2 * Copyright © 2015 Intel Corporation
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:
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
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
30 #include "anv_private.h"
32 #include "genxml/gen_macros.h"
33 #include "genxml/genX_pack.h"
35 /* We reserve GPR 14 and 15 for conditional rendering */
36 #define GEN_MI_BUILDER_NUM_ALLOC_GPRS 14
37 #define __gen_get_batch_dwords anv_batch_emit_dwords
38 #define __gen_address_offset anv_address_add
39 #include "common/gen_mi_builder.h"
40 #include "perf/gen_perf.h"
41 #include "perf/gen_perf_mdapi.h"
43 #define OA_REPORT_N_UINT64 (256 / sizeof(uint64_t))
45 VkResult
genX(CreateQueryPool
)(
47 const VkQueryPoolCreateInfo
* pCreateInfo
,
48 const VkAllocationCallbacks
* pAllocator
,
49 VkQueryPool
* pQueryPool
)
51 ANV_FROM_HANDLE(anv_device
, device
, _device
);
52 const struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
53 struct anv_query_pool
*pool
;
56 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO
);
58 /* Query pool slots are made up of some number of 64-bit values packed
59 * tightly together. For most query types have the first 64-bit value is
60 * the "available" bit which is 0 when the query is unavailable and 1 when
61 * it is available. The 64-bit values that follow are determined by the
64 * For performance queries, we have a requirement to align OA reports at
65 * 64bytes so we put those first and have the "available" bit behind
66 * together with some other counters.
68 uint32_t uint64s_per_slot
= 1;
70 VkQueryPipelineStatisticFlags pipeline_statistics
= 0;
71 switch (pCreateInfo
->queryType
) {
72 case VK_QUERY_TYPE_OCCLUSION
:
73 /* Occlusion queries have two values: begin and end. */
74 uint64s_per_slot
+= 2;
76 case VK_QUERY_TYPE_TIMESTAMP
:
77 /* Timestamps just have the one timestamp value */
78 uint64s_per_slot
+= 1;
80 case VK_QUERY_TYPE_PIPELINE_STATISTICS
:
81 pipeline_statistics
= pCreateInfo
->pipelineStatistics
;
82 /* We're going to trust this field implicitly so we need to ensure that
83 * no unhandled extension bits leak in.
85 pipeline_statistics
&= ANV_PIPELINE_STATISTICS_MASK
;
87 /* Statistics queries have a min and max for every statistic */
88 uint64s_per_slot
+= 2 * util_bitcount(pipeline_statistics
);
90 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
:
91 /* Transform feedback queries are 4 values, begin/end for
94 uint64s_per_slot
+= 4;
96 case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL
: {
97 uint64s_per_slot
= 72; /* 576 bytes, see layout below */
101 assert(!"Invalid query type");
104 pool
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pool
), 8,
105 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
107 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
109 pool
->type
= pCreateInfo
->queryType
;
110 pool
->pipeline_statistics
= pipeline_statistics
;
111 pool
->stride
= uint64s_per_slot
* sizeof(uint64_t);
112 pool
->slots
= pCreateInfo
->queryCount
;
114 uint32_t bo_flags
= 0;
115 if (pdevice
->supports_48bit_addresses
)
116 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
118 if (pdevice
->use_softpin
)
119 bo_flags
|= EXEC_OBJECT_PINNED
;
121 if (pdevice
->has_exec_async
)
122 bo_flags
|= EXEC_OBJECT_ASYNC
;
124 uint64_t size
= pool
->slots
* pool
->stride
;
125 result
= anv_device_alloc_bo(device
, size
,
126 ANV_BO_ALLOC_MAPPED
|
127 ANV_BO_ALLOC_SNOOPED
,
129 if (result
!= VK_SUCCESS
)
132 *pQueryPool
= anv_query_pool_to_handle(pool
);
137 vk_free2(&device
->alloc
, pAllocator
, pool
);
142 void genX(DestroyQueryPool
)(
145 const VkAllocationCallbacks
* pAllocator
)
