src/intel/vulkan/anv_cmd_buffer.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 "vk_format_info.h"
  33
  34 /** \file anv_cmd_buffer.c
  35  *
  36  * This file contains all of the stuff for emitting commands into a command
  37  * buffer.  This includes implementations of most of the vkCmd*
  38  * entrypoints.  This file is concerned entirely with state emission and
  39  * not with the command buffer data structure itself.  As far as this file
  40  * is concerned, most of anv_cmd_buffer is magic.
  41  */
  42
  43 /* TODO: These are taken from GLES.  We should check the Vulkan spec */
  44 const struct anv_dynamic_state default_dynamic_state = {
  45    .viewport = {
  46       .count = 0,
  47    },
  48    .scissor = {
  49       .count = 0,
  50    },
  51    .line_width = 1.0f,
  52    .depth_bias = {
  53       .bias = 0.0f,
  54       .clamp = 0.0f,
  55       .slope = 0.0f,
  56    },
  57    .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
  58    .depth_bounds = {
  59       .min = 0.0f,
  60       .max = 1.0f,
  61    },
  62    .stencil_compare_mask = {
  63       .front = ~0u,
  64       .back = ~0u,
  65    },
  66    .stencil_write_mask = {
  67       .front = ~0u,
  68       .back = ~0u,
  69    },
  70    .stencil_reference = {
  71       .front = 0u,
  72       .back = 0u,
  73    },
  74 };
  75
  76 void
  77 anv_dynamic_state_copy(struct anv_dynamic_state *dest,
  78                        const struct anv_dynamic_state *src,
  79                        uint32_t copy_mask)
  80 {
  81    if (copy_mask & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
  82       dest->viewport.count = src->viewport.count;
  83       typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
  84                    src->viewport.count);
  85    }
  86
  87    if (copy_mask & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
  88       dest->scissor.count = src->scissor.count;
  89       typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
  90                    src->scissor.count);
  91    }
  92
  93    if (copy_mask & (1 << VK_DYNAMIC_STATE_LINE_WIDTH))
  94       dest->line_width = src->line_width;
  95
  96    if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS))
  97       dest->depth_bias = src->depth_bias;
  98
  99    if (copy_mask & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS))
 100       typed_memcpy(dest->blend_constants, src->blend_constants, 4);
 101
 102    if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS))
 103       dest->depth_bounds = src->depth_bounds;
 104
 105    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK))
 106       dest->stencil_compare_mask = src->stencil_compare_mask;
 107
 108    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK))
 109       dest->stencil_write_mask = src->stencil_write_mask;
 110
 111    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE))
 112       dest->stencil_reference = src->stencil_reference;
 113 }
 114
 115 static void
 116 anv_cmd_state_reset(struct anv_cmd_buffer *cmd_buffer)
 117 {
 118    struct anv_cmd_state *state = &cmd_buffer->state;
 119
 120    cmd_buffer->batch.status = VK_SUCCESS;
 121
 122    memset(&state->descriptors, 0, sizeof(state->descriptors));
 123    for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
 124       if (state->push_constants[i] != NULL) {
 125          vk_free(&cmd_buffer->pool->alloc, state->push_constants[i]);
 126          state->push_constants[i] = NULL;
 127       }
 128    }
 129    memset(state->binding_tables, 0, sizeof(state->binding_tables));
 130    memset(state->samplers, 0, sizeof(state->samplers));
 131
 132    /* 0 isn't a valid config.  This ensures that we always configure L3$. */
 133    cmd_buffer->state.current_l3_config = 0;
 134
 135    state->dirty = 0;
 136    state->vb_dirty = 0;
 137    state->pending_pipe_bits = 0;
 138    state->descriptors_dirty = 0;
 139    state->push_constants_dirty = 0;
 140    state->pipeline = NULL;
 141    state->framebuffer = NULL;
 142    state->pass = NULL;
 143    state->subpass = NULL;
 144    state->push_constant_stages = 0;
 145    state->restart_index = UINT32_MAX;
 146    state->dynamic = default_dynamic_state;
 147    state->need_query_wa = true;
 148    state->pma_fix_enabled = false;
 149    state->hiz_enabled = false;
 150
 151    if (state->attachments != NULL) {
