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 "radv_private.h"
33 #include "radv_cs.h"
34 #include "sid.h"
35
get_max_db(struct radv_device * device)36 static unsigned get_max_db(struct radv_device *device)
37 {
38 unsigned num_db = device->physical_device->rad_info.num_render_backends;
39 MAYBE_UNUSED unsigned rb_mask = device->physical_device->rad_info.enabled_rb_mask;
40
41 if (device->physical_device->rad_info.chip_class == SI)
42 num_db = 8;
43 else
44 num_db = MAX2(8, num_db);
45
46 /* Otherwise we need to change the query reset procedure */
47 assert(rb_mask == ((1ull << num_db) - 1));
48
49 return num_db;
50 }
51
radv_CreateQueryPool(VkDevice _device,const VkQueryPoolCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkQueryPool * pQueryPool)52 VkResult radv_CreateQueryPool(
53 VkDevice _device,
54 const VkQueryPoolCreateInfo* pCreateInfo,
55 const VkAllocationCallbacks* pAllocator,
56 VkQueryPool* pQueryPool)
57 {
58 RADV_FROM_HANDLE(radv_device, device, _device);
59 uint64_t size;
60 struct radv_query_pool *pool = vk_alloc2(&device->alloc, pAllocator,
61 sizeof(*pool), 8,
62 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
63
64 if (!pool)
65 return VK_ERROR_OUT_OF_HOST_MEMORY;
66
67
68 switch(pCreateInfo->queryType) {
69 case VK_QUERY_TYPE_OCCLUSION:
70 /* 16 bytes tmp. buffer as the compute packet writes 64 bits, but
71 * the app. may have 32 bits of space. */
72 pool->stride = 16 * get_max_db(device) + 16;
73 break;
74 case VK_QUERY_TYPE_PIPELINE_STATISTICS:
75 pool->stride = 16 * 11;
76 break;
77 case VK_QUERY_TYPE_TIMESTAMP:
78 pool->stride = 8;
79 break;
80 default:
81 unreachable("creating unhandled query type");
82 }
83
84 pool->type = pCreateInfo->queryType;
85 pool->availability_offset = pool->stride * pCreateInfo->queryCount;
86 size = pool->availability_offset + 4 * pCreateInfo->queryCount;
87
88 pool->bo = device->ws->buffer_create(device->ws, size,
89 64, RADEON_DOMAIN_GTT, 0);
90
91 if (!pool->bo) {
92 vk_free2(&device->alloc, pAllocator, pool);
93 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
94 }
95
96 pool->ptr = device->ws->buffer_map(pool->bo);
97
98 if (!pool->ptr) {
99 device->ws->buffer_destroy(pool->bo);
100 vk_free2(&device->alloc, pAllocator, pool);
101 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
102 }
103 memset(pool->ptr, 0, size);
104
105 *pQueryPool = radv_query_pool_to_handle(pool);
106 return VK_SUCCESS;
107 }
108
radv_DestroyQueryPool(VkDevice _device,VkQueryPool _pool,const VkAllocationCallbacks * pAllocator)109 void radv_DestroyQueryPool(
110 VkDevice _device,
111 VkQueryPool _pool,
112 const VkAllocationCallbacks* pAllocator)
113 {
114 RADV_FROM_HANDLE(radv_device, device, _device);
115 RADV_FROM_HANDLE(radv_query_pool, pool, _pool);
116
117 if (!pool)
118 return;
119
120 device->ws->buffer_destroy(pool->bo);
121 vk_free2(&device->alloc, pAllocator, pool);
122 }
123
radv_GetQueryPoolResults(VkDevice _device,VkQueryPool queryPool,uint32_t firstQuery,uint32_t queryCount,size_t dataSize,void * pData,VkDeviceSize stride,VkQueryResultFlags flags)124 VkResult radv_GetQueryPoolResults(
125 VkDevice _device,
126 VkQueryPool queryPool,
127 uint32_t firstQuery,
128 uint32_t queryCount,
129 size_t dataSize,
130 void* pData,
131 VkDeviceSize stride,
132 VkQueryResultFlags flags)
133 {
134 RADV_FROM_HANDLE(radv_device, device, _device);
135 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
136 char *data = pData;
137 VkResult result = VK_SUCCESS;
138
139 for(unsigned i = 0; i < queryCount; ++i, data += stride) {
140 char *dest = data;
141 unsigned query = firstQuery + i;