147 ANV_FROM_HANDLE(anv_device
, device
, _device
);
148 ANV_FROM_HANDLE(anv_query_pool
, pool
, _pool
);
153 anv_device_release_bo(device
, pool
->bo
);
154 vk_free2(&device
->alloc
, pAllocator
, pool
);
157 static struct anv_address
158 anv_query_address(struct anv_query_pool
*pool
, uint32_t query
)
160 return (struct anv_address
) {
162 .offset
= query
* pool
->stride
,
167 * VK_INTEL_performance_query layout (576 bytes) :
169 * ------------------------------
170 * | availability (8b) |
171 * |----------------------------|
173 * |----------------------------|
174 * | begin RPSTAT register (4b) |
175 * |----------------------------|
176 * | end RPSTAT register (4b) |
177 * |----------------------------|
178 * | begin perfcntr 1 & 2 (16b) |
179 * |----------------------------|
180 * | end perfcntr 1 & 2 (16b) |
181 * |----------------------------|
183 * |----------------------------|
184 * | begin MI_RPC (256b) |
185 * |----------------------------|
186 * | end MI_RPC (256b) |
187 * ------------------------------
191 intel_perf_marker_offset(void)
197 intel_perf_rpstart_offset(bool end
)
199 return 16 + (end
? sizeof(uint32_t) : 0);
203 intel_perf_counter(bool end
)
205 return 24 + (end
? (2 * sizeof(uint64_t)) : 0);
209 intel_perf_mi_rpc_offset(bool end
)
211 return 64 + (end
? 256 : 0);
215 cpu_write_query_result(void *dst_slot
, VkQueryResultFlags flags
,
216 uint32_t value_index
, uint64_t result
)
218 if (flags
& VK_QUERY_RESULT_64_BIT
) {
219 uint64_t *dst64
= dst_slot
;
220 dst64
[value_index
] = result
;
222 uint32_t *dst32
= dst_slot
;
223 dst32
[value_index
] = result
;
228 query_slot(struct anv_query_pool
*pool
, uint32_t query
)
230 return pool
->bo
->map
+ query
* pool
->stride
;
234 query_is_available(struct anv_query_pool
*pool
, uint32_t query
)
236 return *(volatile uint64_t *)query_slot(pool
, query
);
240 wait_for_available(struct anv_device
*device
,
241 struct anv_query_pool
*pool
, uint32_t query
)
243 uint64_t abs_timeout
= anv_get_absolute_timeout(5 * NSEC_PER_SEC
);
245 while (anv_gettime_ns() < abs_timeout
) {
246 if (query_is_available(pool
, query
))
248 VkResult status
= anv_device_query_status(device
);
249 if (status
!= VK_SUCCESS
)
253 return anv_device_set_lost(device
, "query timeout");
256 VkResult
genX(GetQueryPoolResults
)(
258 VkQueryPool queryPool
,
264 VkQueryResultFlags flags
)
266 ANV_FROM_HANDLE(anv_device
, device
, _device
);
267 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
269 assert(pool
->type
== VK_QUERY_TYPE_OCCLUSION
||
270 pool
->type
== VK_QUERY_TYPE_PIPELINE_STATISTICS
||
271 pool
->type
== VK_QUERY_TYPE_TIMESTAMP
||
272 pool
->type
== VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
||
273 pool
->type
== VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL
);
275 if (anv_device_is_lost(device
))
276 return VK_ERROR_DEVICE_LOST
;
281 void *data_end
= pData
+ dataSize
;
283 VkResult status
= VK_SUCCESS
;
284 for (uint32_t i
= 0; i
< queryCount
; i
++) {
285 bool available
= query_is_available(pool
, firstQuery
+ i
);
287 if (!available
&& (flags
& VK_QUERY_RESULT_WAIT_BIT
)) {
288 status
= wait_for_available(device
, pool
, firstQuery
+ i
);
289 if (status
!= VK_SUCCESS
)
295 /* From the Vulkan 1.0.42 spec:
297 * "If VK_QUERY_RESULT_WAIT_BIT and VK_QUERY_RESULT_PARTIAL_BIT are
298 * both not set then no result values are written to pData for
299 * queries that are in the unavailable state at the time of the call,
300 * and vkGetQueryPoolResults returns VK_NOT_READY. However,
301 * availability state is still written to pData for those queries if
302 * VK_QUERY_RESULT_WITH_AVAILABILITY_BIT is set."