 152       vk_free(&cmd_buffer->pool->alloc, state->attachments);
 153       state->attachments = NULL;
 154    }
 155
 156    state->gen7.index_buffer = NULL;
 157 }
 158
 159 VkResult
 160 anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer,
 161                                           gl_shader_stage stage, uint32_t size)
 162 {
 163    struct anv_push_constants **ptr = &cmd_buffer->state.push_constants[stage];
 164
 165    if (*ptr == NULL) {
 166       *ptr = vk_alloc(&cmd_buffer->pool->alloc, size, 8,
 167                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 168       if (*ptr == NULL) {
 169          anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
 170          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 171       }
 172    } else if ((*ptr)->size < size) {
 173       *ptr = vk_realloc(&cmd_buffer->pool->alloc, *ptr, size, 8,
 174                          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 175       if (*ptr == NULL) {
 176          anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
 177          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 178       }
 179    }
 180    (*ptr)->size = size;
 181
 182    return VK_SUCCESS;
 183 }
 184
 185 static VkResult anv_create_cmd_buffer(
 186     struct anv_device *                         device,
 187     struct anv_cmd_pool *                       pool,
 188     VkCommandBufferLevel                        level,
 189     VkCommandBuffer*                            pCommandBuffer)
 190 {
 191    struct anv_cmd_buffer *cmd_buffer;
 192    VkResult result;
 193
 194    cmd_buffer = vk_alloc(&pool->alloc, sizeof(*cmd_buffer), 8,
 195                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 196    if (cmd_buffer == NULL)
 197       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 198
 199    cmd_buffer->batch.status = VK_SUCCESS;
 200
 201    for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
 202       cmd_buffer->state.push_constants[i] = NULL;
 203    }
 204    cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 205    cmd_buffer->device = device;
 206    cmd_buffer->pool = pool;
 207    cmd_buffer->level = level;
 208    cmd_buffer->state.attachments = NULL;
 209
 210    result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer);
 211    if (result != VK_SUCCESS)
 212       goto fail;
 213
 214    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 215                          &device->surface_state_pool, 4096);
 216    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 217                          &device->dynamic_state_pool, 16384);
 218
 219    memset(&cmd_buffer->state.push_descriptor, 0,
 220           sizeof(cmd_buffer->state.push_descriptor));
 221
 222    if (pool) {
 223       list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
 224    } else {
 225       /* Init the pool_link so we can safefly call list_del when we destroy
 226        * the command buffer
 227        */
 228       list_inithead(&cmd_buffer->pool_link);
 229    }
 230
 231    *pCommandBuffer = anv_cmd_buffer_to_handle(cmd_buffer);
 232
 233    return VK_SUCCESS;
 234
 235  fail:
 236    vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
 237
 238    return result;
 239 }
 240
 241 VkResult anv_AllocateCommandBuffers(
 242     VkDevice                                    _device,
 243     const VkCommandBufferAllocateInfo*          pAllocateInfo,
 244     VkCommandBuffer*                            pCommandBuffers)
 245 {
 246    ANV_FROM_HANDLE(anv_device, device, _device);
 247    ANV_FROM_HANDLE(anv_cmd_pool, pool, pAllocateInfo->commandPool);
 248
 249    VkResult result = VK_SUCCESS;
 250    uint32_t i;
 251
 252    for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
 253       result = anv_create_cmd_buffer(device, pool, pAllocateInfo->level,
 254                                      &pCommandBuffers[i]);
 255       if (result != VK_SUCCESS)
 256          break;
 257    }
 258
 259    if (result != VK_SUCCESS) {
 260       anv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
 261                              i, pCommandBuffers);
 262       for (i = 0; i < pAllocateInfo->commandBufferCount; i++)
 263          pCommandBuffers[i] = VK_NULL_HANDLE;
 264    }
 265
 266    return result;
 267 }
 268
