIAL_FLUSH | RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2;
2612 si_emit_cache_flush(cmd_buffer);
2613 }
2614
2615 vk_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
2616
2617 if (!cmd_buffer->device->ws->cs_finalize(cmd_buffer->cs))
2618 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2619
2620 cmd_buffer->status = RADV_CMD_BUFFER_STATUS_EXECUTABLE;
2621
2622 return cmd_buffer->record_result;
2623 }
2624
2625 static void
2626 radv_emit_compute_pipeline(struct radv_cmd_buffer *cmd_buffer)
2627 {
2628 struct radv_shader_variant *compute_shader;
2629 struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
2630 struct radv_device *device = cmd_buffer->device;
2631 unsigned compute_resource_limits;
2632 unsigned waves_per_threadgroup;
2633 uint64_t va;
2634
2635 if (!pipeline || pipeline == cmd_buffer->state.emitted_compute_pipeline)
2636 return;
2637
2638 cmd_buffer->state.emitted_compute_pipeline = pipeline;
2639
2640 compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
2641 va = radv_buffer_get_va(compute_shader->bo) + compute_shader->bo_offset;
2642
2643 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
2644 cmd_buffer->cs, 19);
2645
2646 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B830_COMPUTE_PGM_LO, 2);
2647 radeon_emit(cmd_buffer->cs, va >> 8);
2648 radeon_emit(cmd_buffer->cs, va >> 40);
2649
2650 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
2651 radeon_emit(cmd_buffer->cs, compute_shader->rsrc1);
2652 radeon_emit(cmd_buffer->cs, compute_shader->rsrc2);
2653
2654
2655 cmd_buffer->compute_scratch_size_needed =
2656 MAX2(cmd_buffer->compute_scratch_size_needed,
2657 pipeline->max_waves * pipeline->scratch_bytes_per_wave);
2658
2659 /* change these once we have scratch support */
2660 radeon_set_sh_reg(cmd_buffer->cs, R_00B860_COMPUTE_TMPRING_SIZE,
2661 S_00B860_WAVES(pipeline->max_waves) |
2662 S_00B860_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
2663
2664 /* Calculate best compute resource limits. */
2665 waves_per_threadgroup =
2666 DIV_ROUND_UP(compute_shader->info.cs.block_size[0] *
2667 compute_shader->info.cs.block_size[1] *
2668 compute_shader->info.cs.block_size[2], 64);
2669 compute_resource_limits =
2670 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
2671
2672 if (device->physical_device->rad_info.chip_class >= CIK) {
2673 unsigned num_cu_per_se =
2674 device->physical_device->rad_info.num_good_compute_units /
2675 device->physical_device->rad_info.max_se;
2676
2677 /* Force even distribution on all SIMDs in CU if the workgroup
2678 * size is 64. This has shown some good improvements if # of
2679 * CUs per SE is not a multiple of 4.
2680 */
2681 if (num_cu_per_se % 4 && waves_per_threadgroup == 1)
2682 compute_resource_limits |= S_00B854_FORCE_SIMD_DIST(1);
2683 }
2684
2685 radeon_set_sh_reg(cmd_buffer->cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
2686 compute_resource_limits);
2687
2688 radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
2689 radeon_emit(cmd_buffer->cs,
2690 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[0]));
2691 radeon_emit(cmd_buffer->cs,
2692 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[1]));
2693 radeon_emit(cmd_buffer->cs,
2694 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[2]));
2695
2696 assert(cmd_buffer->cs->cdw <= cdw_max);
2697
2698 if (unlikely(cmd_buffer->device->trace_bo))
2699 radv_save_pipeline(cmd_buffer, pipeline, RING_COMPUTE);
2700 }
2701
2702 static void radv_mark_descriptor_sets_dirty(struct radv_cmd_buffer *cmd_buffer)
2703 {
2704 cmd_buffer->state.descriptors_dirty |= cmd_buffer->state.valid_descriptors;
2705 }
2706
2707 void radv_CmdBindPipeline(
2708 VkCommandBuffer commandBuffer,
2709 VkPipelineBindPoint pipelineBindPoint,
2710 VkPipeline _pipeline)
2711 {
2712 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2713 RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
2714
2715 switch (pipelineBindPoint) {
2716 case VK_PIPELINE_BIND_POINT_COMPUTE:
2717 if (cmd_buffer->state.compute_pipeline == pipeline)
2718 return;
2719 radv_mark_descriptor_sets_dirty(cmd_buffer);
2720
2721 cmd_buffer->state.compute_pipeline = pipeline;
2722 cmd_buffer->push_constant_stages |= VK_SHADER_STAGE_COMPUTE_BIT;
2723 break;
2724 case VK_PIPELINE_BIND_POINT_GRAPHICS:
2725 if (cmd_buffer->state.pipeline == pipeline)
2726 return;
2727 radv_mark_descriptor_sets_dirty(cmd_buffer);
2728
2729 cmd_buffer->state.pipeline = pipeline;
2730 if (!pipeline)
2731 break;
2732
2733 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_PIPELINE;
2734 cmd_buffer->push_constant_stages |= pipeline->active_stages;
2735
2736 radv_bind_dynamic_state(cmd_buffer, &pipeline->dynamic_state);
2737
2738 if (pipeline->graphics.esgs_ring_size > cmd_buffer->esgs_ring_size_needed)
2739 cmd_buffer->esgs_ring_size_needed = pipeline->graphics.esgs_ring_size;
2740 if (pipeline->graphics.gsvs_ring_size > cmd_buffer->gsvs_ring_size_needed)
2741 cmd_buffer->gsvs_ring_size_needed = pipeline->graphics.gsvs_ring_size;
2742
2743 if (radv_pipeline_has_tess(pipeline))
2744 cmd_buffer->tess_rings_needed = true;
2745
2746 if (radv_pipeline_has_gs(pipeline)) {
2747 struct ac_userdata_info *loc = radv_lookup_user_sgpr(cmd_buffer->state.pipeline, MESA_SHADER_GEOMETRY,
2748 AC_UD_SCRATCH_RING_OFFSETS);
2749 if (cmd_buffer->ring_offsets_idx == -1)
2750 cmd_buffer->ring_offsets_idx = loc->sgpr_idx;
2751 else if (loc->sgpr_idx != -1)
2752 assert(loc->sgpr_idx == cmd_buffer->ring_offsets_idx);
2753 }
2754 break;
2755 default:
2756 assert(!"invalid bind point");
2757 break;
2758 }
2759 }
2760
2761 void radv_CmdSetViewport(
2762 VkCommandBuffer commandBuffer,
2763 uint32_t firstViewport,
2764 uint32_t viewportCount,
2765 const VkViewport* pViewports)
2766 {
2767 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2768 struct radv_cmd_state *state = &cmd_buffer->state;
2769 MAYBE_UNUSED const uint32_t total_count = firstViewport + viewportCount;
2770
2771 assert(firstViewport < MAX_VIEWPORTS);
2772 assert(total_count >= 1 && total_count <= MAX_VIEWPORTS);
2773
2774 if (cmd_buffer->device->physical_device->has_scissor_bug) {
2775 /* Try to skip unnecessary PS partial flushes when the viewports
2776 * don't change.
2777 */
2778 if (!(state->dirty & (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT |
2779 RADV_CMD_DIRTY_DYNAMIC_SCISSOR)) &&
2780 !memcmp(state->dynamic.viewport.viewports + firstViewport,
2781 pViewports, viewportCount * sizeof(*pViewports))) {
2782 return;
2783 }
2784 }
2785
2786 memcpy(state->dynamic.viewport.viewports + firstViewport, pViewports,
2787 viewportCount * sizeof(*pViewports));
2788
2789 state->dirty |= RADV_CMD_DIRTY_DYNAMIC_VIEWPORT;
2790 }
2791
2792 void radv_CmdSetScissor(
2793 VkCommandBuffer commandBuffer,
2794 uint32_t firstScissor,
2795 uint32_t scissorCount,
2796 const VkRect2D* pScissors)
2797 {
2798 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2799 struct radv_cmd_state *state = &cmd_buffer->state;
2800 MAYBE_UNUSED const uint32_t total_count = firstScissor + scissorCount;
2801
2802 assert(firstScissor < MAX_SCISSORS);
2803 assert(total_count >= 1 && total_count <= MAX_SCISSORS);
2804
2805 if (cmd_buffer->device->physical_device->has_scissor_bug) {
2806 /* Try to skip unnecessary PS partial flushes when the scissors
2807 * don't change.
