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       anv_invalidate_range(slot, MIN2(queryCount * sizeof(*slot), pool->bo.size));
 137
 138    for (uint32_t i = 0; i < queryCount; i++) {
 139       switch (pool->type) {
 140       case VK_QUERY_TYPE_OCCLUSION: {
 141          result = slot[firstQuery + i].end - slot[firstQuery + i].begin;
 142          break;
 143       }
 144       case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 145          unreachable("pipeline stats not supported");
 146       case VK_QUERY_TYPE_TIMESTAMP: {
 147          result = slot[firstQuery + i].begin;
 148          break;
 149       }
 150       default:
 151          unreachable("invalid pool type");
 152       }
 153
 154       if (flags & VK_QUERY_RESULT_64_BIT) {
 155          uint64_t *dst = pData;
 156          dst[0] = result;
 157          if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
 158             dst[1] = slot[firstQuery + i].available;
 159       } else {
 160          uint32_t *dst = pData;
 161          if (result > UINT32_MAX)
 162             result = UINT32_MAX;
 163          dst[0] = result;
 164          if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
 165             dst[1] = slot[firstQuery + i].available;
 166       }
 167
 168       pData += stride;
 169       if (pData >= data_end)
 170          break;
 171    }
 172
 173    return VK_SUCCESS;
 174 }
 175
 176 static void
 177 emit_ps_depth_count(struct anv_cmd_buffer *cmd_buffer,
 178                     struct anv_bo *bo, uint32_t offset)
 179 {
 180    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 181       pc.DestinationAddressType  = DAT_PPGTT;
 182       pc.PostSyncOperation       = WritePSDepthCount;
 183       pc.DepthStallEnable        = true;
 184       pc.Address                 = (struct anv_address) { bo, offset };
 185
 186       if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 187          pc.CommandStreamerStallEnable = true;
 188    }
 189 }
 190
 191 static void
 192 emit_query_availability(struct anv_cmd_buffer *cmd_buffer,
 193                         struct anv_bo *bo, uint32_t offset)
 194 {
 195    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 196       pc.DestinationAddressType  = DAT_PPGTT;
 197       pc.PostSyncOperation       = WriteImmediateData;
 198       pc.Address                 = (struct anv_address) { bo, offset };
 199       pc.ImmediateData           = 1;
 200    }
 201 }
 202
 203 void genX(CmdResetQueryPool)(
 204     VkCommandBuffer                             commandBuffer,
 205     VkQueryPool                                 queryPool,
 206     uint32_t                                    firstQuery,
 207     uint32_t                                    queryCount)
 208 {
 209    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 210    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 211
 212    for (uint32_t i = 0; i < queryCount; i++) {
 213       switch (pool->type) {
 214       case VK_QUERY_TYPE_OCCLUSION:
 215       case VK_QUERY_TYPE_TIMESTAMP: {
 216          anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdm) {
 217             sdm.Address = (struct anv_address) {
 218                .bo = &pool->bo,
 219                .offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot) +
 220                          offsetof(struct anv_query_pool_slot, available),
 221             };
 222             sdm.DataDWord0 = 0;
 223             sdm.DataDWord1 = 0;
 224          }
 225          break;
 226       }
 227       default:
 228          assert(!"Invalid query type");
 229       }
 230    }
 231 }
 232
 233 void genX(CmdBeginQuery)(
 234     VkCommandBuffer                             commandBuffer,
 235     VkQueryPool                                 queryPool,
 236     uint32_t                                    query,
 237     VkQueryControlFlags                         flags)
 238 {
 239    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 240    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 241
 242    /* Workaround: When meta uses the pipeline with the VS disabled, it seems
 243     * that the pipelining of the depth write breaks. What we see is that
 244     * samples from the render pass clear leaks into the first query
 245     * immediately after the clear. Doing a pipecontrol with a post-sync
 246     * operation and DepthStallEnable seems to work around the issue.
