1/*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24#include <assert.h>
25#include <stdbool.h>
26#include <string.h>
27#include <unistd.h>
28#include <fcntl.h>
29
30#include "anv_private.h"
31
32#include "genxml/gen_macros.h"
33#include "genxml/genX_pack.h"
34
35/* We reserve :
36 *    - GPR 14 for perf queries
37 *    - GPR 15 for conditional rendering
38 */
39#define MI_BUILDER_NUM_ALLOC_GPRS 14
40#define MI_BUILDER_CAN_WRITE_BATCH GFX_VER >= 8
41#define __gen_get_batch_dwords anv_batch_emit_dwords
42#define __gen_address_offset anv_address_add
43#define __gen_get_batch_address(b, a) anv_batch_address(b, a)
44#include "common/mi_builder.h"
45#include "perf/intel_perf.h"
46#include "perf/intel_perf_mdapi.h"
47#include "perf/intel_perf_regs.h"
48
49#include "vk_util.h"
50
51static struct anv_address
52anv_query_address(struct anv_query_pool *pool, uint32_t query)
53{
54   return (struct anv_address) {
55      .bo = pool->bo,
56      .offset = query * pool->stride,
57   };
58}
59
60VkResult genX(CreateQueryPool)(
61    VkDevice                                    _device,
62    const VkQueryPoolCreateInfo*                pCreateInfo,
63    const VkAllocationCallbacks*                pAllocator,
64    VkQueryPool*                                pQueryPool)
65{
66   ANV_FROM_HANDLE(anv_device, device, _device);
67   const struct anv_physical_device *pdevice = device->physical;
68#if GFX_VER >= 8
69   const VkQueryPoolPerformanceCreateInfoKHR *perf_query_info = NULL;
70   struct intel_perf_counter_pass *counter_pass;
71   struct intel_perf_query_info **pass_query;
72   uint32_t n_passes = 0;
73#endif
74   uint32_t data_offset = 0;
75   VK_MULTIALLOC(ma);
76   VkResult result;
77
78   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO);
79
80   /* Query pool slots are made up of some number of 64-bit values packed
81    * tightly together. For most query types have the first 64-bit value is
82    * the "available" bit which is 0 when the query is unavailable and 1 when
83    * it is available. The 64-bit values that follow are determined by the
84    * type of query.
85    *
86    * For performance queries, we have a requirement to align OA reports at
87    * 64bytes so we put those first and have the "available" bit behind
88    * together with some other counters.
89    */
90   uint32_t uint64s_per_slot = 0;
91
92   VK_MULTIALLOC_DECL(&ma, struct anv_query_pool, pool, 1);
93
94   VkQueryPipelineStatisticFlags pipeline_statistics = 0;
95   switch (pCreateInfo->queryType) {
96   case VK_QUERY_TYPE_OCCLUSION:
97      /* Occlusion queries have two values: begin and end. */
98      uint64s_per_slot = 1 + 2;
99      break;
100   case VK_QUERY_TYPE_TIMESTAMP:
101      /* Timestamps just have the one timestamp value */
102      uint64s_per_slot = 1 + 1;
103      break;
104   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
105      pipeline_statistics = pCreateInfo->pipelineStatistics;
106      /* We're going to trust this field implicitly so we need to ensure that
107       * no unhandled extension bits leak in.
108       */
109      pipeline_statistics &= ANV_PIPELINE_STATISTICS_MASK;
110
111      /* Statistics queries have a min and max for every statistic */
112      uint64s_per_slot = 1 + 2 * util_bitcount(pipeline_statistics);
113      break;
114   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
115      /* Transform feedback queries are 4 values, begin/end for
116       * written/available.
117       */
118      uint64s_per_slot = 1 + 4;
119      break;
120   case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL: {
121      const struct intel_perf_query_field_layout *layout =
122         &pdevice->perf->query_layout;
123
124      uint64s_per_slot = 2; /* availability + marker */
125      /* Align to the requirement of the layout */
126      uint64s_per_slot = align_u32(uint64s_per_slot,
127                                   DIV_ROUND_UP(layout->alignment, sizeof(uint64_t)));
128      data_offset = uint64s_per_slot * sizeof(uint64_t);
129      /* Add the query data for begin & end commands */
130      uint64s_per_slot += 2 * DIV_ROUND_UP(layout->size, sizeof(uint64_t));
131      break;
132   }
133#if GFX_VER >= 8
134   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR: {
135      const struct intel_perf_query_field_layout *layout =
136         &pdevice->perf->query_layout;
137
138      perf_query_info = vk_find_struct_const(pCreateInfo->pNext,
139                                             QUERY_POOL_PERFORMANCE_CREATE_INFO_KHR);
140      n_passes = intel_perf_get_n_passes(pdevice->perf,
141                                         perf_query_info->pCounterIndices,
142                                         perf_query_info->counterIndexCount,
143                                         NULL);
144      vk_multialloc_add(&ma, &counter_pass, struct intel_perf_counter_pass,
145                             perf_query_info->counterIndexCount);
146      vk_multialloc_add(&ma, &pass_query, struct intel_perf_query_info *,
147                             n_passes);
148      uint64s_per_slot = 4 /* availability + small batch */;
149      /* Align to the requirement of the layout */
150      uint64s_per_slot = align_u32(uint64s_per_slot,
151                                   DIV_ROUND_UP(layout->alignment, sizeof(uint64_t)));
152      data_offset = uint64s_per_slot * sizeof(uint64_t);
153      /* Add the query data for begin & end commands */
154      uint64s_per_slot += 2 * DIV_ROUND_UP(layout->size, sizeof(uint64_t));
155      /* Multiply by the number of passes */
156      uint64s_per_slot *= n_passes;
157      break;
158   }
159#endif
160   default:
161      assert(!"Invalid query type");
162   }
163
164   if (!vk_object_multialloc(&device->vk, &ma, pAllocator,
165                             VK_OBJECT_TYPE_QUERY_POOL))
166      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
167
168   pool->type = pCreateInfo->queryType;
169   pool->pipeline_statistics = pipeline_statistics;
170   pool->stride = uint64s_per_slot * sizeof(uint64_t);
171   pool->slots = pCreateInfo->queryCount;
172
173   if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL) {
174      pool->data_offset = data_offset;
175      pool->snapshot_size = (pool->stride - data_offset) / 2;
176   }
177#if GFX_VER >= 8
178   else if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
179      pool->pass_size = pool->stride / n_passes;
180      pool->data_offset = data_offset;
181      pool->snapshot_size = (pool->pass_size - data_offset) / 2;
182      pool->n_counters = perf_query_info->counterIndexCount;
183      pool->counter_pass = counter_pass;
184      intel_perf_get_counters_passes(pdevice->perf,
185                                     perf_query_info->pCounterIndices,
186                                     perf_query_info->counterIndexCount,
187                                     pool->counter_pass);
188      pool->n_passes = n_passes;
189      pool->pass_query = pass_query;
190      intel_perf_get_n_passes(pdevice->perf,
191                              perf_query_info->pCounterIndices,
192                              perf_query_info->counterIndexCount,
193                              pool->pass_query);
194   }
195#endif
196
197   uint64_t size = pool->slots * (uint64_t)pool->stride;
198   result = anv_device_alloc_bo(device, "query-pool", size,
199                                ANV_BO_ALLOC_MAPPED |
200                                ANV_BO_ALLOC_SNOOPED,
201                                0 /* explicit_address */,
202                                &pool->bo);
203   if (result != VK_SUCCESS)
204      goto fail;
205
206#if GFX_VER >= 8
207   if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
208      for (uint32_t p = 0; p < pool->n_passes; p++) {
209         struct mi_builder b;
210         struct anv_batch batch = {
211            .start = pool->bo->map + khr_perf_query_preamble_offset(pool, p),
212            .end = pool->bo->map + khr_perf_query_preamble_offset(pool, p) + pool->data_offset,
213         };
214         batch.next = batch.start;
215
216         mi_builder_init(&b, &device->info, &batch);
217         mi_store(&b, mi_reg64(ANV_PERF_QUERY_OFFSET_REG),
