src/intel/vulkan/genX_query.c

   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 VkResult genX(CreateQueryPool)(
  36     VkDevice                                    _device,
  37     const VkQueryPoolCreateInfo*                pCreateInfo,
  38     const VkAllocationCallbacks*                pAllocator,
  39     VkQueryPool*                                pQueryPool)
  40 {
  41    ANV_FROM_HANDLE(anv_device, device, _device);
  42    struct anv_query_pool *pool;
  43    VkResult result;
  44    uint32_t slot_size;
  45    uint64_t size;
  46
  47    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO);
  48
  49    switch (pCreateInfo->queryType) {
  50    case VK_QUERY_TYPE_OCCLUSION:
  51    case VK_QUERY_TYPE_TIMESTAMP:
  52       break;
  53    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
  54       return VK_ERROR_INCOMPATIBLE_DRIVER;
  55    default:
  56       assert(!"Invalid query type");
  57    }
  58
  59    slot_size = sizeof(struct anv_query_pool_slot);
  60    pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
  61                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
  62    if (pool == NULL)
  63       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
  64
  65    pool->type = pCreateInfo->queryType;
  66    pool->slots = pCreateInfo->queryCount;
  67
  68    size = pCreateInfo->queryCount * slot_size;
  69    result = anv_bo_init_new(&pool->bo, device, size);
  70    if (result != VK_SUCCESS)
  71       goto fail;
  72
  73    pool->bo.map = anv_gem_mmap(device, pool->bo.gem_handle, 0, size, 0);
  74
  75    *pQueryPool = anv_query_pool_to_handle(pool);
  76
  77    return VK_SUCCESS;
  78
  79  fail:
  80    vk_free2(&device->alloc, pAllocator, pool);
  81
  82    return result;
  83 }
  84
  85 void genX(DestroyQueryPool)(
  86     VkDevice                                    _device,
  87     VkQueryPool                                 _pool,
  88     const VkAllocationCallbacks*                pAllocator)
  89 {
  90    ANV_FROM_HANDLE(anv_device, device, _device);
  91    ANV_FROM_HANDLE(anv_query_pool, pool, _pool);
  92
  93    if (!pool)
  94       return;
  95
  96    anv_gem_munmap(pool->bo.map, pool->bo.size);
  97    anv_gem_close(device, pool->bo.gem_handle);
  98    vk_free2(&device->alloc, pAllocator, pool);
  99 }
 100
 101 VkResult genX(GetQueryPoolResults)(
 102     VkDevice                                    _device,
 103     VkQueryPool                                 queryPool,
 104     uint32_t                                    firstQuery,
 105     uint32_t                                    queryCount,
 106     size_t                                      dataSize,
 107     void*                                       pData,
 108     VkDeviceSize                                stride,
 109     VkQueryResultFlags                          flags)
 110 {
 111    ANV_FROM_HANDLE(anv_device, device, _device);
 112    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 113    int64_t timeout = INT64_MAX;
 114    uint64_t result;
 115    int ret;
 116
 117    assert(pool->type == VK_QUERY_TYPE_OCCLUSION ||
 118           pool->type == VK_QUERY_TYPE_TIMESTAMP);
 119
 120    if (pData == NULL)
 121       return VK_SUCCESS;
 122
 123    if (flags & VK_QUERY_RESULT_WAIT_BIT) {
 124       ret = anv_gem_wait(device, pool->bo.gem_handle, &timeout);
 125       if (ret == -1) {
 126          /* We don't know the real error. */
 127          return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
 128                           "gem_wait failed %m");
 129       }
 130    }
 131
 132    void *data_end = pData + dataSize;
 133    struct anv_query_pool_slot *slot = pool->bo.map;
 134
 135    if (!device->info.has_llc) {
 136       uint64_t offset = firstQuery * sizeof(*slot);
 137       uint64_t size = queryCount * sizeof(*slot);
 138       anv_invalidate_range(pool->bo.map + offset,
 139                            MIN2(size, pool->bo.size - offset));
 140    }
 141
 142    VkResult status = VK_SUCCESS;
 143    for (uint32_t i = 0; i < queryCount; i++) {
 144       bool available = slot[firstQuery + i].available;
 145
 146       /* From the Vulkan 1.0.42 spec:
 147        *
 148        *    "If VK_QUERY_RESULT_WAIT_BIT and VK_QUERY_RESULT_PARTIAL_BIT are
 149        *    both not set then no result values are written to pData for
 150        *    queries that are in the unavailable state at the time of the call,
 151        *    and vkGetQueryPoolResults returns VK_NOT_READY. However,
 152        *    availability state is still written to pData for those queries if
 153        *    VK_QUERY_RESULT_WITH_AVAILABILITY_BIT is set."
