projects / mesa.git / blob
? search:


1bdcc998c65570bedb0c9f4c69e7a8c210c93b7d
[mesa.git] / src / amd / vulkan / radv_query.c
1 /*
2 * Copyrigh 2016 Red Hat Inc.
3 * Based on anv:
4 * Copyright © 2015 Intel Corporation
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
23 * IN THE SOFTWARE.
24 */
25
26 #include <assert.h>
27 #include <stdbool.h>
28 #include <string.h>
29 #include <unistd.h>
30 #include <fcntl.h>
31
32 #include "nir/nir_builder.h"
33 #include "radv_meta.h"
34 #include "radv_private.h"
35 #include "radv_cs.h"
36 #include "sid.h"
37
38 #define TIMESTAMP_NOT_READY UINT64_MAX
39
40 static const int pipelinestat_block_size = 11 * 8;
41 static const unsigned pipeline_statistics_indices[] = {7, 6, 3, 4, 5, 2, 1, 0, 8, 9, 10};
42
43 static nir_ssa_def *nir_test_flag(nir_builder *b, nir_ssa_def *flags, uint32_t flag)
44 {
45 return nir_i2b(b, nir_iand(b, flags, nir_imm_int(b, flag)));
46 }
47
48 static void radv_break_on_count(nir_builder *b, nir_variable *var, nir_ssa_def *count)
49 {
50 nir_ssa_def *counter = nir_load_var(b, var);
51
52 nir_if *if_stmt = nir_if_create(b->shader);
53 if_stmt->condition = nir_src_for_ssa(nir_uge(b, counter, count));
54 nir_cf_node_insert(b->cursor, &if_stmt->cf_node);
55
56 b->cursor = nir_after_cf_list(&if_stmt->then_list);
57
58 nir_jump_instr *instr = nir_jump_instr_create(b->shader, nir_jump_break);
59 nir_builder_instr_insert(b, &instr->instr);
60
61 b->cursor = nir_after_cf_node(&if_stmt->cf_node);
62 counter = nir_iadd(b, counter, nir_imm_int(b, 1));
63 nir_store_var(b, var, counter, 0x1);
64 }
65
66 static struct nir_ssa_def *
67 radv_load_push_int(nir_builder *b, unsigned offset, const char *name)
68 {
69 nir_intrinsic_instr *flags = nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_push_constant);
70 nir_intrinsic_set_base(flags, 0);
71 nir_intrinsic_set_range(flags, 16);
72 flags->src[0] = nir_src_for_ssa(nir_imm_int(b, offset));
73 flags->num_components = 1;
74 nir_ssa_dest_init(&flags->instr, &flags->dest, 1, 32, name);
75 nir_builder_instr_insert(b, &flags->instr);
76 return &flags->dest.ssa;
77 }
78
79 static nir_shader *
80 build_occlusion_query_shader(struct radv_device *device) {
81 /* the shader this builds is roughly
82 *
83 * push constants {
84 * uint32_t flags;
85 * uint32_t dst_stride;
86 * };
87 *
88 * uint32_t src_stride = 16 * db_count;
89 *
90 * location(binding = 0) buffer dst_buf;
91 * location(binding = 1) buffer src_buf;
92 *
93 * void main() {
94 * uint64_t result = 0;
95 * uint64_t src_offset = src_stride * global_id.x;
96 * uint64_t dst_offset = dst_stride * global_id.x;
97 * bool available = true;
98 * for (int i = 0; i < db_count; ++i) {
99 * if (enabled_rb_mask & (1 << i)) {
100 * uint64_t start = src_buf[src_offset + 16 * i];
101 * uint64_t end = src_buf[src_offset + 16 * i + 8];
102 * if ((start & (1ull << 63)) && (end & (1ull << 63)))
103 * result += end - start;
104 * else
105 * available = false;
106 * }
107 * }
108 * uint32_t elem_size = flags & VK_QUERY_RESULT_64_BIT ? 8 : 4;
109 * if ((flags & VK_QUERY_RESULT_PARTIAL_BIT) || available) {
110 * if (flags & VK_QUERY_RESULT_64_BIT)
111 * dst_buf[dst_offset] = result;
112 * else
113 * dst_buf[dst_offset] = (uint32_t)result.
114 * }
115 * if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
116 * dst_buf[dst_offset + elem_size] = available;
117 * }
118 * }
119 */
120 nir_builder b;
121 nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
122 b.shader->info.name = ralloc_strdup(b.shader, "occlusion_query");
123 b.shader->info.cs.local_size[0] = 64;
124 b.shader->info.cs.local_size[1] = 1;
125 b.shader->info.cs.local_size[2] = 1;
126
127 nir_variable *result = nir_local_variable_create(b.impl, glsl_uint64_t_type(), "result");
128 nir_variable *outer_counter = nir_local_variable_create(b.impl, glsl_int_type(), "outer_counter");
129 nir_variable *start = nir_local_variable_create(b.impl, glsl_uint64_t_type(), "start");
130 nir_variable *end = nir_local_variable_create(b.impl, glsl_uint64_t_type(), "end");
131 nir_variable *available = nir_local_variable_create(b.impl, glsl_bool_type(), "available");
132 unsigned enabled_rb_mask = device->physical_device->rad_info.enabled_rb_mask;
133 unsigned db_count = device->physical_device->rad_info.num_render_backends;
134
135 nir_ssa_def *flags = radv_load_push_int(&b, 0, "flags");
136
137 nir_intrinsic_instr *dst_buf = nir_intrinsic_instr_create(b.shader,
138 nir_intrinsic_vulkan_resource_index);
139 dst_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
140 dst_buf->num_components = 1;
141 nir_intrinsic_set_desc_set(dst_buf, 0);
142 nir_intrinsic_set_binding(dst_buf, 0);
143 nir_ssa_dest_init(&dst_buf->instr, &dst_buf->dest, dst_buf->num_components, 32, NULL);
144 nir_builder_instr_insert(&b, &dst_buf->instr);
145
146 nir_intrinsic_instr *src_buf = nir_intrinsic_instr_create(b.shader,
147 nir_intrinsic_vulkan_resource_index);
148 src_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
149 src_buf->num_components = 1;
150 nir_intrinsic_set_desc_set(src_buf, 0);
151 nir_intrinsic_set_binding(src_buf, 1);
152 nir_ssa_dest_init(&src_buf->instr, &src_buf->dest, src_buf->num_components, 32, NULL);
153 nir_builder_instr_insert(&b, &src_buf->instr);
154
155 nir_ssa_def *invoc_id = nir_load_local_invocation_id(&b);
156 nir_ssa_def *wg_id = nir_load_work_group_id(&b);
157 nir_ssa_def *block_size = nir_imm_ivec4(&b,
158 b.shader->info.cs.local_size[0],
159 b.shader->info.cs.local_size[1],
160 b.shader->info.cs.local_size[2], 0);
161 nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
162 global_id = nir_channel(&b, global_id, 0); // We only care about x here.