218                      mi_imm(p * (uint64_t)pool->pass_size));
219         anv_batch_emit(&batch, GENX(MI_BATCH_BUFFER_END), bbe);
220      }
221   }
222#endif
223
224   *pQueryPool = anv_query_pool_to_handle(pool);
225
226   return VK_SUCCESS;
227
228 fail:
229   vk_free2(&device->vk.alloc, pAllocator, pool);
230
231   return result;
232}
233
234void genX(DestroyQueryPool)(
235    VkDevice                                    _device,
236    VkQueryPool                                 _pool,
237    const VkAllocationCallbacks*                pAllocator)
238{
239   ANV_FROM_HANDLE(anv_device, device, _device);
240   ANV_FROM_HANDLE(anv_query_pool, pool, _pool);
241
242   if (!pool)
243      return;
244
245   anv_device_release_bo(device, pool->bo);
246   vk_object_free(&device->vk, pAllocator, pool);
247}
248
249#if GFX_VER >= 8
250/**
251 * VK_KHR_performance_query layout  :
252 *
253 * --------------------------------------------
254 * |       availability (8b)       | |        |
255 * |-------------------------------| |        |
256 * |      Small batch loading      | |        |
257 * |   ANV_PERF_QUERY_OFFSET_REG   | |        |
258 * |            (24b)              | | Pass 0 |
259 * |-------------------------------| |        |
260 * |       some padding (see       | |        |
261 * | query_field_layout:alignment) | |        |
262 * |-------------------------------| |        |
263 * |           query data          | |        |
264 * | (2 * query_field_layout:size) | |        |
265 * |-------------------------------|--        | Query 0
266 * |       availability (8b)       | |        |
267 * |-------------------------------| |        |
268 * |      Small batch loading      | |        |
269 * |   ANV_PERF_QUERY_OFFSET_REG   | |        |
270 * |            (24b)              | | Pass 1 |
271 * |-------------------------------| |        |
272 * |       some padding (see       | |        |
273 * | query_field_layout:alignment) | |        |
274 * |-------------------------------| |        |
275 * |           query data          | |        |
276 * | (2 * query_field_layout:size) | |        |
277 * |-------------------------------|-----------
278 * |       availability (8b)       | |        |
279 * |-------------------------------| |        |
280 * |      Small batch loading      | |        |
281 * |   ANV_PERF_QUERY_OFFSET_REG   | |        |
282 * |            (24b)              | | Pass 0 |
283 * |-------------------------------| |        |
284 * |       some padding (see       | |        |
285 * | query_field_layout:alignment) | |        |
286 * |-------------------------------| |        |
287 * |           query data          | |        |
288 * | (2 * query_field_layout:size) | |        |
289 * |-------------------------------|--        | Query 1
290 * |               ...             | |        |
291 * --------------------------------------------
292 */
293
294static uint64_t
295khr_perf_query_availability_offset(struct anv_query_pool *pool, uint32_t query, uint32_t pass)
296{
297   return query * (uint64_t)pool->stride + pass * (uint64_t)pool->pass_size;
298}
299
300static uint64_t
301khr_perf_query_data_offset(struct anv_query_pool *pool, uint32_t query, uint32_t pass, bool end)
302{
303   return query * (uint64_t)pool->stride + pass * (uint64_t)pool->pass_size +
304      pool->data_offset + (end ? pool->snapshot_size : 0);
305}
306
307static struct anv_address
308khr_perf_query_availability_address(struct anv_query_pool *pool, uint32_t query, uint32_t pass)
309{
310   return anv_address_add(
311      (struct anv_address) { .bo = pool->bo, },
312      khr_perf_query_availability_offset(pool, query, pass));
313}
314
315static struct anv_address
316khr_perf_query_data_address(struct anv_query_pool *pool, uint32_t query, uint32_t pass, bool end)
317{
318   return anv_address_add(
319      (struct anv_address) { .bo = pool->bo, },
320      khr_perf_query_data_offset(pool, query, pass, end));
321}
322
323static bool
324khr_perf_query_ensure_relocs(struct anv_cmd_buffer *cmd_buffer)
325{
326   if (anv_batch_has_error(&cmd_buffer->batch))
327      return false;
328
329   if (cmd_buffer->self_mod_locations)
330      return true;
331
332   struct anv_device *device = cmd_buffer->device;
333   const struct anv_physical_device *pdevice = device->physical;
334
335   cmd_buffer->self_mod_locations =
336      vk_alloc(&cmd_buffer->pool->alloc,
337               pdevice->n_perf_query_commands * sizeof(*cmd_buffer->self_mod_locations), 8,
338               VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
339
340   if (!cmd_buffer->self_mod_locations) {
341      anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
342      return false;
343   }
344
345   return true;
346}
347#endif
348
349/**
350 * VK_INTEL_performance_query layout :
351 *
352 * ---------------------------------
353 * |       availability (8b)       |
354 * |-------------------------------|
355 * |          marker (8b)          |
356 * |-------------------------------|
357 * |       some padding (see       |
358 * | query_field_layout:alignment) |
359 * |-------------------------------|
360 * |           query data          |
361 * | (2 * query_field_layout:size) |
362 * ---------------------------------
363 */
364
365static uint32_t
366intel_perf_marker_offset(void)
367{
368   return 8;
369}
370
371static uint32_t
372intel_perf_query_data_offset(struct anv_query_pool *pool, bool end)
373{
374   return pool->data_offset + (end ? pool->snapshot_size : 0);
375}
376
377static void
378cpu_write_query_result(void *dst_slot, VkQueryResultFlags flags,
379                       uint32_t value_index, uint64_t result)
380{
381   if (flags & VK_QUERY_RESULT_64_BIT) {
382      uint64_t *dst64 = dst_slot;
383      dst64[value_index] = result;
384   } else {
385      uint32_t *dst32 = dst_slot;
386      dst32[value_index] = result;
387   }
388}
389
390static void *
391query_slot(struct anv_query_pool *pool, uint32_t query)
392{
393   return pool->bo->map + query * pool->stride;
394}
395
396static bool
397query_is_available(struct anv_query_pool *pool, uint32_t query)
398{
399#if GFX_VER >= 8
400   if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
401      for (uint32_t p = 0; p < pool->n_passes; p++) {
402         volatile uint64_t *slot =
403            pool->bo->map + khr_perf_query_availability_offset(pool, query, p);
404         if (!slot[0])
405            return false;
406      }
407      return true;
408   }
409#endif
410
411   return *(volatile uint64_t *)query_slot(pool, query);
412}
413
414static VkResult
415wait_for_available(struct anv_device *device,
416                   struct anv_query_pool *pool, uint32_t query)
417{
418   uint64_t abs_timeout = anv_get_absolute_timeout(2 * NSEC_PER_SEC);
419
420   while (anv_gettime_ns() < abs_timeout) {
421      if (query_is_available(pool, query))
422         return VK_SUCCESS;
423      VkResult status = anv_device_query_status(device);
424      if (status != VK_SUCCESS)
425         return status;
426   }
427
428   return anv_device_set_lost(device, "query timeout");
429}
430
431VkResult genX(GetQueryPoolResults)(
432    VkDevice                                    _device,
433    VkQueryPool                                 queryPool,
434    uint32_t                                    firstQuery,
435    uint32_t                                    queryCount,
436    size_t                                      dataSize,
437    void*                                       pData,
438    VkDeviceSize                                stride,
439    VkQueryResultFlags                          flags)
440{
441   ANV_FROM_HANDLE(anv_device, device, _device);
442   ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
443
444   assert(pool->type == VK_QUERY_TYPE_OCCLUSION ||
445          pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS ||
446          pool->type == VK_QUERY_TYPE_TIMESTAMP ||
447          pool->type == VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT ||
448          pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR ||
449          pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL);
450
451   if (anv_device_is_lost(device))
