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    case 10:
 339       return gen10_cmd_buffer_emit_state_base_address(cmd_buffer);
 340    default:
 341       unreachable("unsupported gen\n");
 342    }
 343 }
 344
 345 void anv_CmdBindPipeline(
 346     VkCommandBuffer                             commandBuffer,
 347     VkPipelineBindPoint                         pipelineBindPoint,
 348     VkPipeline                                  _pipeline)
 349 {
 350    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 351    ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 352
 353    switch (pipelineBindPoint) {
 354    case VK_PIPELINE_BIND_POINT_COMPUTE:
 355       cmd_buffer->state.compute_pipeline = pipeline;
 356       cmd_buffer->state.compute_dirty |= ANV_CMD_DIRTY_PIPELINE;
 357       cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 358       cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 359       break;
 360
 361    case VK_PIPELINE_BIND_POINT_GRAPHICS:
 362       cmd_buffer->state.pipeline = pipeline;
 363       cmd_buffer->state.vb_dirty |= pipeline->vb_used;
 364       cmd_buffer->state.dirty |= ANV_CMD_DIRTY_PIPELINE;
 365       cmd_buffer->state.push_constants_dirty |= pipeline->active_stages;
 366       cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;
 367
 368       /* Apply the dynamic state from the pipeline */
 369       cmd_buffer->state.dirty |= pipeline->dynamic_state_mask;
 370       anv_dynamic_state_copy(&cmd_buffer->state.dynamic,
 371                              &pipeline->dynamic_state,
 372                              pipeline->dynamic_state_mask);
 373       break;
 374
 375    default:
 376       assert(!"invalid bind point");
 377       break;
 378    }
 379 }
 380
 381 void anv_CmdSetViewport(
 382     VkCommandBuffer                             commandBuffer,
 383     uint32_t                                    firstViewport,
 384     uint32_t                                    viewportCount,
 385     const VkViewport*                           pViewports)
 386 {
 387    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 388
 389    const uint32_t total_count = firstViewport + viewportCount;
 390    if (cmd_buffer->state.dynamic.viewport.count < total_count)
 391       cmd_buffer->state.dynamic.viewport.count = total_count;
 392
 393    memcpy(cmd_buffer->state.dynamic.viewport.viewports + firstViewport,
 394           pViewports, viewportCount * sizeof(*pViewports));
 395
 396    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT;
 397 }
 398
 399 void anv_CmdSetScissor(
 400     VkCommandBuffer                             commandBuffer,
 401     uint32_t                                    firstScissor,
 402     uint32_t                                    scissorCount,
 403     const VkRect2D*                             pScissors)
 404 {
 405    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 406
 407    const uint32_t total_count = firstScissor + scissorCount;
 408    if (cmd_buffer->state.dynamic.scissor.count < total_count)
 409       cmd_buffer->state.dynamic.scissor.count = total_count;
 410
 411    memcpy(cmd_buffer->state.dynamic.scissor.scissors + firstScissor,
 412           pScissors, scissorCount * sizeof(*pScissors));
 413
 414    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR;
 415 }
 416
 417 void anv_CmdSetLineWidth(
 418     VkCommandBuffer                             commandBuffer,
 419     float                                       lineWidth)
 420 {
 421    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 422
 423    cmd_buffer->state.dynamic.line_width = lineWidth;
 424    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
 425 }
 426
 427 void anv_CmdSetDepthBias(
 428     VkCommandBuffer                             commandBuffer,
 429     float                                       depthBiasConstantFactor,
 430     float                                       depthBiasClamp,
 431     float                                       depthBiasSlopeFactor)
 432 {
 433    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 434
 435    cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
 436    cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
 437    cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
 438
 439    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
 440 }
 441
 442 void anv_CmdSetBlendConstants(
 443     VkCommandBuffer                             commandBuffer,
 444     const float                                 blendConstants[4])
 445 {
 446    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 447
 448    memcpy(cmd_buffer->state.dynamic.blend_constants,
 449           blendConstants, sizeof(float) * 4);
 450
 451    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
 452 }
 453