163
164 nir_ssa_def *input_stride = nir_imm_int(&b, db_count * 16);
165 nir_ssa_def *input_base = nir_imul(&b, input_stride, global_id);
166 nir_ssa_def *output_stride = radv_load_push_int(&b, 4, "output_stride");
167 nir_ssa_def *output_base = nir_imul(&b, output_stride, global_id);
168
169
170 nir_store_var(&b, result, nir_imm_int64(&b, 0), 0x1);
171 nir_store_var(&b, outer_counter, nir_imm_int(&b, 0), 0x1);
172 nir_store_var(&b, available, nir_imm_true(&b), 0x1);
173
174 nir_loop *outer_loop = nir_loop_create(b.shader);
175 nir_builder_cf_insert(&b, &outer_loop->cf_node);
176 b.cursor = nir_after_cf_list(&outer_loop->body);
177
178 nir_ssa_def *current_outer_count = nir_load_var(&b, outer_counter);
179 radv_break_on_count(&b, outer_counter, nir_imm_int(&b, db_count));
180
181 nir_ssa_def *enabled_cond =
182 nir_iand(&b, nir_imm_int(&b, enabled_rb_mask),
183 nir_ishl(&b, nir_imm_int(&b, 1), current_outer_count));
184
185 nir_if *enabled_if = nir_if_create(b.shader);
186 enabled_if->condition = nir_src_for_ssa(nir_i2b(&b, enabled_cond));
187 nir_cf_node_insert(b.cursor, &enabled_if->cf_node);
188
189 b.cursor = nir_after_cf_list(&enabled_if->then_list);
190
191 nir_ssa_def *load_offset = nir_imul(&b, current_outer_count, nir_imm_int(&b, 16));
192 load_offset = nir_iadd(&b, input_base, load_offset);
193
194 nir_intrinsic_instr *load = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
195 load->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
196 load->src[1] = nir_src_for_ssa(load_offset);
197 nir_ssa_dest_init(&load->instr, &load->dest, 2, 64, NULL);
198 load->num_components = 2;
199 nir_builder_instr_insert(&b, &load->instr);
200
201 nir_store_var(&b, start, nir_channel(&b, &load->dest.ssa, 0), 0x1);
202 nir_store_var(&b, end, nir_channel(&b, &load->dest.ssa, 1), 0x1);
203
204 nir_ssa_def *start_done = nir_ilt(&b, nir_load_var(&b, start), nir_imm_int64(&b, 0));
205 nir_ssa_def *end_done = nir_ilt(&b, nir_load_var(&b, end), nir_imm_int64(&b, 0));
206
207 nir_if *update_if = nir_if_create(b.shader);
208 update_if->condition = nir_src_for_ssa(nir_iand(&b, start_done, end_done));
209 nir_cf_node_insert(b.cursor, &update_if->cf_node);
210
211 b.cursor = nir_after_cf_list(&update_if->then_list);
212
213 nir_store_var(&b, result,
214 nir_iadd(&b, nir_load_var(&b, result),
215 nir_isub(&b, nir_load_var(&b, end),
216 nir_load_var(&b, start))), 0x1);
217
218 b.cursor = nir_after_cf_list(&update_if->else_list);
219
220 nir_store_var(&b, available, nir_imm_false(&b), 0x1);
221
222 b.cursor = nir_after_cf_node(&outer_loop->cf_node);
223
224 /* Store the result if complete or if partial results have been requested. */
225
226 nir_ssa_def *result_is_64bit = nir_test_flag(&b, flags, VK_QUERY_RESULT_64_BIT);
227 nir_ssa_def *result_size = nir_bcsel(&b, result_is_64bit, nir_imm_int(&b, 8), nir_imm_int(&b, 4));
228
229 nir_if *store_if = nir_if_create(b.shader);
230 store_if->condition = nir_src_for_ssa(nir_ior(&b, nir_test_flag(&b, flags, VK_QUERY_RESULT_PARTIAL_BIT), nir_load_var(&b, available)));
231 nir_cf_node_insert(b.cursor, &store_if->cf_node);
232
233 b.cursor = nir_after_cf_list(&store_if->then_list);
234
235 nir_if *store_64bit_if = nir_if_create(b.shader);
236 store_64bit_if->condition = nir_src_for_ssa(result_is_64bit);
237 nir_cf_node_insert(b.cursor, &store_64bit_if->cf_node);
238
239 b.cursor = nir_after_cf_list(&store_64bit_if->then_list);
240
241 nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
242 store->src[0] = nir_src_for_ssa(nir_load_var(&b, result));
243 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
244 store->src[2] = nir_src_for_ssa(output_base);
245 nir_intrinsic_set_write_mask(store, 0x1);
246 store->num_components = 1;
247 nir_builder_instr_insert(&b, &store->instr);
248
249 b.cursor = nir_after_cf_list(&store_64bit_if->else_list);
250
251 store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
252 store->src[0] = nir_src_for_ssa(nir_u2u32(&b, nir_load_var(&b, result)));
253 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
254 store->src[2] = nir_src_for_ssa(output_base);
255 nir_intrinsic_set_write_mask(store, 0x1);
256 store->num_components = 1;
257 nir_builder_instr_insert(&b, &store->instr);
258
259 b.cursor = nir_after_cf_node(&store_if->cf_node);
260
261 /* Store the availability bit if requested. */
262
263 nir_if *availability_if = nir_if_create(b.shader);
264 availability_if->condition = nir_src_for_ssa(nir_test_flag(&b, flags, VK_QUERY_RESULT_WITH_AVAILABILITY_BIT));
265 nir_cf_node_insert(b.cursor, &availability_if->cf_node);
266
267 b.cursor = nir_after_cf_list(&availability_if->then_list);
268
269 store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
270 store->src[0] = nir_src_for_ssa(nir_b2i32(&b, nir_load_var(&b, available)));
271 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
272 store->src[2] = nir_src_for_ssa(nir_iadd(&b, result_size, output_base));
273 nir_intrinsic_set_write_mask(store, 0x1);
274 store->num_components = 1;
275 nir_builder_instr_insert(&b, &store->instr);
276
277 return b.shader;
278 }
279
280 static nir_shader *
281 build_pipeline_statistics_query_shader(struct radv_device *device) {
282 /* the shader this builds is roughly
283 *
284 * push constants {
285 * uint32_t flags;
286 * uint32_t dst_stride;
287 * uint32_t stats_mask;
288 * uint32_t avail_offset;
289 * };
290 *
291 * uint32_t src_stride = pipelinestat_block_size * 2;
292 *
293 * location(binding = 0) buffer dst_buf;
294 * location(binding = 1) buffer src_buf;
295 *
296 * void main() {
297 * uint64_t src_offset = src_stride * global_id.x;
298 * uint64_t dst_base = dst_stride * global_id.x;
299 * uint64_t dst_offset = dst_base;
300 * uint32_t elem_size = flags & VK_QUERY_RESULT_64_BIT ? 8 : 4;
301 * uint32_t elem_count = stats_mask >> 16;
302 * uint32_t available32 = src_buf[avail_offset + 4 * global_id.x];
303 * if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
304 * dst_buf[dst_offset + elem_count * elem_size] = available32;
305 * }
306 * if ((bool)available32) {
307 * // repeat 11 times:
308 * if (stats_mask & (1 << 0)) {
309 * uint64_t start = src_buf[src_offset + 8 * indices[0]];
310 * uint64_t end = src_buf[src_offset + 8 * indices[0] + pipelinestat_block_size];
311 * uint64_t result = end - start;
312 * if (flags & VK_QUERY_RESULT_64_BIT)
313 * dst_buf[dst_offset] = result;
314 * else
315 * dst_buf[dst_offset] = (uint32_t)result.
316 * dst_offset += elem_size;
317 * }
318 * } else if (flags & VK_QUERY_RESULT_PARTIAL_BIT) {
319 * // Set everything to 0 as we don't know what is valid.
320 * for (int i = 0; i < elem_count; ++i)
321 * dst_buf[dst_base + elem_size * i] = 0;
322 * }
323 * }
324 */
325 nir_builder b;
326 nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
327 b.shader->info.name = ralloc_strdup(b.shader, "pipeline_statistics_query");
328 b.shader->info.cs.local_size[0] = 64;
329 b.shader->info.cs.local_size[1] = 1;
330 b.shader->info.cs.local_size[2] = 1;
331
332 nir_variable *output_offset = nir_local_variable_create(b.impl, glsl_int_type(), "output_offset");
333
334 nir_ssa_def *flags = radv_load_push_int(&b, 0, "flags");
335 nir_ssa_def *stats_mask = radv_load_push_int(&b, 8, "stats_mask");
336 nir_ssa_def *avail_offset = radv_load_push_int(&b, 12, "avail_offset");
337
338 nir_intrinsic_instr *dst_buf = nir_intrinsic_instr_create(b.shader,
339 nir_intrinsic_vulkan_resource_index);
340 dst_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
341 dst_buf->num_components = 1;;
342 nir_intrinsic_set_desc_set(dst_buf, 0);
343 nir_intrinsic_set_binding(dst_buf, 0);
344 nir_ssa_dest_init(&dst_buf->instr, &dst_buf->dest, dst_buf->num_components, 32, NULL);
345 nir_builder_instr_insert(&b, &dst_buf->instr);
346
347 nir_intrinsic_instr *src_buf = nir_intrinsic_instr_create(b.shader,
348 nir_intrinsic_vulkan_resource_index);
349 src_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
350 src_buf->num_components = 1;
351 nir_intrinsic_set_desc_set(src_buf, 0);
352 nir_intrinsic_set_binding(src_buf, 1);
353 nir_ssa_dest_init(&src_buf->instr, &src_buf->dest, src_buf->num_components, 32, NULL);
354 nir_builder_instr_insert(&b, &src_buf->instr);
355
356 nir_ssa_def *invoc_id = nir_load_local_invocation_id(&b);
357 nir_ssa_def *wg_id = nir_load_work_group_id(&b);
358 nir_ssa_def *block_size = nir_imm_ivec4(&b,
359 b.shader->info.cs.local_size[0],
360 b.shader->info.cs.local_size[1],
361 b.shader->info.cs.local_size[2], 0);
362 nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
363 global_id = nir_channel(&b, global_id, 0); // We only care about x here.