 269 static void
 270 anv_cmd_buffer_destroy(struct anv_cmd_buffer *cmd_buffer)
 271 {
 272    list_del(&cmd_buffer->pool_link);
 273
 274    anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer);
 275
 276    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 277    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 278
 279    vk_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
 280    vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
 281 }
 282
 283 void anv_FreeCommandBuffers(
 284     VkDevice                                    device,
 285     VkCommandPool                               commandPool,
 286     uint32_t                                    commandBufferCount,
 287     const VkCommandBuffer*                      pCommandBuffers)
 288 {
 289    for (uint32_t i = 0; i < commandBufferCount; i++) {
 290       ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
 291
 292       if (!cmd_buffer)
 293          continue;
 294
 295       anv_cmd_buffer_destroy(cmd_buffer);
 296    }
 297 }
 298
 299 VkResult
 300 anv_cmd_buffer_reset(struct anv_cmd_buffer *cmd_buffer)
 301 {
 302    cmd_buffer->usage_flags = 0;
 303    cmd_buffer->state.current_pipeline = UINT32_MAX;
 304    anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer);
 305    anv_cmd_state_reset(cmd_buffer);
 306
 307    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 308    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 309                          &cmd_buffer->device->surface_state_pool, 4096);
 310
 311    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 312    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 313                          &cmd_buffer->device->dynamic_state_pool, 16384);
 314    return VK_SUCCESS;
 315 }
 316
 317 VkResult anv_ResetCommandBuffer(
 318     VkCommandBuffer                             commandBuffer,
 319     VkCommandBufferResetFlags                   flags)
 320 {
 321    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 322    return anv_cmd_buffer_reset(cmd_buffer);
 323 }
 324
 325 void
 326 anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer)
 327 {
 328    switch (cmd_buffer->device->info.gen) {
 329    case 7:
 330       if (cmd_buffer->device->info.is_haswell)
 331          return gen75_cmd_buffer_emit_state_base_address(cmd_buffer);
 332       else
 333          return gen7_cmd_buffer_emit_state_base_address(cmd_buffer);
 334    case 8:
 335       return gen8_cmd_buffer_emit_state_base_address(cmd_buffer);
 336    case 9:
 337       return gen9_cmd_buffer_emit_state_base_address(cmd_buffer);
 338    default:
 339       unreachable("unsupported gen\n");
 340    }
 341 }
 342
 343 void anv_CmdBindPipeline(
 344     VkCommandBuffer                             commandBuffer,
 345     VkPipelineBindPoint                         pipelineBindPoint,
 346     VkPipeline                                  _pipeline)
 347 {
 348    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 349    ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 350
 351    switch (pipelineBindPoint) {
 352    case VK_PIPELINE_BIND_POINT_COMPUTE:
 353       cmd_buffer->state.compute_pipeline = pipeline;
 354       cmd_buffer->state.compute_dirty |= ANV_CMD_DIRTY_PIPELINE;
 355       cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 356       cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 357       break;
 358
 359    case VK_PIPELINE_BIND_POINT_GRAPHICS:
 360       cmd_buffer->state.pipeline = pipeline;
 361       cmd_buffer->state.vb_dirty |= pipeline->vb_used;
 362       cmd_buffer->state.dirty |= ANV_CMD_DIRTY_PIPELINE;
 363       cmd_buffer->state.push_constants_dirty |= pipeline->active_stages;
 364       cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;
 365
 366       /* Apply the dynamic state from the pipeline */
 367       cmd_buffer->state.dirty |= pipeline->dynamic_state_mask;
 368       anv_dynamic_state_copy(&cmd_buffer->state.dynamic,
 369                              &pipeline->dynamic_state,
 370                              pipeline->dynamic_state_mask);
 371       break;
 372
 373    default:
 374       assert(!"invalid bind point");
 375       break;
 376    }
 377 }
 378
 379 void anv_CmdSetViewport(
 380     VkCommandBuffer                             commandBuffer,
 381     uint32_t                                    firstViewport,