 154        */
 155       bool write_results = available || (flags & VK_QUERY_RESULT_PARTIAL_BIT);
 156
 157       if (write_results) {
 158          switch (pool->type) {
 159          case VK_QUERY_TYPE_OCCLUSION: {
 160             result = slot[firstQuery + i].end - slot[firstQuery + i].begin;
 161             break;
 162          }
 163          case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 164             unreachable("pipeline stats not supported");
 165          case VK_QUERY_TYPE_TIMESTAMP: {
 166             result = slot[firstQuery + i].begin;
 167             break;
 168          }
 169          default:
 170             unreachable("invalid pool type");
 171          }
 172       } else {
 173          status = VK_NOT_READY;
 174       }
 175
 176       if (flags & VK_QUERY_RESULT_64_BIT) {
 177          uint64_t *dst = pData;
 178          if (write_results)
 179             dst[0] = result;
 180          if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
 181             dst[1] = slot[firstQuery + i].available;
 182       } else {
 183          uint32_t *dst = pData;
 184          if (result > UINT32_MAX)
 185             result = UINT32_MAX;
 186          if (write_results)
 187             dst[0] = result;
 188          if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
 189             dst[1] = slot[firstQuery + i].available;
 190       }
 191
 192       pData += stride;
 193       if (pData >= data_end)
 194          break;
 195    }
 196
 197    return status;
 198 }
 199
 200 static void
 201 emit_ps_depth_count(struct anv_cmd_buffer *cmd_buffer,
 202                     struct anv_bo *bo, uint32_t offset)
 203 {
 204    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 205       pc.DestinationAddressType  = DAT_PPGTT;
 206       pc.PostSyncOperation       = WritePSDepthCount;
 207       pc.DepthStallEnable        = true;
 208       pc.Address                 = (struct anv_address) { bo, offset };
 209
 210       if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 211          pc.CommandStreamerStallEnable = true;
 212    }
 213 }
 214
 215 static void
 216 emit_query_availability(struct anv_cmd_buffer *cmd_buffer,
 217                         struct anv_bo *bo, uint32_t offset)
 218 {
 219    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 220       pc.DestinationAddressType  = DAT_PPGTT;
 221       pc.PostSyncOperation       = WriteImmediateData;
 222       pc.Address                 = (struct anv_address) { bo, offset };
 223       pc.ImmediateData           = 1;
 224    }
 225 }
 226
 227 void genX(CmdResetQueryPool)(
 228     VkCommandBuffer                             commandBuffer,
 229     VkQueryPool                                 queryPool,
 230     uint32_t                                    firstQuery,
 231     uint32_t                                    queryCount)
 232 {
 233    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 234    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 235
 236    for (uint32_t i = 0; i < queryCount; i++) {
 237       switch (pool->type) {
 238       case VK_QUERY_TYPE_OCCLUSION:
 239       case VK_QUERY_TYPE_TIMESTAMP: {
 240          anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdm) {
 241             sdm.Address = (struct anv_address) {
 242                .bo = &pool->bo,
 243                .offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot) +
 244                          offsetof(struct anv_query_pool_slot, available),
 245             };
 246             sdm.DataDWord0 = 0;
 247             sdm.DataDWord1 = 0;
 248          }
 249          break;
 250       }
 251       default:
 252          assert(!"Invalid query type");
 253       }
 254    }
 255 }
 256
 257 void genX(CmdBeginQuery)(
 258     VkCommandBuffer                             commandBuffer,
 259     VkQueryPool                                 queryPool,
 260     uint32_t                                    query,
 261     VkQueryControlFlags                         flags)
 262 {
 263    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 264    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 265
 266    /* Workaround: When meta uses the pipeline with the VS disabled, it seems
 267     * that the pipelining of the depth write breaks. What we see is that
 268     * samples from the render pass clear leaks into the first query
 269     * immediately after the clear. Doing a pipecontrol with a post-sync
 270     * operation and DepthStallEnable seems to work around the issue.