364
365 nir_ssa_def *input_stride = nir_imm_int(&b, pipelinestat_block_size * 2);
366 nir_ssa_def *input_base = nir_imul(&b, input_stride, global_id);
367 nir_ssa_def *output_stride = radv_load_push_int(&b, 4, "output_stride");
368 nir_ssa_def *output_base = nir_imul(&b, output_stride, global_id);
369
370
371 avail_offset = nir_iadd(&b, avail_offset,
372 nir_imul(&b, global_id, nir_imm_int(&b, 4)));
373
374 nir_intrinsic_instr *load = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
375 load->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
376 load->src[1] = nir_src_for_ssa(avail_offset);
377 nir_ssa_dest_init(&load->instr, &load->dest, 1, 32, NULL);
378 load->num_components = 1;
379 nir_builder_instr_insert(&b, &load->instr);
380 nir_ssa_def *available32 = &load->dest.ssa;
381
382 nir_ssa_def *result_is_64bit = nir_test_flag(&b, flags, VK_QUERY_RESULT_64_BIT);
383 nir_ssa_def *elem_size = nir_bcsel(&b, result_is_64bit, nir_imm_int(&b, 8), nir_imm_int(&b, 4));
384 nir_ssa_def *elem_count = nir_ushr(&b, stats_mask, nir_imm_int(&b, 16));
385
386 /* Store the availability bit if requested. */
387
388 nir_if *availability_if = nir_if_create(b.shader);
389 availability_if->condition = nir_src_for_ssa(nir_test_flag(&b, flags, VK_QUERY_RESULT_WITH_AVAILABILITY_BIT));
390 nir_cf_node_insert(b.cursor, &availability_if->cf_node);
391
392 b.cursor = nir_after_cf_list(&availability_if->then_list);
393
394 nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
395 store->src[0] = nir_src_for_ssa(available32);
396 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
397 store->src[2] = nir_src_for_ssa(nir_iadd(&b, output_base, nir_imul(&b, elem_count, elem_size)));
398 nir_intrinsic_set_write_mask(store, 0x1);
399 store->num_components = 1;
400 nir_builder_instr_insert(&b, &store->instr);
401
402 b.cursor = nir_after_cf_node(&availability_if->cf_node);
403
404 nir_if *available_if = nir_if_create(b.shader);
405 available_if->condition = nir_src_for_ssa(nir_i2b(&b, available32));
406 nir_cf_node_insert(b.cursor, &available_if->cf_node);
407
408 b.cursor = nir_after_cf_list(&available_if->then_list);
409
410 nir_store_var(&b, output_offset, output_base, 0x1);
411 for (int i = 0; i < 11; ++i) {
412 nir_if *store_if = nir_if_create(b.shader);
413 store_if->condition = nir_src_for_ssa(nir_test_flag(&b, stats_mask, 1u << i));
414 nir_cf_node_insert(b.cursor, &store_if->cf_node);
415
416 b.cursor = nir_after_cf_list(&store_if->then_list);
417
418 load = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
419 load->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
420 load->src[1] = nir_src_for_ssa(nir_iadd(&b, input_base,
421 nir_imm_int(&b, pipeline_statistics_indices[i] * 8)));
422 nir_ssa_dest_init(&load->instr, &load->dest, 1, 64, NULL);
423 load->num_components = 1;
424 nir_builder_instr_insert(&b, &load->instr);
425 nir_ssa_def *start = &load->dest.ssa;
426
427 load = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
428 load->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
429 load->src[1] = nir_src_for_ssa(nir_iadd(&b, input_base,
430 nir_imm_int(&b, pipeline_statistics_indices[i] * 8 + pipelinestat_block_size)));
431 nir_ssa_dest_init(&load->instr, &load->dest, 1, 64, NULL);
432 load->num_components = 1;
433 nir_builder_instr_insert(&b, &load->instr);
434 nir_ssa_def *end = &load->dest.ssa;
435
436 nir_ssa_def *result = nir_isub(&b, end, start);
437
438 /* Store result */
439 nir_if *store_64bit_if = nir_if_create(b.shader);
440 store_64bit_if->condition = nir_src_for_ssa(result_is_64bit);
441 nir_cf_node_insert(b.cursor, &store_64bit_if->cf_node);
442
443 b.cursor = nir_after_cf_list(&store_64bit_if->then_list);
444
445 nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
446 store->src[0] = nir_src_for_ssa(result);
447 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
448 store->src[2] = nir_src_for_ssa(nir_load_var(&b, output_offset));
449 nir_intrinsic_set_write_mask(store, 0x1);
450 store->num_components = 1;
451 nir_builder_instr_insert(&b, &store->instr);
452
453 b.cursor = nir_after_cf_list(&store_64bit_if->else_list);
454
455 store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
456 store->src[0] = nir_src_for_ssa(nir_u2u32(&b, result));
457 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
458 store->src[2] = nir_src_for_ssa(nir_load_var(&b, output_offset));
459 nir_intrinsic_set_write_mask(store, 0x1);
460 store->num_components = 1;
461 nir_builder_instr_insert(&b, &store->instr);
462
463 b.cursor = nir_after_cf_node(&store_64bit_if->cf_node);
464
465 nir_store_var(&b, output_offset,
466 nir_iadd(&b, nir_load_var(&b, output_offset),
467 elem_size), 0x1);
468
469 b.cursor = nir_after_cf_node(&store_if->cf_node);
470 }
471
472 b.cursor = nir_after_cf_list(&available_if->else_list);
473
474 available_if = nir_if_create(b.shader);
475 available_if->condition = nir_src_for_ssa(nir_test_flag(&b, flags, VK_QUERY_RESULT_PARTIAL_BIT));
476 nir_cf_node_insert(b.cursor, &available_if->cf_node);
477
478 b.cursor = nir_after_cf_list(&available_if->then_list);
479
480 /* Stores zeros in all outputs. */
481
482 nir_variable *counter = nir_local_variable_create(b.impl, glsl_int_type(), "counter");
483 nir_store_var(&b, counter, nir_imm_int(&b, 0), 0x1);
484
485 nir_loop *loop = nir_loop_create(b.shader);
486 nir_builder_cf_insert(&b, &loop->cf_node);
487 b.cursor = nir_after_cf_list(&loop->body);
488
489 nir_ssa_def *current_counter = nir_load_var(&b, counter);
490 radv_break_on_count(&b, counter, elem_count);
491
492 nir_ssa_def *output_elem = nir_iadd(&b, output_base,
493 nir_imul(&b, elem_size, current_counter));
494
495 nir_if *store_64bit_if = nir_if_create(b.shader);
496 store_64bit_if->condition = nir_src_for_ssa(result_is_64bit);
497 nir_cf_node_insert(b.cursor, &store_64bit_if->cf_node);
498
499 b.cursor = nir_after_cf_list(&store_64bit_if->then_list);
500
501 store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
502 store->src[0] = nir_src_for_ssa(nir_imm_int64(&b, 0));
503 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
504 store->src[2] = nir_src_for_ssa(output_elem);
505 nir_intrinsic_set_write_mask(store, 0x1);
506 store->num_components = 1;
507 nir_builder_instr_insert(&b, &store->instr);
508
509 b.cursor = nir_after_cf_list(&store_64bit_if->else_list);
510
511 store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
512 store->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
513 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
514 store->src[2] = nir_src_for_ssa(output_elem);
515 nir_intrinsic_set_write_mask(store, 0x1);
516 store->num_components = 1;
517 nir_builder_instr_insert(&b, &store->instr);
518
519 b.cursor = nir_after_cf_node(&loop->cf_node);
520 return b.shader;
521 }
522
523 static nir_shader *
524 build_tfb_query_shader(struct radv_device *device)
525 {
526 /* the shader this builds is roughly
527 *
528 * uint32_t src_stride = 32;
529 *
530 * location(binding = 0) buffer dst_buf;
531 * location(binding = 1) buffer src_buf;
532 *
533 * void main() {
534 * uint64_t result[2] = {};
535 * bool available = false;
536 * uint64_t src_offset = src_stride * global_id.x;
537 * uint64_t dst_offset = dst_stride * global_id.x;
538 * uint64_t *src_data = src_buf[src_offset];
539 * uint32_t avail = (src_data[0] >> 32) &
540 * (src_data[1] >> 32) &
541 * (src_data[2] >> 32) &
542 * (src_data[3] >> 32);
543 * if (avail & 0x80000000) {
544 * result[0] = src_data[3] - src_data[1];
545 * result[1] = src_data[2] - src_data[0];
546 * available = true;
547 * }
548 * uint32_t result_size = flags & VK_QUERY_RESULT_64_BIT ? 16 : 8;
549 * if ((flags & VK_QUERY_RESULT_PARTIAL_BIT) || available) {
550 * if (flags & VK_QUERY_RESULT_64_BIT) {
551 * dst_buf[dst_offset] = result;
552 * } else {
553 * dst_buf[dst_offset] = (uint32_t)result;
554 * }
555 * }
556 * if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
557 * dst_buf[dst_offset + result_size] = available;
558 * }
559 * }
560 */
561 nir_builder b;
562 nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
563 b.shader->info.name = ralloc_strdup(b.shader, "tfb_query");
564 b.shader->info.cs.local_size[0] = 64;
565 b.shader->info.cs.local_size[1] = 1;
566 b.shader->info.cs.local_size[2] = 1;
567
568 /* Create and initialize local variables. */
569 nir_variable *result =
570 nir_local_variable_create(b.impl,
571 glsl_vector_type(GLSL_TYPE_UINT64, 2),
572 "result");
573 nir_variable *available =
574 nir_local_variable_create(b.impl, glsl_bool_type(), "available");
575
576 nir_store_var(&b, result,
577 nir_vec2(&b, nir_imm_int64(&b, 0),
578 nir_imm_int64(&b, 0)), 0x3);
579 nir_store_var(&b, available, nir_imm_false(&b), 0x1);
580
581 nir_ssa_def *flags = radv_load_push_int(&b, 0, "flags");
582
583 /* Load resources. */
584 nir_intrinsic_instr *dst_buf = nir_intrinsic_instr_create(b.shader,
585 nir_intrinsic_vulkan_resource_index);
586 dst_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
587 dst_buf->num_components = 1;
588 nir_intrinsic_set_desc_set(dst_buf, 0);
589 nir_intrinsic_set_binding(dst_buf, 0);
590 nir_ssa_dest_init(&dst_buf->instr, &dst_buf->dest, dst_buf->num_components, 32, NULL);
591 nir_builder_instr_insert(&b, &dst_buf->instr);
592
593 nir_intrinsic_instr *src_buf = nir_intrinsic_instr_create(b.shader,
594 nir_intrinsic_vulkan_resource_index);
595 src_buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
596 src_buf->num_components = 1;
597 nir_intrinsic_set_desc_set(src_buf, 0);
598 nir_intrinsic_set_binding(src_buf, 1);
599 nir_ssa_dest_init(&src_buf->instr, &src_buf->dest, src_buf->num_components, 32, NULL);
600 nir_builder_instr_insert(&b, &src_buf->instr);
601
602 /* Compute global ID. */
603 nir_ssa_def *invoc_id = nir_load_local_invocation_id(&b);
604 nir_ssa_def *wg_id = nir_load_work_group_id(&b);
605 nir_ssa_def *block_size = nir_imm_ivec4(&b,
606 b.shader->info.cs.local_size[0],
607 b.shader->info.cs.local_size[1],
608 b.shader->info.cs.local_size[2], 0);
609 nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
610 global_id = nir_channel(&b, global_id, 0); // We only care about x here.
