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 (write_results)
 185             dst[0] = result;
 186          if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
 187             dst[1] = slot[firstQuery + i].available;
 188       }
 189
 190       pData += stride;
 191       if (pData >= data_end)
 192          break;
 193    }
 194
 195    return status;
 196 }
 197
 198 static void
 199 emit_ps_depth_count(struct anv_cmd_buffer *cmd_buffer,
 200                     struct anv_bo *bo, uint32_t offset)
 201 {
 202    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 203       pc.DestinationAddressType  = DAT_PPGTT;
 204       pc.PostSyncOperation       = WritePSDepthCount;
 205       pc.DepthStallEnable        = true;
 206       pc.Address                 = (struct anv_address) { bo, offset };
 207
 208       if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 209          pc.CommandStreamerStallEnable = true;
 210    }
 211 }
 212
 213 static void
 214 emit_query_availability(struct anv_cmd_buffer *cmd_buffer,
 215                         struct anv_bo *bo, uint32_t offset)
 216 {
 217    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 218       pc.DestinationAddressType  = DAT_PPGTT;
 219       pc.PostSyncOperation       = WriteImmediateData;
 220       pc.Address                 = (struct anv_address) { bo, offset };
 221       pc.ImmediateData           = 1;
 222    }
 223 }
 224
 225 void genX(CmdResetQueryPool)(
 226     VkCommandBuffer                             commandBuffer,
 227     VkQueryPool                                 queryPool,
 228     uint32_t                                    firstQuery,
 229     uint32_t                                    queryCount)
 230 {
 231    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 232    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 233
 234    for (uint32_t i = 0; i < queryCount; i++) {
 235       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdm) {
 236          sdm.Address = (struct anv_address) {
 237             .bo = &pool->bo,
 238             .offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot),
 239          };
 240          sdm.DataDWord0 = 0;
 241          sdm.DataDWord1 = 0;
 242       }
 243    }
 244 }
 245
 246 void genX(CmdBeginQuery)(
 247     VkCommandBuffer                             commandBuffer,
 248     VkQueryPool                                 queryPool,
 249     uint32_t                                    query,
 250     VkQueryControlFlags                         flags)
 251 {
 252    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 253    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 254
 255    /* Workaround: When meta uses the pipeline with the VS disabled, it seems
 256     * that the pipelining of the depth write breaks. What we see is that
 257     * samples from the render pass clear leaks into the first query
 258     * immediately after the clear. Doing a pipecontrol with a post-sync
 259     * operation and DepthStallEnable seems to work around the issue.
 260     */
 261    if (cmd_buffer->state.need_query_wa) {
 262       cmd_buffer->state.need_query_wa = false;
 263       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 264          pc.DepthCacheFlushEnable   = true;
 265          pc.DepthStallEnable        = true;
 266       }
 267    }
 268
 269    switch (pool->type) {
 270    case VK_QUERY_TYPE_OCCLUSION:
 271       emit_ps_depth_count(cmd_buffer, &pool->bo,
 272                           query * sizeof(struct anv_query_pool_slot) + 8);
 273       break;
 274
 275    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 276    default:
 277       unreachable("");
 278    }
 279 }
 280
 281 void genX(CmdEndQuery)(
 282     VkCommandBuffer                             commandBuffer,
 283     VkQueryPool                                 queryPool,
 284     uint32_t                                    query)
 285 {
 286    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 287    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 288
 289    switch (pool->type) {
 290    case VK_QUERY_TYPE_OCCLUSION:
 291       emit_ps_depth_count(cmd_buffer, &pool->bo,
 292                           query * sizeof(struct anv_query_pool_slot) + 16);
 293
 294       emit_query_availability(cmd_buffer, &pool->bo,
 295                               query * sizeof(struct anv_query_pool_slot));
 296       break;
 297
 298    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 299    default:
 300       unreachable("");
 301    }
 302 }
 303
 304 #define TIMESTAMP 0x2358
 305
 306 void genX(CmdWriteTimestamp)(
 307     VkCommandBuffer                             commandBuffer,
 308     VkPipelineStageFlagBits                     pipelineStage,
 309     VkQueryPool                                 queryPool,
 310     uint32_t                                    query)
 311 {
 312    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 313    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 314    uint32_t offset = query * sizeof(struct anv_query_pool_slot);
 315
 316    assert(pool->type == VK_QUERY_TYPE_TIMESTAMP);
 317
 318    switch (pipelineStage) {
 319    case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT:
 320       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 321          srm.RegisterAddress  = TIMESTAMP;
 322          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset + 8 };
 323       }
 324       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 325          srm.RegisterAddress  = TIMESTAMP + 4;
 326          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset + 12 };
 327       }
 328       break;
 329
 330    default:
 331       /* Everything else is bottom-of-pipe */
 332       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 333          pc.DestinationAddressType  = DAT_PPGTT;
 334          pc.PostSyncOperation       = WriteTimestamp;
 335          pc.Address = (struct anv_address) { &pool->bo, offset + 8 };
 336
 337          if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 338             pc.CommandStreamerStallEnable = true;