304 bool write_results
= available
|| (flags
& VK_QUERY_RESULT_PARTIAL_BIT
);
307 switch (pool
->type
) {
308 case VK_QUERY_TYPE_OCCLUSION
: {
309 uint64_t *slot
= query_slot(pool
, firstQuery
+ i
);
311 cpu_write_query_result(pData
, flags
, idx
, slot
[2] - slot
[1]);
316 case VK_QUERY_TYPE_PIPELINE_STATISTICS
: {
317 uint64_t *slot
= query_slot(pool
, firstQuery
+ i
);
318 uint32_t statistics
= pool
->pipeline_statistics
;
320 uint32_t stat
= u_bit_scan(&statistics
);
322 uint64_t result
= slot
[idx
* 2 + 2] - slot
[idx
* 2 + 1];
324 /* WaDividePSInvocationCountBy4:HSW,BDW */
325 if ((device
->info
.gen
== 8 || device
->info
.is_haswell
) &&
326 (1 << stat
) == VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT
)
329 cpu_write_query_result(pData
, flags
, idx
, result
);
333 assert(idx
== util_bitcount(pool
->pipeline_statistics
));
337 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
: {
338 uint64_t *slot
= query_slot(pool
, firstQuery
+ i
);
340 cpu_write_query_result(pData
, flags
, idx
, slot
[2] - slot
[1]);
343 cpu_write_query_result(pData
, flags
, idx
, slot
[4] - slot
[3]);
348 case VK_QUERY_TYPE_TIMESTAMP
: {
349 uint64_t *slot
= query_slot(pool
, firstQuery
+ i
);
351 cpu_write_query_result(pData
, flags
, idx
, slot
[1]);
356 case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL
: {
359 const void *query_data
= query_slot(pool
, firstQuery
+ i
);
360 const uint32_t *oa_begin
= query_data
+ intel_perf_mi_rpc_offset(false);
361 const uint32_t *oa_end
= query_data
+ intel_perf_mi_rpc_offset(true);
362 const uint32_t *rpstat_begin
= query_data
+ intel_perf_rpstart_offset(false);
363 const uint32_t *rpstat_end
= query_data
+ intel_perf_mi_rpc_offset(true);
364 struct gen_perf_query_result result
;
365 struct gen_perf_query_info metric
= {
366 .oa_format
= (GEN_GEN
>= 8 ?
367 I915_OA_FORMAT_A32u40_A4u32_B8_C8
:
368 I915_OA_FORMAT_A45_B8_C8
),
370 uint32_t core_freq
[2];
372 core_freq
[0] = ((*rpstat_begin
>> 7) & 0x7f) * 1000000ULL;
373 core_freq
[1] = ((*rpstat_end
>> 7) & 0x7f) * 1000000ULL;
375 core_freq
[0] = ((*rpstat_begin
>> 23) & 0x1ff) * 1000000ULL;
376 core_freq
[1] = ((*rpstat_end
>> 23) & 0x1ff) * 1000000ULL;
378 gen_perf_query_result_clear(&result
);
379 gen_perf_query_result_accumulate(&result
, &metric
,
381 gen_perf_query_result_read_frequencies(&result
, &device
->info
,
383 gen_perf_query_result_write_mdapi(pData
, stride
,
386 core_freq
[0], core_freq
[1]);
387 gen_perf_query_mdapi_write_perfcntr(pData
, stride
, &device
->info
,
388 query_data
+ intel_perf_counter(false),
389 query_data
+ intel_perf_counter(true));
390 const uint64_t *marker
= query_data
+ intel_perf_marker_offset();
391 gen_perf_query_mdapi_write_marker(pData
, stride
, &device
->info
, *marker
);
396 unreachable("invalid pool type");
400 status
= VK_NOT_READY
;
402 if (flags
& VK_QUERY_RESULT_WITH_AVAILABILITY_BIT
)
403 cpu_write_query_result(pData
, flags
, idx
, available
);
406 if (pData
>= data_end
)
414 emit_ps_depth_count(struct anv_cmd_buffer
*cmd_buffer
,
415 struct anv_address addr
)
417 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
418 pc
.DestinationAddressType
= DAT_PPGTT
;
419 pc
.PostSyncOperation
= WritePSDepthCount
;
420 pc
.DepthStallEnable
= true;
423 if (GEN_GEN
== 9 && cmd_buffer
->device
->info
.gt
== 4)
424 pc
.CommandStreamerStallEnable
= true;
429 emit_query_mi_availability(struct gen_mi_builder
*b
,
430 struct anv_address addr
,
433 gen_mi_store(b
, gen_mi_mem64(addr
), gen_mi_imm(available
));
437 emit_query_pc_availability(struct anv_cmd_buffer
*cmd_buffer
,
438 struct anv_address addr
,
441 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
442 pc
.DestinationAddressType
= DAT_PPGTT
;
443 pc
.PostSyncOperation
= WriteImmediateData
;
445 pc
.ImmediateData
= available
;
450 * Goes through a series of consecutive query indices in the given pool
451 * setting all element values to 0 and emitting them as available.