611
612 /* Compute src/dst strides. */
613 nir_ssa_def *input_stride = nir_imm_int(&b, 32);
614 nir_ssa_def *input_base = nir_imul(&b, input_stride, global_id);
615 nir_ssa_def *output_stride = radv_load_push_int(&b, 4, "output_stride");
616 nir_ssa_def *output_base = nir_imul(&b, output_stride, global_id);
617
618 /* Load data from the query pool. */
619 nir_intrinsic_instr *load1 = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
620 load1->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
621 load1->src[1] = nir_src_for_ssa(input_base);
622 nir_ssa_dest_init(&load1->instr, &load1->dest, 4, 32, NULL);
623 load1->num_components = 4;
624 nir_builder_instr_insert(&b, &load1->instr);
625
626 nir_intrinsic_instr *load2 = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
627 load2->src[0] = nir_src_for_ssa(&src_buf->dest.ssa);
628 load2->src[1] = nir_src_for_ssa(nir_iadd(&b, input_base, nir_imm_int(&b, 16)));
629 nir_ssa_dest_init(&load2->instr, &load2->dest, 4, 32, NULL);
630 load2->num_components = 4;
631 nir_builder_instr_insert(&b, &load2->instr);
632
633 /* Check if result is available. */
634 nir_ssa_def *avails[2];
635 avails[0] = nir_iand(&b, nir_channel(&b, &load1->dest.ssa, 1),
636 nir_channel(&b, &load1->dest.ssa, 3));
637 avails[1] = nir_iand(&b, nir_channel(&b, &load2->dest.ssa, 1),
638 nir_channel(&b, &load2->dest.ssa, 3));
639 nir_ssa_def *result_is_available =
640 nir_i2b(&b, nir_iand(&b, nir_iand(&b, avails[0], avails[1]),
641 nir_imm_int(&b, 0x80000000)));
642
643 /* Only compute result if available. */
644 nir_if *available_if = nir_if_create(b.shader);
645 available_if->condition = nir_src_for_ssa(result_is_available);
646 nir_cf_node_insert(b.cursor, &available_if->cf_node);
647
648 b.cursor = nir_after_cf_list(&available_if->then_list);
649
650 /* Pack values. */
651 nir_ssa_def *packed64[4];
652 packed64[0] = nir_pack_64_2x32(&b, nir_vec2(&b,
653 nir_channel(&b, &load1->dest.ssa, 0),
654 nir_channel(&b, &load1->dest.ssa, 1)));
655 packed64[1] = nir_pack_64_2x32(&b, nir_vec2(&b,
656 nir_channel(&b, &load1->dest.ssa, 2),
657 nir_channel(&b, &load1->dest.ssa, 3)));
658 packed64[2] = nir_pack_64_2x32(&b, nir_vec2(&b,
659 nir_channel(&b, &load2->dest.ssa, 0),
660 nir_channel(&b, &load2->dest.ssa, 1)));
661 packed64[3] = nir_pack_64_2x32(&b, nir_vec2(&b,
662 nir_channel(&b, &load2->dest.ssa, 2),
663 nir_channel(&b, &load2->dest.ssa, 3)));
664
665 /* Compute result. */
666 nir_ssa_def *num_primitive_written =
667 nir_isub(&b, packed64[3], packed64[1]);
668 nir_ssa_def *primitive_storage_needed =
669 nir_isub(&b, packed64[2], packed64[0]);
670
671 nir_store_var(&b, result,
672 nir_vec2(&b, num_primitive_written,
673 primitive_storage_needed), 0x3);
674 nir_store_var(&b, available, nir_imm_true(&b), 0x1);
675
676 b.cursor = nir_after_cf_node(&available_if->cf_node);
677
678 /* Determine if result is 64 or 32 bit. */
679 nir_ssa_def *result_is_64bit =
680 nir_test_flag(&b, flags, VK_QUERY_RESULT_64_BIT);
681 nir_ssa_def *result_size =
682 nir_bcsel(&b, result_is_64bit, nir_imm_int(&b, 16),
683 nir_imm_int(&b, 8));
684
685 /* Store the result if complete or partial results have been requested. */
686 nir_if *store_if = nir_if_create(b.shader);
687 store_if->condition =
688 nir_src_for_ssa(nir_ior(&b, nir_test_flag(&b, flags, VK_QUERY_RESULT_PARTIAL_BIT),
689 nir_load_var(&b, available)));
690 nir_cf_node_insert(b.cursor, &store_if->cf_node);
691
692 b.cursor = nir_after_cf_list(&store_if->then_list);
693
694 /* Store result. */
695 nir_if *store_64bit_if = nir_if_create(b.shader);
696 store_64bit_if->condition = nir_src_for_ssa(result_is_64bit);
697 nir_cf_node_insert(b.cursor, &store_64bit_if->cf_node);
698
699 b.cursor = nir_after_cf_list(&store_64bit_if->then_list);
700
701 nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
702 store->src[0] = nir_src_for_ssa(nir_load_var(&b, result));
703 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
704 store->src[2] = nir_src_for_ssa(output_base);
705 nir_intrinsic_set_write_mask(store, 0x3);
706 store->num_components = 2;
707 nir_builder_instr_insert(&b, &store->instr);
708
709 b.cursor = nir_after_cf_list(&store_64bit_if->else_list);
710
711 store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
712 store->src[0] = nir_src_for_ssa(nir_u2u32(&b, nir_load_var(&b, result)));
713 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
714 store->src[2] = nir_src_for_ssa(output_base);
715 nir_intrinsic_set_write_mask(store, 0x3);
716 store->num_components = 2;
717 nir_builder_instr_insert(&b, &store->instr);
718
719 b.cursor = nir_after_cf_node(&store_64bit_if->cf_node);
720
721 b.cursor = nir_after_cf_node(&store_if->cf_node);
722
723 /* Store the availability bit if requested. */
724 nir_if *availability_if = nir_if_create(b.shader);
725 availability_if->condition =
726 nir_src_for_ssa(nir_test_flag(&b, flags, VK_QUERY_RESULT_WITH_AVAILABILITY_BIT));
727 nir_cf_node_insert(b.cursor, &availability_if->cf_node);
728
729 b.cursor = nir_after_cf_list(&availability_if->then_list);
730
731 store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
732 store->src[0] = nir_src_for_ssa(nir_b2i32(&b, nir_load_var(&b, available)));
733 store->src[1] = nir_src_for_ssa(&dst_buf->dest.ssa);
734 store->src[2] = nir_src_for_ssa(nir_iadd(&b, result_size, output_base));
735 nir_intrinsic_set_write_mask(store, 0x1);
736 store->num_components = 1;
737 nir_builder_instr_insert(&b, &store->instr);
738
739 b.cursor = nir_after_cf_node(&availability_if->cf_node);
740
741 return b.shader;
742 }
743
744 static VkResult radv_device_init_meta_query_state_internal(struct radv_device *device)
745 {
746 VkResult result;
747 struct radv_shader_module occlusion_cs = { .nir = NULL };
748 struct radv_shader_module pipeline_statistics_cs = { .nir = NULL };
749 struct radv_shader_module tfb_cs = { .nir = NULL };
750
751 mtx_lock(&device->meta_state.mtx);
752 if (device->meta_state.query.pipeline_statistics_query_pipeline) {
753 mtx_unlock(&device->meta_state.mtx);
754 return VK_SUCCESS;
755 }
756 occlusion_cs.nir = build_occlusion_query_shader(device);
757 pipeline_statistics_cs.nir = build_pipeline_statistics_query_shader(device);
758 tfb_cs.nir = build_tfb_query_shader(device);
759
760 VkDescriptorSetLayoutCreateInfo occlusion_ds_create_info = {
761 .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
762 .flags = VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR,
763 .bindingCount = 2,
764 .pBindings = (VkDescriptorSetLayoutBinding[]) {
765 {
766 .binding = 0,
767 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
768 .descriptorCount = 1,
769 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
770 .pImmutableSamplers = NULL
771 },
772 {
773 .binding = 1,
774 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
775 .descriptorCount = 1,
776 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
777 .pImmutableSamplers = NULL
778 },
779 }
780 };
781
782 result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
783 &occlusion_ds_create_info,
784 &device->meta_state.alloc,
785 &device->meta_state.query.ds_layout);
786 if (result != VK_SUCCESS)
787 goto fail;
788
789 VkPipelineLayoutCreateInfo occlusion_pl_create_info = {
790 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
791 .setLayoutCount = 1,
792 .pSetLayouts = &device->meta_state.query.ds_layout,
793 .pushConstantRangeCount = 1,
794 .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_COMPUTE_BIT, 0, 16},
795 };
796
797 result = radv_CreatePipelineLayout(radv_device_to_handle(device),
798 &occlusion_pl_create_info,
799 &device->meta_state.alloc,
800 &device->meta_state.query.p_layout);
801 if (result != VK_SUCCESS)
802 goto fail;
803
804 VkPipelineShaderStageCreateInfo occlusion_pipeline_shader_stage = {
805 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
806 .stage = VK_SHADER_STAGE_COMPUTE_BIT,
807 .module = radv_shader_module_to_handle(&occlusion_cs),
808 .pName = "main",
809 .pSpecializationInfo = NULL,
810 };
811
812 VkComputePipelineCreateInfo occlusion_vk_pipeline_info = {
813 .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
814 .stage = occlusion_pipeline_shader_stage,
815 .flags = 0,
816 .layout = device->meta_state.query.p_layout,