 247     */
 248    if (cmd_buffer->state.need_query_wa) {
 249       cmd_buffer->state.need_query_wa = false;
 250       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 251          pc.DepthCacheFlushEnable   = true;
 252          pc.DepthStallEnable        = true;
 253       }
 254    }
 255
 256    switch (pool->type) {
 257    case VK_QUERY_TYPE_OCCLUSION:
 258       emit_ps_depth_count(cmd_buffer, &pool->bo,
 259                           query * sizeof(struct anv_query_pool_slot));
 260       break;
 261
 262    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 263    default:
 264       unreachable("");
 265    }
 266 }
 267
 268 void genX(CmdEndQuery)(
 269     VkCommandBuffer                             commandBuffer,
 270     VkQueryPool                                 queryPool,
 271     uint32_t                                    query)
 272 {
 273    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 274    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 275
 276    switch (pool->type) {
 277    case VK_QUERY_TYPE_OCCLUSION:
 278       emit_ps_depth_count(cmd_buffer, &pool->bo,
 279                           query * sizeof(struct anv_query_pool_slot) + 8);
 280
 281       emit_query_availability(cmd_buffer, &pool->bo,
 282                               query * sizeof(struct anv_query_pool_slot) + 16);
 283       break;
 284
 285    case VK_QUERY_TYPE_PIPELINE_STATISTICS:
 286    default:
 287       unreachable("");
 288    }
 289 }
 290
 291 #define TIMESTAMP 0x2358
 292
 293 void genX(CmdWriteTimestamp)(
 294     VkCommandBuffer                             commandBuffer,
 295     VkPipelineStageFlagBits                     pipelineStage,
 296     VkQueryPool                                 queryPool,
 297     uint32_t                                    query)
 298 {
 299    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 300    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 301    uint32_t offset = query * sizeof(struct anv_query_pool_slot);
 302
 303    assert(pool->type == VK_QUERY_TYPE_TIMESTAMP);
 304
 305    switch (pipelineStage) {
 306    case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT:
 307       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 308          srm.RegisterAddress  = TIMESTAMP;
 309          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset };
 310       }
 311       anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 312          srm.RegisterAddress  = TIMESTAMP + 4;
 313          srm.MemoryAddress    = (struct anv_address) { &pool->bo, offset + 4 };
 314       }
 315       break;
 316
 317    default:
 318       /* Everything else is bottom-of-pipe */
 319       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 320          pc.DestinationAddressType  = DAT_PPGTT;
 321          pc.PostSyncOperation       = WriteTimestamp;
 322          pc.Address = (struct anv_address) { &pool->bo, offset };
 323
 324          if (GEN_GEN == 9 && cmd_buffer->device->info.gt == 4)
 325             pc.CommandStreamerStallEnable = true;
 326       }
 327       break;
 328    }
 329
 330    emit_query_availability(cmd_buffer, &pool->bo, query + 16);
 331 }
 332
 333 #if GEN_GEN > 7 || GEN_IS_HASWELL
 334
 335 #define alu_opcode(v)   __gen_uint((v),  20, 31)
 336 #define alu_operand1(v) __gen_uint((v),  10, 19)
 337 #define alu_operand2(v) __gen_uint((v),   0,  9)
 338 #define alu(opcode, operand1, operand2) \
 339    alu_opcode(opcode) | alu_operand1(operand1) | alu_operand2(operand2)
 340
 341 #define OPCODE_NOOP      0x000
 342 #define OPCODE_LOAD      0x080
 343 #define OPCODE_LOADINV   0x480
 344 #define OPCODE_LOAD0     0x081
 345 #define OPCODE_LOAD1     0x481
 346 #define OPCODE_ADD       0x100
 347 #define OPCODE_SUB       0x101
 348 #define OPCODE_AND       0x102
 349 #define OPCODE_OR        0x103
 350 #define OPCODE_XOR       0x104
 351 #define OPCODE_STORE     0x180
 352 #define OPCODE_STOREINV  0x580
 353
 354 #define OPERAND_R0   0x00
 355 #define OPERAND_R1   0x01
 356 #define OPERAND_R2   0x02
 357 #define OPERAND_R3   0x03
 358 #define OPERAND_R4   0x04
 359 #define OPERAND_SRCA 0x20
 360 #define OPERAND_SRCB 0x21
 361 #define OPERAND_ACCU 0x31
 362 #define OPERAND_ZF   0x32
 363 #define OPERAND_CF   0x33
 364
 365 #define CS_GPR(n) (0x2600 + (n) * 8)
 366
 367 static void
 368 emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t reg,
 369                       struct anv_bo *bo, uint32_t offset)
 370 {
 371    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 372       lrm.RegisterAddress  = reg,