454 emit_zero_queries(struct anv_cmd_buffer
*cmd_buffer
,
455 struct gen_mi_builder
*b
, struct anv_query_pool
*pool
,
456 uint32_t first_index
, uint32_t num_queries
)
458 switch (pool
->type
) {
459 case VK_QUERY_TYPE_OCCLUSION
:
460 case VK_QUERY_TYPE_TIMESTAMP
:
461 /* These queries are written with a PIPE_CONTROL so clear them using the
462 * PIPE_CONTROL as well so we don't have to synchronize between 2 types
465 assert((pool
->stride
% 8) == 0);
466 for (uint32_t i
= 0; i
< num_queries
; i
++) {
467 struct anv_address slot_addr
=
468 anv_query_address(pool
, first_index
+ i
);
470 for (uint32_t qword
= 1; qword
< (pool
->stride
/ 8); qword
++) {
471 emit_query_pc_availability(cmd_buffer
,
472 anv_address_add(slot_addr
, qword
* 8),
475 emit_query_pc_availability(cmd_buffer
, slot_addr
, true);
479 case VK_QUERY_TYPE_PIPELINE_STATISTICS
:
480 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
:
481 for (uint32_t i
= 0; i
< num_queries
; i
++) {
482 struct anv_address slot_addr
=
483 anv_query_address(pool
, first_index
+ i
);
484 gen_mi_memset(b
, anv_address_add(slot_addr
, 8), 0, pool
->stride
- 8);
485 emit_query_mi_availability(b
, slot_addr
, true);
489 case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL
:
490 for (uint32_t i
= 0; i
< num_queries
; i
++) {
491 struct anv_address slot_addr
=
492 anv_query_address(pool
, first_index
+ i
);
493 gen_mi_memset(b
, slot_addr
, 0, pool
->stride
- 8);
494 emit_query_mi_availability(b
, anv_address_add(slot_addr
,
495 pool
->stride
- 8), true);
500 unreachable("Unsupported query type");
504 void genX(CmdResetQueryPool
)(
505 VkCommandBuffer commandBuffer
,
506 VkQueryPool queryPool
,
510 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
511 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
513 switch (pool
->type
) {
514 case VK_QUERY_TYPE_OCCLUSION
:
515 case VK_QUERY_TYPE_TIMESTAMP
:
516 for (uint32_t i
= 0; i
< queryCount
; i
++) {
517 emit_query_pc_availability(cmd_buffer
,
518 anv_query_address(pool
, firstQuery
+ i
),
523 case VK_QUERY_TYPE_PIPELINE_STATISTICS
:
524 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
: {
525 struct gen_mi_builder b
;
526 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
528 for (uint32_t i
= 0; i
< queryCount
; i
++)
529 emit_query_mi_availability(&b
, anv_query_address(pool
, firstQuery
+ i
), false);
533 case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL
: {
534 struct gen_mi_builder b
;
535 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
537 for (uint32_t i
= 0; i
< queryCount
; i
++) {
538 emit_query_mi_availability(
541 anv_query_address(pool
, firstQuery
+ i
),
549 unreachable("Unsupported query type");
553 void genX(ResetQueryPoolEXT
)(
555 VkQueryPool queryPool
,
559 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
561 for (uint32_t i
= 0; i
< queryCount
; i
++) {
562 uint64_t *slot
= query_slot(pool
, firstQuery
+ i
);
567 static const uint32_t vk_pipeline_stat_to_reg
[] = {
568 GENX(IA_VERTICES_COUNT_num
),
569 GENX(IA_PRIMITIVES_COUNT_num