817 };
818
819 result = radv_CreateComputePipelines(radv_device_to_handle(device),
820 radv_pipeline_cache_to_handle(&device->meta_state.cache),
821 1, &occlusion_vk_pipeline_info, NULL,
822 &device->meta_state.query.occlusion_query_pipeline);
823 if (result != VK_SUCCESS)
824 goto fail;
825
826 VkPipelineShaderStageCreateInfo pipeline_statistics_pipeline_shader_stage = {
827 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
828 .stage = VK_SHADER_STAGE_COMPUTE_BIT,
829 .module = radv_shader_module_to_handle(&pipeline_statistics_cs),
830 .pName = "main",
831 .pSpecializationInfo = NULL,
832 };
833
834 VkComputePipelineCreateInfo pipeline_statistics_vk_pipeline_info = {
835 .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
836 .stage = pipeline_statistics_pipeline_shader_stage,
837 .flags = 0,
838 .layout = device->meta_state.query.p_layout,
839 };
840
841 result = radv_CreateComputePipelines(radv_device_to_handle(device),
842 radv_pipeline_cache_to_handle(&device->meta_state.cache),
843 1, &pipeline_statistics_vk_pipeline_info, NULL,
844 &device->meta_state.query.pipeline_statistics_query_pipeline);
845 if (result != VK_SUCCESS)
846 goto fail;
847
848 VkPipelineShaderStageCreateInfo tfb_pipeline_shader_stage = {
849 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
850 .stage = VK_SHADER_STAGE_COMPUTE_BIT,
851 .module = radv_shader_module_to_handle(&tfb_cs),
852 .pName = "main",
853 .pSpecializationInfo = NULL,
854 };
855
856 VkComputePipelineCreateInfo tfb_pipeline_info = {
857 .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
858 .stage = tfb_pipeline_shader_stage,
859 .flags = 0,
860 .layout = device->meta_state.query.p_layout,
861 };
862
863 result = radv_CreateComputePipelines(radv_device_to_handle(device),
864 radv_pipeline_cache_to_handle(&device->meta_state.cache),
865 1, &tfb_pipeline_info, NULL,
866 &device->meta_state.query.tfb_query_pipeline);
867 fail:
868 if (result != VK_SUCCESS)
869 radv_device_finish_meta_query_state(device);
870 ralloc_free(occlusion_cs.nir);
871 ralloc_free(pipeline_statistics_cs.nir);
872 ralloc_free(tfb_cs.nir);
873 mtx_unlock(&device->meta_state.mtx);
874 return result;
875 }
876
877 VkResult radv_device_init_meta_query_state(struct radv_device *device, bool on_demand)
878 {
879 if (on_demand)
880 return VK_SUCCESS;
881
882 return radv_device_init_meta_query_state_internal(device);
883 }
884
885 void radv_device_finish_meta_query_state(struct radv_device *device)
886 {
887 if (device->meta_state.query.tfb_query_pipeline)
888 radv_DestroyPipeline(radv_device_to_handle(device),
889 device->meta_state.query.tfb_query_pipeline,
890 &device->meta_state.alloc);
891
892 if (device->meta_state.query.pipeline_statistics_query_pipeline)
893 radv_DestroyPipeline(radv_device_to_handle(device),
894 device->meta_state.query.pipeline_statistics_query_pipeline,
895 &device->meta_state.alloc);
896
897 if (device->meta_state.query.occlusion_query_pipeline)
898 radv_DestroyPipeline(radv_device_to_handle(device),
899 device->meta_state.query.occlusion_query_pipeline,
900 &device->meta_state.alloc);
901
902 if (device->meta_state.query.p_layout)
903 radv_DestroyPipelineLayout(radv_device_to_handle(device),
904 device->meta_state.query.p_layout,
905 &device->meta_state.alloc);
906
907 if (device->meta_state.query.ds_layout)
908 radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
909 device->meta_state.query.ds_layout,
910 &device->meta_state.alloc);
911 }
912
913 static void radv_query_shader(struct radv_cmd_buffer *cmd_buffer,
914 VkPipeline *pipeline,
915 struct radeon_winsys_bo *src_bo,
916 struct radeon_winsys_bo *dst_bo,
917 uint64_t src_offset, uint64_t dst_offset,
918 uint32_t src_stride, uint32_t dst_stride,
919 uint32_t count, uint32_t flags,
920 uint32_t pipeline_stats_mask, uint32_t avail_offset)
921 {
922 struct radv_device *device = cmd_buffer->device;
923 struct radv_meta_saved_state saved_state;
924 bool old_predicating;
925
926 if (!*pipeline) {
927 VkResult ret = radv_device_init_meta_query_state_internal(device);
928 if (ret != VK_SUCCESS) {
929 cmd_buffer->record_result = ret;
930 return;
931 }
932 }
933
934 radv_meta_save(&saved_state, cmd_buffer,
935 RADV_META_SAVE_COMPUTE_PIPELINE |
936 RADV_META_SAVE_CONSTANTS |
937 RADV_META_SAVE_DESCRIPTORS);
938
939 /* VK_EXT_conditional_rendering says that copy commands should not be
940 * affected by conditional rendering.
941 */
942 old_predicating = cmd_buffer->state.predicating;
943 cmd_buffer->state.predicating = false;
944
945 struct radv_buffer dst_buffer = {
946 .bo = dst_bo,
947 .offset = dst_offset,
948 .size = dst_stride * count
949 };
950
951 struct radv_buffer src_buffer = {
952 .bo = src_bo,
953 .offset = src_offset,
954 .size = MAX2(src_stride * count, avail_offset + 4 * count - src_offset)
955 };
956
957 radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
958 VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
959
960 radv_meta_push_descriptor_set(cmd_buffer,
961 VK_PIPELINE_BIND_POINT_COMPUTE,
962 device->meta_state.query.p_layout,
963 0, /* set */
964 2, /* descriptorWriteCount */
965 (VkWriteDescriptorSet[]) {
966 {
967 .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
968 .dstBinding = 0,
969 .dstArrayElement = 0,
970 .descriptorCount = 1,
971 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
972 .pBufferInfo = &(VkDescriptorBufferInfo) {
973 .buffer = radv_buffer_to_handle(&dst_buffer),
974 .offset = 0,
975 .range = VK_WHOLE_SIZE
976 }
977 },
978 {
979 .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
980 .dstBinding = 1,
981 .dstArrayElement = 0,
982 .descriptorCount = 1,
983 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
984 .pBufferInfo = &(VkDescriptorBufferInfo) {
985 .buffer = radv_buffer_to_handle(&src_buffer),
986 .offset = 0,
987 .range = VK_WHOLE_SIZE
988 }
989 }
990 });
991
992 /* Encode the number of elements for easy access by the shader. */
993 pipeline_stats_mask &= 0x7ff;
994 pipeline_stats_mask |= util_bitcount(pipeline_stats_mask) << 16;
995
996 avail_offset -= src_offset;
997
998 struct {
999 uint32_t flags;
1000 uint32_t dst_stride;
1001 uint32_t pipeline_stats_mask;
1002 uint32_t avail_offset;
1003 } push_constants = {
1004 flags,
1005 dst_stride,
1006 pipeline_stats_mask,
1007 avail_offset
1008 };
1009
1010 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
1011 device->meta_state.query.p_layout,
1012 VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(push_constants),
1013 &push_constants);
1014
1015 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_INV_L2 |
1016 RADV_CMD_FLAG_INV_VCACHE;
1017
1018 if (flags & VK_QUERY_RESULT_WAIT_BIT)
1019 cmd_buffer->state.flush_bits |= RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER;
1020
1021 radv_unaligned_dispatch(cmd_buffer, count, 1, 1);
1022
1023 /* Restore conditional rendering. */
1024 cmd_buffer->state.predicating = old_predicating;
1025
1026 radv_meta_restore(&saved_state, cmd_buffer);
1027 }
1028
1029 VkResult radv_CreateQueryPool(
1030 VkDevice _device,
1031 const VkQueryPoolCreateInfo* pCreateInfo,
1032 const VkAllocationCallbacks* pAllocator,
1033 VkQueryPool* pQueryPool)
1034 {
1035 RADV_FROM_HANDLE(radv_device, device, _device);
1036 struct radv_query_pool *pool = vk_alloc2(&device->alloc, pAllocator,
1037 sizeof(*pool), 8,
1038 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1039
1040 if (!pool)
1041 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1042
1043
1044 switch(pCreateInfo->queryType) {
1045 case VK_QUERY_TYPE_OCCLUSION:
1046 pool->stride = 16 * device->physical_device->rad_info.num_render_backends;
1047 break;
1048 case VK_QUERY_TYPE_PIPELINE_STATISTICS:
1049 pool->stride = pipelinestat_block_size * 2;
1050 break;
1051 case VK_QUERY_TYPE_TIMESTAMP:
1052 pool->stride = 8;
1053 break;
1054 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
1055 pool->stride = 32;
1056 break;
1057 default:
1058 unreachable("creating unhandled query type");
1059 }
1060
1061 pool->type = pCreateInfo->queryType;
1062 pool->pipeline_stats_mask = pCreateInfo->pipelineStatistics;
1063 pool->availability_offset = pool->stride * pCreateInfo->queryCount;
1064 pool->size = pool->availability_offset;