452      return VK_ERROR_DEVICE_LOST;
453
454   if (pData == NULL)
455      return VK_SUCCESS;
456
457   void *data_end = pData + dataSize;
458
459   VkResult status = VK_SUCCESS;
460   for (uint32_t i = 0; i < queryCount; i++) {
461      bool available = query_is_available(pool, firstQuery + i);
462
463      if (!available && (flags & VK_QUERY_RESULT_WAIT_BIT)) {
464         status = wait_for_available(device, pool, firstQuery + i);
465         if (status != VK_SUCCESS) {
466            return status;
467         }
468
469         available = true;
470      }
471
472      /* From the Vulkan 1.0.42 spec:
473       *
474       *    "If VK_QUERY_RESULT_WAIT_BIT and VK_QUERY_RESULT_PARTIAL_BIT are
475       *    both not set then no result values are written to pData for
476       *    queries that are in the unavailable state at the time of the call,
477       *    and vkGetQueryPoolResults returns VK_NOT_READY. However,
478       *    availability state is still written to pData for those queries if
479       *    VK_QUERY_RESULT_WITH_AVAILABILITY_BIT is set."
480       *
481       * From VK_KHR_performance_query :
482       *
483       *    "VK_QUERY_RESULT_PERFORMANCE_QUERY_RECORDED_COUNTERS_BIT_KHR specifies
484       *     that the result should contain the number of counters that were recorded
485       *     into a query pool of type ename:VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR"
486       */
487      bool write_results = available || (flags & VK_QUERY_RESULT_PARTIAL_BIT);
488
489      uint32_t idx = 0;
490      switch (pool->type) {
491      case VK_QUERY_TYPE_OCCLUSION: {
492         uint64_t *slot = query_slot(pool, firstQuery + i);
493         if (write_results) {
494            /* From the Vulkan 1.2.132 spec:
495             *
496             *    "If VK_QUERY_RESULT_PARTIAL_BIT is set,
497             *    VK_QUERY_RESULT_WAIT_BIT is not set, and the query’s status
498             *    is unavailable, an intermediate result value between zero and
499             *    the final result value is written to pData for that query."
500             */
501            uint64_t result = available ? slot[2] - slot[1] : 0;
502            cpu_write_query_result(pData, flags, idx, result);
503         }
504         idx++;
505         break;
506      }
507
508      case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
509         uint64_t *slot = query_slot(pool, firstQuery + i);
510         uint32_t statistics = pool->pipeline_statistics;
511         while (statistics) {
512            uint32_t stat = u_bit_scan(&statistics);
513            if (write_results) {
514               uint64_t result = slot[idx * 2 + 2] - slot[idx * 2 + 1];
515
516               /* WaDividePSInvocationCountBy4:HSW,BDW */
517               if ((device->info.ver == 8 || device->info.is_haswell) &&
518                   (1 << stat) == VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT)
519                  result >>= 2;
520
521               cpu_write_query_result(pData, flags, idx, result);
522            }
523            idx++;
524         }
525         assert(idx == util_bitcount(pool->pipeline_statistics));
526         break;
527      }
528
529      case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT: {
530         uint64_t *slot = query_slot(pool, firstQuery + i);
531         if (write_results)
532            cpu_write_query_result(pData, flags, idx, slot[2] - slot[1]);
533         idx++;
534         if (write_results)
535            cpu_write_query_result(pData, flags, idx, slot[4] - slot[3]);
536         idx++;
537         break;
538      }
539
540      case VK_QUERY_TYPE_TIMESTAMP: {
541         uint64_t *slot = query_slot(pool, firstQuery + i);
542         if (write_results)
543            cpu_write_query_result(pData, flags, idx, slot[1]);
544         idx++;
545         break;
546      }
547
548#if GFX_VER >= 8
549      case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR: {
550         const struct anv_physical_device *pdevice = device->physical;
551         assert((flags & (VK_QUERY_RESULT_WITH_AVAILABILITY_BIT |
552                          VK_QUERY_RESULT_PARTIAL_BIT)) == 0);
553         for (uint32_t p = 0; p < pool->n_passes; p++) {
554            const struct intel_perf_query_info *query = pool->pass_query[p];
555            struct intel_perf_query_result result;
556            intel_perf_query_result_clear(&result);
557            intel_perf_query_result_accumulate_fields(&result, query, &device->info,
558                                                      pool->bo->map + khr_perf_query_data_offset(pool, firstQuery + i, p, false),
559                                                      pool->bo->map + khr_perf_query_data_offset(pool, firstQuery + i, p, true),
560                                                      false /* no_oa_accumulate */);
561            anv_perf_write_pass_results(pdevice->perf, pool, p, &result, pData);
562         }
563         break;
564      }
565#endif
566
567      case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL: {
568         if (!write_results)
569            break;
570         const void *query_data = query_slot(pool, firstQuery + i);
571         const struct intel_perf_query_info *query = &device->physical->perf->queries[0];
572         struct intel_perf_query_result result;
573         intel_perf_query_result_clear(&result);
574         intel_perf_query_result_accumulate_fields(&result, query, &device->info,
575                                                   query_data + intel_perf_query_data_offset(pool, false),
576                                                   query_data + intel_perf_query_data_offset(pool, true),
577                                                   false /* no_oa_accumulate */);
578         intel_perf_query_result_write_mdapi(pData, stride,
579                                             &device->info,
580                                             query, &result);
581         const uint64_t *marker = query_data + intel_perf_marker_offset();
582         intel_perf_query_mdapi_write_marker(pData, stride, &device->info, *marker);
583         break;
584      }
585
586      default:
587         unreachable("invalid pool type");
588      }
589
590      if (!write_results)
591         status = VK_NOT_READY;
592
593      if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
594         cpu_write_query_result(pData, flags, idx, available);
595
596      pData += stride;
597      if (pData >= data_end)
598         break;
599   }
600
601   return status;
602}
603
604static void
605emit_ps_depth_count(struct anv_cmd_buffer *cmd_buffer,
606                    struct anv_address addr)
607{
608   cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
609   genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
610
611   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
612      pc.DestinationAddressType  = DAT_PPGTT;
613      pc.PostSyncOperation       = WritePSDepthCount;
614      pc.DepthStallEnable        = true;
615      pc.Address                 = addr;
616
617      if (GFX_VER == 9 && cmd_buffer->device->info.gt == 4)
618         pc.CommandStreamerStallEnable = true;
619   }
620}
621
622static void
623emit_query_mi_availability(struct mi_builder *b,
624                           struct anv_address addr,
625                           bool available)
626{
627   mi_store(b, mi_mem64(addr), mi_imm(available));
628}
629
630static void
631emit_query_pc_availability(struct anv_cmd_buffer *cmd_buffer,
632                           struct anv_address addr,
633                           bool available)
634{
635   cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
636   genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
637
638   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
639      pc.DestinationAddressType  = DAT_PPGTT;
640      pc.PostSyncOperation       = WriteImmediateData;
641      pc.Address                 = addr;
642      pc.ImmediateData           = available;
643   }
644}
645
646/**
647 * Goes through a series of consecutive query indices in the given pool
648 * setting all element values to 0 and emitting them as available.