142 char *src = pool->ptr + query * pool->stride;
143 uint32_t available;
144
145 switch (pool->type) {
146 case VK_QUERY_TYPE_TIMESTAMP: {
147 if (flags & VK_QUERY_RESULT_WAIT_BIT) {
148 while(!*(volatile uint32_t*)(pool->ptr + pool->availability_offset + 4 * query))
149 ;
150 }
151
152 available = *(uint32_t*)(pool->ptr + pool->availability_offset + 4 * query);
153 if (!available && !(flags & VK_QUERY_RESULT_PARTIAL_BIT)) {
154 result = VK_NOT_READY;
155 break;
156
157 }
158
159 if (flags & VK_QUERY_RESULT_64_BIT) {
160 *(uint64_t*)dest = *(uint64_t*)src;
161 dest += 8;
162 } else {
163 *(uint32_t*)dest = *(uint32_t*)src;
164 dest += 4;
165 }
166 break;
167 }
168 case VK_QUERY_TYPE_OCCLUSION: {
169 volatile uint64_t const *src64 = (volatile uint64_t const *)src;
170 uint64_t result = 0;
171 int db_count = get_max_db(device);
172 available = 1;
173
174 for (int i = 0; i < db_count; ++i) {
175 uint64_t start, end;
176 do {
177 start = src64[2 * i];
178 end = src64[2 * i + 1];
179 } while ((!(start & (1ull << 63)) || !(end & (1ull << 63))) && (flags & VK_QUERY_RESULT_WAIT_BIT));
180
181 if (!(start & (1ull << 63)) || !(end & (1ull << 63)))
182 available = 0;
183 else {
184 result += end - start;
185 }
186 }
187
188 if (!available && !(flags & VK_QUERY_RESULT_PARTIAL_BIT)) {
189 result = VK_NOT_READY;
190 break;
191
192 }
193
194 if (flags & VK_QUERY_RESULT_64_BIT) {
195 *(uint64_t*)dest = result;
196 dest += 8;
197 } else {
198 *(uint32_t*)dest = result;
199 dest += 4;
200 }
201 break;
202 default:
203 unreachable("trying to get results of unhandled query type");
204 }
205 }
206
207 if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
208 if (flags & VK_QUERY_RESULT_64_BIT) {
209 *(uint64_t*)dest = available;
210 } else {
211 *(uint32_t*)dest = available;
212 }
213 }
214 }
215
216 return result;
217 }
218
radv_CmdCopyQueryPoolResults(VkCommandBuffer commandBuffer,VkQueryPool queryPool,uint32_t firstQuery,uint32_t queryCount,VkBuffer dstBuffer,VkDeviceSize dstOffset,VkDeviceSize stride,VkQueryResultFlags flags)219 void radv_CmdCopyQueryPoolResults(
220 VkCommandBuffer commandBuffer,
221 VkQueryPool queryPool,
222 uint32_t firstQuery,
223 uint32_t queryCount,
224 VkBuffer dstBuffer,
225 VkDeviceSize dstOffset,
226 VkDeviceSize stride,
227 VkQueryResultFlags flags)
228 {
229 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
230 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
231 RADV_FROM_HANDLE(radv_buffer, dst_buffer, dstBuffer);
232 struct radeon_winsys_cs *cs = cmd_buffer->cs;
233 uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
234 uint64_t dest_va = cmd_buffer->device->ws->buffer_get_va(dst_buffer->bo);
235 dest_va += dst_buffer->offset + dstOffset;
236
237 cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, pool->bo, 8);
238 cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, dst_buffer->bo, 8);
239
240 for(unsigned i = 0; i < queryCount; ++i, dest_va += stride) {
241 unsigned query = firstQuery + i;
242 uint64_t local_src_va = va + query * pool->stride;
243 unsigned elem_size = (flags & VK_QUERY_RESULT_64_BIT) ? 8 : 4;
244
245 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 26);
246
247 if (flags & VK_QUERY_RESULT_WAIT_BIT) {
248 /* TODO, not sure if there is any case where we won't always be ready yet */
249 uint64_t avail_va = va + pool->availability_offset + 4 * query;
250
251
252 /* This waits on the ME. All copies below are done on the ME */
253 radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
254 radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
255 radeon_emit(cs, avail_va);