1065 if (pCreateInfo->queryType == VK_QUERY_TYPE_PIPELINE_STATISTICS)
1066 pool->size += 4 * pCreateInfo->queryCount;
1067
1068 pool->bo = device->ws->buffer_create(device->ws, pool->size,
1069 64, RADEON_DOMAIN_GTT, RADEON_FLAG_NO_INTERPROCESS_SHARING,
1070 RADV_BO_PRIORITY_QUERY_POOL);
1071
1072 if (!pool->bo) {
1073 vk_free2(&device->alloc, pAllocator, pool);
1074 return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
1075 }
1076
1077 pool->ptr = device->ws->buffer_map(pool->bo);
1078
1079 if (!pool->ptr) {
1080 device->ws->buffer_destroy(pool->bo);
1081 vk_free2(&device->alloc, pAllocator, pool);
1082 return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
1083 }
1084
1085 *pQueryPool = radv_query_pool_to_handle(pool);
1086 return VK_SUCCESS;
1087 }
1088
1089 void radv_DestroyQueryPool(
1090 VkDevice _device,
1091 VkQueryPool _pool,
1092 const VkAllocationCallbacks* pAllocator)
1093 {
1094 RADV_FROM_HANDLE(radv_device, device, _device);
1095 RADV_FROM_HANDLE(radv_query_pool, pool, _pool);
1096
1097 if (!pool)
1098 return;
1099
1100 device->ws->buffer_destroy(pool->bo);
1101 vk_free2(&device->alloc, pAllocator, pool);
1102 }
1103
1104 VkResult radv_GetQueryPoolResults(
1105 VkDevice _device,
1106 VkQueryPool queryPool,
1107 uint32_t firstQuery,
1108 uint32_t queryCount,
1109 size_t dataSize,
1110 void* pData,
1111 VkDeviceSize stride,
1112 VkQueryResultFlags flags)
1113 {
1114 RADV_FROM_HANDLE(radv_device, device, _device);
1115 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
1116 char *data = pData;
1117 VkResult result = VK_SUCCESS;
1118
1119 for(unsigned i = 0; i < queryCount; ++i, data += stride) {
1120 char *dest = data;
1121 unsigned query = firstQuery + i;
1122 char *src = pool->ptr + query * pool->stride;
1123 uint32_t available;
1124
1125 switch (pool->type) {
1126 case VK_QUERY_TYPE_TIMESTAMP: {
1127 available = *(uint64_t *)src != TIMESTAMP_NOT_READY;
1128
1129 if (flags & VK_QUERY_RESULT_WAIT_BIT) {
1130 while (*(volatile uint64_t *)src == TIMESTAMP_NOT_READY)
1131 ;
1132 available = true;
1133 }
1134
1135 if (!available && !(flags & VK_QUERY_RESULT_PARTIAL_BIT))
1136 result = VK_NOT_READY;
1137
1138 if (flags & VK_QUERY_RESULT_64_BIT) {
1139 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1140 *(uint64_t*)dest = *(uint64_t*)src;
1141 dest += 8;
1142 } else {
1143 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1144 *(uint32_t*)dest = *(uint32_t*)src;
1145 dest += 4;
1146 }
1147 break;
1148 }
1149 case VK_QUERY_TYPE_OCCLUSION: {
1150 volatile uint64_t const *src64 = (volatile uint64_t const *)src;
1151 uint32_t db_count = device->physical_device->rad_info.num_render_backends;
1152 uint32_t enabled_rb_mask = device->physical_device->rad_info.enabled_rb_mask;
1153 uint64_t sample_count = 0;
1154 available = 1;
1155
1156 for (int i = 0; i < db_count; ++i) {
1157 uint64_t start, end;
1158
1159 if (!(enabled_rb_mask & (1 << i)))
1160 continue;
1161
1162 do {
1163 start = src64[2 * i];
1164 end = src64[2 * i + 1];
1165 } while ((!(start & (1ull << 63)) || !(end & (1ull << 63))) && (flags & VK_QUERY_RESULT_WAIT_BIT));
1166
1167 if (!(start & (1ull << 63)) || !(end & (1ull << 63)))
1168 available = 0;
1169 else {
1170 sample_count += end - start;
1171 }
1172 }
1173
1174 if (!available && !(flags & VK_QUERY_RESULT_PARTIAL_BIT))
1175 result = VK_NOT_READY;
1176
1177 if (flags & VK_QUERY_RESULT_64_BIT) {
1178 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1179 *(uint64_t*)dest = sample_count;
1180 dest += 8;
1181 } else {
1182 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1183 *(uint32_t*)dest = sample_count;
1184 dest += 4;
1185 }
1186 break;
1187 }
1188 case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
1189 if (flags & VK_QUERY_RESULT_WAIT_BIT)
1190 while(!*(volatile uint32_t*)(pool->ptr + pool->availability_offset + 4 * query))
1191 ;
1192 available = *(uint32_t*)(pool->ptr + pool->availability_offset + 4 * query);
1193
1194 if (!available && !(flags & VK_QUERY_RESULT_PARTIAL_BIT))
1195 result = VK_NOT_READY;
1196
1197 const uint64_t *start = (uint64_t*)src;
1198 const uint64_t *stop = (uint64_t*)(src + pipelinestat_block_size);
1199 if (flags & VK_QUERY_RESULT_64_BIT) {
1200 uint64_t *dst = (uint64_t*)dest;
1201 dest += util_bitcount(pool->pipeline_stats_mask) * 8;
1202 for(int i = 0; i < 11; ++i) {
1203 if(pool->pipeline_stats_mask & (1u << i)) {
1204 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1205 *dst = stop[pipeline_statistics_indices[i]] -
1206 start[pipeline_statistics_indices[i]];
1207 dst++;
1208 }
1209 }
1210
1211 } else {
1212 uint32_t *dst = (uint32_t*)dest;
1213 dest += util_bitcount(pool->pipeline_stats_mask) * 4;
1214 for(int i = 0; i < 11; ++i) {
1215 if(pool->pipeline_stats_mask & (1u << i)) {
1216 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1217 *dst = stop[pipeline_statistics_indices[i]] -
1218 start[pipeline_statistics_indices[i]];
1219 dst++;
1220 }
1221 }
1222 }
1223 break;
1224 }
1225 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT: {
1226 volatile uint64_t const *src64 = (volatile uint64_t const *)src;
1227 uint64_t num_primitives_written;
1228 uint64_t primitive_storage_needed;
1229
1230 /* SAMPLE_STREAMOUTSTATS stores this structure:
1231 * {
1232 * u64 NumPrimitivesWritten;
1233 * u64 PrimitiveStorageNeeded;
1234 * }
1235 */
1236 available = 1;
1237 for (int j = 0; j < 4; j++) {
1238 if (!(src64[j] & 0x8000000000000000UL))
1239 available = 0;
1240 }
1241
1242 if (!available && !(flags & VK_QUERY_RESULT_PARTIAL_BIT))
1243 result = VK_NOT_READY;
1244
1245 num_primitives_written = src64[3] - src64[1];
1246 primitive_storage_needed = src64[2] - src64[0];
1247
1248 if (flags & VK_QUERY_RESULT_64_BIT) {
1249 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1250 *(uint64_t *)dest = num_primitives_written;
1251 dest += 8;
1252 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1253 *(uint64_t *)dest = primitive_storage_needed;
1254 dest += 8;
1255 } else {
1256 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1257 *(uint32_t *)dest = num_primitives_written;
1258 dest += 4;
1259 if (available || (flags & VK_QUERY_RESULT_PARTIAL_BIT))
1260 *(uint32_t *)dest = primitive_storage_needed;
1261 dest += 4;
1262 }
1263 break;
1264 }
1265 default:
1266 unreachable("trying to get results of unhandled query type");
1267 }
1268
1269 if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
1270 if (flags & VK_QUERY_RESULT_64_BIT) {
1271 *(uint64_t*)dest = available;
1272 } else {
1273 *(uint32_t*)dest = available;
1274 }
1275 }
1276 }
1277
1278 return result;
1279 }
1280
1281 static void emit_query_flush(struct radv_cmd_buffer *cmd_buffer,
1282 struct radv_query_pool *pool)
1283 {
1284 if (cmd_buffer->pending_reset_query) {
1285 if (pool->size >= RADV_BUFFER_OPS_CS_THRESHOLD) {
1286 /* Only need to flush caches if the query pool size is
1287 * large enough to be resetted using the compute shader
1288 * path. Small pools don't need any cache flushes
1289 * because we use a CP dma clear.
1290 */
1291 si_emit_cache_flush(cmd_buffer);
1292 }
1293 }
1294 }
1295
1296 void radv_CmdCopyQueryPoolResults(
1297 VkCommandBuffer commandBuffer,
1298 VkQueryPool queryPool,
1299 uint32_t firstQuery,
1300 uint32_t queryCount,
1301 VkBuffer dstBuffer,
1302 VkDeviceSize dstOffset,
1303 VkDeviceSize stride,
1304 VkQueryResultFlags flags)
1305 {
1306 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1307 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
1308 RADV_FROM_HANDLE(radv_buffer, dst_buffer, dstBuffer);
1309 struct radeon_cmdbuf *cs = cmd_buffer->cs;
1310 unsigned elem_size = (flags & VK_QUERY_RESULT_64_BIT) ? 8 : 4;
1311 uint64_t va = radv_buffer_get_va(pool->bo);
1312 uint64_t dest_va = radv_buffer_get_va(dst_buffer->bo);
1313 dest_va += dst_buffer->offset + dstOffset;
1314
1315 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, pool->bo);
1316 radv_cs_add_buffer(cmd_buffer->device->ws, cmd_buffer->cs, dst_buffer->bo);
1317
1318 /* From the Vulkan spec 1.1.108:
1319 *
1320 * "vkCmdCopyQueryPoolResults is guaranteed to see the effect of
1321 * previous uses of vkCmdResetQueryPool in the same queue, without any
1322 * additional synchronization."