),
570 GENX(VS_INVOCATION_COUNT_num
),
571 GENX(GS_INVOCATION_COUNT_num
),
572 GENX(GS_PRIMITIVES_COUNT_num
),
573 GENX(CL_INVOCATION_COUNT_num
),
574 GENX(CL_PRIMITIVES_COUNT_num
),
575 GENX(PS_INVOCATION_COUNT_num
),
576 GENX(HS_INVOCATION_COUNT_num
),
577 GENX(DS_INVOCATION_COUNT_num
),
578 GENX(CS_INVOCATION_COUNT_num
),
582 emit_pipeline_stat(struct gen_mi_builder
*b
, uint32_t stat
,
583 struct anv_address addr
)
585 STATIC_ASSERT(ANV_PIPELINE_STATISTICS_MASK
==
586 (1 << ARRAY_SIZE(vk_pipeline_stat_to_reg
)) - 1);
588 assert(stat
< ARRAY_SIZE(vk_pipeline_stat_to_reg
));
589 gen_mi_store(b
, gen_mi_mem64(addr
),
590 gen_mi_reg64(vk_pipeline_stat_to_reg
[stat
]));
594 emit_xfb_query(struct gen_mi_builder
*b
, uint32_t stream
,
595 struct anv_address addr
)
597 assert(stream
< MAX_XFB_STREAMS
);
599 gen_mi_store(b
, gen_mi_mem64(anv_address_add(addr
, 0)),
600 gen_mi_reg64(GENX(SO_NUM_PRIMS_WRITTEN0_num
) + stream
* 8));
601 gen_mi_store(b
, gen_mi_mem64(anv_address_add(addr
, 16)),
602 gen_mi_reg64(GENX(SO_PRIM_STORAGE_NEEDED0_num
) + stream
* 8));
605 void genX(CmdBeginQuery
)(
606 VkCommandBuffer commandBuffer
,
607 VkQueryPool queryPool
,
609 VkQueryControlFlags flags
)
611 genX(CmdBeginQueryIndexedEXT
)(commandBuffer
, queryPool
, query
, flags
, 0);
614 void genX(CmdBeginQueryIndexedEXT
)(
615 VkCommandBuffer commandBuffer
,
616 VkQueryPool queryPool
,
618 VkQueryControlFlags flags
,
621 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
622 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
623 struct anv_address query_addr
= anv_query_address(pool
, query
);
625 struct gen_mi_builder b
;
626 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
628 switch (pool
->type
) {
629 case VK_QUERY_TYPE_OCCLUSION
:
630 emit_ps_depth_count(cmd_buffer
, anv_address_add(query_addr
, 8));
633 case VK_QUERY_TYPE_PIPELINE_STATISTICS
: {
634 /* TODO: This might only be necessary for certain stats */
635 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
636 pc
.CommandStreamerStallEnable
= true;
637 pc
.StallAtPixelScoreboard
= true;
640 uint32_t statistics
= pool
->pipeline_statistics
;
643 uint32_t stat
= u_bit_scan(&statistics
);
644 emit_pipeline_stat(&b
, stat
, anv_address_add(query_addr
, offset
));
650 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
:
651 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
652 pc
.CommandStreamerStallEnable
= true;
653 pc
.StallAtPixelScoreboard
= true;
655 emit_xfb_query(&b
, index
, anv_address_add(query_addr
, 8));
658 case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL
: {
659 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
660 pc
.CommandStreamerStallEnable
= true;
661 pc
.StallAtPixelScoreboard
= true;
663 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_REPORT_PERF_COUNT
), rpc
) {
665 anv_address_add(query_addr
, intel_perf_mi_rpc_offset(false));
669 gen_mi_mem32(anv_address_add(query_addr
,
670 intel_perf_rpstart_offset(false))),
671 gen_mi_reg32(GENX(RPSTAT1_num
)));
674 gen_mi_mem32(anv_address_add(query_addr
