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
  45    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO);
  46
  47    /* Query pool slots are made up of some number of 64-bit values packed
  48     * tightly together.  The first 64-bit value is always the "available" bit
  49     * which is 0 when the query is unavailable and 1 when it is available.
  50     * The 64-bit values that follow are determined by the type of query.
  51     */
  52    uint32_t uint64s_per_slot = 1;
  53
  54    switch (pCreateInfo->queryType) {
  55    case VK_QUERY_TYPE_OCCLUSION:
  56       /* Occlusion queries have two values: begin and end. */
  57       uint64s_per_slot += 2;
  58       break;
  59    case VK_QUERY_TYPE_TIMESTAMP:
  60       /* Timestamps just have the one timestamp value */
  61       uint64s_per_slot += 1;
  62       break;
  63    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
  64       return VK_ERROR_INCOMPATIBLE_DRIVER;
  65    default:
  66       assert(!"Invalid query type");
  67    }
  68
  69    pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
  70                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
  71    if (pool == NULL)
  72       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
  73
  74    pool->type = pCreateInfo->queryType;
  75    pool->stride = uint64s_per_slot * sizeof(uint64_t);
  76    pool->slots = pCreateInfo->queryCount;
  77
  78    uint64_t size = pool->slots * pool->stride;
  79    result = anv_bo_init_new(&pool->bo, device, size);
  80    if (result != VK_SUCCESS)
  81       goto fail;
  82
  83    pool->bo.map = anv_gem_mmap(device, pool->bo.gem_handle, 0, size, 0);
  84
  85    *pQueryPool = anv_query_pool_to_handle(pool);
  86
  87    return VK_SUCCESS;
  88
  89  fail:
  90    vk_free2(&device->alloc, pAllocator, pool);
  91
  92    return result;
  93 }
  94
  95 void genX(DestroyQueryPool)(
  96     VkDevice                                    _device,
  97     VkQueryPool                                 _pool,
  98     const VkAllocationCallbacks*                pAllocator)
  99 {
 100    ANV_FROM_HANDLE(anv_device, device, _device);
 101    ANV_FROM_HANDLE(anv_query_pool, pool, _pool);
 102
 103    if (!pool)
 104       return;
 105
 106    anv_gem_munmap(pool->bo.map, pool->bo.size);
 107    anv_gem_close(device, pool->bo.gem_handle);
 108    vk_free2(&device->alloc, pAllocator, pool);
 109 }
 110
 111 VkResult genX(GetQueryPoolResults)(
 112     VkDevice                                    _device,
 113     VkQueryPool                                 queryPool,
 114     uint32_t                                    firstQuery,
 115     uint32_t                                    queryCount,
 116     size_t                                      dataSize,
 117     void*                                       pData,
 118     VkDeviceSize                                stride,
 119     VkQueryResultFlags                          flags)
 120 {
 121    ANV_FROM_HANDLE(anv_device, device, _device);
 122    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 123    int64_t timeout = INT64_MAX;
 124    uint64_t result;
 125    int ret;
 126
 127    assert(pool->type == VK_QUERY_TYPE_OCCLUSION ||
 128           pool->type == VK_QUERY_TYPE_TIMESTAMP);
 129
 130    if (pData == NULL)
 131       return VK_SUCCESS;
 132
 133    if (flags & VK_QUERY_RESULT_WAIT_BIT) {
 134       ret = anv_gem_wait(device, pool->bo.gem_handle, &timeout);
 135       if (ret == -1) {
 136          /* We don't know the real error. */
 137          return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
 138                           "gem_wait failed %m");
 139       }
 140    }
 141
 142    void *data_end = pData + dataSize;
 143
 144    if (!device->info.has_llc) {
 145       uint64_t offset = firstQuery * pool->stride;
 146       uint64_t size = queryCount * pool->stride;
 147       anv_invalidate_range(pool->bo.map + offset,
 148                            MIN2(size, pool->bo.size - offset));
 149    }
 150
 151    VkResult status = VK_SUCCESS;
 152    for (uint32_t i = 0; i < queryCount; i++) {
 153       uint64_t *slot = pool->bo.map + (firstQuery + i) * pool->stride;
 154
 155       /* Availability is always at the start of the slot */
 156       bool available = slot[0];
 157
 158       /* From the Vulkan 1.0.42 spec:
 159        *
 160        *    "If VK_QUERY_RESULT_WAIT_BIT and VK_QUERY_RESULT_PARTIAL_BIT are
 161        *    both not set then no result values are written to pData for
 162        *    queries that are in the unavailable state at the time of the call,
 163        *    and vkGetQueryPoolResults returns VK_NOT_READY. However,
 164        *    availability state is still written to pData for those queries if
 165        *    VK_QUERY_RESULT_WITH_AVAILABILITY_BIT is set."