1323 *
1324 * So, we have to flush the caches if the compute shader path was used.
1325 */
1326 emit_query_flush(cmd_buffer, pool);
1327
1328 switch (pool->type) {
1329 case VK_QUERY_TYPE_OCCLUSION:
1330 if (flags & VK_QUERY_RESULT_WAIT_BIT) {
1331 for(unsigned i = 0; i < queryCount; ++i, dest_va += stride) {
1332 unsigned query = firstQuery + i;
1333 uint64_t src_va = va + query * pool->stride + pool->stride - 4;
1334
1335 radeon_check_space(cmd_buffer->device->ws, cs, 7);
1336
1337 /* Waits on the upper word of the last DB entry */
1338 radv_cp_wait_mem(cs, WAIT_REG_MEM_GREATER_OR_EQUAL,
1339 src_va, 0x80000000, 0xffffffff);
1340 }
1341 }
1342 radv_query_shader(cmd_buffer, &cmd_buffer->device->meta_state.query.occlusion_query_pipeline,
1343 pool->bo, dst_buffer->bo, firstQuery * pool->stride,
1344 dst_buffer->offset + dstOffset,
1345 pool->stride, stride,
1346 queryCount, flags, 0, 0);
1347 break;
1348 case VK_QUERY_TYPE_PIPELINE_STATISTICS:
1349 if (flags & VK_QUERY_RESULT_WAIT_BIT) {
1350 for(unsigned i = 0; i < queryCount; ++i, dest_va += stride) {
1351 unsigned query = firstQuery + i;
1352
1353 radeon_check_space(cmd_buffer->device->ws, cs, 7);
1354
1355 uint64_t avail_va = va + pool->availability_offset + 4 * query;
1356
1357 /* This waits on the ME. All copies below are done on the ME */
1358 radv_cp_wait_mem(cs, WAIT_REG_MEM_EQUAL,
1359 avail_va, 1, 0xffffffff);
1360 }
1361 }
1362 radv_query_shader(cmd_buffer, &cmd_buffer->device->meta_state.query.pipeline_statistics_query_pipeline,
1363 pool->bo, dst_buffer->bo, firstQuery * pool->stride,
1364 dst_buffer->offset + dstOffset,
1365 pool->stride, stride, queryCount, flags,
1366 pool->pipeline_stats_mask,
1367 pool->availability_offset + 4 * firstQuery);
1368 break;
1369 case VK_QUERY_TYPE_TIMESTAMP:
1370 for(unsigned i = 0; i < queryCount; ++i, dest_va += stride) {
1371 unsigned query = firstQuery + i;
1372 uint64_t local_src_va = va + query * pool->stride;
1373
1374 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 19);
1375
1376
1377 if (flags & VK_QUERY_RESULT_WAIT_BIT) {
1378 /* Wait on the high 32 bits of the timestamp in
1379 * case the low part is 0xffffffff.
1380 */
1381 radv_cp_wait_mem(cs, WAIT_REG_MEM_NOT_EQUAL,
1382 local_src_va + 4,
1383 TIMESTAMP_NOT_READY >> 32,
1384 0xffffffff);
1385 }
1386 if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
1387 uint64_t avail_dest_va = dest_va + elem_size;
1388
1389 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
1390 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_SRC_MEM) |
1391 COPY_DATA_DST_SEL(COPY_DATA_DST_MEM_GRBM));
1392 radeon_emit(cs, local_src_va);
1393 radeon_emit(cs, local_src_va >> 32);
1394 radeon_emit(cs, avail_dest_va);
1395 radeon_emit(cs, avail_dest_va >> 32);
1396 }
1397
1398 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
1399 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_SRC_MEM) |
1400 COPY_DATA_DST_SEL(COPY_DATA_DST_MEM_GRBM) |
1401 ((flags & VK_QUERY_RESULT_64_BIT) ? COPY_DATA_COUNT_SEL : 0));
1402 radeon_emit(cs, local_src_va);
1403 radeon_emit(cs, local_src_va >> 32);
1404 radeon_emit(cs, dest_va);
1405 radeon_emit(cs, dest_va >> 32);
1406
1407
1408 assert(cs->cdw <= cdw_max);
1409 }
1410 break;
1411 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
1412 if (flags & VK_QUERY_RESULT_WAIT_BIT) {
1413 for(unsigned i = 0; i < queryCount; i++) {
1414 unsigned query = firstQuery + i;
1415 uint64_t src_va = va + query * pool->stride;
1416
1417 radeon_check_space(cmd_buffer->device->ws, cs, 7 * 4);
1418
1419 /* Wait on the upper word of all results. */
1420 for (unsigned j = 0; j < 4; j++, src_va += 8) {
1421 radv_cp_wait_mem(cs, WAIT_REG_MEM_GREATER_OR_EQUAL,
1422 src_va + 4, 0x80000000,
1423 0xffffffff);
1424 }
1425 }
1426 }
1427
1428 radv_query_shader(cmd_buffer, &cmd_buffer->device->meta_state.query.tfb_query_pipeline,
1429 pool->bo, dst_buffer->bo,
1430 firstQuery * pool->stride,
1431 dst_buffer->offset + dstOffset,
1432 pool->stride, stride,
1433 queryCount, flags, 0, 0);
1434 break;
1435 default:
1436 unreachable("trying to get results of unhandled query type");
1437 }
1438
1439 }
1440
1441 void radv_CmdResetQueryPool(
1442 VkCommandBuffer commandBuffer,
1443 VkQueryPool queryPool,
1444 uint32_t firstQuery,
1445 uint32_t queryCount)
1446 {
1447 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1448 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
1449 uint32_t value = pool->type == VK_QUERY_TYPE_TIMESTAMP
1450 ? TIMESTAMP_NOT_READY : 0;
1451 uint32_t flush_bits = 0;
1452
1453 /* Make sure to sync all previous work if the given command buffer has
1454 * pending active queries. Otherwise the GPU might write queries data
1455 * after the reset operation.
1456 */
1457 cmd_buffer->state.flush_bits |= cmd_buffer->active_query_flush_bits;
1458
1459 flush_bits |= radv_fill_buffer(cmd_buffer, pool->bo,
1460 firstQuery * pool->stride,
1461 queryCount * pool->stride, value);
1462
1463 if (pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS) {
1464 flush_bits |= radv_fill_buffer(cmd_buffer, pool->bo,
1465 pool->availability_offset + firstQuery * 4,
1466 queryCount * 4, 0);
1467 }
1468
1469 if (flush_bits) {
1470 /* Only need to flush caches for the compute shader path. */
1471 cmd_buffer->pending_reset_query = true;
1472 cmd_buffer->state.flush_bits |= flush_bits;
1473 }
1474 }
1475
1476 void radv_ResetQueryPoolEXT(
1477 VkDevice _device,
1478 VkQueryPool queryPool,
1479 uint32_t firstQuery,
1480 uint32_t queryCount)
1481 {
1482 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
1483
1484 uint32_t value = pool->type == VK_QUERY_TYPE_TIMESTAMP
1485 ? TIMESTAMP_NOT_READY : 0;
1486 uint32_t *data = (uint32_t*)(pool->ptr + firstQuery * pool->stride);
1487 uint32_t *data_end = (uint32_t*)(pool->ptr + (firstQuery + queryCount) * pool->stride);
1488
1489 for(uint32_t *p = data; p != data_end; ++p)
1490 *p = value;
1491
1492 if (pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS) {
1493 memset(pool->ptr + pool->availability_offset + firstQuery * 4,
1494 0, queryCount * 4);
1495 }
1496 }
1497
1498 static unsigned event_type_for_stream(unsigned stream)
1499 {
1500 switch (stream) {
1501 default:
1502 case 0: return V_028A90_SAMPLE_STREAMOUTSTATS;
1503 case 1: return V_028A90_SAMPLE_STREAMOUTSTATS1;
1504 case 2: return V_028A90_SAMPLE_STREAMOUTSTATS2;
1505 case 3: return V_028A90_SAMPLE_STREAMOUTSTATS3;
1506 }
1507 }
1508
1509 static void emit_begin_query(struct radv_cmd_buffer *cmd_buffer,
1510 uint64_t va,
1511 VkQueryType query_type,
1512 VkQueryControlFlags flags,
1513 uint32_t index)
1514 {
1515 struct radeon_cmdbuf *cs = cmd_buffer->cs;
1516 switch (query_type) {
1517 case VK_QUERY_TYPE_OCCLUSION:
1518 radeon_check_space(cmd_buffer->device->ws, cs, 7);
1519
1520 ++cmd_buffer->state.active_occlusion_queries;
1521 if (cmd_buffer->state.active_occlusion_queries == 1) {
1522 if (flags & VK_QUERY_CONTROL_PRECISE_BIT) {
1523 /* This is the first occlusion query, enable
1524 * the hint if the precision bit is set.
1525 */
1526 cmd_buffer->state.perfect_occlusion_queries_enabled = true;
1527 }
1528
1529 radv_set_db_count_control(cmd_buffer);
1530 } else {
1531 if ((flags & VK_QUERY_CONTROL_PRECISE_BIT) &&
1532 !cmd_buffer->state.perfect_occlusion_queries_enabled) {
1533 /* This is not the first query, but this one
1534 * needs to enable precision, DB_COUNT_CONTROL
1535 * has to be updated accordingly.