,
675 intel_perf_rpstart_offset(false))),
676 gen_mi_reg32(GENX(RPSTAT0_num
)));
678 #if GEN_GEN >= 8 && GEN_GEN <= 11
679 gen_mi_store(&b
, gen_mi_mem64(anv_address_add(query_addr
,
680 intel_perf_counter(false))),
681 gen_mi_reg64(GENX(PERFCNT1_num
)));
682 gen_mi_store(&b
, gen_mi_mem64(anv_address_add(query_addr
,
683 intel_perf_counter(false) + 8)),
684 gen_mi_reg64(GENX(PERFCNT2_num
)));
694 void genX(CmdEndQuery
)(
695 VkCommandBuffer commandBuffer
,
696 VkQueryPool queryPool
,
699 genX(CmdEndQueryIndexedEXT
)(commandBuffer
, queryPool
, query
, 0);
702 void genX(CmdEndQueryIndexedEXT
)(
703 VkCommandBuffer commandBuffer
,
704 VkQueryPool queryPool
,
708 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
709 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
710 struct anv_address query_addr
= anv_query_address(pool
, query
);
712 struct gen_mi_builder b
;
713 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
715 switch (pool
->type
) {
716 case VK_QUERY_TYPE_OCCLUSION
:
717 emit_ps_depth_count(cmd_buffer
, anv_address_add(query_addr
, 16));
718 emit_query_pc_availability(cmd_buffer
, query_addr
, true);
721 case VK_QUERY_TYPE_PIPELINE_STATISTICS
: {
722 /* TODO: This might only be necessary for certain stats */
723 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
724 pc
.CommandStreamerStallEnable
= true;
725 pc
.StallAtPixelScoreboard
= true;
728 uint32_t statistics
= pool
->pipeline_statistics
;
729 uint32_t offset
= 16;
731 uint32_t stat
= u_bit_scan(&statistics
);
732 emit_pipeline_stat(&b
, stat
, anv_address_add(query_addr
, offset
));
736 emit_query_mi_availability(&b
, query_addr
, true);
740 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
:
741 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
742 pc
.CommandStreamerStallEnable
= true;
743 pc
.StallAtPixelScoreboard
= true;
746 emit_xfb_query(&b
, index
, anv_address_add(query_addr
, 16));
747 emit_query_mi_availability(&b
, query_addr
, true);
750 case VK_QUERY_TYPE_PERFORMANCE_QUERY_INTEL
: {
751 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
752 pc
.CommandStreamerStallEnable
= true;
753 pc
.StallAtPixelScoreboard
= true;
755 uint32_t marker_offset
= intel_perf_marker_offset();
756 gen_mi_store(&b
, gen_mi_mem64(anv_address_add(query_addr
, marker_offset
)),
757 gen_mi_imm(cmd_buffer
->intel_perf_marker
));
758 #if GEN_GEN >= 8 && GEN_GEN <= 11
759 gen_mi_store(&b
, gen_mi_mem64(anv_address_add(query_addr
, intel_perf_counter(true))),
760 gen_mi_reg64(GENX(PERFCNT1_num
)));
761 gen_mi_store(&b
, gen_mi_mem64(anv_address_add(query_addr
, intel_perf_counter(true) + 8)),
762 gen_mi_reg64(GENX(PERFCNT2_num
)));
766 gen_mi_mem32(anv_address_add(query_addr
,
767 intel_perf_rpstart_offset(true))),
768 gen_mi_reg32(GENX(RPSTAT1_num
)));
771 gen_mi_mem32(anv_address_add(query_addr
,
772 intel_perf_rpstart_offset(true))),
773 gen_mi_reg32(GENX(RPSTAT0_num
)));
775 /* Position the last OA snapshot at the beginning of the query so that
776 * we can tell whether it's ready.