 454 void anv_CmdSetDepthBounds(
 455     VkCommandBuffer                             commandBuffer,
 456     float                                       minDepthBounds,
 457     float                                       maxDepthBounds)
 458 {
 459    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 460
 461    cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
 462    cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
 463
 464    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
 465 }
 466
 467 void anv_CmdSetStencilCompareMask(
 468     VkCommandBuffer                             commandBuffer,
 469     VkStencilFaceFlags                          faceMask,
 470     uint32_t                                    compareMask)
 471 {
 472    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 473
 474    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 475       cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
 476    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 477       cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
 478
 479    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
 480 }
 481
 482 void anv_CmdSetStencilWriteMask(
 483     VkCommandBuffer                             commandBuffer,
 484     VkStencilFaceFlags                          faceMask,
 485     uint32_t                                    writeMask)
 486 {
 487    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 488
 489    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 490       cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
 491    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 492       cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
 493
 494    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
 495 }
 496
 497 void anv_CmdSetStencilReference(
 498     VkCommandBuffer                             commandBuffer,
 499     VkStencilFaceFlags                          faceMask,
 500     uint32_t                                    reference)
 501 {
 502    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 503
 504    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 505       cmd_buffer->state.dynamic.stencil_reference.front = reference;
 506    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 507       cmd_buffer->state.dynamic.stencil_reference.back = reference;
 508
 509    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
 510 }
 511
 512 void anv_CmdBindDescriptorSets(
 513     VkCommandBuffer                             commandBuffer,
 514     VkPipelineBindPoint                         pipelineBindPoint,
 515     VkPipelineLayout                            _layout,
 516     uint32_t                                    firstSet,
 517     uint32_t                                    descriptorSetCount,
 518     const VkDescriptorSet*                      pDescriptorSets,
 519     uint32_t                                    dynamicOffsetCount,
 520     const uint32_t*                             pDynamicOffsets)
 521 {
 522    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 523    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 524    struct anv_descriptor_set_layout *set_layout;
 525
 526    assert(firstSet + descriptorSetCount < MAX_SETS);
 527
 528    uint32_t dynamic_slot = 0;
 529    for (uint32_t i = 0; i < descriptorSetCount; i++) {
 530       ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
 531       set_layout = layout->set[firstSet + i].layout;
 532
 533       cmd_buffer->state.descriptors[firstSet + i] = set;
 534
 535       if (set_layout->dynamic_offset_count > 0) {
 536          uint32_t dynamic_offset_start =
 537             layout->set[firstSet + i].dynamic_offset_start;
 538
 539          /* Assert that everything is in range */
 540          assert(dynamic_offset_start + set_layout->dynamic_offset_count <=
 541                 ARRAY_SIZE(cmd_buffer->state.dynamic_offsets));
 542          assert(dynamic_slot + set_layout->dynamic_offset_count <=
 543                 dynamicOffsetCount);
 544
 545          typed_memcpy(&cmd_buffer->state.dynamic_offsets[dynamic_offset_start],
 546                       &pDynamicOffsets[dynamic_slot],
 547                       set_layout->dynamic_offset_count);
 548
 549          dynamic_slot += set_layout->dynamic_offset_count;
 550       }
 551
 552       cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 553    }
 554 }
 555
 556 void anv_CmdBindVertexBuffers(
 557     VkCommandBuffer                             commandBuffer,
 558     uint32_t                                    firstBinding,
 559     uint32_t                                    bindingCount,
 560     const VkBuffer*                             pBuffers,
 561     const VkDeviceSize*                         pOffsets)
 562 {