 373       lrm.MemoryAddress    = (struct anv_address) { bo, offset };
 374    }
 375    anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
 376       lrm.RegisterAddress  = reg + 4;
 377       lrm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 378    }
 379 }
 380
 381 static void
 382 store_query_result(struct anv_batch *batch, uint32_t reg,
 383                    struct anv_bo *bo, uint32_t offset, VkQueryResultFlags flags)
 384 {
 385    anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 386       srm.RegisterAddress  = reg;
 387       srm.MemoryAddress    = (struct anv_address) { bo, offset };
 388    }
 389
 390    if (flags & VK_QUERY_RESULT_64_BIT) {
 391       anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
 392          srm.RegisterAddress  = reg + 4;
 393          srm.MemoryAddress    = (struct anv_address) { bo, offset + 4 };
 394       }
 395    }
 396 }
 397
 398 void genX(CmdCopyQueryPoolResults)(
 399     VkCommandBuffer                             commandBuffer,
 400     VkQueryPool                                 queryPool,
 401     uint32_t                                    firstQuery,
 402     uint32_t                                    queryCount,
 403     VkBuffer                                    destBuffer,
 404     VkDeviceSize                                destOffset,
 405     VkDeviceSize                                destStride,
 406     VkQueryResultFlags                          flags)
 407 {
 408    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 409    ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
 410    ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer);
 411    uint32_t slot_offset, dst_offset;
 412
 413    if (flags & VK_QUERY_RESULT_WAIT_BIT) {
 414       anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
 415          pc.CommandStreamerStallEnable = true;
 416          pc.StallAtPixelScoreboard     = true;
 417       }
 418    }
 419
 420    dst_offset = buffer->offset + destOffset;
 421    for (uint32_t i = 0; i < queryCount; i++) {
 422
 423       slot_offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot);
 424       switch (pool->type) {
 425       case VK_QUERY_TYPE_OCCLUSION:
 426          emit_load_alu_reg_u64(&cmd_buffer->batch,
 427                                CS_GPR(0), &pool->bo, slot_offset);
 428          emit_load_alu_reg_u64(&cmd_buffer->batch,
 429                                CS_GPR(1), &pool->bo, slot_offset + 8);
 430
 431          /* FIXME: We need to clamp the result for 32 bit. */
 432
 433          uint32_t *dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
 434          dw[1] = alu(OPCODE_LOAD, OPERAND_SRCA, OPERAND_R1);
 435          dw[2] = alu(OPCODE_LOAD, OPERAND_SRCB, OPERAND_R0);
 436          dw[3] = alu(OPCODE_SUB, 0, 0);
 437          dw[4] = alu(OPCODE_STORE, OPERAND_R2, OPERAND_ACCU);
 438          break;
 439
 440       case VK_QUERY_TYPE_TIMESTAMP:
 441          emit_load_alu_reg_u64(&cmd_buffer->batch,
 442                                CS_GPR(2), &pool->bo, slot_offset);
 443          break;
 444
 445       default:
 446          unreachable("unhandled query type");
 447       }
 448
 449       store_query_result(&cmd_buffer->batch,
 450                          CS_GPR(2), buffer->bo, dst_offset, flags);
 451
 452       if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
 453          emit_load_alu_reg_u64(&cmd_buffer->batch, CS_GPR(0),
 454                                &pool->bo, slot_offset + 16);
 455          if (flags & VK_QUERY_RESULT_64_BIT)
 456             store_query_result(&cmd_buffer->batch,
 457                                CS_GPR(0), buffer->bo, dst_offset + 8, flags);
 458          else
 459             store_query_result(&cmd_buffer->batch,
 460                                CS_GPR(0), buffer->bo, dst_offset + 4, flags);
 461       }
 462
 463       dst_offset += destStride;
 464    }
 465 }
 466
 467 #else
 468 void genX(CmdCopyQueryPoolResults)(
 469     VkCommandBuffer                             commandBuffer,
 470     VkQueryPool                                 queryPool,
 471     uint32_t                                    firstQuery,
 472     uint32_t                                    queryCount,
 473     VkBuffer                                    destBuffer,
 474     VkDeviceSize                                destOffset,
 475     VkDeviceSize                                destStride,
 476     VkQueryResultFlags                          flags)
 477 {
 478    anv_finishme("Queries not yet supported on Ivy Bridge");
 479 }
 480 #endif