2808 */
2809 if (!(state->dirty & (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT |
2810 RADV_CMD_DIRTY_DYNAMIC_SCISSOR)) &&
2811 !memcmp(state->dynamic.scissor.scissors + firstScissor,
2812 pScissors, scissorCount * sizeof(*pScissors))) {
2813 return;
2814 }
2815 }
2816
2817 memcpy(state->dynamic.scissor.scissors + firstScissor, pScissors,
2818 scissorCount * sizeof(*pScissors));
2819
2820 state->dirty |= RADV_CMD_DIRTY_DYNAMIC_SCISSOR;
2821 }
2822
2823 void radv_CmdSetLineWidth(
2824 VkCommandBuffer commandBuffer,
2825 float lineWidth)
2826 {
2827 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2828 cmd_buffer->state.dynamic.line_width = lineWidth;
2829 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
2830 }
2831
2832 void radv_CmdSetDepthBias(
2833 VkCommandBuffer commandBuffer,
2834 float depthBiasConstantFactor,
2835 float depthBiasClamp,
2836 float depthBiasSlopeFactor)
2837 {
2838 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2839
2840 cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
2841 cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
2842 cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
2843
2844 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
2845 }
2846
2847 void radv_CmdSetBlendConstants(
2848 VkCommandBuffer commandBuffer,
2849 const float blendConstants[4])
2850 {
2851 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2852
2853 memcpy(cmd_buffer->state.dynamic.blend_constants,
2854 blendConstants, sizeof(float) * 4);
2855
2856 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
2857 }
2858
2859 void radv_CmdSetDepthBounds(
2860 VkCommandBuffer commandBuffer,
2861 float minDepthBounds,
2862 float maxDepthBounds)
2863 {
2864 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2865
2866 cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
2867 cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
2868
2869 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
2870 }
2871
2872 void radv_CmdSetStencilCompareMask(
2873 VkCommandBuffer commandBuffer,
2874 VkStencilFaceFlags faceMask,
2875 uint32_t compareMask)
2876 {
2877 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2878
2879 if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2880 cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
2881 if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2882 cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
2883
2884 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
2885 }
2886
2887 void radv_CmdSetStencilWriteMask(
2888 VkCommandBuffer commandBuffer,
2889 VkStencilFaceFlags faceMask,
2890 uint32_t writeMask)
2891 {
2892 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2893
2894 if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2895 cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
2896 if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2897 cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
2898
2899 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
2900 }
2901
2902 void radv_CmdSetStencilReference(
2903 VkCommandBuffer commandBuffer,
2904 VkStencilFaceFlags faceMask,
2905 uint32_t reference)
2906 {
2907 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2908
2909 if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
2910 cmd_buffer->state.dynamic.stencil_reference.front = reference;
2911 if (faceMask & VK_STENCIL_FACE_BACK_BIT)
2912 cmd_buffer->state.dynamic.stencil_reference.back = reference;
2913
2914 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
2915 }
2916
2917 void radv_CmdSetDiscardRectangleEXT(
2918 VkCommandBuffer commandBuffer,
2919 uint32_t firstDiscardRectangle,
2920 uint32_t discardRectangleCount,
2921 const VkRect2D* pDiscardRectangles)
2922 {
2923 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2924 struct radv_cmd_state *state = &cmd_buffer->state;
2925 MAYBE_UNUSED const uint32_t total_count = firstDiscardRectangle + discardRectangleCount;
2926
2927 assert(firstDiscardRectangle < MAX_DISCARD_RECTANGLES);
2928 assert(total_count >= 1 && total_count <= MAX_DISCARD_RECTANGLES);
2929
2930 typed_memcpy(&state->dynamic.discard_rectangle.rectangles[firstDiscardRectangle],
2931 pDiscardRectangles, discardRectangleCount);
2932
2933 state->dirty |= RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE;
2934 }
2935
2936 void radv_CmdExecuteCommands(
2937 VkCommandBuffer commandBuffer,
2938 uint32_t commandBufferCount,
2939 const VkCommandBuffer* pCmdBuffers)
2940 {
2941 RADV_FROM_HANDLE(radv_cmd_buffer, primary, commandBuffer);
2942
2943 assert(commandBufferCount > 0);
2944
2945 /* Emit pending flushes on primary prior to executing secondary */
2946 si_emit_cache_flush(primary);
2947
2948 for (uint32_t i = 0; i < commandBufferCount; i++) {
2949 RADV_FROM_HANDLE(radv_cmd_buffer, secondary, pCmdBuffers[i]);
2950
2951 primary->scratch_size_needed = MAX2(primary->scratch_size_needed,
2952 secondary->scratch_size_needed);
2953 primary->compute_scratch_size_needed = MAX2(primary->compute_scratch_size_needed,
2954 secondary->compute_scratch_size_needed);
2955
2956 if (secondary->esgs_ring_size_needed > primary->esgs_ring_size_needed)
2957 primary->esgs_ring_size_needed = secondary->esgs_ring_size_needed;
2958 if (secondary->gsvs_ring_size_needed > primary->gsvs_ring_size_needed)
2959 primary->gsvs_ring_size_needed = secondary->gsvs_ring_size_needed;
2960 if (secondary->tess_rings_needed)
2961 primary->tess_rings_needed = true;
2962 if (secondary->sample_positions_needed)
2963 primary->sample_positions_needed = true;
2964
2965 if (secondary->ring_offsets_idx != -1) {
2966 if (primary->ring_offsets_idx == -1)
2967 primary->ring_offsets_idx = secondary->ring_offsets_idx;
2968 else
2969 assert(secondary->ring_offsets_idx == primary->ring_offsets_idx);
2970 }
2971 primary->device->ws->cs_execute_secondary(primary->cs, secondary->cs);
2972
2973
2974 /* When the secondary command buffer is compute only we don't
2975 * need to re-emit the current graphics pipeline.
2976 */
2977 if (secondary->state.emitted_pipeline) {
2978 primary->state.emitted_pipeline =
2979 secondary->state.emitted_pipeline;
2980 }
2981
2982 /* When the secondary command buffer is graphics only we don't
2983 * need to re-emit the current compute pipeline.
2984 */
2985 if (secondary->state.emitted_compute_pipeline) {
2986 primary->state.emitted_compute_pipeline =
2987 secondary->state.emitted_compute_pipeline;
2988 }
2989
2990 /* Only re-emit the draw packets when needed. */
2991 if (secondary->state.last_primitive_reset_en != -1) {
2992 primary->state.last_primitive_reset_en =
2993 secondary->state.last_primitive_reset_en;
2994 }
2995
2996 if (secondary->state.last_primitive_reset_index) {
2997 primary->state.last_primitive_reset_index =
2998 secondary->state.last_primitive_reset_index;
2999 }
3000
3001 if (secondary->state.last_ia_multi_vgt_param) {
3002 primary->state.last_ia_multi_vgt_param =
3003 secondary->state.last_ia_multi_vgt_param;
3004 }
3005
3006 if (secondary->state.last_index_type != -1) {
3007 primary->state.last_index_type =
3008 secondary->state.last_index_type;
3009 }
3010 }
3011
3012 /* After executing commands from secondary buffers we have to dirty
3013 * some states.