 166        */
 167       bool write_results = available || (flags & VK_QUERY_RESULT_PARTIAL_BIT);
 168
 169       if (write_results) {
 170          switch (pool->type) {
 171          case VK_QUERY_TYPE_OCCLUSION: {
 172             result = slot[2] - slot[1];
 173             break;
 174          }
 175          case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 176             unreachable("pipeline stats not supported");
 177          case VK_QUERY_TYPE_TIMESTAMP: {
 178             result = slot[1];
 179             break;
 180          }
 181          default:
 182             unreachable("invalid pool type");
 183          }
 184       } else {
 185          status = VK_NOT_READY;
 186       }
 187
 188       if (flags & VK_QUERY_RESULT_64_BIT) {
 189          uint64_t *dst = pData;
 190          if (write_results)
 191             dst[0] = result;
 192          if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
 193             dst[1] = available;
 194       } else {
 195          uint32_t *dst = pData;
 196          if (write_results)
 197             dst[0] = result;
 198          if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
 199             dst[1] = available;
 200       }
 201
 202       pData += stride;
 203       if (pData >= data_end)
 204          break;
 205    }
 206
 207    return status;
 208 }
 209
 210 static void
 211 emit_ps_depth_count(struct anv_cmd_buffer *cmd_buffer,
 212                     struct anv_bo *bo, uint32_t offset)
 213 {
 214    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 215       pc.DestinationAddressType  = DAT_PPGTT;
 216       pc.PostSyncOperation       = WritePSDepthCount;
 217       pc.DepthStallEnable        = true;
 218       pc.Address                 = (struct anv_address) { bo, offset };
 219
 220       if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 221          pc.CommandStreamerStallEnable = true;
 222    }
 223 }
 224
 225 static void
 226 emit_query_availability(struct anv_cmd_buffer *cmd_buffer,
 227                         struct anv_bo *bo, uint32_t offset)
 228 {
 229    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 230       pc.DestinationAddressType  = DAT_PPGTT;
 231       pc.PostSyncOperation       = WriteImmediateData;
 232       pc.Address                 = (struct anv_address) { bo, offset };
 233       pc.ImmediateData           = 1;
 234    }
 235 }
 236
 237 void genX(CmdResetQueryPool)(
 238     VkCommandBuffer                             commandBuffer,
 239     VkQueryPool                                 queryPool,
 240     uint32_t                                    firstQuery,
 241     uint32_t                                    queryCount)
 242 {
 243    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 244    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 245
 246    for (uint32_t i = 0; i < queryCount; i++) {
 247       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdm) {
 248          sdm.Address = (struct anv_address) {
 249             .bo = &pool->bo,
 250             .offset = (firstQuery + i) * pool->stride,
 251          };
 252          sdm.DataDWord0 = 0;
 253          sdm.DataDWord1 = 0;
 254       }
 255    }
 256 }
 257
 258 void genX(CmdBeginQuery)(
 259     VkCommandBuffer                             commandBuffer,
 260     VkQueryPool                                 queryPool,
 261     uint32_t                                    query,
 262     VkQueryControlFlags                         flags)
 263 {
 264    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 265    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 266
 267    /* Workaround: When meta uses the pipeline with the VS disabled, it seems
 268     * that the pipelining of the depth write breaks. What we see is that
 269     * samples from the render pass clear leaks into the first query
 270     * immediately after the clear. Doing a pipecontrol with a post-sync
 271     * operation and DepthStallEnable seems to work around the issue.