 271     */
 272    if (cmd_buffer->state.need_query_wa) {
 273       cmd_buffer->state.need_query_wa = false;
 274       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 275          pc.DepthCacheFlushEnable   = true;
 276          pc.DepthStallEnable        = true;
 277       }
 278    }
 279
 280    switch (pool->type) {
 281    case VK_QUERY_TYPE_OCCLUSION:
 282       emit_ps_depth_count(cmd_buffer, &pool->bo,
 283                           query * sizeof(struct anv_query_pool_slot));
 284       break;
 285
 286    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 287    default:
 288       unreachable("");
 289    }
 290 }
 291
 292 void genX(CmdEndQuery)(
 293     VkCommandBuffer                             commandBuffer,
 294     VkQueryPool                                 queryPool,
 295     uint32_t                                    query)
 296 {
 297    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 298    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 299
 300    switch (pool->type) {
 301    case VK_QUERY_TYPE_OCCLUSION:
 302       emit_ps_depth_count(cmd_buffer, &pool->bo,
 303                           query * sizeof(struct anv_query_pool_slot) + 8);
 304
 305       emit_query_availability(cmd_buffer, &pool->bo,
 306                               query * sizeof(struct anv_query_pool_slot) + 16);
 307       break;
 308
 309    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 310    default:
 311       unreachable("");
 312    }
 313 }
 314
 315 #define TIMESTAMP 0x2358
 316
 317 void genX(CmdWriteTimestamp)(
 318     VkCommandBuffer                             commandBuffer,
 319     VkPipelineStageFlagBits                     pipelineStage,
 320     VkQueryPool                                 queryPool,
 321     uint32_t                                    query)
 322 {
 323    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 324    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 325    uint32_t offset = query * sizeof(struct anv_query_pool_slot);
 326
 327    assert(pool->type == VK_QUERY_TYPE_TIMESTAMP);
 328
 329    switch (pipelineStage) {
 330    case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT:
 331       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 332          srm.RegisterAddress  = TIMESTAMP;
 333          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset };
 334       }
 335       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 336          srm.RegisterAddress  = TIMESTAMP + 4;
 337          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset + 4 };
 338       }
 339       break;
 340
 341    default:
 342       /* Everything else is bottom-of-pipe */
 343       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 344          pc.DestinationAddressType  = DAT_PPGTT;
 345          pc.PostSyncOperation       = WriteTimestamp;
 346          pc.Address = (struct anv_address) { &pool->bo, offset };
 347
 348          if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 349             pc.CommandStreamerStallEnable = true;
 350       }
 351       break;
 352    }
 353
 354    emit_query_availability(cmd_buffer, &pool->bo, offset + 16);
 355 }
 356
 357 #if GEN_GEN > 7 || GEN_IS_HASWELL
 358
 359 #define alu_opcode(v)   __gen_uint((v),  20, 31)
 360 #define alu_operand1(v) __gen_uint((v),  10, 19)
 361 #define alu_operand2(v) __gen_uint((v),   0,  9)
 362 #define alu(opcode, operand1, operand2) \
 363    alu_opcode(opcode) | alu_operand1(operand1) | alu_operand2(operand2)
 364
 365 #define OPCODE_NOOP      0x000
 366 #define OPCODE_LOAD      0x080
 367 #define OPCODE_LOADINV   0x480
 368 #define OPCODE_LOAD0     0x081
 369 #define OPCODE_LOAD1     0x481
 370 #define OPCODE_ADD       0x100
 371 #define OPCODE_SUB       0x101
 372 #define OPCODE_AND       0x102
 373 #define OPCODE_OR        0x103
 374 #define OPCODE_XOR       0x104
 375 #define OPCODE_STORE     0x180
 376 #define OPCODE_STOREINV  0x580
 377
 378 #define OPERAND_R0   0x00
 379 #define OPERAND_R1   0x01
 380 #define OPERAND_R2   0x02
 381 #define OPERAND_R3   0x03
 382 #define OPERAND_R4   0x04
 383 #define OPERAND_SRCA 0x20
 384 #define OPERAND_SRCB 0x21
 385 #define OPERAND_ACCU 0x31
 386 #define OPERAND_ZF   0x32
 387 #define OPERAND_CF   0x33
 388
 389 #define CS_GPR(n) (0x2600 + (n) * 8)
 390
 391 static void
 392 emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t reg,
 393                       struct anv_bo *bo, uint32_t offset)
 394 {
 395    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 396       lrm.RegisterAddress  = reg,