649 */
650static void
651emit_zero_queries(struct anv_cmd_buffer *cmd_buffer,
652                  struct mi_builder *b, struct anv_query_pool *pool,
653                  uint32_t first_index, uint32_t num_queries)
654{
655   switch (pool->type) {
656   case VK_QUERY_TYPE_OCCLUSION:
657   case VK_QUERY_TYPE_TIMESTAMP:
658      /* These queries are written with a PIPE_CONTROL so clear them using the
659       * PIPE_CONTROL as well so we don't have to synchronize between 2 types
660       * of operations.
661       */
662      assert((pool->stride % 8) == 0);
663      for (uint32_t i = 0; i < num_queries; i++) {
664         struct anv_address slot_addr =
665            anv_query_address(pool, first_index + i);
666
667         for (uint32_t qword = 1; qword < (pool->stride / 8); qword++) {
668            emit_query_pc_availability(cmd_buffer,
669                                       anv_address_add(slot_addr, qword * 8),
670                                       false);
671         }
672         emit_query_pc_availability(cmd_buffer, slot_addr, true);
673      }
674      break;
675
676   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
677   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
678      for (uint32_t i = 0; i < num_queries; i++) {
679         struct anv_address slot_addr =
680            anv_query_address(pool, first_index + i);
681         mi_memset(b, anv_address_add(slot_addr, 8), 0, pool->stride - 8);
682         emit_query_mi_availability(b, slot_addr, true);
683      }
684      break;
685
686#if GFX_VER >= 8
687   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR: {
688      for (uint32_t i = 0; i < num_queries; i++) {
689         for (uint32_t p = 0; p < pool->n_passes; p++) {
690            mi_memset(b, khr_perf_query_data_address(pool, first_index + i, p, false),
691                         0, 2 * pool->snapshot_size);
692            emit_query_mi_availability(b,
693                                       khr_perf_query_availability_address(pool, first_index + i, p),
694                                       true);
695         }
696      }
697      break;
698   }
699#endif
700
701   case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL:
702      for (uint32_t i = 0; i < num_queries; i++) {
703         struct anv_address slot_addr =
704            anv_query_address(pool, first_index + i);
705         mi_memset(b, anv_address_add(slot_addr, 8), 0, pool->stride - 8);
706         emit_query_mi_availability(b, slot_addr, true);
707      }
708      break;
709
710   default:
711      unreachable("Unsupported query type");
712   }
713}
714
715void genX(CmdResetQueryPool)(
716    VkCommandBuffer                             commandBuffer,
717    VkQueryPool                                 queryPool,
718    uint32_t                                    firstQuery,
719    uint32_t                                    queryCount)
720{
721   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
722   ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
723
724   switch (pool->type) {
725   case VK_QUERY_TYPE_OCCLUSION:
726   case VK_QUERY_TYPE_TIMESTAMP:
727      for (uint32_t i = 0; i < queryCount; i++) {
728         emit_query_pc_availability(cmd_buffer,
729                                    anv_query_address(pool, firstQuery + i),
730                                    false);
731      }
732      break;
733
734   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
735   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT: {
736      struct mi_builder b;
737      mi_builder_init(&b, &cmd_buffer->device->info, &cmd_buffer->batch);
738
739      for (uint32_t i = 0; i < queryCount; i++)
740         emit_query_mi_availability(&b, anv_query_address(pool, firstQuery + i), false);
741      break;
742   }
743
744#if GFX_VER >= 8
745   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR: {
746      struct mi_builder b;
747      mi_builder_init(&b, &cmd_buffer->device->info, &cmd_buffer->batch);
748
749      for (uint32_t i = 0; i < queryCount; i++) {
750         for (uint32_t p = 0; p < pool->n_passes; p++) {
751            emit_query_mi_availability(
752               &b,
753               khr_perf_query_availability_address(pool, firstQuery + i, p),
754               false);
755         }
756      }
757      break;
758   }
759#endif
760
761   case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL: {
762      struct mi_builder b;
763      mi_builder_init(&b, &cmd_buffer->device->info, &cmd_buffer->batch);
764
765      for (uint32_t i = 0; i < queryCount; i++)
766         emit_query_mi_availability(&b, anv_query_address(pool, firstQuery + i), false);
767      break;
768   }
769
770   default:
771      unreachable("Unsupported query type");
772   }
773}
774
775void genX(ResetQueryPool)(
776    VkDevice                                    _device,
777    VkQueryPool                                 queryPool,
778    uint32_t                                    firstQuery,
779    uint32_t                                    queryCount)
780{
781   ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
782
783   for (uint32_t i = 0; i < queryCount; i++) {
784      if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
785#if GFX_VER >= 8
786         for (uint32_t p = 0; p < pool->n_passes; p++) {
787            uint64_t *pass_slot = pool->bo->map +
788               khr_perf_query_availability_offset(pool, firstQuery + i, p);
789            *pass_slot = 0;
790         }
791#endif
792      } else {
793         uint64_t *slot = query_slot(pool, firstQuery + i);
794         *slot = 0;
795      }
796   }
797}
798
799static const uint32_t vk_pipeline_stat_to_reg[] = {
800   GENX(IA_VERTICES_COUNT_num),
801   GENX(IA_PRIMITIVES_COUNT_num),
802   GENX(VS_INVOCATION_COUNT_num),
803   GENX(GS_INVOCATION_COUNT_num),
804   GENX(GS_PRIMITIVES_COUNT_num),
805   GENX(CL_INVOCATION_COUNT_num),
806   GENX(CL_PRIMITIVES_COUNT_num),
807   GENX(PS_INVOCATION_COUNT_num),
808   GENX(HS_INVOCATION_COUNT_num),
809   GENX(DS_INVOCATION_COUNT_num),
810   GENX(CS_INVOCATION_COUNT_num),
811};
812
813static void
814emit_pipeline_stat(struct mi_builder *b, uint32_t stat,
815                   struct anv_address addr)
816{
817   STATIC_ASSERT(ANV_PIPELINE_STATISTICS_MASK ==
818                 (1 << ARRAY_SIZE(vk_pipeline_stat_to_reg)) - 1);
819
820   assert(stat < ARRAY_SIZE(vk_pipeline_stat_to_reg));
821   mi_store(b, mi_mem64(addr), mi_reg64(vk_pipeline_stat_to_reg[stat]));
822}
823
824static void
825emit_xfb_query(struct mi_builder *b, uint32_t stream,
826               struct anv_address addr)
827{
828   assert(stream < MAX_XFB_STREAMS);
829
830   mi_store(b, mi_mem64(anv_address_add(addr, 0)),