 382     uint32_t                                    viewportCount,
 383     const VkViewport*                           pViewports)
 384 {
 385    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 386
 387    const uint32_t total_count = firstViewport + viewportCount;
 388    if (cmd_buffer->state.dynamic.viewport.count < total_count)
 389       cmd_buffer->state.dynamic.viewport.count = total_count;
 390
 391    memcpy(cmd_buffer->state.dynamic.viewport.viewports + firstViewport,
 392           pViewports, viewportCount * sizeof(*pViewports));
 393
 394    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT;
 395 }
 396
 397 void anv_CmdSetScissor(
 398     VkCommandBuffer                             commandBuffer,
 399     uint32_t                                    firstScissor,
 400     uint32_t                                    scissorCount,
 401     const VkRect2D*                             pScissors)
 402 {
 403    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 404
 405    const uint32_t total_count = firstScissor + scissorCount;
 406    if (cmd_buffer->state.dynamic.scissor.count < total_count)
 407       cmd_buffer->state.dynamic.scissor.count = total_count;
 408
 409    memcpy(cmd_buffer->state.dynamic.scissor.scissors + firstScissor,
 410           pScissors, scissorCount * sizeof(*pScissors));
 411
 412    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR;
 413 }
 414
 415 void anv_CmdSetLineWidth(
 416     VkCommandBuffer                             commandBuffer,
 417     float                                       lineWidth)
 418 {
 419    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 420
 421    cmd_buffer->state.dynamic.line_width = lineWidth;
 422    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
 423 }
 424
 425 void anv_CmdSetDepthBias(
 426     VkCommandBuffer                             commandBuffer,
 427     float                                       depthBiasConstantFactor,
 428     float                                       depthBiasClamp,
 429     float                                       depthBiasSlopeFactor)
 430 {
 431    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 432
 433    cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
 434    cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
 435    cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
 436
 437    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
 438 }
 439
 440 void anv_CmdSetBlendConstants(
 441     VkCommandBuffer                             commandBuffer,
 442     const float                                 blendConstants[4])
 443 {
 444    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 445
 446    memcpy(cmd_buffer->state.dynamic.blend_constants,
 447           blendConstants, sizeof(float) * 4);
 448
 449    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
 450 }
 451
 452 void anv_CmdSetDepthBounds(
 453     VkCommandBuffer                             commandBuffer,
 454     float                                       minDepthBounds,
 455     float                                       maxDepthBounds)
 456 {
 457    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 458
 459    cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
 460    cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
 461
 462    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
 463 }
 464
 465 void anv_CmdSetStencilCompareMask(
 466     VkCommandBuffer                             commandBuffer,
 467     VkStencilFaceFlags                          faceMask,
 468     uint32_t                                    compareMask)
 469 {
 470    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 471
 472    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 473       cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
 474    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 475       cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
 476
 477    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
 478 }
 479
 480 void anv_CmdSetStencilWriteMask(
 481     VkCommandBuffer                             commandBuffer,
 482     VkStencilFaceFlags                          faceMask,
 483     uint32_t                                    writeMask)
 484 {