 397       lrm.MemoryAddress    = (struct anv_address) { bo, offset };
 398    }
 399    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 400       lrm.RegisterAddress  = reg + 4;
 401       lrm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 402    }
 403 }
 404
 405 static void
 406 store_query_result(struct anv_batch *batch, uint32_t reg,
 407                    struct anv_bo *bo, uint32_t offset, VkQueryResultFlags flags)
 408 {
 409    anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 410       srm.RegisterAddress  = reg;
 411       srm.MemoryAddress    = (struct anv_address) { bo, offset };
 412    }
 413
 414    if (flags & VK_QUERY_RESULT_64_BIT) {
 415       anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 416          srm.RegisterAddress  = reg + 4;
 417          srm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 418       }
 419    }
 420 }
 421
 422 void genX(CmdCopyQueryPoolResults)(
 423     VkCommandBuffer                             commandBuffer,
 424     VkQueryPool                                 queryPool,
 425     uint32_t                                    firstQuery,
 426     uint32_t                                    queryCount,
 427     VkBuffer                                    destBuffer,
 428     VkDeviceSize                                destOffset,
 429     VkDeviceSize                                destStride,
 430     VkQueryResultFlags                          flags)
 431 {
 432    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 433    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 434    ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer);
 435    uint32_t slot_offset, dst_offset;
 436
 437    if (flags & VK_QUERY_RESULT_WAIT_BIT) {
 438       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 439          pc.CommandStreamerStallEnable = true;
 440          pc.StallAtPixelScoreboard     = true;
 441       }
 442    }
 443
 444    dst_offset = buffer->offset + destOffset;
 445    for (uint32_t i = 0; i < queryCount; i++) {
 446
 447       slot_offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot);
 448       switch (pool->type) {
 449       case VK_QUERY_TYPE_OCCLUSION:
 450          emit_load_alu_reg_u64(&cmd_buffer->batch,
 451                                CS_GPR(0), &pool->bo, slot_offset);
 452          emit_load_alu_reg_u64(&cmd_buffer->batch,
 453                                CS_GPR(1), &pool->bo, slot_offset + 8);
 454
 455          /* FIXME: We need to clamp the result for 32 bit. */
 456
 457          uint32_t *dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
 458          dw[1] = alu(OPCODE_LOAD, OPERAND_SRCA, OPERAND_R1);
 459          dw[2] = alu(OPCODE_LOAD, OPERAND_SRCB, OPERAND_R0);
 460          dw[3] = alu(OPCODE_SUB, 0, 0);
 461          dw[4] = alu(OPCODE_STORE, OPERAND_R2, OPERAND_ACCU);
 462          break;
 463
 464       case VK_QUERY_TYPE_TIMESTAMP:
 465          emit_load_alu_reg_u64(&cmd_buffer->batch,
 466                                CS_GPR(2), &pool->bo, slot_offset);
 467          break;
 468
 469       default:
 470          unreachable("unhandled query type");
 471       }
 472
 473       store_query_result(&cmd_buffer->batch,
 474                          CS_GPR(2), buffer->bo, dst_offset, flags);
 475
 476       if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
 477          emit_load_alu_reg_u64(&cmd_buffer->batch, CS_GPR(0),
 478                                &pool->bo, slot_offset + 16);
 479          if (flags & VK_QUERY_RESULT_64_BIT)
 480             store_query_result(&cmd_buffer->batch,
 481                                CS_GPR(0), buffer->bo, dst_offset + 8, flags);
 482          else
 483             store_query_result(&cmd_buffer->batch,
 484                                CS_GPR(0), buffer->bo, dst_offset + 4, flags);
 485       }
 486
 487       dst_offset += destStride;
 488    }
 489 }
 490
 491 #else
 492 void genX(CmdCopyQueryPoolResults)(
 493     VkCommandBuffer                             commandBuffer,
 494     VkQueryPool                                 queryPool,
 495     uint32_t                                    firstQuery,
 496     uint32_t                                    queryCount,
 497     VkBuffer                                    destBuffer,
 498     VkDeviceSize                                destOffset,
 499     VkDeviceSize                                destStride,
 500     VkQueryResultFlags                          flags)
 501 {
 502    anv_finishme("Queries not yet supported on Ivy Bridge");
 503 }
 504 #endif