 563    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 564    struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 565
 566    /* We have to defer setting up vertex buffer since we need the buffer
 567     * stride from the pipeline. */
 568
 569    assert(firstBinding + bindingCount <= MAX_VBS);
 570    for (uint32_t i = 0; i < bindingCount; i++) {
 571       vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
 572       vb[firstBinding + i].offset = pOffsets[i];
 573       cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i);
 574    }
 575 }
 576
 577 enum isl_format
 578 anv_isl_format_for_descriptor_type(VkDescriptorType type)
 579 {
 580    switch (type) {
 581    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 582    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 583       return ISL_FORMAT_R32G32B32A32_FLOAT;
 584
 585    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 586    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 587       return ISL_FORMAT_RAW;
 588
 589    default:
 590       unreachable("Invalid descriptor type");
 591    }
 592 }
 593
 594 struct anv_state
 595 anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
 596                             const void *data, uint32_t size, uint32_t alignment)
 597 {
 598    struct anv_state state;
 599
 600    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment);
 601    memcpy(state.map, data, size);
 602
 603    anv_state_flush(cmd_buffer->device, state);
 604
 605    VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size));
 606
 607    return state;
 608 }
 609
 610 struct anv_state
 611 anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
 612                              uint32_t *a, uint32_t *b,
 613                              uint32_t dwords, uint32_t alignment)
 614 {
 615    struct anv_state state;
 616    uint32_t *p;
 617
 618    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 619                                               dwords * 4, alignment);
 620    p = state.map;
 621    for (uint32_t i = 0; i < dwords; i++)
 622       p[i] = a[i] | b[i];
 623
 624    anv_state_flush(cmd_buffer->device, state);
 625
 626    VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
 627
 628    return state;
 629 }
 630
 631 struct anv_state
 632 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
 633                               gl_shader_stage stage)
 634 {
 635    /* If we don't have this stage, bail. */
 636    if (!anv_pipeline_has_stage(cmd_buffer->state.pipeline, stage))
 637       return (struct anv_state) { .offset = 0 };
 638
 639    struct anv_push_constants *data =
 640       cmd_buffer->state.push_constants[stage];
 641    const struct brw_stage_prog_data *prog_data =
 642       cmd_buffer->state.pipeline->shaders[stage]->prog_data;
 643
 644    /* If we don't actually have any push constants, bail. */
 645    if (data == NULL || prog_data == NULL || prog_data->nr_params == 0)
 646       return (struct anv_state) { .offset = 0 };
 647
 648    struct anv_state state =
 649       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 650                                          prog_data->nr_params * sizeof(float),
 651                                          32 /* bottom 5 bits MBZ */);
 652
 653    /* Walk through the param array and fill the buffer with data */
 654    uint32_t *u32_map = state.map;
 655    for (unsigned i = 0; i < prog_data->nr_params; i++) {
 656       uint32_t offset = (uintptr_t)prog_data->param[i];
 657       u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
 658    }
 659
 660    anv_state_flush(cmd_buffer->device, state);
 661
 662    return state;
 663 }
 664
 665 struct anv_state
 666 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
 667 {
 668    struct anv_push_constants *data =
 669       cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
 670    struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
 671    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
 672    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
 673
 674    /* If we don't actually have any push constants, bail. */
 675    if (cs_prog_data->push.total.size == 0)
 676       return (struct anv_state) { .offset = 0 };
 677
 678    const unsigned push_constant_alignment =
 679       cmd_buffer->device->info.gen < 8 ? 32 : 64;
 680    const unsigned aligned_total_push_constants_size =
 681       ALIGN(cs_prog_data->push.total.size, push_constant_alignment);
 682    struct anv_state state =
 683       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 684                                          aligned_total_push_constants_size,
 685                                          push_constant_alignment);
 686
 687    /* Walk through the param array and fill the buffer with data */