 485    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 486
 487    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 488       cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
 489    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 490       cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
 491
 492    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
 493 }
 494
 495 void anv_CmdSetStencilReference(
 496     VkCommandBuffer                             commandBuffer,
 497     VkStencilFaceFlags                          faceMask,
 498     uint32_t                                    reference)
 499 {
 500    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 501
 502    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 503       cmd_buffer->state.dynamic.stencil_reference.front = reference;
 504    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 505       cmd_buffer->state.dynamic.stencil_reference.back = reference;
 506
 507    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
 508 }
 509
 510 void anv_CmdBindDescriptorSets(
 511     VkCommandBuffer                             commandBuffer,
 512     VkPipelineBindPoint                         pipelineBindPoint,
 513     VkPipelineLayout                            _layout,
 514     uint32_t                                    firstSet,
 515     uint32_t                                    descriptorSetCount,
 516     const VkDescriptorSet*                      pDescriptorSets,
 517     uint32_t                                    dynamicOffsetCount,
 518     const uint32_t*                             pDynamicOffsets)
 519 {
 520    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 521    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 522    struct anv_descriptor_set_layout *set_layout;
 523
 524    assert(firstSet + descriptorSetCount < MAX_SETS);
 525
 526    uint32_t dynamic_slot = 0;
 527    for (uint32_t i = 0; i < descriptorSetCount; i++) {
 528       ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
 529       set_layout = layout->set[firstSet + i].layout;
 530
 531       cmd_buffer->state.descriptors[firstSet + i] = set;
 532
 533       if (set_layout->dynamic_offset_count > 0) {
 534          uint32_t dynamic_offset_start =
 535             layout->set[firstSet + i].dynamic_offset_start;
 536
 537          /* Assert that everything is in range */
 538          assert(dynamic_offset_start + set_layout->dynamic_offset_count <=
 539                 ARRAY_SIZE(cmd_buffer->state.dynamic_offsets));
 540          assert(dynamic_slot + set_layout->dynamic_offset_count <=
 541                 dynamicOffsetCount);
 542
 543          typed_memcpy(&cmd_buffer->state.dynamic_offsets[dynamic_offset_start],
 544                       &pDynamicOffsets[dynamic_slot],
 545                       set_layout->dynamic_offset_count);
 546
 547          dynamic_slot += set_layout->dynamic_offset_count;
 548       }
 549
 550       cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 551    }
 552 }
 553
 554 void anv_CmdBindVertexBuffers(
 555     VkCommandBuffer                             commandBuffer,
 556     uint32_t                                    firstBinding,
 557     uint32_t                                    bindingCount,
 558     const VkBuffer*                             pBuffers,
 559     const VkDeviceSize*                         pOffsets)
 560 {
 561    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 562    struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 563
 564    /* We have to defer setting up vertex buffer since we need the buffer
 565     * stride from the pipeline. */
 566
 567    assert(firstBinding + bindingCount < MAX_VBS);
 568    for (uint32_t i = 0; i < bindingCount; i++) {
 569       vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
 570       vb[firstBinding + i].offset = pOffsets[i];
 571       cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i);
 572    }
 573 }
 574
 575 enum isl_format
 576 anv_isl_format_for_descriptor_type(VkDescriptorType type)
 577 {
 578    switch (type) {
 579    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 580    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 581       return ISL_FORMAT_R32G32B32A32_FLOAT;
 582
 583    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 584    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 585       return ISL_FORMAT_RAW;