3014 */
3015 primary->state.dirty |= RADV_CMD_DIRTY_PIPELINE |
3016 RADV_CMD_DIRTY_INDEX_BUFFER |
3017 RADV_CMD_DIRTY_DYNAMIC_ALL;
3018 radv_mark_descriptor_sets_dirty(primary);
3019 }
3020
3021 VkResult radv_CreateCommandPool(
3022 VkDevice _device,
3023 const VkCommandPoolCreateInfo* pCreateInfo,
3024 const VkAllocationCallbacks* pAllocator,
3025 VkCommandPool* pCmdPool)
3026 {
3027 RADV_FROM_HANDLE(radv_device, device, _device);
3028 struct radv_cmd_pool *pool;
3029
3030 pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
3031 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3032 if (pool == NULL)
3033 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3034
3035 if (pAllocator)
3036 pool->alloc = *pAllocator;
3037 else
3038 pool->alloc = device->alloc;
3039
3040 list_inithead(&pool->cmd_buffers);
3041 list_inithead(&pool->free_cmd_buffers);
3042
3043 pool->queue_family_index = pCreateInfo->queueFamilyIndex;
3044
3045 *pCmdPool = radv_cmd_pool_to_handle(pool);
3046
3047 return VK_SUCCESS;
3048
3049 }
3050
3051 void radv_DestroyCommandPool(
3052 VkDevice _device,
3053 VkCommandPool commandPool,
3054 const VkAllocationCallbacks* pAllocator)
3055 {
3056 RADV_FROM_HANDLE(radv_device, device, _device);
3057 RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
3058
3059 if (!pool)
3060 return;
3061
3062 list_for_each_entry_safe(struct radv_cmd_buffer, cmd_buffer,
3063 &pool->cmd_buffers, pool_link) {
3064 radv_cmd_buffer_destroy(cmd_buffer);
3065 }
3066
3067 list_for_each_entry_safe(struct radv_cmd_buffer, cmd_buffer,
3068 &pool->free_cmd_buffers, pool_link) {
3069 radv_cmd_buffer_destroy(cmd_buffer);
3070 }
3071
3072 vk_free2(&device->alloc, pAllocator, pool);
3073 }
3074
3075 VkResult radv_ResetCommandPool(
3076 VkDevice device,
3077 VkCommandPool commandPool,
3078 VkCommandPoolResetFlags flags)
3079 {
3080 RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
3081 VkResult result;
3082
3083 list_for_each_entry(struct radv_cmd_buffer, cmd_buffer,
3084 &pool->cmd_buffers, pool_link) {
3085 result = radv_reset_cmd_buffer(cmd_buffer);
3086 if (result != VK_SUCCESS)
3087 return result;
3088 }
3089
3090 return VK_SUCCESS;
3091 }
3092
3093 void radv_TrimCommandPoolKHR(
3094 VkDevice device,
3095 VkCommandPool commandPool,
3096 VkCommandPoolTrimFlagsKHR flags)
3097 {
3098 RADV_FROM_HANDLE(radv_cmd_pool, pool, commandPool);
3099
3100 if (!pool)
3101 return;
3102
3103 list_for_each_entry_safe(struct radv_cmd_buffer, cmd_buffer,
3104 &pool->free_cmd_buffers, pool_link) {
3105 radv_cmd_buffer_destroy(cmd_buffer);
3106 }
3107 }
3108
3109 void radv_CmdBeginRenderPass(
3110 VkCommandBuffer commandBuffer,
3111 const VkRenderPassBeginInfo* pRenderPassBegin,
3112 VkSubpassContents contents)
3113 {
3114 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3115 RADV_FROM_HANDLE(radv_render_pass, pass, pRenderPassBegin->renderPass);
3116 RADV_FROM_HANDLE(radv_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
3117
3118 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
3119 cmd_buffer->cs, 2048);
3120 MAYBE_UNUSED VkResult result;
3121
3122 cmd_buffer->state.framebuffer = framebuffer;
3123 cmd_buffer->state.pass = pass;
3124 cmd_buffer->state.render_area = pRenderPassBegin->renderArea;
3125
3126 result = radv_cmd_state_setup_attachments(cmd_buffer, pass, pRenderPassBegin);
3127 if (result != VK_SUCCESS)
3128 return;
3129
3130 radv_cmd_buffer_set_subpass(cmd_buffer, pass->subpasses, true);
3131 assert(cmd_buffer->cs->cdw <= cdw_max);
3132
3133 radv_cmd_buffer_clear_subpass(cmd_buffer);
3134 }
3135
3136 void radv_CmdNextSubpass(
3137 VkCommandBuffer commandBuffer,
3138 VkSubpassContents contents)
3139 {
3140 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3141
3142 radv_cmd_buffer_resolve_subpass(cmd_buffer);
3143
3144 radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs,
3145 2048);
3146
3147 radv_cmd_buffer_set_subpass(cmd_buffer, cmd_buffer->state.subpass + 1, true);
3148 radv_cmd_buffer_clear_subpass(cmd_buffer);
3149 }
3150
3151 static void radv_emit_view_index(struct radv_cmd_buffer *cmd_buffer, unsigned index)
3152 {
3153 struct radv_pipeline *pipeline = cmd_buffer->state.pipeline;
3154 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; ++stage) {
3155 if (!pipeline->shaders[stage])
3156 continue;
3157 struct ac_userdata_info *loc = radv_lookup_user_sgpr(pipeline, stage, AC_UD_VIEW_INDEX);
3158 if (loc->sgpr_idx == -1)
3159 continue;
3160 uint32_t base_reg = pipeline->user_data_0[stage];
3161 radeon_set_sh_reg(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, index);
3162
3163 }
3164 if (pipeline->gs_copy_shader) {
3165 struct ac_userdata_info *loc = &pipeline->gs_copy_shader->info.user_sgprs_locs.shader_data[AC_UD_VIEW_INDEX];
3166 if (loc->sgpr_idx != -1) {
3167 uint32_t base_reg = R_00B130_SPI_SHADER_USER_DATA_VS_0;
3168 radeon_set_sh_reg(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, index);
3169 }
3170 }
3171 }
3172
3173 static void
3174 radv_cs_emit_draw_packet(struct radv_cmd_buffer *cmd_buffer,
3175 uint32_t vertex_count)
3176 {
3177 radeon_emit(cmd_buffer->cs, PKT3(PKT3_DRAW_INDEX_AUTO, 1, cmd_buffer->state.predicating));
3178 radeon_emit(cmd_buffer->cs, vertex_count);
3179 radeon_emit(cmd_buffer->cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX |
3180 S_0287F0_USE_OPAQUE(0));
3181 }
3182
3183 static void
3184 radv_cs_emit_draw_indexed_packet(struct radv_cmd_buffer *cmd_buffer,
3185 uint64_t index_va,
3186 uint32_t index_count)
3187 {
3188 radeon_emit(cmd_buffer->cs, PKT3(PKT3_DRAW_INDEX_2, 4, false));
3189 radeon_emit(cmd_buffer->cs, cmd_buffer->state.max_index_count);
3190 radeon_emit(cmd_buffer->cs, index_va);
3191 radeon_emit(cmd_buffer->cs, index_va >> 32);
3192 radeon_emit(cmd_buffer->cs, index_count);
3193 radeon_emit(cmd_buffer->cs, V_0287F0_DI_SRC_SEL_DMA);
3194 }
3195
3196 static void
3197 radv_cs_emit_indirect_draw_packet(struct radv_cmd_buffer *cmd_buffer,
3198 bool indexed,
3199 uint32_t draw_count,
3200 uint64_t count_va,
3201 uint32_t stride)
3202 {
3203 struct radeon_winsys_cs *cs = cmd_buffer->cs;
3204 unsigned di_src_sel = indexed ? V_0287F0_DI_SRC_SEL_DMA
3205 : V_0287F0_DI_SRC_SEL_AUTO_INDEX;
3206 bool draw_id_enable = radv_get_vertex_shader(cmd_buffer->state.pipeline)->info.info.vs.needs_draw_id;
3207 uint32_t base_reg = cmd_buffer->state.pipeline->graphics.vtx_base_sgpr;
3208 assert(base_reg);
3209
3210 if (draw_count == 1 && !count_va && !draw_id_enable) {
3211 radeon_emit(cs, PKT3(indexed ? PKT3_DRAW_INDEX_INDIRECT :
3212 PKT3_DRAW_INDIRECT, 3, false));
3213 radeon_emit(cs, 0);
3214 radeon_emit(cs, (base_reg - SI_SH_REG_OFFSET) >> 2);
3215 radeon_emit(cs, ((base_reg + 4) - SI_SH_REG_OFFSET) >> 2);
3216 radeon_emit(cs, di_src_sel);
3217 } else {
3218 radeon_emit(cs, PKT3(indexed ? PKT3_DRAW_INDEX_INDIRECT_MULTI :
3219 PKT3_DRAW_INDIRECT_MULTI,
3220 8, false));
3221 radeon_emit(cs, 0);
3222 radeon_emit(cs, (base_reg - SI_SH_REG_OFFSET) >> 2);
3223 radeon_emit(cs, ((base_reg + 4) - SI_SH_REG_OFFSET) >> 2);
3224 radeon_emit(cs, (((base_reg + 8) - SI_SH_REG_OFFSET) >> 2) |
3225 S_2C3_DRAW_INDEX_ENABLE(draw_id_enable) |
3226 S_2C3_COUNT_INDIRECT_ENABLE(!!count_va));
3227 radeon_emit(cs, draw_count); /* count */
3228 radeon_emit(cs, count_va); /* count_addr */
3229 radeon_emit(cs, count_va >> 32);
3230 radeon_emit(cs, stride); /* stride */
3231 radeon_emit(cs, di_src_sel);
3232 }
3233 }
3234
3235 struct radv_draw_info {
3236 /**
3237 * Number of vertices.