778 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_REPORT_PERF_COUNT
), rpc
) {
779 rpc
.MemoryAddress
= anv_address_add(query_addr
,
780 intel_perf_mi_rpc_offset(true));
781 rpc
.ReportID
= 0xdeadbeef; /* This goes in the first dword */
783 emit_query_mi_availability(&b
,
784 anv_address_add(query_addr
, pool
->stride
- 8),
793 /* When multiview is active the spec requires that N consecutive query
794 * indices are used, where N is the number of active views in the subpass.
795 * The spec allows that we only write the results to one of the queries
796 * but we still need to manage result availability for all the query indices.
797 * Since we only emit a single query for all active views in the
798 * first index, mark the other query indices as being already available
801 if (cmd_buffer
->state
.subpass
&& cmd_buffer
->state
.subpass
->view_mask
) {
802 const uint32_t num_queries
=
803 util_bitcount(cmd_buffer
->state
.subpass
->view_mask
);
805 emit_zero_queries(cmd_buffer
, &b
, pool
, query
+ 1, num_queries
- 1);
809 #define TIMESTAMP 0x2358
811 void genX(CmdWriteTimestamp
)(
812 VkCommandBuffer commandBuffer
,
813 VkPipelineStageFlagBits pipelineStage
,
814 VkQueryPool queryPool
,
817 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
818 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
819 struct anv_address query_addr
= anv_query_address(pool
, query
);
821 assert(pool
->type
== VK_QUERY_TYPE_TIMESTAMP
);
823 struct gen_mi_builder b
;
824 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
826 switch (pipelineStage
) {
827 case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT
:
828 gen_mi_store(&b
, gen_mi_mem64(anv_address_add(query_addr
, 8)),
829 gen_mi_reg64(TIMESTAMP
));
833 /* Everything else is bottom-of-pipe */
834 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
835 pc
.DestinationAddressType
= DAT_PPGTT
;
836 pc
.PostSyncOperation
= WriteTimestamp
;
837 pc
.Address
= anv_address_add(query_addr
, 8);
839 if (GEN_GEN
== 9 && cmd_buffer
->device
->info
.gt
== 4)
840 pc
.CommandStreamerStallEnable
= true;
845 emit_query_pc_availability(cmd_buffer
, query_addr
, true);
847 /* When multiview is active the spec requires that N consecutive query
848 * indices are used, where N is the number of active views in the subpass.
849 * The spec allows that we only write the results to one of the queries
850 * but we still need to manage result availability for all the query indices.
851 * Since we only emit a single query for all active views in the
852 * first index, mark the other query indices as being already available
855 if (cmd_buffer
->state
.subpass
&& cmd_buffer
->state
.subpass
->view_mask
) {
856 const uint32_t num_queries
=
857 util_bitcount(cmd_buffer
->state
.subpass
->view_mask
);
859 emit_zero_queries(cmd_buffer
, &b
, pool
, query
+ 1, num_queries
- 1);
863 #if GEN_GEN > 7 || GEN_IS_HASWELL
866 gpu_write_query_result(struct gen_mi_builder
*b
,
867 struct anv_address dst_addr
,
868 VkQueryResultFlags flags
,
869 uint32_t value_index
,
870 struct gen_mi_value query_result
)
872 if (flags
& VK_QUERY_RESULT_64_BIT
) {
873 struct anv_address res_addr
= anv_address_add(dst_addr
, value_index
* 8);
874 gen_mi_store(b
, gen_mi_mem64(res_addr
), query_result
);
876 struct anv_address res_addr
= anv_address_add(dst_addr
, value_index
* 4);
877 gen_mi_store(b
, gen_mi_mem32(res_addr
), query_result
);
881 static struct gen_mi_value
882 compute_query_result(struct gen_mi_builder
*b
, struct anv_address addr
)
884 return gen_mi_isub(b
, gen_mi_mem64(anv_address_add(addr
, 8)),
885 gen_mi_mem64(anv_address_add(addr
, 0)));
888 void genX(CmdCopyQueryPoolResults
)(
889 VkCommandBuffer commandBuffer
,
890 VkQueryPool queryPool
,
894 VkDeviceSize destOffset
,
895 VkDeviceSize destStride
,
896 VkQueryResultFlags flags
)
898 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
899 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
900 ANV_FROM_HANDLE(anv_buffer
, buffer
, destBuffer
);
902 struct gen_mi_builder b
;
903 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
904 struct gen_mi_value result
;
906 /* If render target writes are ongoing, request a render target cache flush
907 * to ensure proper ordering of the commands from the 3d pipe and the
910 if (cmd_buffer
->state
.pending_pipe_bits
& ANV_PIPE_RENDER_TARGET_BUFFER_WRITES
) {
911 cmd_buffer
->state
.pending_pipe_bits
|=
912 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
;
915 if ((flags
& VK_QUERY_RESULT_WAIT_BIT
) ||
916 (cmd_buffer
->state
.pending_pipe_bits
& ANV_PIPE_FLUSH_BITS
) ||
917 /* Occlusion & timestamp queries are written using a PIPE_CONTROL and
918 * because we're about to copy values from MI commands, we need to
919 * stall the command streamer to make sure the PIPE_CONTROL values have
920 * landed, otherwise we could see inconsistent values & availability.