 272     */
 273    if (cmd_buffer->state.need_query_wa) {
 274       cmd_buffer->state.need_query_wa = false;
 275       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 276          pc.DepthCacheFlushEnable   = true;
 277          pc.DepthStallEnable        = true;
 278       }
 279    }
 280
 281    switch (pool->type) {
 282    case VK_QUERY_TYPE_OCCLUSION:
 283       emit_ps_depth_count(cmd_buffer, &pool->bo, query * pool->stride + 8);
 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, query * pool->stride + 16);
 303       emit_query_availability(cmd_buffer, &pool->bo, query * pool->stride);
 304       break;
 305
 306    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 307    default:
 308       unreachable("");
 309    }
 310 }
 311
 312 #define TIMESTAMP 0x2358
 313
 314 void genX(CmdWriteTimestamp)(
 315     VkCommandBuffer                             commandBuffer,
 316     VkPipelineStageFlagBits                     pipelineStage,
 317     VkQueryPool                                 queryPool,
 318     uint32_t                                    query)
 319 {
 320    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 321    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 322    uint32_t offset = query * pool->stride;
 323
 324    assert(pool->type == VK_QUERY_TYPE_TIMESTAMP);
 325
 326    switch (pipelineStage) {
 327    case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT:
 328       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 329          srm.RegisterAddress  = TIMESTAMP;
 330          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset + 8 };
 331       }
 332       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 333          srm.RegisterAddress  = TIMESTAMP + 4;
 334          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset + 12 };
 335       }
 336       break;
 337
 338    default:
 339       /* Everything else is bottom-of-pipe */
 340       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 341          pc.DestinationAddressType  = DAT_PPGTT;
 342          pc.PostSyncOperation       = WriteTimestamp;
 343          pc.Address = (struct anv_address) { &pool->bo, offset + 8 };
 344
 345          if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 346             pc.CommandStreamerStallEnable = true;
 347       }
 348       break;
 349    }
 350
 351    emit_query_availability(cmd_buffer, &pool->bo, offset);
 352 }
 353
 354 #if GEN_GEN > 7 || GEN_IS_HASWELL
 355
 356 #define alu_opcode(v)   __gen_uint((v),  20, 31)
 357 #define alu_operand1(v) __gen_uint((v),  10, 19)
 358 #define alu_operand2(v) __gen_uint((v),   0,  9)
 359 #define alu(opcode, operand1, operand2) \
 360    alu_opcode(opcode) | alu_operand1(operand1) | alu_operand2(operand2)
 361
 362 #define OPCODE_NOOP      0x000
 363 #define OPCODE_LOAD      0x080
 364 #define OPCODE_LOADINV   0x480
 365 #define OPCODE_LOAD0     0x081
 366 #define OPCODE_LOAD1     0x481
 367 #define OPCODE_ADD       0x100
 368 #define OPCODE_SUB       0x101
 369 #define OPCODE_AND       0x102
 370 #define OPCODE_OR        0x103
 371 #define OPCODE_XOR       0x104
 372 #define OPCODE_STORE     0x180
 373 #define OPCODE_STOREINV  0x580
 374
 375 #define OPERAND_R0   0x00
 376 #define OPERAND_R1   0x01
 377 #define OPERAND_R2   0x02
 378 #define OPERAND_R3   0x03
 379 #define OPERAND_R4   0x04
 380 #define OPERAND_SRCA 0x20
 381 #define OPERAND_SRCB 0x21
 382 #define OPERAND_ACCU 0x31
 383 #define OPERAND_ZF   0x32
 384 #define OPERAND_CF   0x33
 385
 386 #define CS_GPR(n) (0x2600 + (n) * 8)
 387
 388 static void
 389 emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t reg,
 390                       struct anv_bo *bo, uint32_t offset)
 391 {
 392    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 393       lrm.RegisterAddress  = reg,
 394       lrm.MemoryAddress    = (struct anv_address) { bo, offset };
 395    }
 396    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 397       lrm.RegisterAddress  = reg + 4;
 398       lrm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 399    }
 400 }
 401
 402 static void
 403 store_query_result(struct anv_batch *batch, uint32_t reg,