256 radeon_emit(cs, avail_va >> 32);
257 radeon_emit(cs, 1); /* reference value */
258 radeon_emit(cs, 0xffffffff); /* mask */
259 radeon_emit(cs, 4); /* poll interval */
260 }
261
262 switch (pool->type) {
263 case VK_QUERY_TYPE_OCCLUSION:
264 local_src_va += pool->stride - 16;
265
266 case VK_QUERY_TYPE_TIMESTAMP:
267 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
268 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
269 COPY_DATA_DST_SEL(COPY_DATA_MEM) |
270 ((flags & VK_QUERY_RESULT_64_BIT) ? COPY_DATA_COUNT_SEL : 0));
271 radeon_emit(cs, local_src_va);
272 radeon_emit(cs, local_src_va >> 32);
273 radeon_emit(cs, dest_va);
274 radeon_emit(cs, dest_va >> 32);
275 break;
276 default:
277 unreachable("trying to get results of unhandled query type");
278 }
279
280 /* The flag could be still changed while the data copy is busy and we
281 * then might have invalid data, but a ready flag. However, the availability
282 * writes happen on the ME too, so they should be synchronized. Might need to
283 * revisit this with multiple queues.
284 */
285 if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
286 uint64_t avail_va = va + pool->availability_offset + 4 * query;
287 uint64_t avail_dest_va = dest_va;
288 if (pool->type != VK_QUERY_TYPE_PIPELINE_STATISTICS)
289 avail_dest_va += elem_size;
290 else
291 abort();
292
293 radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
294 radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
295 COPY_DATA_DST_SEL(COPY_DATA_MEM));
296 radeon_emit(cs, avail_va);
297 radeon_emit(cs, avail_va >> 32);
298 radeon_emit(cs, avail_dest_va);
299 radeon_emit(cs, avail_dest_va >> 32);
300 }
301
302 assert(cs->cdw <= cdw_max);
303 }
304
305 }
306
radv_CmdResetQueryPool(VkCommandBuffer commandBuffer,VkQueryPool queryPool,uint32_t firstQuery,uint32_t queryCount)307 void radv_CmdResetQueryPool(
308 VkCommandBuffer commandBuffer,
309 VkQueryPool queryPool,
310 uint32_t firstQuery,
311 uint32_t queryCount)
312 {
313 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
314 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
315 uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
316
317 cmd_buffer->device->ws->cs_add_buffer(cmd_buffer->cs, pool->bo, 8);
318
319 si_cp_dma_clear_buffer(cmd_buffer, va + firstQuery * pool->stride,
320 queryCount * pool->stride, 0);
321 si_cp_dma_clear_buffer(cmd_buffer, va + pool->availability_offset + firstQuery * 4,
322 queryCount * 4, 0);
323 }
324
radv_CmdBeginQuery(VkCommandBuffer commandBuffer,VkQueryPool queryPool,uint32_t query,VkQueryControlFlags flags)325 void radv_CmdBeginQuery(
326 VkCommandBuffer commandBuffer,
327 VkQueryPool queryPool,
328 uint32_t query,
329 VkQueryControlFlags flags)
330 {
331 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
332 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
333 struct radeon_winsys_cs *cs = cmd_buffer->cs;
334 uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
335 va += pool->stride * query;
336
337 cmd_buffer->device->ws->cs_add_buffer(cs, pool->bo, 8);
338
339 switch (pool->type) {
340 case VK_QUERY_TYPE_OCCLUSION:
341 radeon_check_space(cmd_buffer->device->ws, cs, 7);
342
343 ++cmd_buffer->state.active_occlusion_queries;
344 if (cmd_buffer->state.active_occlusion_queries == 1)
345 radv_set_db_count_control(cmd_buffer);
346
347 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
348 radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
349 radeon_emit(cs, va);
350 radeon_emit(cs, va >> 32);
351 break;
352 default:
353 unreachable("beginning unhandled query type");
354 }
355 }
356
357