 339       }
 340       break;
 341    }
 342
 343    emit_query_availability(cmd_buffer, &pool->bo, offset);
 344 }
 345
 346 #if GEN_GEN > 7 || GEN_IS_HASWELL
 347
 348 #define alu_opcode(v)   __gen_uint((v),  20, 31)
 349 #define alu_operand1(v) __gen_uint((v),  10, 19)
 350 #define alu_operand2(v) __gen_uint((v),   0,  9)
 351 #define alu(opcode, operand1, operand2) \
 352    alu_opcode(opcode) | alu_operand1(operand1) | alu_operand2(operand2)
 353
 354 #define OPCODE_NOOP      0x000
 355 #define OPCODE_LOAD      0x080
 356 #define OPCODE_LOADINV   0x480
 357 #define OPCODE_LOAD0     0x081
 358 #define OPCODE_LOAD1     0x481
 359 #define OPCODE_ADD       0x100
 360 #define OPCODE_SUB       0x101
 361 #define OPCODE_AND       0x102
 362 #define OPCODE_OR        0x103
 363 #define OPCODE_XOR       0x104
 364 #define OPCODE_STORE     0x180
 365 #define OPCODE_STOREINV  0x580
 366
 367 #define OPERAND_R0   0x00
 368 #define OPERAND_R1   0x01
 369 #define OPERAND_R2   0x02
 370 #define OPERAND_R3   0x03
 371 #define OPERAND_R4   0x04
 372 #define OPERAND_SRCA 0x20
 373 #define OPERAND_SRCB 0x21
 374 #define OPERAND_ACCU 0x31
 375 #define OPERAND_ZF   0x32
 376 #define OPERAND_CF   0x33
 377
 378 #define CS_GPR(n) (0x2600 + (n) * 8)
 379
 380 static void
 381 emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t reg,
 382                       struct anv_bo *bo, uint32_t offset)
 383 {
 384    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 385       lrm.RegisterAddress  = reg,
 386       lrm.MemoryAddress    = (struct anv_address) { bo, offset };
 387    }
 388    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 389       lrm.RegisterAddress  = reg + 4;
 390       lrm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 391    }
 392 }
 393
 394 static void
 395 store_query_result(struct anv_batch *batch, uint32_t reg,
 396                    struct anv_bo *bo, uint32_t offset, VkQueryResultFlags flags)
 397 {
 398    anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 399       srm.RegisterAddress  = reg;
 400       srm.MemoryAddress    = (struct anv_address) { bo, offset };
 401    }
 402
 403    if (flags & VK_QUERY_RESULT_64_BIT) {
 404       anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 405          srm.RegisterAddress  = reg + 4;
 406          srm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 407       }
 408    }
 409 }
 410
 411 void genX(CmdCopyQueryPoolResults)(
 412     VkCommandBuffer                             commandBuffer,
 413     VkQueryPool                                 queryPool,
 414     uint32_t                                    firstQuery,
 415     uint32_t                                    queryCount,
 416     VkBuffer                                    destBuffer,
 417     VkDeviceSize                                destOffset,
 418     VkDeviceSize                                destStride,
 419     VkQueryResultFlags                          flags)
 420 {
 421    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 422    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 423    ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer);
 424    uint32_t slot_offset, dst_offset;
 425
 426    if (flags & VK_QUERY_RESULT_WAIT_BIT) {
 427       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 428          pc.CommandStreamerStallEnable = true;
 429          pc.StallAtPixelScoreboard     = true;
 430       }
 431    }
 432
 433    dst_offset = buffer->offset + destOffset;
 434    for (uint32_t i = 0; i < queryCount; i++) {
 435
 436       slot_offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot);
 437       switch (pool->type) {
 438       case VK_QUERY_TYPE_OCCLUSION:
 439          emit_load_alu_reg_u64(&cmd_buffer->batch,
 440                                CS_GPR(0), &pool->bo, slot_offset + 8);
 441          emit_load_alu_reg_u64(&cmd_buffer->batch,
 442                                CS_GPR(1), &pool->bo, slot_offset + 16);
 443
 444          /* FIXME: We need to clamp the result for 32 bit. */
 445
 446          uint32_t *dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
 447          dw[1] = alu(OPCODE_LOAD, OPERAND_SRCA, OPERAND_R1);
 448          dw[2] = alu(OPCODE_LOAD, OPERAND_SRCB, OPERAND_R0);
 449          dw[3] = alu(OPCODE_SUB, 0, 0);
 450          dw[4] = alu(OPCODE_STORE, OPERAND_R2, OPERAND_ACCU);
 451          break;
 452
 453       case VK_QUERY_TYPE_TIMESTAMP:
 454          emit_load_alu_reg_u64(&cmd_buffer->batch,
 455                                CS_GPR(2), &pool->bo, slot_offset + 8);
 456          break;
 457
 458       default:
 459          unreachable("unhandled query type");
 460       }
 461
 462       store_query_result(&cmd_buffer->batch,
 463                          CS_GPR(2), buffer->bo, dst_offset, flags);
 464
 465       if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
 466          emit_load_alu_reg_u64(&cmd_buffer->batch, CS_GPR(0),
 467                                &pool->bo, slot_offset);
 468          if (flags & VK_QUERY_RESULT_64_BIT)
 469             store_query_result(&cmd_buffer->batch,
 470                                CS_GPR(0), buffer->bo, dst_offset + 8, flags);
 471          else
 472             store_query_result(&cmd_buffer->batch,
 473                                CS_GPR(0), buffer->bo, dst_offset + 4, flags);
 474       }
 475
 476       dst_offset += destStride;
 477    }
 478 }
 479
 480 #else
 481 void genX(CmdCopyQueryPoolResults)(
 482     VkCommandBuffer                             commandBuffer,
 483     VkQueryPool                                 queryPool,
 484     uint32_t                                    firstQuery,
 485     uint32_t                                    queryCount,
 486     VkBuffer                                    destBuffer,
 487     VkDeviceSize                                destOffset,
 488     VkDeviceSize                                destStride,
 489     VkQueryResultFlags                          flags)
 490 {
 491    anv_finishme("Queries not yet supported on Ivy Bridge");
 492 }
 493 #endif