831               mi_reg64(GENX(SO_NUM_PRIMS_WRITTEN0_num) + stream * 8));
832   mi_store(b, mi_mem64(anv_address_add(addr, 16)),
833               mi_reg64(GENX(SO_PRIM_STORAGE_NEEDED0_num) + stream * 8));
834}
835
836static void
837emit_perf_intel_query(struct anv_cmd_buffer *cmd_buffer,
838                      struct anv_query_pool *pool,
839                      struct mi_builder *b,
840                      struct anv_address query_addr,
841                      bool end)
842{
843   const struct intel_perf_query_field_layout *layout =
844      &cmd_buffer->device->physical->perf->query_layout;
845   struct anv_address data_addr =
846      anv_address_add(query_addr, intel_perf_query_data_offset(pool, end));
847
848   for (uint32_t f = 0; f < layout->n_fields; f++) {
849      const struct intel_perf_query_field *field =
850         &layout->fields[end ? f : (layout->n_fields - 1 - f)];
851
852      switch (field->type) {
853      case INTEL_PERF_QUERY_FIELD_TYPE_MI_RPC:
854         anv_batch_emit(&cmd_buffer->batch, GENX(MI_REPORT_PERF_COUNT), rpc) {
855            rpc.MemoryAddress = anv_address_add(data_addr, field->location);
856         }
857         break;
858
859      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_PERFCNT:
860      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_RPSTAT:
861      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_B:
862      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_C: {
863         struct anv_address addr = anv_address_add(data_addr, field->location);
864         struct mi_value src = field->size == 8 ?
865            mi_reg64(field->mmio_offset) :
866            mi_reg32(field->mmio_offset);
867         struct mi_value dst = field->size == 8 ?
868            mi_mem64(addr) : mi_mem32(addr);
869         mi_store(b, dst, src);
870         break;
871      }
872
873      default:
874         unreachable("Invalid query field");
875         break;
876      }
877   }
878}
879
880void genX(CmdBeginQuery)(
881    VkCommandBuffer                             commandBuffer,
882    VkQueryPool                                 queryPool,
883    uint32_t                                    query,
884    VkQueryControlFlags                         flags)
885{
886   genX(CmdBeginQueryIndexedEXT)(commandBuffer, queryPool, query, flags, 0);
887}
888
889void genX(CmdBeginQueryIndexedEXT)(
890    VkCommandBuffer                             commandBuffer,
891    VkQueryPool                                 queryPool,
892    uint32_t                                    query,
893    VkQueryControlFlags                         flags,
894    uint32_t                                    index)
895{
896   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
897   ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
898   struct anv_address query_addr = anv_query_address(pool, query);
899
900   struct mi_builder b;
901   mi_builder_init(&b, &cmd_buffer->device->info, &cmd_buffer->batch);
902
903   switch (pool->type) {
904   case VK_QUERY_TYPE_OCCLUSION:
905      emit_ps_depth_count(cmd_buffer, anv_address_add(query_addr, 8));
906      break;
907
908   case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
909      /* TODO: This might only be necessary for certain stats */
910      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
911         pc.CommandStreamerStallEnable = true;
912         pc.StallAtPixelScoreboard = true;
913      }
914
915      uint32_t statistics = pool->pipeline_statistics;
916      uint32_t offset = 8;
917      while (statistics) {
918         uint32_t stat = u_bit_scan(&statistics);
919         emit_pipeline_stat(&b, stat, anv_address_add(query_addr, offset));
920         offset += 16;
921      }
922      break;
923   }
924
925   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
926      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
927         pc.CommandStreamerStallEnable = true;
928         pc.StallAtPixelScoreboard = true;
929      }
930      emit_xfb_query(&b, index, anv_address_add(query_addr, 8));
931      break;
932
933#if GFX_VER >= 8
934   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR: {
935      if (!khr_perf_query_ensure_relocs(cmd_buffer))
936         return;
937
938      const struct anv_physical_device *pdevice = cmd_buffer->device->physical;
939      const struct intel_perf_query_field_layout *layout = &pdevice->perf->query_layout;
940
941      uint32_t reloc_idx = 0;
942      for (uint32_t end = 0; end < 2; end++) {
943         for (uint32_t r = 0; r < layout->n_fields; r++) {
944            const struct intel_perf_query_field *field =
945               &layout->fields[end ? r : (layout->n_fields - 1 - r)];
946            struct mi_value reg_addr =
947               mi_iadd(
948                  &b,
949                  mi_imm(intel_canonical_address(pool->bo->offset +
950                                                 khr_perf_query_data_offset(pool, query, 0, end) +
951                                                 field->location)),
952                  mi_reg64(ANV_PERF_QUERY_OFFSET_REG));
953            cmd_buffer->self_mod_locations[reloc_idx++] = mi_store_address(&b, reg_addr);
954
955            if (field->type != INTEL_PERF_QUERY_FIELD_TYPE_MI_RPC &&
956                field->size == 8) {
957               reg_addr =
958                  mi_iadd(
959                     &b,
960                     mi_imm(intel_canonical_address(pool->bo->offset +
961                                                    khr_perf_query_data_offset(pool, query, 0, end) +
962                                                    field->location + 4)),
963                     mi_reg64(ANV_PERF_QUERY_OFFSET_REG));
964               cmd_buffer->self_mod_locations[reloc_idx++] = mi_store_address(&b, reg_addr);
965            }
966         }
967      }
968
969      struct mi_value availability_write_offset =
970         mi_iadd(
971            &b,
972            mi_imm(
973               intel_canonical_address(
974                  pool->bo->offset +
975                  khr_perf_query_availability_offset(pool, query, 0 /* pass */))),
976            mi_reg64(ANV_PERF_QUERY_OFFSET_REG));
977      cmd_buffer->self_mod_locations[reloc_idx++] =
978         mi_store_address(&b, availability_write_offset);
979
980      assert(reloc_idx == pdevice->n_perf_query_commands);
981
982      mi_self_mod_barrier(&b);
983
984      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
985         pc.CommandStreamerStallEnable = true;
986         pc.StallAtPixelScoreboard = true;
987      }
988      cmd_buffer->perf_query_pool = pool;
989
990      cmd_buffer->perf_reloc_idx = 0;
991      for (uint32_t r = 0; r < layout->n_fields; r++) {