3238 */
3239 uint32_t count;
3240
3241 /**
3242 * Index of the first vertex.
3243 */
3244 int32_t vertex_offset;
3245
3246 /**
3247 * First instance id.
3248 */
3249 uint32_t first_instance;
3250
3251 /**
3252 * Number of instances.
3253 */
3254 uint32_t instance_count;
3255
3256 /**
3257 * First index (indexed draws only).
3258 */
3259 uint32_t first_index;
3260
3261 /**
3262 * Whether it's an indexed draw.
3263 */
3264 bool indexed;
3265
3266 /**
3267 * Indirect draw parameters resource.
3268 */
3269 struct radv_buffer *indirect;
3270 uint64_t indirect_offset;
3271 uint32_t stride;
3272
3273 /**
3274 * Draw count parameters resource.
3275 */
3276 struct radv_buffer *count_buffer;
3277 uint64_t count_buffer_offset;
3278 };
3279
3280 static void
3281 radv_emit_draw_packets(struct radv_cmd_buffer *cmd_buffer,
3282 const struct radv_draw_info *info)
3283 {
3284 struct radv_cmd_state *state = &cmd_buffer->state;
3285 struct radeon_winsys *ws = cmd_buffer->device->ws;
3286 struct radeon_winsys_cs *cs = cmd_buffer->cs;
3287
3288 if (info->indirect) {
3289 uint64_t va = radv_buffer_get_va(info->indirect->bo);
3290 uint64_t count_va = 0;
3291
3292 va += info->indirect->offset + info->indirect_offset;
3293
3294 radv_cs_add_buffer(ws, cs, info->indirect->bo, 8);
3295
3296 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
3297 radeon_emit(cs, 1);
3298 radeon_emit(cs, va);
3299 radeon_emit(cs, va >> 32);
3300
3301 if (info->count_buffer) {
3302 count_va = radv_buffer_get_va(info->count_buffer->bo);
3303 count_va += info->count_buffer->offset +
3304 info->count_buffer_offset;
3305
3306 radv_cs_add_buffer(ws, cs, info->count_buffer->bo, 8);
3307 }
3308
3309 if (!state->subpass->view_mask) {
3310 radv_cs_emit_indirect_draw_packet(cmd_buffer,
3311 info->indexed,
3312 info->count,
3313 count_va,
3314 info->stride);
3315 } else {
3316 unsigned i;
3317 for_each_bit(i, state->subpass->view_mask) {
3318 radv_emit_view_index(cmd_buffer, i);
3319
3320 radv_cs_emit_indirect_draw_packet(cmd_buffer,
3321 info->indexed,
3322 info->count,
3323 count_va,
3324 info->stride);
3325 }
3326 }
3327 } else {
3328 assert(state->pipeline->graphics.vtx_base_sgpr);
3329 radeon_set_sh_reg_seq(cs, state->pipeline->graphics.vtx_base_sgpr,
3330 state->pipeline->graphics.vtx_emit_num);
3331 radeon_emit(cs, info->vertex_offset);
3332 radeon_emit(cs, info->first_instance);
3333 if (state->pipeline->graphics.vtx_emit_num == 3)
3334 radeon_emit(cs, 0);
3335
3336 radeon_emit(cs, PKT3(PKT3_NUM_INSTANCES, 0, state->predicating));
3337 radeon_emit(cs, info->instance_count);
3338
3339 if (info->indexed) {
3340 int index_size = state->index_type ? 4 : 2;
3341 uint64_t index_va;
3342
3343 index_va = state->index_va;
3344 index_va += info->first_index * index_size;
3345
3346 if (!state->subpass->view_mask) {
3347 radv_cs_emit_draw_indexed_packet(cmd_buffer,
3348 index_va,
3349 info->count);
3350 } else {
3351 unsigned i;
3352 for_each_bit(i, state->subpass->view_mask) {
3353 radv_emit_view_index(cmd_buffer, i);
3354
3355 radv_cs_emit_draw_indexed_packet(cmd_buffer,
3356 index_va,
3357 info->count);
3358 }
3359 }
3360 } else {
3361 if (!state->subpass->view_mask) {
3362 radv_cs_emit_draw_packet(cmd_buffer, info->count);
3363 } else {
3364 unsigned i;
3365 for_each_bit(i, state->subpass->view_mask) {
3366 radv_emit_view_index(cmd_buffer, i);
3367
3368 radv_cs_emit_draw_packet(cmd_buffer,
3369 info->count);
3370 }
3371 }
3372 }
3373 }
3374 }
3375
3376 static void
3377 radv_emit_all_graphics_states(struct radv_cmd_buffer *cmd_buffer,
3378 const struct radv_draw_info *info)
3379 {
3380 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_PIPELINE)
3381 radv_emit_graphics_pipeline(cmd_buffer);
3382
3383 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_FRAMEBUFFER)
3384 radv_emit_framebuffer_state(cmd_buffer);
3385
3386 if (info->indexed) {
3387 if (cmd_buffer->state.dirty & RADV_CMD_DIRTY_INDEX_BUFFER)
3388 radv_emit_index_buffer(cmd_buffer);
3389 } else {
3390 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
3391 * so the state must be re-emitted before the next indexed
3392 * draw.
3393 */
3394 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
3395 cmd_buffer->state.last_index_type = -1;
3396 cmd_buffer->state.dirty |= RADV_CMD_DIRTY_INDEX_BUFFER;
3397 }
3398 }
3399
3400 radv_cmd_buffer_flush_dynamic_state(cmd_buffer);
3401
3402 radv_emit_draw_registers(cmd_buffer, info->indexed,
3403 info->instance_count > 1, info->indirect,
3404 info->indirect ? 0 : info->count);
3405 }
3406
3407 static void
3408 radv_draw(struct radv_cmd_buffer *cmd_buffer,
3409 const struct radv_draw_info *info)
3410 {
3411 bool pipeline_is_dirty =
3412 (cmd_buffer->state.dirty & RADV_CMD_DIRTY_PIPELINE) &&
3413 cmd_buffer->state.pipeline &&
3414 cmd_buffer->state.pipeline != cmd_buffer->state.emitted_pipeline;
3415
3416 MAYBE_UNUSED unsigned cdw_max =
3417 radeon_check_space(cmd_buffer->device->ws,
3418 cmd_buffer->cs, 4096);
3419
3420 /* Use optimal packet order based on whether we need to sync the
3421 * pipeline.