radv_CmdEndQuery(VkCommandBuffer commandBuffer,VkQueryPool queryPool,uint32_t query)358 void radv_CmdEndQuery(
359 VkCommandBuffer commandBuffer,
360 VkQueryPool queryPool,
361 uint32_t query)
362 {
363 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
364 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
365 struct radeon_winsys_cs *cs = cmd_buffer->cs;
366 uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
367 uint64_t avail_va = va + pool->availability_offset + 4 * query;
368 va += pool->stride * query;
369
370 cmd_buffer->device->ws->cs_add_buffer(cs, pool->bo, 8);
371
372 switch (pool->type) {
373 case VK_QUERY_TYPE_OCCLUSION:
374 radeon_check_space(cmd_buffer->device->ws, cs, 14);
375
376 cmd_buffer->state.active_occlusion_queries--;
377 if (cmd_buffer->state.active_occlusion_queries == 0)
378 radv_set_db_count_control(cmd_buffer);
379
380 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
381 radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
382 radeon_emit(cs, va + 8);
383 radeon_emit(cs, (va + 8) >> 32);
384
385 /* hangs for VK_COMMAND_BUFFER_LEVEL_SECONDARY. */
386 if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
387 radeon_emit(cs, PKT3(PKT3_OCCLUSION_QUERY, 3, 0));
388 radeon_emit(cs, va);
389 radeon_emit(cs, va >> 32);
390 radeon_emit(cs, va + pool->stride - 16);
391 radeon_emit(cs, (va + pool->stride - 16) >> 32);
392 }
393
394 break;
395 default:
396 unreachable("ending unhandled query type");
397 }
398
399 radeon_check_space(cmd_buffer->device->ws, cs, 5);
400
401 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
402 radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
403 S_370_WR_CONFIRM(1) |
404 S_370_ENGINE_SEL(V_370_ME));
405 radeon_emit(cs, avail_va);
406 radeon_emit(cs, avail_va >> 32);
407 radeon_emit(cs, 1);
408 }
409
radv_CmdWriteTimestamp(VkCommandBuffer commandBuffer,VkPipelineStageFlagBits pipelineStage,VkQueryPool queryPool,uint32_t query)410 void radv_CmdWriteTimestamp(
411 VkCommandBuffer commandBuffer,
412 VkPipelineStageFlagBits pipelineStage,
413 VkQueryPool queryPool,
414 uint32_t query)
415 {
416 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
417 RADV_FROM_HANDLE(radv_query_pool, pool, queryPool);
418 bool mec = radv_cmd_buffer_uses_mec(cmd_buffer);
419 struct radeon_winsys_cs *cs = cmd_buffer->cs;
420 uint64_t va = cmd_buffer->device->ws->buffer_get_va(pool->bo);
421 uint64_t avail_va = va + pool->availability_offset + 4 * query;
422 uint64_t query_va = va + pool->stride * query;
423
424 cmd_buffer->device->ws->cs_add_buffer(cs, pool->bo, 5);
425
426 MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws, cs, 12);
427
428 if (mec) {
429 radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, 5, 0));
430 radeon_emit(cs, EVENT_TYPE(V_028A90_BOTTOM_OF_PIPE_TS) | EVENT_INDEX(5));
431 radeon_emit(cs, 3 << 29);
432 radeon_emit(cs, query_va);
433 radeon_emit(cs, query_va >> 32);
434 radeon_emit(cs, 0);
435 radeon_emit(cs, 0);
436 } else {
437 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
438 radeon_emit(cs, EVENT_TYPE(V_028A90_BOTTOM_OF_PIPE_TS) | EVENT_INDEX(5));
439 radeon_emit(cs, query_va);
440 radeon_emit(cs, (3 << 29) | ((query_va >> 32) & 0xFFFF));
441 radeon_emit(cs, 0);
442 radeon_emit(cs, 0);
443 }
444
445 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
446 radeon_emit(cs, S_370_DST_SEL(mec ? V_370_MEM_ASYNC : V_370_MEMORY_SYNC) |
447 S_370_WR_CONFIRM(1) |
448 S_370_ENGINE_SEL(V_370_ME));
449 radeon_emit(cs, avail_va);
450 radeon_emit(cs, avail_va >> 32);
451 radeon_emit(cs, 1);
452
453 assert(cmd_buffer->cs->cdw <= cdw_max);
454 }
455