992         const struct intel_perf_query_field *field =
993            &layout->fields[layout->n_fields - 1 - r];
994         void *dws;
995
996         switch (field->type) {
997         case INTEL_PERF_QUERY_FIELD_TYPE_MI_RPC:
998            dws = anv_batch_emitn(&cmd_buffer->batch,
999                                  GENX(MI_REPORT_PERF_COUNT_length),
1000                                  GENX(MI_REPORT_PERF_COUNT),
1001                                  .MemoryAddress = query_addr /* Will be overwritten */);
1002            _mi_resolve_address_token(&b,
1003                                      cmd_buffer->self_mod_locations[cmd_buffer->perf_reloc_idx++],
1004                                      dws +
1005                                      GENX(MI_REPORT_PERF_COUNT_MemoryAddress_start) / 8);
1006            break;
1007
1008         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_PERFCNT:
1009         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_RPSTAT:
1010         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_B:
1011         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_C:
1012            dws =
1013               anv_batch_emitn(&cmd_buffer->batch,
1014                               GENX(MI_STORE_REGISTER_MEM_length),
1015                               GENX(MI_STORE_REGISTER_MEM),
1016                               .RegisterAddress = field->mmio_offset,
1017                               .MemoryAddress = query_addr /* Will be overwritten */ );
1018            _mi_resolve_address_token(&b,
1019                                      cmd_buffer->self_mod_locations[cmd_buffer->perf_reloc_idx++],
1020                                      dws +
1021                                      GENX(MI_STORE_REGISTER_MEM_MemoryAddress_start) / 8);
1022            if (field->size == 8) {
1023               dws =
1024                  anv_batch_emitn(&cmd_buffer->batch,
1025                                  GENX(MI_STORE_REGISTER_MEM_length),
1026                                  GENX(MI_STORE_REGISTER_MEM),
1027                                  .RegisterAddress = field->mmio_offset + 4,
1028                                  .MemoryAddress = query_addr /* Will be overwritten */ );
1029               _mi_resolve_address_token(&b,
1030                                         cmd_buffer->self_mod_locations[cmd_buffer->perf_reloc_idx++],
1031                                         dws +
1032                                         GENX(MI_STORE_REGISTER_MEM_MemoryAddress_start) / 8);
1033            }
1034            break;
1035
1036         default:
1037            unreachable("Invalid query field");
1038            break;
1039         }
1040      }
1041      break;
1042   }
1043#endif
1044
1045   case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL: {
1046      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
1047         pc.CommandStreamerStallEnable = true;
1048         pc.StallAtPixelScoreboard = true;
1049      }
1050      emit_perf_intel_query(cmd_buffer, pool, &b, query_addr, false);
1051      break;
1052   }
1053
1054   default:
1055      unreachable("");
1056   }
1057}
1058
1059void genX(CmdEndQuery)(
1060    VkCommandBuffer                             commandBuffer,
1061    VkQueryPool                                 queryPool,
1062    uint32_t                                    query)
1063{
1064   genX(CmdEndQueryIndexedEXT)(commandBuffer, queryPool, query, 0);
1065}
1066
1067void genX(CmdEndQueryIndexedEXT)(
1068    VkCommandBuffer                             commandBuffer,
1069    VkQueryPool                                 queryPool,
1070    uint32_t                                    query,
1071    uint32_t                                    index)
1072{
1073   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1074   ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
1075   struct anv_address query_addr = anv_query_address(pool, query);
1076
1077   struct mi_builder b;
1078   mi_builder_init(&b, &cmd_buffer->device->info, &cmd_buffer->batch);
1079
1080   switch (pool->type) {
1081   case VK_QUERY_TYPE_OCCLUSION:
1082      emit_ps_depth_count(cmd_buffer, anv_address_add(query_addr, 16));
1083      emit_query_pc_availability(cmd_buffer, query_addr, true);
1084      break;
1085
1086   case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
1087      /* TODO: This might only be necessary for certain stats */
1088      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
1089         pc.CommandStreamerStallEnable = true;
1090         pc.StallAtPixelScoreboard = true;
1091      }
1092
1093      uint32_t statistics = pool->pipeline_statistics;
1094      uint32_t offset = 16;
1095      while (statistics) {
1096         uint32_t stat = u_bit_scan(&statistics);
1097         emit_pipeline_stat(&b, stat, anv_address_add(query_addr, offset));
1098         offset += 16;
1099      }
1100
1101      emit_query_mi_availability(&b, query_addr, true);
1102      break;
1103   }
1104
1105   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
1106      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
1107         pc.CommandStreamerStallEnable = true;
1108         pc.StallAtPixelScoreboard = true;
1109      }
1110
1111      emit_xfb_query(&b, index, anv_address_add(query_addr, 16));
1112      emit_query_mi_availability(&b, query_addr, true);
1113      break;
1114
1115#if GFX_VER >= 8
1116   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR: {
1117      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
1118         pc.CommandStreamerStallEnable = true;
1119         pc.StallAtPixelScoreboard = true;
1120      }
1121      cmd_buffer->perf_query_pool = pool;
1122
1123      if (!khr_perf_query_ensure_relocs(cmd_buffer))
1124         return;
1125
1126      const struct anv_physical_device *pdevice = cmd_buffer->device->physical;
1127      const struct intel_perf_query_field_layout *layout = &pdevice->perf->query_layout;
1128
1129      void *dws;
1130      for (uint32_t r = 0; r < layout->n_fields; r++) {
1131         const struct intel_perf_query_field *field = &layout->fields[r];
1132
1133         switch (field->type) {
1134         case INTEL_PERF_QUERY_FIELD_TYPE_MI_RPC:
1135            dws = anv_batch_emitn(&cmd_buffer->batch,
1136                                  GENX(MI_REPORT_PERF_COUNT_length),
1137                                  GENX(MI_REPORT_PERF_COUNT),
1138                                  .MemoryAddress = query_addr /* Will be overwritten */);
1139            _mi_resolve_address_token(&b,
1140                                      cmd_buffer->self_mod_locations[cmd_buffer->perf_reloc_idx++],
1141                                      dws +
1142                                      GENX(MI_REPORT_PERF_COUNT_MemoryAddress_start) / 8);