3422 */
3423 if (cmd_buffer->state.flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3424 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3425 RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
3426 RADV_CMD_FLAG_CS_PARTIAL_FLUSH)) {
3427 /* If we have to wait for idle, set all states first, so that
3428 * all SET packets are processed in parallel with previous draw
3429 * calls. Then upload descriptors, set shader pointers, and
3430 * draw, and prefetch at the end. This ensures that the time
3431 * the CUs are idle is very short. (there are only SET_SH
3432 * packets between the wait and the draw)
3433 */
3434 radv_emit_all_graphics_states(cmd_buffer, info);
3435 si_emit_cache_flush(cmd_buffer);
3436 /* <-- CUs are idle here --> */
3437
3438 if (!radv_upload_graphics_shader_descriptors(cmd_buffer, pipeline_is_dirty))
3439 return;
3440
3441 radv_emit_draw_packets(cmd_buffer, info);
3442 /* <-- CUs are busy here --> */
3443
3444 /* Start prefetches after the draw has been started. Both will
3445 * run in parallel, but starting the draw first is more
3446 * important.
3447 */
3448 if (pipeline_is_dirty) {
3449 radv_emit_prefetch(cmd_buffer,
3450 cmd_buffer->state.pipeline);
3451 }
3452 } else {
3453 /* If we don't wait for idle, start prefetches first, then set
3454 * states, and draw at the end.
3455 */
3456 si_emit_cache_flush(cmd_buffer);
3457
3458 if (pipeline_is_dirty) {
3459 radv_emit_prefetch(cmd_buffer,
3460 cmd_buffer->state.pipeline);
3461 }
3462
3463 if (!radv_upload_graphics_shader_descriptors(cmd_buffer, pipeline_is_dirty))
3464 return;
3465
3466 radv_emit_all_graphics_states(cmd_buffer, info);
3467 radv_emit_draw_packets(cmd_buffer, info);
3468 }
3469
3470 assert(cmd_buffer->cs->cdw <= cdw_max);
3471 radv_cmd_buffer_after_draw(cmd_buffer);
3472 }
3473
3474 void radv_CmdDraw(
3475 VkCommandBuffer commandBuffer,
3476 uint32_t vertexCount,
3477 uint32_t instanceCount,
3478 uint32_t firstVertex,
3479 uint32_t firstInstance)
3480 {
3481 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3482 struct radv_draw_info info = {};
3483
3484 info.count = vertexCount;
3485 info.instance_count = instanceCount;
3486 info.first_instance = firstInstance;
3487 info.vertex_offset = firstVertex;
3488
3489 radv_draw(cmd_buffer, &info);
3490 }
3491
3492 void radv_CmdDrawIndexed(
3493 VkCommandBuffer commandBuffer,
3494 uint32_t indexCount,
3495 uint32_t instanceCount,
3496 uint32_t firstIndex,
3497 int32_t vertexOffset,
3498 uint32_t firstInstance)
3499 {
3500 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3501 struct radv_draw_info info = {};
3502
3503 info.indexed = true;
3504 info.count = indexCount;
3505 info.instance_count = instanceCount;
3506 info.first_index = firstIndex;
3507 info.vertex_offset = vertexOffset;
3508 info.first_instance = firstInstance;
3509
3510 radv_draw(cmd_buffer, &info);
3511 }
3512
3513 void radv_CmdDrawIndirect(
3514 VkCommandBuffer commandBuffer,
3515 VkBuffer _buffer,
3516 VkDeviceSize offset,
3517 uint32_t drawCount,
3518 uint32_t stride)
3519 {
3520 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3521 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3522 struct radv_draw_info info = {};
3523
3524 info.count = drawCount;
3525 info.indirect = buffer;
3526 info.indirect_offset = offset;
3527 info.stride = stride;
3528
3529 radv_draw(cmd_buffer, &info);
3530 }
3531
3532 void radv_CmdDrawIndexedIndirect(
3533 VkCommandBuffer commandBuffer,
3534 VkBuffer _buffer,
3535 VkDeviceSize offset,
3536 uint32_t drawCount,
3537 uint32_t stride)
3538 {
3539 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3540 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3541 struct radv_draw_info info = {};
3542
3543 info.indexed = true;
3544 info.count = drawCount;
3545 info.indirect = buffer;
3546 info.indirect_offset = offset;
3547 info.stride = stride;
3548
3549 radv_draw(cmd_buffer, &info);
3550 }
3551
3552 void radv_CmdDrawIndirectCountAMD(
3553 VkCommandBuffer commandBuffer,
3554 VkBuffer _buffer,
3555 VkDeviceSize offset,
3556 VkBuffer _countBuffer,
3557 VkDeviceSize countBufferOffset,
3558 uint32_t maxDrawCount,
3559 uint32_t stride)
3560 {
3561 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3562 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3563 RADV_FROM_HANDLE(radv_buffer, count_buffer, _countBuffer);
3564 struct radv_draw_info info = {};
3565
3566 info.count = maxDrawCount;
3567 info.indirect = buffer;
3568 info.indirect_offset = offset;
3569 info.count_buffer = count_buffer;
3570 info.count_buffer_offset = countBufferOffset;
3571 info.stride = stride;
3572
3573 radv_draw(cmd_buffer, &info);
3574 }
3575
3576 void radv_CmdDrawIndexedIndirectCountAMD(
3577 VkCommandBuffer commandBuffer,
3578 VkBuffer _buffer,
3579 VkDeviceSize offset,
3580 VkBuffer _countBuffer,
3581 VkDeviceSize countBufferOffset,
3582 uint32_t maxDrawCount,
3583 uint32_t stride)
3584 {
3585 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3586 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3587 RADV_FROM_HANDLE(radv_buffer, count_buffer, _countBuffer);
3588 struct radv_draw_info info = {};
3589
3590 info.indexed = true;
3591 info.count = maxDrawCount;
3592 info.indirect = buffer;
3593 info.indirect_offset = offset;
3594 info.count_buffer = count_buffer;
3595 info.count_buffer_offset = countBufferOffset;
3596 info.stride = stride;
3597
3598 radv_draw(cmd_buffer, &info);
3599 }
3600
3601 struct radv_dispatch_info {
3602 /**
3603 * Determine the layout of the grid (in block units) to be used.
3604 */
3605 uint32_t blocks[3];
3606
3607 /**
3608 * Whether it's an unaligned compute dispatch.
3609 */
3610 bool unaligned;
3611
3612 /**
3613 * Indirect compute parameters resource.