1536 */
1537 cmd_buffer->state.perfect_occlusion_queries_enabled = true;
1538
1539 radv_set_db_count_control(cmd_buffer);
1540 }
1541 }
1542
1543 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
1544 radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
1545 radeon_emit(cs, va);
1546 radeon_emit(cs, va >> 32);
1547 break;
1548 case VK_QUERY_TYPE_PIPELINE_STATISTICS:
1549 radeon_check_space(cmd_buffer->device->ws, cs, 4);
1550
1551 ++cmd_buffer->state.active_pipeline_queries;
1552 if (cmd_buffer->state.active_pipeline_queries == 1) {
1553 cmd_buffer->state.flush_bits &= ~RADV_CMD_FLAG_STOP_PIPELINE_STATS;
1554 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_START_PIPELINE_STATS;
1555 }
1556
1557 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
1558 radeon_emit(cs, EVENT_TYPE(V_028A90_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
1559 radeon_emit(cs, va);
1560 radeon_emit(cs, va >> 32);
1561 break;
1562 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
1563 radeon_check_space(cmd_buffer->device->ws, cs, 4);
1564
1565 assert(index < MAX_SO_STREAMS);
1566
1567 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
1568 radeon_emit(cs, EVENT_TYPE(event_type_for_stream(index)) | EVENT_INDEX(3));
1569 radeon_emit(cs, va);
1570 radeon_emit(cs, va >> 32);
1571 break;
1572 default:
1573 unreachable("beginning unhandled query type");
1574 }
1575
1576 }
1577
1578 static void emit_end_query(struct radv_cmd_buffer *cmd_buffer,
1579 uint64_t va, uint64_t avail_va,
1580 VkQueryType query_type, uint32_t index)
1581 {
1582 struct radeon_cmdbuf *cs = cmd_buffer->cs;
1583 switch (query_type) {
1584 case VK_QUERY_TYPE_OCCLUSION:
1585 radeon_check_space(cmd_buffer->device->ws, cs, 14);
1586
1587 cmd_buffer->state.active_occlusion_queries--;
1588 if (cmd_buffer->state.active_occlusion_queries == 0) {
1589 radv_set_db_count_control(cmd_buffer);
1590
1591 /* Reset the perfect occlusion queries hint now that no
1592 * queries are active.
1593 */
1594 cmd_buffer->state.perfect_occlusion_queries_enabled = false;
1595 }
1596
1597 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
1598 radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
1599 radeon_emit(cs, va + 8);
1600 radeon_emit(cs, (va + 8) >> 32);
1601
1602 break;
1603 case VK_QUERY_TYPE_PIPELINE_STATISTICS:
1604 radeon_check_space(cmd_buffer->device->ws, cs, 16);
1605
1606 cmd_buffer->state.active_pipeline_queries--;
1607 if (cmd_buffer->state.active_pipeline_queries == 0) {
1608 cmd_buffer->state.flush_bits &= ~RADV_CMD_FLAG_START_PIPELINE_STATS;
1609 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_STOP_PIPELINE_STATS;
1610 }
1611 va += pipelinestat_block_size;
1612
1613 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
1614 radeon_emit(cs, EVENT_TYPE(V_028A90_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
1615 radeon_emit(cs, va);
1616 radeon_emit(cs, va >> 32);
1617
1618 si_cs_emit_write_event_eop(cs,
1619 cmd_buffer->device->physical_device->rad_info.chip_class,
1620 radv_cmd_buffer_uses_mec(cmd_buffer),
1621 V_028A90_BOTTOM_OF_PIPE_TS, 0,
1622 EOP_DST_SEL_MEM,
1623 EOP_DATA_SEL_VALUE_32BIT,
1624 avail_va, 1,
1625 cmd_buffer->gfx9_eop_bug_va);
1626 break;
1627 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
1628 radeon_check_space(cmd_buffer->device->ws, cs, 4);
1629
1630 assert(index < MAX_SO_STREAMS);
1631
1632 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
1633 radeon_emit(cs, EVENT_TYPE(event_type_for_stream(index)) | EVENT_INDEX(3));
1634 radeon_emit(cs, (va + 16));
1635 radeon_emit(cs, (va + 16) >> 32);
1636 break;
1637 default:
1638 unreachable("ending unhandled query type");
1639 }
1640
1641 cmd_buffer->active_query_flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
1642 RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
1643 RADV_CMD_FLAG_INV_L2 |
1644 RADV_CMD_FLAG_INV_VCACHE;
1645 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1646 cmd_buffer->active_query_flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1647 RADV_CMD_FLAG_FLUSH_AND_INV_DB;
1648 }
1649 }
1650
1651 void radv_CmdBeginQueryIndexedEXT(
1652 VkCommandBuffer commandBuffer,
1653 VkQueryPool queryPool,
1654 uint32_t query,
1655 VkQueryControlFlags flags,
1656 uint32_t index)
1657 {
1658 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1659 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
1660 struct radeon_cmdbuf *cs = cmd_buffer->cs;
1661 uint64_t va = radv_buffer_get_va(pool->bo);
1662
1663 radv_cs_add_buffer(cmd_buffer->device->ws, cs, pool->bo);
1664
1665 emit_query_flush(cmd_buffer, pool);
1666
1667 va += pool->stride * query;
1668
1669 emit_begin_query(cmd_buffer, va, pool->type, flags, index);
1670 }
1671
1672 void radv_CmdBeginQuery(
1673 VkCommandBuffer commandBuffer,
1674 VkQueryPool queryPool,
1675 uint32_t query,
1676 VkQueryControlFlags flags)
1677 {
1678 radv_CmdBeginQueryIndexedEXT(commandBuffer, queryPool, query, flags, 0);
1679 }
1680
1681 void radv_CmdEndQueryIndexedEXT(
1682 VkCommandBuffer commandBuffer,
1683 VkQueryPool queryPool,
1684 uint32_t query,
1685 uint32_t index)
1686 {
1687 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1688 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
1689 uint64_t va = radv_buffer_get_va(pool->bo);
1690 uint64_t avail_va = va + pool->availability_offset + 4 * query;
1691 va += pool->stride * query;
1692
1693 /* Do not need to add the pool BO to the list because the query must
1694 * currently be active, which means the BO is already in the list.
1695 */
1696 emit_end_query(cmd_buffer, va, avail_va, pool->type, index);
1697
1698 /*
1699 * For multiview we have to emit a query for each bit in the mask,
1700 * however the first query we emit will get the totals for all the
1701 * operations, so we don't want to get a real value in the other
1702 * queries. This emits a fake begin/end sequence so the waiting
1703 * code gets a completed query value and doesn't hang, but the
1704 * query returns 0.
1705 */
1706 if (cmd_buffer->state.subpass && cmd_buffer->state.subpass->view_mask) {
1707 uint64_t avail_va = va + pool->availability_offset + 4 * query;
1708
1709
1710 for (unsigned i = 1; i < util_bitcount(cmd_buffer->state.subpass->view_mask); i++) {
1711 va += pool->stride;
1712 avail_va += 4;
1713 emit_begin_query(cmd_buffer, va, pool->type, 0, 0);
1714 emit_end_query(cmd_buffer, va, avail_va, pool->type, 0);
1715 }
1716 }
1717 }
1718
1719 void radv_CmdEndQuery(
1720 VkCommandBuffer commandBuffer,
1721 VkQueryPool queryPool,
1722 uint32_t query)
1723 {
1724 radv_CmdEndQueryIndexedEXT(commandBuffer, queryPool, query, 0);
1725 }
1726
1727 void radv_CmdWriteTimestamp(
1728 VkCommandBuffer commandBuffer,
1729 VkPipelineStageFlagBits pipelineStage,
1730 VkQueryPool queryPool,
1731 uint32_t query)
1732 {
1733 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1734 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
1735 bool mec = radv_cmd_buffer_uses_mec(cmd_buffer);
1736 struct radeon_cmdbuf *cs = cmd_buffer->cs;
1737 uint64_t va = radv_buffer_get_va(pool->bo);
1738 uint64_t query_va = va + pool->stride * query;
1739
1740 radv_cs_add_buffer(cmd_buffer->device->ws, cs, pool->bo);
1741
1742 emit_query_flush(cmd_buffer, pool);
1743
1744 int num_queries = 1;
1745 if (cmd_buffer->state.subpass && cmd_buffer->state.subpass->view_mask)
1746 num_queries = util_bitcount(cmd_buffer->state.subpass->view_mask);
1747
1748 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 28 * num_queries);
1749
1750 for (unsigned i = 0; i < num_queries; i++) {
1751 switch(pipelineStage) {
1752 case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT:
1753 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
1754 radeon_emit(cs, COPY_DATA_COUNT_SEL | COPY_DATA_WR_CONFIRM |
1755 COPY_DATA_SRC_SEL(COPY_DATA_TIMESTAMP) |
1756 COPY_DATA_DST_SEL(V_370_MEM));
1757 radeon_emit(cs, 0);
1758 radeon_emit(cs, 0);
1759 radeon_emit(cs, query_va);
1760 radeon_emit(cs, query_va >> 32);
1761 break;
1762 default:
1763 si_cs_emit_write_event_eop(cs,
1764 cmd_buffer->device->physical_device->rad_info.chip_class,
1765 mec,
1766 V_028A90_BOTTOM_OF_PIPE_TS, 0,
1767 EOP_DST_SEL_MEM,
1768 EOP_DATA_SEL_TIMESTAMP,
1769 query_va, 0,
1770 cmd_buffer->gfx9_eop_bug_va);
1771 break;
1772 }
1773 query_va += pool->stride;
1774 }
1775 assert(cmd_buffer->cs->cdw <= cdw_max);
1776 }