1143            break;
1144
1145         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_PERFCNT:
1146         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_RPSTAT:
1147         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_B:
1148         case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_C:
1149            dws =
1150               anv_batch_emitn(&cmd_buffer->batch,
1151                               GENX(MI_STORE_REGISTER_MEM_length),
1152                               GENX(MI_STORE_REGISTER_MEM),
1153                               .RegisterAddress = field->mmio_offset,
1154                               .MemoryAddress = query_addr /* Will be overwritten */ );
1155            _mi_resolve_address_token(&b,
1156                                      cmd_buffer->self_mod_locations[cmd_buffer->perf_reloc_idx++],
1157                                      dws +
1158                                      GENX(MI_STORE_REGISTER_MEM_MemoryAddress_start) / 8);
1159            if (field->size == 8) {
1160               dws =
1161                  anv_batch_emitn(&cmd_buffer->batch,
1162                                  GENX(MI_STORE_REGISTER_MEM_length),
1163                                  GENX(MI_STORE_REGISTER_MEM),
1164                                  .RegisterAddress = field->mmio_offset + 4,
1165                                  .MemoryAddress = query_addr /* Will be overwritten */ );
1166               _mi_resolve_address_token(&b,
1167                                         cmd_buffer->self_mod_locations[cmd_buffer->perf_reloc_idx++],
1168                                         dws +
1169                                         GENX(MI_STORE_REGISTER_MEM_MemoryAddress_start) / 8);
1170            }
1171            break;
1172
1173         default:
1174            unreachable("Invalid query field");
1175            break;
1176         }
1177      }
1178
1179      dws =
1180         anv_batch_emitn(&cmd_buffer->batch,
1181                         GENX(MI_STORE_DATA_IMM_length),
1182                         GENX(MI_STORE_DATA_IMM),
1183                         .ImmediateData = true);
1184      _mi_resolve_address_token(&b,
1185                                cmd_buffer->self_mod_locations[cmd_buffer->perf_reloc_idx++],
1186                                dws +
1187                                GENX(MI_STORE_DATA_IMM_Address_start) / 8);
1188
1189      assert(cmd_buffer->perf_reloc_idx == pdevice->n_perf_query_commands);
1190      break;
1191   }
1192#endif
1193
1194   case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL: {
1195      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
1196         pc.CommandStreamerStallEnable = true;
1197         pc.StallAtPixelScoreboard = true;
1198      }
1199      uint32_t marker_offset = intel_perf_marker_offset();
1200      mi_store(&b, mi_mem64(anv_address_add(query_addr, marker_offset)),
1201                   mi_imm(cmd_buffer->intel_perf_marker));
1202      emit_perf_intel_query(cmd_buffer, pool, &b, query_addr, true);
1203      emit_query_mi_availability(&b, query_addr, true);
1204      break;
1205   }
1206
1207   default:
1208      unreachable("");
1209   }
1210
1211   /* When multiview is active the spec requires that N consecutive query
1212    * indices are used, where N is the number of active views in the subpass.
1213    * The spec allows that we only write the results to one of the queries
1214    * but we still need to manage result availability for all the query indices.
1215    * Since we only emit a single query for all active views in the
1216    * first index, mark the other query indices as being already available
1217    * with result 0.
1218    */
1219   if (cmd_buffer->state.subpass && cmd_buffer->state.subpass->view_mask) {
1220      const uint32_t num_queries =
1221         util_bitcount(cmd_buffer->state.subpass->view_mask);
1222      if (num_queries > 1)
1223         emit_zero_queries(cmd_buffer, &b, pool, query + 1, num_queries - 1);
1224   }
1225}
1226
1227#define TIMESTAMP 0x2358
1228
1229void genX(CmdWriteTimestamp2KHR)(
1230    VkCommandBuffer                             commandBuffer,
1231    VkPipelineStageFlags2KHR                    stage,
1232    VkQueryPool                                 queryPool,
1233    uint32_t                                    query)
1234{
1235   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1236   ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
1237   struct anv_address query_addr = anv_query_address(pool, query);
1238
1239   assert(pool->type == VK_QUERY_TYPE_TIMESTAMP);
1240
1241   struct mi_builder b;
1242   mi_builder_init(&b, &cmd_buffer->device->info, &cmd_buffer->batch);
1243
1244   if (stage == VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT_KHR) {
1245      mi_store(&b, mi_mem64(anv_address_add(query_addr, 8)),
1246                   mi_reg64(TIMESTAMP));
1247   } else {
1248      /* Everything else is bottom-of-pipe */
1249      cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
1250      genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
1251
1252      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
1253         pc.DestinationAddressType  = DAT_PPGTT;
1254         pc.PostSyncOperation       = WriteTimestamp;
1255         pc.Address                 = anv_address_add(query_addr, 8);
1256
1257         if (GFX_VER == 9 && cmd_buffer->device->info.gt == 4)
1258            pc.CommandStreamerStallEnable = true;
1259      }
1260   }
1261
1262   emit_query_pc_availability(cmd_buffer, query_addr, true);
1263
1264   /* When multiview is active the spec requires that N consecutive query
1265    * indices are used, where N is the number of active views in the subpass.
1266    * The spec allows that we only write the results to one of the queries
1267    * but we still need to manage result availability for all the query indices.
1268    * Since we only emit a single query for all active views in the
1269    * first index, mark the other query indices as being already available
1270    * with result 0.
1271    */
1272   if (cmd_buffer->state.subpass && cmd_buffer->state.subpass->view_mask) {
1273      const uint32_t num_queries =
1274         util_bitcount(cmd_buffer->state.subpass->view_mask);
1275      if (num_queries > 1)
1276         emit_zero_queries(cmd_buffer, &b, pool, query + 1, num_queries - 1);
1277   }
1278}
1279
1280#if GFX_VERx10 >= 75
1281
1282#define MI_PREDICATE_SRC0    0x2400
1283#define MI_PREDICATE_SRC1    0x2408
1284#define MI_PREDICATE_RESULT  0x2418
1285
1286/**
1287 * Writes the results of a query to dst_addr is the value at poll_addr is equal
1288 * to the reference value.