3614 */
3615 struct radv_buffer *indirect;
3616 uint64_t indirect_offset;
3617 };
3618
3619 static void
3620 radv_emit_dispatch_packets(struct radv_cmd_buffer *cmd_buffer,
3621 const struct radv_dispatch_info *info)
3622 {
3623 struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
3624 struct radv_shader_variant *compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
3625 unsigned dispatch_initiator = cmd_buffer->device->dispatch_initiator;
3626 struct radeon_winsys *ws = cmd_buffer->device->ws;
3627 struct radeon_winsys_cs *cs = cmd_buffer->cs;
3628 struct ac_userdata_info *loc;
3629
3630 loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_COMPUTE,
3631 AC_UD_CS_GRID_SIZE);
3632
3633 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(ws, cs, 25);
3634
3635 if (info->indirect) {
3636 uint64_t va = radv_buffer_get_va(info->indirect->bo);
3637
3638 va += info->indirect->offset + info->indirect_offset;
3639
3640 radv_cs_add_buffer(ws, cs, info->indirect->bo, 8);
3641
3642 if (loc->sgpr_idx != -1) {
3643 for (unsigned i = 0; i < 3; ++i) {
3644 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
3645 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
3646 COPY_DATA_DST_SEL(COPY_DATA_REG));
3647 radeon_emit(cs, (va + 4 * i));
3648 radeon_emit(cs, (va + 4 * i) >> 32);
3649 radeon_emit(cs, ((R_00B900_COMPUTE_USER_DATA_0
3650 + loc->sgpr_idx * 4) >> 2) + i);
3651 radeon_emit(cs, 0);
3652 }
3653 }
3654
3655 if (radv_cmd_buffer_uses_mec(cmd_buffer)) {
3656 radeon_emit(cs, PKT3(PKT3_DISPATCH_INDIRECT, 2, 0) |
3657 PKT3_SHADER_TYPE_S(1));
3658 radeon_emit(cs, va);
3659 radeon_emit(cs, va >> 32);
3660 radeon_emit(cs, dispatch_initiator);
3661 } else {
3662 radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0) |
3663 PKT3_SHADER_TYPE_S(1));
3664 radeon_emit(cs, 1);
3665 radeon_emit(cs, va);
3666 radeon_emit(cs, va >> 32);
3667
3668 radeon_emit(cs, PKT3(PKT3_DISPATCH_INDIRECT, 1, 0) |
3669 PKT3_SHADER_TYPE_S(1));
3670 radeon_emit(cs, 0);
3671 radeon_emit(cs, dispatch_initiator);
3672 }
3673 } else {
3674 unsigned blocks[3] = { info->blocks[0], info->blocks[1], info->blocks[2] };
3675
3676 if (info->unaligned) {
3677 unsigned *cs_block_size = compute_shader->info.cs.block_size;
3678 unsigned remainder[3];
3679
3680 /* If aligned, these should be an entire block size,
3681 * not 0.
3682 */
3683 remainder[0] = blocks[0] + cs_block_size[0] -
3684 align_u32_npot(blocks[0], cs_block_size[0]);
3685 remainder[1] = blocks[1] + cs_block_size[1] -
3686 align_u32_npot(blocks[1], cs_block_size[1]);
3687 remainder[2] = blocks[2] + cs_block_size[2] -
3688 align_u32_npot(blocks[2], cs_block_size[2]);
3689
3690 blocks[0] = round_up_u32(blocks[0], cs_block_size[0]);
3691 blocks[1] = round_up_u32(blocks[1], cs_block_size[1]);
3692 blocks[2] = round_up_u32(blocks[2], cs_block_size[2]);
3693
3694 radeon_set_sh_reg_seq(cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
3695 radeon_emit(cs,
3696 S_00B81C_NUM_THREAD_FULL(cs_block_size[0]) |
3697 S_00B81C_NUM_THREAD_PARTIAL(remainder[0]));
3698 radeon_emit(cs,
3699 S_00B81C_NUM_THREAD_FULL(cs_block_size[1]) |
3700 S_00B81C_NUM_THREAD_PARTIAL(remainder[1]));
3701 radeon_emit(cs,
3702 S_00B81C_NUM_THREAD_FULL(cs_block_size[2]) |
3703 S_00B81C_NUM_THREAD_PARTIAL(remainder[2]));
3704
3705 dispatch_initiator |= S_00B800_PARTIAL_TG_EN(1);
3706 }
3707
3708 if (loc->sgpr_idx != -1) {
3709 assert(!loc->indirect);
3710 assert(loc->num_sgprs == 3);
3711
3712 radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0 +
3713 loc->sgpr_idx * 4, 3);
3714 radeon_emit(cs, blocks[0]);
3715 radeon_emit(cs, blocks[1]);
3716 radeon_emit(cs, blocks[2]);
3717 }
3718
3719 radeon_emit(cs, PKT3(PKT3_DISPATCH_DIRECT, 3, 0) |
3720 PKT3_SHADER_TYPE_S(1));
3721 radeon_emit(cs, blocks[0]);
3722 radeon_emit(cs, blocks[1]);
3723 radeon_emit(cs, blocks[2]);
3724 radeon_emit(cs, dispatch_initiator);
3725 }
3726
3727 assert(cmd_buffer->cs->cdw <= cdw_max);
3728 }
3729
3730 static void
3731 radv_upload_compute_shader_descriptors(struct radv_cmd_buffer *cmd_buffer)
3732 {
3733 radv_flush_descriptors(cmd_buffer, VK_SHADER_STAGE_COMPUTE_BIT);
3734 radv_flush_constants(cmd_buffer, cmd_buffer->state.compute_pipeline,
3735 VK_SHADER_STAGE_COMPUTE_BIT);
3736 }
3737
3738 static void
3739 radv_dispatch(struct radv_cmd_buffer *cmd_buffer,
3740 const struct radv_dispatch_info *info)
3741 {
3742 struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
3743 bool pipeline_is_dirty = pipeline &&
3744 pipeline != cmd_buffer->state.emitted_compute_pipeline;
3745
3746 if (cmd_buffer->state.flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3747 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3748 RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
3749 RADV_CMD_FLAG_CS_PARTIAL_FLUSH)) {
3750 /* If we have to wait for idle, set all states first, so that
3751 * all SET packets are processed in parallel with previous draw
3752 * calls. Then upload descriptors, set shader pointers, and
3753 * dispatch, and prefetch at the end. This ensures that the
3754 * time the CUs are idle is very short. (there are only SET_SH
3755 * packets between the wait and the draw)
3756 */
3757 radv_emit_compute_pipeline(cmd_buffer);
3758 si_emit_cache_flush(cmd_buffer);
3759 /* <-- CUs are idle here --> */
3760
3761 radv_upload_compute_shader_descriptors(cmd_buffer);
3762
3763 radv_emit_dispatch_packets(cmd_buffer, info);
3764 /* <-- CUs are busy here --> */
3765
3766 /* Start prefetches after the dispatch has been started. Both
3767 * will run in parallel, but starting the dispatch first is
3768 * more important.
3769 */
3770 if (pipeline_is_dirty) {
3771 radv_emit_shader_prefetch(cmd_buffer,
3772 pipeline->shaders[MESA_SHADER_COMPUTE]);
3773 }
3774 } else {
3775 /* If we don't wait for idle, start prefetches first, then set
3776 * states, and dispatch at the end.
3777 */
3778 si_emit_cache_flush(cmd_buffer);
3779
3780 if (pipeline_is_dirty) {
3781 radv_emit_shader_prefetch(cmd_buffer,
3782 pipeline->shaders[MESA_SHADER_COMPUTE]);
3783 }
3784
3785 radv_upload_compute_shader_descriptors(cmd_buffer);
3786
3787 radv_emit_compute_pipeline(cmd_buffer);
3788 radv_emit_dispatch_packets(cmd_buffer, info);
3789 }
3790
3791 radv_cmd_buffer_after_draw(cmd_buffer);
3792 }
3793
3794 void radv_CmdDispatch(
3795 VkCommandBuffer commandBuffer,
3796 uint32_t x,
3797 uint32_t y,
3798 uint32_t z)
3799 {
3800 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3801 struct radv_dispatch_info info = {};
3802
3803 info.blocks[0] = x;
3804 info.blocks[1] = y;
3805 info.blocks[2] = z;
3806
3807 radv_dispatch(cmd_buffer, &info);
3808 }
3809
3810 void radv_CmdDispatchIndirect(
3811 VkCommandBuffer commandBuffer,
3812 VkBuffer _buffer,
3813 VkDeviceSize offset)
3814 {
3815 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3816 RADV_FROM_HANDLE(radv_buffer, buffer, _buffer);
3817 struct radv_dispatch_info info = {};
3818
3819 info.indirect = buffer;
3820 info.indirect_offset = offset;
3821
3822 radv_dispatch(cmd_buffer, &info);
3823 }
3824
3825 void radv_unaligned_dispatch(
3826 struct radv_cmd_buffer *cmd_buffer,
3827 uint32_t x,
3828 uint32_t y,
3829 uint32_t z)
3830 {
3831 struct radv_dispatch_info info = {};
3832
3833 info.blocks[0] = x;
3834 info.blocks[1] = y;
3835 info.blocks[2] = z;
3836 info.unaligned = 1;
3837
3838 radv_dispatch(cmd_buffer, &info);
3839 }
3840
3841 void radv_CmdEndRenderPass(
3842 VkCommandBuffer commandBuffer)
3843 {
3844 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
3845
3846 radv_subpass_barrier(cmd_buffer, &cmd_buffer->state.pass->end_barrier);
3847
3848 radv_cmd_buffer_resolve_subpass(cmd_buffer);
3849
3850 for (unsigned i = 0; i < cmd_buffer->state.framebuffer->attachment_count; ++i) {
3851 VkImageLayout layout = cmd_buffer->state.pass->attachments[i].final_layout;
3852 radv_handle_subpass_image_transition(cmd_buffer,
3853 (VkAttachmentReference){i, layout});
3854 }
3855
3856 vk_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
3857
3858 cmd_buffer->state.pass = NULL;
3859 cmd_buffer->state.subpass = NULL;
3860 cmd_buffer->state.attachments = NULL;
3861 cmd_buffer->state.framebuffer = NULL;
3862 }
3863
3864 /*
3865 * For HTILE we have the following interesting clear words:
3866 * 0x0000030f: Uncompressed for depth+stencil HTILE.