 404                    struct anv_bo *bo, uint32_t offset, VkQueryResultFlags flags)
 405 {
 406    anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 407       srm.RegisterAddress  = reg;
 408       srm.MemoryAddress    = (struct anv_address) { bo, offset };
 409    }
 410
 411    if (flags & VK_QUERY_RESULT_64_BIT) {
 412       anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 413          srm.RegisterAddress  = reg + 4;
 414          srm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 415       }
 416    }
 417 }
 418
 419 void genX(CmdCopyQueryPoolResults)(
 420     VkCommandBuffer                             commandBuffer,
 421     VkQueryPool                                 queryPool,
 422     uint32_t                                    firstQuery,
 423     uint32_t                                    queryCount,
 424     VkBuffer                                    destBuffer,
 425     VkDeviceSize                                destOffset,
 426     VkDeviceSize                                destStride,
 427     VkQueryResultFlags                          flags)
 428 {
 429    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 430    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 431    ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer);
 432    uint32_t slot_offset, dst_offset;
 433
 434    if (flags & VK_QUERY_RESULT_WAIT_BIT) {
 435       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 436          pc.CommandStreamerStallEnable = true;
 437          pc.StallAtPixelScoreboard     = true;
 438       }
 439    }
 440
 441    dst_offset = buffer->offset + destOffset;
 442    for (uint32_t i = 0; i < queryCount; i++) {
 443
 444       slot_offset = (firstQuery + i) * pool->stride;
 445       switch (pool->type) {
 446       case VK_QUERY_TYPE_OCCLUSION:
 447          emit_load_alu_reg_u64(&cmd_buffer->batch,
 448                                CS_GPR(0), &pool->bo, slot_offset + 8);
 449          emit_load_alu_reg_u64(&cmd_buffer->batch,
 450                                CS_GPR(1), &pool->bo, slot_offset + 16);
 451
 452          /* FIXME: We need to clamp the result for 32 bit. */
 453
 454          uint32_t *dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
 455          dw[1] = alu(OPCODE_LOAD, OPERAND_SRCA, OPERAND_R1);
 456          dw[2] = alu(OPCODE_LOAD, OPERAND_SRCB, OPERAND_R0);
 457          dw[3] = alu(OPCODE_SUB, 0, 0);
 458          dw[4] = alu(OPCODE_STORE, OPERAND_R2, OPERAND_ACCU);
 459          break;
 460
 461       case VK_QUERY_TYPE_TIMESTAMP:
 462          emit_load_alu_reg_u64(&cmd_buffer->batch,
 463                                CS_GPR(2), &pool->bo, slot_offset + 8);
 464          break;
 465
 466       default:
 467          unreachable("unhandled query type");
 468       }
 469
 470       store_query_result(&cmd_buffer->batch,
 471                          CS_GPR(2), buffer->bo, dst_offset, flags);
 472
 473       if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
 474          emit_load_alu_reg_u64(&cmd_buffer->batch, CS_GPR(0),
 475                                &pool->bo, slot_offset);
 476          if (flags & VK_QUERY_RESULT_64_BIT)
 477             store_query_result(&cmd_buffer->batch,
 478                                CS_GPR(0), buffer->bo, dst_offset + 8, flags);
 479          else
 480             store_query_result(&cmd_buffer->batch,
 481                                CS_GPR(0), buffer->bo, dst_offset + 4, flags);
 482       }
 483
 484       dst_offset += destStride;
 485    }
 486 }
 487
 488 #else
 489 void genX(CmdCopyQueryPoolResults)(
 490     VkCommandBuffer                             commandBuffer,
 491     VkQueryPool                                 queryPool,
 492     uint32_t                                    firstQuery,
 493     uint32_t                                    queryCount,
 494     VkBuffer                                    destBuffer,
 495     VkDeviceSize                                destOffset,
 496     VkDeviceSize                                destStride,
 497     VkQueryResultFlags                          flags)
 498 {
 499    anv_finishme("Queries not yet supported on Ivy Bridge");
 500 }
 501 #endif