 586
 587    default:
 588       unreachable("Invalid descriptor type");
 589    }
 590 }
 591
 592 struct anv_state
 593 anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
 594                             const void *data, uint32_t size, uint32_t alignment)
 595 {
 596    struct anv_state state;
 597
 598    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment);
 599    memcpy(state.map, data, size);
 600
 601    anv_state_flush(cmd_buffer->device, state);
 602
 603    VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size));
 604
 605    return state;
 606 }
 607
 608 struct anv_state
 609 anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
 610                              uint32_t *a, uint32_t *b,
 611                              uint32_t dwords, uint32_t alignment)
 612 {
 613    struct anv_state state;
 614    uint32_t *p;
 615
 616    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 617                                               dwords * 4, alignment);
 618    p = state.map;
 619    for (uint32_t i = 0; i < dwords; i++)
 620       p[i] = a[i] | b[i];
 621
 622    anv_state_flush(cmd_buffer->device, state);
 623
 624    VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
 625
 626    return state;
 627 }
 628
 629 struct anv_state
 630 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
 631                               gl_shader_stage stage)
 632 {
 633    /* If we don't have this stage, bail. */
 634    if (!anv_pipeline_has_stage(cmd_buffer->state.pipeline, stage))
 635       return (struct anv_state) { .offset = 0 };
 636
 637    struct anv_push_constants *data =
 638       cmd_buffer->state.push_constants[stage];
 639    const struct brw_stage_prog_data *prog_data =
 640       cmd_buffer->state.pipeline->shaders[stage]->prog_data;
 641
 642    /* If we don't actually have any push constants, bail. */
 643    if (data == NULL || prog_data == NULL || prog_data->nr_params == 0)
 644       return (struct anv_state) { .offset = 0 };
 645
 646    struct anv_state state =
 647       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 648                                          prog_data->nr_params * sizeof(float),
 649                                          32 /* bottom 5 bits MBZ */);
 650
 651    /* Walk through the param array and fill the buffer with data */
 652    uint32_t *u32_map = state.map;
 653    for (unsigned i = 0; i < prog_data->nr_params; i++) {
 654       uint32_t offset = (uintptr_t)prog_data->param[i];
 655       u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
 656    }
 657
 658    anv_state_flush(cmd_buffer->device, state);
 659
 660    return state;
 661 }
 662
 663 struct anv_state
 664 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
 665 {
 666    struct anv_push_constants *data =
 667       cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
 668    struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
 669    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
 670    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
 671
 672    /* If we don't actually have any push constants, bail. */
 673    if (cs_prog_data->push.total.size == 0)
 674       return (struct anv_state) { .offset = 0 };
 675
 676    const unsigned push_constant_alignment =
 677       cmd_buffer->device->info.gen < 8 ? 32 : 64;
 678    const unsigned aligned_total_push_constants_size =
 679       ALIGN(cs_prog_data->push.total.size, push_constant_alignment);
 680    struct anv_state state =
 681       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 682                                          aligned_total_push_constants_size,
 683                                          push_constant_alignment);
 684
 685    /* Walk through the param array and fill the buffer with data */
 686    uint32_t *u32_map = state.map;
 687
 688    if (cs_prog_data->push.cross_thread.size > 0) {
 689       assert(cs_prog_data->thread_local_id_index < 0 ||
 690              cs_prog_data->thread_local_id_index >=
 691                 cs_prog_data->push.cross_thread.dwords);
 692       for (unsigned i = 0;
 693            i < cs_prog_data->push.cross_thread.dwords;
 694            i++) {
 695          uint32_t offset = (uintptr_t)prog_data->param[i];