 688    uint32_t *u32_map = state.map;
 689
 690    if (cs_prog_data->push.cross_thread.size > 0) {
 691       assert(cs_prog_data->thread_local_id_index < 0 ||
 692              cs_prog_data->thread_local_id_index >=
 693                 cs_prog_data->push.cross_thread.dwords);
 694       for (unsigned i = 0;
 695            i < cs_prog_data->push.cross_thread.dwords;
 696            i++) {
 697          uint32_t offset = (uintptr_t)prog_data->param[i];
 698          u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
 699       }
 700    }
 701
 702    if (cs_prog_data->push.per_thread.size > 0) {
 703       for (unsigned t = 0; t < cs_prog_data->threads; t++) {
 704          unsigned dst =
 705             8 * (cs_prog_data->push.per_thread.regs * t +
 706                  cs_prog_data->push.cross_thread.regs);
 707          unsigned src = cs_prog_data->push.cross_thread.dwords;
 708          for ( ; src < prog_data->nr_params; src++, dst++) {
 709             if (src != cs_prog_data->thread_local_id_index) {
 710                uint32_t offset = (uintptr_t)prog_data->param[src];
 711                u32_map[dst] = *(uint32_t *)((uint8_t *)data + offset);
 712             } else {
 713                u32_map[dst] = t * cs_prog_data->simd_size;
 714             }
 715          }
 716       }
 717    }
 718
 719    anv_state_flush(cmd_buffer->device, state);
 720
 721    return state;
 722 }
 723
 724 void anv_CmdPushConstants(
 725     VkCommandBuffer                             commandBuffer,
 726     VkPipelineLayout                            layout,
 727     VkShaderStageFlags                          stageFlags,
 728     uint32_t                                    offset,
 729     uint32_t                                    size,
 730     const void*                                 pValues)
 731 {
 732    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 733
 734    anv_foreach_stage(stage, stageFlags) {
 735       VkResult result =
 736          anv_cmd_buffer_ensure_push_constant_field(cmd_buffer,
 737                                                    stage, client_data);
 738       if (result != VK_SUCCESS)
 739          return;
 740
 741       memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
 742              pValues, size);
 743    }
 744
 745    cmd_buffer->state.push_constants_dirty |= stageFlags;
 746 }
 747
 748 VkResult anv_CreateCommandPool(
 749     VkDevice                                    _device,
 750     const VkCommandPoolCreateInfo*              pCreateInfo,
 751     const VkAllocationCallbacks*                pAllocator,
 752     VkCommandPool*                              pCmdPool)
 753 {
 754    ANV_FROM_HANDLE(anv_device, device, _device);
 755    struct anv_cmd_pool *pool;
 756
 757    pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
 758                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 759    if (pool == NULL)
 760       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 761
 762    if (pAllocator)
 763       pool->alloc = *pAllocator;
 764    else
 765       pool->alloc = device->alloc;
 766
 767    list_inithead(&pool->cmd_buffers);
 768
 769    *pCmdPool = anv_cmd_pool_to_handle(pool);
 770
 771    return VK_SUCCESS;
 772 }
 773
 774 void anv_DestroyCommandPool(
 775     VkDevice                                    _device,
 776     VkCommandPool                               commandPool,
 777     const VkAllocationCallbacks*                pAllocator)
 778 {
 779    ANV_FROM_HANDLE(anv_device, device, _device);
 780    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
 781
 782    if (!pool)
 783       return;
 784
 785    list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
 786                             &pool->cmd_buffers, pool_link) {
 787       anv_cmd_buffer_destroy(cmd_buffer);
 788    }
 789
 790    vk_free2(&device->alloc, pAllocator, pool);
 791 }
 792
 793 VkResult anv_ResetCommandPool(
 794     VkDevice                                    device,
 795     VkCommandPool                               commandPool,
 796     VkCommandPoolResetFlags                     flags)
 797 {
 798    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
 799
 800    list_for_each_entry(struct anv_cmd_buffer, cmd_buffer,
 801                        &pool->cmd_buffers, pool_link) {
 802       anv_cmd_buffer_reset(cmd_buffer);
 803    }
 804
 805    return VK_SUCCESS;
 806 }
 807
 808 void anv_TrimCommandPoolKHR(
 809     VkDevice                                    device,
 810     VkCommandPool                               commandPool,
 811     VkCommandPoolTrimFlagsKHR                   flags)
 812 {
 813    /* Nothing for us to do here.  Our pools stay pretty tidy. */
 814 }
 815
 816 /**
 817  * Return NULL if the current subpass has no depthstencil attachment.