3867 * 0x0000000f: Uncompressed for depth only HTILE.
3868 * 0xfffffff0: Clear depth to 1.0
3869 * 0x00000000: Clear depth to 0.0
3870 */
3871 static void radv_initialize_htile(struct radv_cmd_buffer *cmd_buffer,
3872 struct radv_image *image,
3873 const VkImageSubresourceRange *range,
3874 uint32_t clear_word)
3875 {
3876 assert(range->baseMipLevel == 0);
3877 assert(range->levelCount == 1 || range->levelCount == VK_REMAINING_ARRAY_LAYERS);
3878 unsigned layer_count = radv_get_layerCount(image, range);
3879 uint64_t size = image->surface.htile_slice_size * layer_count;
3880 uint64_t offset = image->offset + image->htile_offset +
3881 image->surface.htile_slice_size * range->baseArrayLayer;
3882 struct radv_cmd_state *state = &cmd_buffer->state;
3883
3884 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3885 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3886
3887 state->flush_bits |= radv_fill_buffer(cmd_buffer, image->bo, offset,
3888 size, clear_word);
3889
3890 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3891 }
3892
3893 static void radv_handle_depth_image_transition(struct radv_cmd_buffer *cmd_buffer,
3894 struct radv_image *image,
3895 VkImageLayout src_layout,
3896 VkImageLayout dst_layout,
3897 unsigned src_queue_mask,
3898 unsigned dst_queue_mask,
3899 const VkImageSubresourceRange *range,
3900 VkImageAspectFlags pending_clears)
3901 {
3902 if (dst_layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL &&
3903 (pending_clears & vk_format_aspects(image->vk_format)) == vk_format_aspects(image->vk_format) &&
3904 cmd_buffer->state.render_area.offset.x == 0 && cmd_buffer->state.render_area.offset.y == 0 &&
3905 cmd_buffer->state.render_area.extent.width == image->info.width &&
3906 cmd_buffer->state.render_area.extent.height == image->info.height) {
3907 /* The clear will initialize htile. */
3908 return;
3909 } else if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED &&
3910 radv_layout_has_htile(image, dst_layout, dst_queue_mask)) {
3911 /* TODO: merge with the clear if applicable */
3912 radv_initialize_htile(cmd_buffer, image, range, 0);
3913 } else if (!radv_layout_is_htile_compressed(image, src_layout, src_queue_mask) &&
3914 radv_layout_is_htile_compressed(image, dst_layout, dst_queue_mask)) {
3915 uint32_t clear_value = vk_format_is_stencil(image->vk_format) ? 0x30f : 0xf;
3916 radv_initialize_htile(cmd_buffer, image, range, clear_value);
3917 } else if (radv_layout_is_htile_compressed(image, src_layout, src_queue_mask) &&
3918 !radv_layout_is_htile_compressed(image, dst_layout, dst_queue_mask)) {
3919 VkImageSubresourceRange local_range = *range;
3920 local_range.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
3921 local_range.baseMipLevel = 0;
3922 local_range.levelCount = 1;
3923
3924 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3925 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3926
3927 radv_decompress_depth_image_inplace(cmd_buffer, image, &local_range);
3928
3929 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
3930 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
3931 }
3932 }
3933
3934 void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
3935 struct radv_image *image, uint32_t value)
3936 {
3937 struct radv_cmd_state *state = &cmd_buffer->state;
3938
3939 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3940 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3941
3942 state->flush_bits |= radv_fill_buffer(cmd_buffer, image->bo,
3943 image->offset + image->cmask.offset,
3944 image->cmask.size, value);
3945
3946 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3947 }
3948
3949 static void radv_handle_cmask_image_transition(struct radv_cmd_buffer *cmd_buffer,
3950 struct radv_image *image,
3951 VkImageLayout src_layout,
3952 VkImageLayout dst_layout,
3953 unsigned src_queue_mask,
3954 unsigned dst_queue_mask,
3955 const VkImageSubresourceRange *range)
3956 {
3957 if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
3958 if (image->fmask.size)
3959 radv_initialise_cmask(cmd_buffer, image, 0xccccccccu);
3960 else
3961 radv_initialise_cmask(cmd_buffer, image, 0xffffffffu);
3962 } else if (radv_layout_can_fast_clear(image, src_layout, src_queue_mask) &&
3963 !radv_layout_can_fast_clear(image, dst_layout, dst_queue_mask)) {
3964 radv_fast_clear_flush_image_inplace(cmd_buffer, image, range);
3965 }
3966 }
3967
3968 void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
3969 struct radv_image *image, uint32_t value)
3970 {
3971 struct radv_cmd_state *state = &cmd_buffer->state;
3972
3973 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3974 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3975
3976 state->flush_bits |= radv_fill_buffer(cmd_buffer, image->bo,
3977 image->offset + image->dcc_offset,
3978 image->surface.dcc_size, value);
3979
3980 state->flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
3981 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
3982 }
3983
3984 static void radv_handle_dcc_image_transition(struct radv_cmd_buffer *cmd_buffer,
3985 struct radv_image *image,
3986 VkImageLayout src_layout,
3987 VkImageLayout dst_layout,
3988 unsigned src_queue_mask,
3989 unsigned dst_queue_mask,
3990 const VkImageSubresourceRange *range)
3991 {
3992 if (src_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
3993 radv_initialize_dcc(cmd_buffer, image,
3994 radv_layout_dcc_compressed(image, dst_layout, dst_queue_mask) ?