 696          u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
 697       }
 698    }
 699
 700    if (cs_prog_data->push.per_thread.size > 0) {
 701       for (unsigned t = 0; t < cs_prog_data->threads; t++) {
 702          unsigned dst =
 703             8 * (cs_prog_data->push.per_thread.regs * t +
 704                  cs_prog_data->push.cross_thread.regs);
 705          unsigned src = cs_prog_data->push.cross_thread.dwords;
 706          for ( ; src < prog_data->nr_params; src++, dst++) {
 707             if (src != cs_prog_data->thread_local_id_index) {
 708                uint32_t offset = (uintptr_t)prog_data->param[src];
 709                u32_map[dst] = *(uint32_t *)((uint8_t *)data + offset);
 710             } else {
 711                u32_map[dst] = t * cs_prog_data->simd_size;
 712             }
 713          }
 714       }
 715    }
 716
 717    anv_state_flush(cmd_buffer->device, state);
 718
 719    return state;
 720 }
 721
 722 void anv_CmdPushConstants(
 723     VkCommandBuffer                             commandBuffer,
 724     VkPipelineLayout                            layout,
 725     VkShaderStageFlags                          stageFlags,
 726     uint32_t                                    offset,
 727     uint32_t                                    size,
 728     const void*                                 pValues)
 729 {
 730    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 731
 732    anv_foreach_stage(stage, stageFlags) {
 733       VkResult result =
 734          anv_cmd_buffer_ensure_push_constant_field(cmd_buffer,
 735                                                    stage, client_data);
 736       if (result != VK_SUCCESS)
 737          return;
 738
 739       memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
 740              pValues, size);
 741    }
 742
 743    cmd_buffer->state.push_constants_dirty |= stageFlags;
 744 }
 745
 746 VkResult anv_CreateCommandPool(
 747     VkDevice                                    _device,
 748     const VkCommandPoolCreateInfo*              pCreateInfo,
 749     const VkAllocationCallbacks*                pAllocator,
 750     VkCommandPool*                              pCmdPool)
 751 {
 752    ANV_FROM_HANDLE(anv_device, device, _device);
 753    struct anv_cmd_pool *pool;
 754
 755    pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
 756                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 757    if (pool == NULL)
 758       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 759
 760    if (pAllocator)
 761       pool->alloc = *pAllocator;
 762    else
 763       pool->alloc = device->alloc;
 764
 765    list_inithead(&pool->cmd_buffers);
 766
 767    *pCmdPool = anv_cmd_pool_to_handle(pool);
 768
 769    return VK_SUCCESS;
 770 }
 771
 772 void anv_DestroyCommandPool(
 773     VkDevice                                    _device,
 774     VkCommandPool                               commandPool,
 775     const VkAllocationCallbacks*                pAllocator)
 776 {
 777    ANV_FROM_HANDLE(anv_device, device, _device);
 778    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
 779
 780    if (!pool)
 781       return;
 782
 783    list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
 784                             &pool->cmd_buffers, pool_link) {
 785       anv_cmd_buffer_destroy(cmd_buffer);
 786    }
 787
 788    vk_free2(&device->alloc, pAllocator, pool);
 789 }
 790
 791 VkResult anv_ResetCommandPool(
 792     VkDevice                                    device,
 793     VkCommandPool                               commandPool,
 794     VkCommandPoolResetFlags                     flags)
 795 {
 796    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
 797
 798    list_for_each_entry(struct anv_cmd_buffer, cmd_buffer,
 799                        &pool->cmd_buffers, pool_link) {
 800       anv_cmd_buffer_reset(cmd_buffer);
 801    }
 802
 803    return VK_SUCCESS;
 804 }
 805
 806 void anv_TrimCommandPoolKHR(
 807     VkDevice                                    device,
 808     VkCommandPool                               commandPool,
 809     VkCommandPoolTrimFlagsKHR                   flags)
 810 {
 811    /* Nothing for us to do here.  Our pools stay pretty tidy. */
 812 }
 813
 814 /**
 815  * Return NULL if the current subpass has no depthstencil attachment.