922 * From the vulkan spec:
924 * "vkCmdCopyQueryPoolResults is guaranteed to see the effect of
925 * previous uses of vkCmdResetQueryPool in the same queue, without
926 * any additional synchronization."
928 pool
->type
== VK_QUERY_TYPE_OCCLUSION
||
929 pool
->type
== VK_QUERY_TYPE_TIMESTAMP
) {
930 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
931 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
934 struct anv_address dest_addr
= anv_address_add(buffer
->address
, destOffset
);
935 for (uint32_t i
= 0; i
< queryCount
; i
++) {
936 struct anv_address query_addr
= anv_query_address(pool
, firstQuery
+ i
);
938 switch (pool
->type
) {
939 case VK_QUERY_TYPE_OCCLUSION
:
940 result
= compute_query_result(&b
, anv_address_add(query_addr
, 8));
941 gpu_write_query_result(&b
, dest_addr
, flags
, idx
++, result
);
944 case VK_QUERY_TYPE_PIPELINE_STATISTICS
: {
945 uint32_t statistics
= pool
->pipeline_statistics
;
947 uint32_t stat
= u_bit_scan(&statistics
);
949 result
= compute_query_result(&b
, anv_address_add(query_addr
,
952 /* WaDividePSInvocationCountBy4:HSW,BDW */
953 if ((cmd_buffer
->device
->info
.gen
== 8 ||
954 cmd_buffer
->device
->info
.is_haswell
) &&
955 (1 << stat
) == VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT
) {
956 result
= gen_mi_ushr32_imm(&b
, result
, 2);
959 gpu_write_query_result(&b
, dest_addr
, flags
, idx
++, result
);
961 assert(idx
== util_bitcount(pool
->pipeline_statistics
));
965 case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT
:
966 result
= compute_query_result(&b
, anv_address_add(query_addr
, 8));
967 gpu_write_query_result(&b
, dest_addr
, flags
, idx
++, result
);
968 result
= compute_query_result(&b
, anv_address_add(query_addr
, 24));
969 gpu_write_query_result(&b
, dest_addr
, flags
, idx
++, result
);
972 case VK_QUERY_TYPE_TIMESTAMP
:
973 result
= gen_mi_mem64(anv_address_add(query_addr
, 8));
974 gpu_write_query_result(&b
, dest_addr
, flags
, 0, result
);
978 unreachable("unhandled query type");
981 if (flags
& VK_QUERY_RESULT_WITH_AVAILABILITY_BIT
) {
982 gpu_write_query_result(&b
, dest_addr
, flags
, idx
,
983 gen_mi_mem64(query_addr
));
986 dest_addr
= anv_address_add(dest_addr
, destStride
);
991 void genX(CmdCopyQueryPoolResults
)(
992 VkCommandBuffer commandBuffer
,
993 VkQueryPool queryPool
,
997 VkDeviceSize destOffset
,
998 VkDeviceSize destStride
,
999 VkQueryResultFlags flags
)
1001 anv_finishme("Queries not yet supported on Ivy Bridge");