3995 0x20202020u : 0xffffffffu);
3996 } else if (radv_layout_dcc_compressed(image, src_layout, src_queue_mask) &&
3997 !radv_layout_dcc_compressed(image, dst_layout, dst_queue_mask)) {
3998 radv_decompress_dcc(cmd_buffer, image, range);
3999 } else if (radv_layout_can_fast_clear(image, src_layout, src_queue_mask) &&
4000 !radv_layout_can_fast_clear(image, dst_layout, dst_queue_mask)) {
4001 radv_fast_clear_flush_image_inplace(cmd_buffer, image, range);
4002 }
4003 }
4004
4005 static void radv_handle_image_transition(struct radv_cmd_buffer *cmd_buffer,
4006 struct radv_image *image,
4007 VkImageLayout src_layout,
4008 VkImageLayout dst_layout,
4009 uint32_t src_family,
4010 uint32_t dst_family,
4011 const VkImageSubresourceRange *range,
4012 VkImageAspectFlags pending_clears)
4013 {
4014 if (image->exclusive && src_family != dst_family) {
4015 /* This is an acquire or a release operation and there will be
4016 * a corresponding release/acquire. Do the transition in the
4017 * most flexible queue. */
4018
4019 assert(src_family == cmd_buffer->queue_family_index ||
4020 dst_family == cmd_buffer->queue_family_index);
4021
4022 if (cmd_buffer->queue_family_index == RADV_QUEUE_TRANSFER)
4023 return;
4024
4025 if (cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE &&
4026 (src_family == RADV_QUEUE_GENERAL ||
4027 dst_family == RADV_QUEUE_GENERAL))
4028 return;
4029 }
4030
4031 unsigned src_queue_mask = radv_image_queue_family_mask(image, src_family, cmd_buffer->queue_family_index);
4032 unsigned dst_queue_mask = radv_image_queue_family_mask(image, dst_family, cmd_buffer->queue_family_index);
4033
4034 if (image->surface.htile_size)
4035 radv_handle_depth_image_transition(cmd_buffer, image, src_layout,
4036 dst_layout, src_queue_mask,
4037 dst_queue_mask, range,
4038 pending_clears);
4039
4040 if (image->cmask.size || image->fmask.size)
4041 radv_handle_cmask_image_transition(cmd_buffer, image, src_layout,
4042 dst_layout, src_queue_mask,
4043 dst_queue_mask, range);
4044
4045 if (image->surface.dcc_size)
4046 radv_handle_dcc_image_transition(cmd_buffer, image, src_layout,
4047 dst_layout, src_queue_mask,
4048 dst_queue_mask, range);
4049 }
4050
4051 void radv_CmdPipelineBarrier(
4052 VkCommandBuffer commandBuffer,
4053 VkPipelineStageFlags srcStageMask,
4054 VkPipelineStageFlags destStageMask,
4055 VkBool32 byRegion,
4056 uint32_t memoryBarrierCount,
4057 const VkMemoryBarrier* pMemoryBarriers,
4058 uint32_t bufferMemoryBarrierCount,
4059 const VkBufferMemoryBarrier* pBufferMemoryBarriers,
4060 uint32_t imageMemoryBarrierCount,
4061 const VkImageMemoryBarrier* pImageMemoryBarriers)
4062 {
4063 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
4064 enum radv_cmd_flush_bits src_flush_bits = 0;
4065 enum radv_cmd_flush_bits dst_flush_bits = 0;
4066
4067 for (uint32_t i = 0; i < memoryBarrierCount; i++) {
4068 src_flush_bits |= radv_src_access_flush(cmd_buffer, pMemoryBarriers[i].srcAccessMask);
4069 dst_flush_bits |= radv_dst_access_flush(cmd_buffer, pMemoryBarriers[i].dstAccessMask,
4070 NULL);
4071 }
4072
4073 for (uint32_t i = 0; i < bufferMemoryBarrierCount; i++) {
4074 src_flush_bits |= radv_src_access_flush(cmd_buffer, pBufferMemoryBarriers[i].srcAccessMask);
4075 dst_flush_bits |= radv_dst_access_flush(cmd_buffer, pBufferMemoryBarriers[i].dstAccessMask,
4076 NULL);
4077 }
4078
4079 for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
4080 RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
4081 src_flush_bits |= radv_src_access_flush(cmd_buffer, pImageMemoryBarriers[i].srcAccessMask);
4082 dst_flush_bits |= radv_dst_access_flush(cmd_buffer, pImageMemoryBarriers[i].dstAccessMask,
4083 image);
4084 }
4085
4086 radv_stage_flush(cmd_buffer, srcStageMask);
4087 cmd_buffer->state.flush_bits |= src_flush_bits;
4088
4089 for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
4090 RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
4091 radv_handle_image_transition(cmd_buffer, image,
4092 pImageMemoryBarriers[i].oldLayout,
4093 pImageMemoryBarriers[i].newLayout,
4094 pImageMemoryBarriers[i].srcQueueFamilyIndex,
4095 pImageMemoryBarriers[i].dstQueueFamilyIndex,
4096 &pImageMemoryBarriers[i].subresourceRange,
4097 0);
4098 }
4099
4100 cmd_buffer->state.flush_bits |= dst_flush_bits;
4101 }
4102
4103
4104 static void write_event(struct radv_cmd_buffer *cmd_buffer,
4105 struct radv_event *event,
4106 VkPipelineStageFlags stageMask,
4107 unsigned value)
4108 {
4109 struct radeon_winsys_cs *cs = cmd_buffer->cs;
4110 uint64_t va = radv_buffer_get_va(event->bo);
4111
4112 radv_cs_add_buffer(cmd_buffer->device->ws, cs, event->bo, 8);
4113
4114 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 18);
4115
4116 /* TODO: this is overkill. Probably should figure something out from
4117 * the stage mask. */
4118
4119 si_cs_emit_write_event_eop(cs,
4120 cmd_buffer->state.predicating,
4121 cmd_buffer->device->physical_device->rad_info.chip_class,
4122 radv_cmd_buffer_uses_mec(cmd_buffer),
4123 V_028A90_BOTTOM_OF_PIPE_TS, 0,
4124 1, va, 2, value);
4125
4126 assert(cmd_buffer->cs->cdw <= cdw_max);
4127 }
4128
4129 void radv_CmdSetEvent(VkCommandBuffer commandBuffer,
4130 VkEvent _event,
4131 VkPipelineStageFlags stageMask)
4132 {
4133 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
4134 RADV_FROM_HANDLE(radv_event, event, _event);
4135
4136 write_event(cmd_buffer, event, stageMask, 1);
4137 }
4138
4139 void radv_CmdResetEvent(VkCommandBuffer commandBuffer,
4140 VkEvent _event,
4141 VkPipelineStageFlags stageMask)
4142 {
4143 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
4144 RADV_FROM_HANDLE(radv_event, event, _event);
4145
4146 write_event(cmd_buffer, event, stageMask, 0);
4147 }
4148
4149 void radv_CmdWaitEvents(VkCommandBuffer commandBuffer,
4150 uint32_t eventCount,
4151 const VkEvent* pEvents,
4152 VkPipelineStageFlags srcStageMask,
4153 VkPipelineStageFlags dstStageMask,
4154 uint32_t memoryBarrierCount,
4155 const VkMemoryBarrier* pMemoryBarriers,
4156 uint32_t bufferMemoryBarrierCount,
4157 const VkBufferMemoryBarrier* pBufferMemoryBarriers,
4158 uint32_t imageMemoryBarrierCount,
4159 const VkImageMemoryBarrier* pImageMemoryBarriers)
4160 {
4161 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
4162 struct radeon_winsys_cs *cs = cmd_buffer->cs;
4163
4164 for (unsigned i = 0; i < eventCount; ++i) {
4165 RADV_FROM_HANDLE(radv_event, event, pEvents[i]);
4166 uint64_t va = radv_buffer_get_va(event->bo);
4167
4168 radv_cs_add_buffer(cmd_buffer->device->ws, cs, event->bo, 8);
4169
4170 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 7);
4171
4172 si_emit_wait_fence(cs, false, va, 1, 0xffffffff);
4173 assert(cmd_buffer->cs->cdw <= cdw_max);
4174 }
4175
4176
4177 for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
4178 RADV_FROM_HANDLE(radv_image, image, pImageMemoryBarriers[i].image);
4179
4180 radv_handle_image_transition(cmd_buffer, image,
4181 pImageMemoryBarriers[i].oldLayout,
4182 pImageMemoryBarriers[i].newLayout,
4183 pImageMemoryBarriers[i].srcQueueFamilyIndex,
4184 pImageMemoryBarriers[i].dstQueueFamilyIndex,
4185 &pImageMemoryBarriers[i].subresourceRange,
4186 0);
4187 }
4188
4189 /* TODO: figure out how to do memory barriers without waiting */
4190 cmd_buffer->state.flush_bits |= RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER |
4191 RADV_CMD_FLAG_INV_GLOBAL_L2 |
4192 RADV_CMD_FLAG_INV_VMEM_L1 |
4193 RADV_CMD_FLAG_INV_SMEM_L1;
4194 }