 816  */
 817 const struct anv_image_view *
 818 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
 819 {
 820    const struct anv_subpass *subpass = cmd_buffer->state.subpass;
 821    const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 822
 823    if (subpass->depth_stencil_attachment.attachment == VK_ATTACHMENT_UNUSED)
 824       return NULL;
 825
 826    const struct anv_image_view *iview =
 827       fb->attachments[subpass->depth_stencil_attachment.attachment];
 828
 829    assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
 830                                 VK_IMAGE_ASPECT_STENCIL_BIT));
 831
 832    return iview;
 833 }
 834
 835 void anv_CmdPushDescriptorSetKHR(
 836     VkCommandBuffer commandBuffer,
 837     VkPipelineBindPoint pipelineBindPoint,
 838     VkPipelineLayout _layout,
 839     uint32_t _set,
 840     uint32_t descriptorWriteCount,
 841     const VkWriteDescriptorSet* pDescriptorWrites)
 842 {
 843    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 844    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 845
 846    assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS ||
 847           pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE);
 848    assert(_set < MAX_SETS);
 849
 850    const struct anv_descriptor_set_layout *set_layout =
 851       layout->set[_set].layout;
 852    struct anv_descriptor_set *set = &cmd_buffer->state.push_descriptor.set;
 853
 854    set->layout = set_layout;
 855    set->size = anv_descriptor_set_layout_size(set_layout);
 856    set->buffer_count = set_layout->buffer_count;
 857    set->buffer_views = cmd_buffer->state.push_descriptor.buffer_views;
 858
 859    /* Go through the user supplied descriptors. */
 860    for (uint32_t i = 0; i < descriptorWriteCount; i++) {
 861       const VkWriteDescriptorSet *write = &pDescriptorWrites[i];
 862
 863       switch (write->descriptorType) {
 864       case VK_DESCRIPTOR_TYPE_SAMPLER:
 865       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 866       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 867       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
 868       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 869          for (uint32_t j = 0; j < write->descriptorCount; j++) {
 870             anv_descriptor_set_write_image_view(set, &cmd_buffer->device->info,
 871                                                 write->pImageInfo + j,
 872                                                 write->descriptorType,
 873                                                 write->dstBinding,
 874                                                 write->dstArrayElement + j);
 875          }
 876          break;
 877
 878       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
 879       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
 880          for (uint32_t j = 0; j < write->descriptorCount; j++) {
 881             ANV_FROM_HANDLE(anv_buffer_view, bview,
 882                             write->pTexelBufferView[j]);
 883
 884             anv_descriptor_set_write_buffer_view(set,
 885                                                  write->descriptorType,
 886                                                  bview,
 887                                                  write->dstBinding,
 888                                                  write->dstArrayElement + j);
 889          }
 890          break;
 891
 892       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 893       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 894       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 895       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 896          for (uint32_t j = 0; j < write->descriptorCount; j++) {
 897             assert(write->pBufferInfo[j].buffer);
 898             ANV_FROM_HANDLE(anv_buffer, buffer, write->pBufferInfo[j].buffer);
 899             assert(buffer);
 900
 901             anv_descriptor_set_write_buffer(set,
 902                                             cmd_buffer->device,
 903                                             &cmd_buffer->surface_state_stream,
 904                                             write->descriptorType,
 905                                             buffer,
 906                                             write->dstBinding,
 907                                             write->dstArrayElement + j,
 908                                             write->pBufferInfo[j].offset,
 909                                             write->pBufferInfo[j].range);
 910          }
 911          break;
 912
 913       default:
 914          break;
 915       }
 916    }
 917
 918    cmd_buffer->state.descriptors[_set] = set;
 919    cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 920 }
 921
 922 void anv_CmdPushDescriptorSetWithTemplateKHR(
 923     VkCommandBuffer                             commandBuffer,
 924     VkDescriptorUpdateTemplateKHR               descriptorUpdateTemplate,
 925     VkPipelineLayout                            _layout,
 926     uint32_t                                    _set,
 927     const void*                                 pData)
 928 {
 929    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 930    ANV_FROM_HANDLE(anv_descriptor_update_template, template,
 931                    descriptorUpdateTemplate);
 932    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 933
 934    assert(_set < MAX_PUSH_DESCRIPTORS);
 935
 936    const struct anv_descriptor_set_layout *set_layout =
 937       layout->set[_set].layout;
 938    struct anv_descriptor_set *set = &cmd_buffer->state.push_descriptor.set;
 939
 940    set->layout = set_layout;
 941    set->size = anv_descriptor_set_layout_size(set_layout);
 942    set->buffer_count = set_layout->buffer_count;
 943    set->buffer_views = cmd_buffer->state.push_descriptor.buffer_views;
 944
 945    anv_descriptor_set_write_template(set,
 946                                      cmd_buffer->device,
 947                                      &cmd_buffer->surface_state_stream,
 948                                      template,
 949                                      pData);
 950
 951    cmd_buffer->state.descriptors[_set] = set;
 952    cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 953 }