1289 */
1290static void
1291gpu_write_query_result_cond(struct anv_cmd_buffer *cmd_buffer,
1292                            struct mi_builder *b,
1293                            struct anv_address poll_addr,
1294                            struct anv_address dst_addr,
1295                            uint64_t ref_value,
1296                            VkQueryResultFlags flags,
1297                            uint32_t value_index,
1298                            struct mi_value query_result)
1299{
1300   mi_store(b, mi_reg64(MI_PREDICATE_SRC0), mi_mem64(poll_addr));
1301   mi_store(b, mi_reg64(MI_PREDICATE_SRC1), mi_imm(ref_value));
1302   anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
1303      mip.LoadOperation    = LOAD_LOAD;
1304      mip.CombineOperation = COMBINE_SET;
1305      mip.CompareOperation = COMPARE_SRCS_EQUAL;
1306   }
1307
1308   if (flags & VK_QUERY_RESULT_64_BIT) {
1309      struct anv_address res_addr = anv_address_add(dst_addr, value_index * 8);
1310      mi_store_if(b, mi_mem64(res_addr), query_result);
1311   } else {
1312      struct anv_address res_addr = anv_address_add(dst_addr, value_index * 4);
1313      mi_store_if(b, mi_mem32(res_addr), query_result);
1314   }
1315}
1316
1317static void
1318gpu_write_query_result(struct mi_builder *b,
1319                       struct anv_address dst_addr,
1320                       VkQueryResultFlags flags,
1321                       uint32_t value_index,
1322                       struct mi_value query_result)
1323{
1324   if (flags & VK_QUERY_RESULT_64_BIT) {
1325      struct anv_address res_addr = anv_address_add(dst_addr, value_index * 8);
1326      mi_store(b, mi_mem64(res_addr), query_result);
1327   } else {
1328      struct anv_address res_addr = anv_address_add(dst_addr, value_index * 4);
1329      mi_store(b, mi_mem32(res_addr), query_result);
1330   }
1331}
1332
1333static struct mi_value
1334compute_query_result(struct mi_builder *b, struct anv_address addr)
1335{
1336   return mi_isub(b, mi_mem64(anv_address_add(addr, 8)),
1337                     mi_mem64(anv_address_add(addr, 0)));
1338}
1339
1340void genX(CmdCopyQueryPoolResults)(
1341    VkCommandBuffer                             commandBuffer,
1342    VkQueryPool                                 queryPool,
1343    uint32_t                                    firstQuery,
1344    uint32_t                                    queryCount,
1345    VkBuffer                                    destBuffer,
1346    VkDeviceSize                                destOffset,
1347    VkDeviceSize                                destStride,
1348    VkQueryResultFlags                          flags)
1349{
1350   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1351   ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
1352   ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer);
1353
1354   struct mi_builder b;
1355   mi_builder_init(&b, &cmd_buffer->device->info, &cmd_buffer->batch);
1356   struct mi_value result;
1357
1358   /* If render target writes are ongoing, request a render target cache flush
1359    * to ensure proper ordering of the commands from the 3d pipe and the
1360    * command streamer.
1361    */
1362   if (cmd_buffer->state.pending_pipe_bits & ANV_PIPE_RENDER_TARGET_BUFFER_WRITES) {
1363      anv_add_pending_pipe_bits(cmd_buffer,
1364                                ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT,
1365                                "CopyQueryPoolResults");
1366   }
1367
1368   if ((flags & VK_QUERY_RESULT_WAIT_BIT) ||
1369       (cmd_buffer->state.pending_pipe_bits & ANV_PIPE_FLUSH_BITS) ||
1370       /* Occlusion & timestamp queries are written using a PIPE_CONTROL and
1371        * because we're about to copy values from MI commands, we need to
1372        * stall the command streamer to make sure the PIPE_CONTROL values have
1373        * landed, otherwise we could see inconsistent values & availability.
1374        *
1375        *  From the vulkan spec:
1376        *
1377        *     "vkCmdCopyQueryPoolResults is guaranteed to see the effect of
1378        *     previous uses of vkCmdResetQueryPool in the same queue, without
1379        *     any additional synchronization."
1380        */
1381       pool->type == VK_QUERY_TYPE_OCCLUSION ||
1382       pool->type == VK_QUERY_TYPE_TIMESTAMP) {
1383      anv_add_pending_pipe_bits(cmd_buffer,
1384                                ANV_PIPE_CS_STALL_BIT,
1385                                "CopyQueryPoolResults");
1386      genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
1387   }
1388
1389   struct anv_address dest_addr = anv_address_add(buffer->address, destOffset);
1390   for (uint32_t i = 0; i < queryCount; i++) {
1391      struct anv_address query_addr = anv_query_address(pool, firstQuery + i);
1392      uint32_t idx = 0;
1393      switch (pool->type) {
1394      case VK_QUERY_TYPE_OCCLUSION:
1395         result = compute_query_result(&b, anv_address_add(query_addr, 8));
1396         /* Like in the case of vkGetQueryPoolResults, if the query is
1397          * unavailable and the VK_QUERY_RESULT_PARTIAL_BIT flag is set,
1398          * conservatively write 0 as the query result. If the
1399          * VK_QUERY_RESULT_PARTIAL_BIT isn't set, don't write any value.
1400          */
1401         gpu_write_query_result_cond(cmd_buffer, &b, query_addr, dest_addr,
1402               1 /* available */, flags, idx, result);
1403         if (flags & VK_QUERY_RESULT_PARTIAL_BIT) {
1404            gpu_write_query_result_cond(cmd_buffer, &b, query_addr, dest_addr,
1405                  0 /* unavailable */, flags, idx, mi_imm(0));
1406         }
1407         idx++;
1408         break;
1409
1410      case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
1411         uint32_t statistics = pool->pipeline_statistics;
1412         while (statistics) {
1413            uint32_t stat = u_bit_scan(&statistics);
1414
1415            result = compute_query_result(&b, anv_address_add(query_addr,
1416                                                              idx * 16 + 8));
1417
1418            /* WaDividePSInvocationCountBy4:HSW,BDW */
1419            if ((cmd_buffer->device->info.ver == 8 ||
1420                 cmd_buffer->device->info.is_haswell) &&
1421                (1 << stat) == VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT) {
1422               result = mi_ushr32_imm(&b, result, 2);
1423            }
1424
1425            gpu_write_query_result(&b, dest_addr, flags, idx++, result);
1426         }
1427         assert(idx == util_bitcount(pool->pipeline_statistics));
1428         break;
1429      }
1430
1431      case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
1432         result = compute_query_result(&b, anv_address_add(query_addr, 8));
1433         gpu_write_query_result(&b, dest_addr, flags, idx++, result);
1434         result = compute_query_result(&b, anv_address_add(query_addr, 24));
1435         gpu_write_query_result(&b, dest_addr, flags, idx++, result);
1436         break;
1437
1438      case VK_QUERY_TYPE_TIMESTAMP:
1439         result = mi_mem64(anv_address_add(query_addr, 8));
1440         gpu_write_query_result(&b, dest_addr, flags, idx++, result);
1441         break;
1442
1443#if GFX_VER >= 8
1444      case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR:
1445         unreachable("Copy KHR performance query results not implemented");
1446         break;
1447#endif
1448
1449      default:
1450         unreachable("unhandled query type");
1451      }
1452
1453      if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
1454         gpu_write_query_result(&b, dest_addr, flags, idx,
1455                                mi_mem64(query_addr));
1456      }
1457
1458      dest_addr = anv_address_add(dest_addr, destStride);
1459   }
1460}
1461
1462#else
1463void genX(CmdCopyQueryPoolResults)(
1464    VkCommandBuffer                             commandBuffer,
1465    VkQueryPool                                 queryPool,
1466    uint32_t                                    firstQuery,
1467    uint32_t                                    queryCount,
1468    VkBuffer                                    destBuffer,
1469    VkDeviceSize                                destOffset,
1470    VkDeviceSize                                destStride,
1471    VkQueryResultFlags                          flags)
1472{
1473   anv_finishme("Queries not yet supported on Ivy Bridge");
1474}
1475#endif
1476