 818  */
 819 const struct anv_image_view *
 820 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
 821 {
 822    const struct anv_subpass *subpass = cmd_buffer->state.subpass;
 823    const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 824
 825    if (subpass->depth_stencil_attachment.attachment == VK_ATTACHMENT_UNUSED)
 826       return NULL;
 827
 828    const struct anv_image_view *iview =
 829       fb->attachments[subpass->depth_stencil_attachment.attachment];
 830
 831    assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
 832                                 VK_IMAGE_ASPECT_STENCIL_BIT));
 833
 834    return iview;
 835 }
 836
 837 void anv_CmdPushDescriptorSetKHR(
 838     VkCommandBuffer commandBuffer,
 839     VkPipelineBindPoint pipelineBindPoint,
 840     VkPipelineLayout _layout,
 841     uint32_t _set,
 842     uint32_t descriptorWriteCount,
 843     const VkWriteDescriptorSet* pDescriptorWrites)
 844 {
 845    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 846    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 847
 848    assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS ||
 849           pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE);
 850    assert(_set < MAX_SETS);
 851
 852    const struct anv_descriptor_set_layout *set_layout =
 853       layout->set[_set].layout;
 854    struct anv_descriptor_set *set = &cmd_buffer->state.push_descriptor.set;
 855
 856    set->layout = set_layout;
 857    set->size = anv_descriptor_set_layout_size(set_layout);
 858    set->buffer_count = set_layout->buffer_count;
 859    set->buffer_views = cmd_buffer->state.push_descriptor.buffer_views;
 860
 861    /* Go through the user supplied descriptors. */
 862    for (uint32_t i = 0; i < descriptorWriteCount; i++) {
 863       const VkWriteDescriptorSet *write = &pDescriptorWrites[i];
 864
 865       switch (write->descriptorType) {
 866       case VK_DESCRIPTOR_TYPE_SAMPLER:
 867       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 868       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 869       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
 870       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 871          for (uint32_t j = 0; j < write->descriptorCount; j++) {
 872             anv_descriptor_set_write_image_view(set, &cmd_buffer->device->info,
 873                                                 write->pImageInfo + j,
 874                                                 write->descriptorType,
 875                                                 write->dstBinding,
 876                                                 write->dstArrayElement + j);
 877          }
 878          break;
 879
 880       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
 881       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
 882          for (uint32_t j = 0; j < write->descriptorCount; j++) {
 883             ANV_FROM_HANDLE(anv_buffer_view, bview,
 884                             write->pTexelBufferView[j]);
 885
 886             anv_descriptor_set_write_buffer_view(set,
 887                                                  write->descriptorType,
 888                                                  bview,
 889                                                  write->dstBinding,
 890                                                  write->dstArrayElement + j);
 891          }
 892          break;
 893
 894       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 895       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 896       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 897       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 898          for (uint32_t j = 0; j < write->descriptorCount; j++) {
 899             assert(write->pBufferInfo[j].buffer);
 900             ANV_FROM_HANDLE(anv_buffer, buffer, write->pBufferInfo[j].buffer);
 901             assert(buffer);
 902
 903             anv_descriptor_set_write_buffer(set,
 904                                             cmd_buffer->device,
 905                                             &cmd_buffer->surface_state_stream,
 906                                             write->descriptorType,
 907                                             buffer,
 908                                             write->dstBinding,
 909                                             write->dstArrayElement + j,
 910                                             write->pBufferInfo[j].offset,
 911                                             write->pBufferInfo[j].range);
 912          }
 913          break;
 914
 915       default:
 916          break;
 917       }
 918    }
 919
 920    cmd_buffer->state.descriptors[_set] = set;
 921    cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 922 }
 923
 924 void anv_CmdPushDescriptorSetWithTemplateKHR(
 925     VkCommandBuffer                             commandBuffer,
 926     VkDescriptorUpdateTemplateKHR               descriptorUpdateTemplate,
 927     VkPipelineLayout                            _layout,
 928     uint32_t                                    _set,
 929     const void*                                 pData)
 930 {
 931    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 932    ANV_FROM_HANDLE(anv_descriptor_update_template, template,
 933                    descriptorUpdateTemplate);
 934    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 935
 936    assert(_set < MAX_PUSH_DESCRIPTORS);
 937
 938    const struct anv_descriptor_set_layout *set_layout =
 939       layout->set[_set].layout;
 940    struct anv_descriptor_set *set = &cmd_buffer->state.push_descriptor.set;
 941
 942    set->layout = set_layout;
 943    set->size = anv_descriptor_set_layout_size(set_layout);
 944    set->buffer_count = set_layout->buffer_count;
 945    set->buffer_views = cmd_buffer->state.push_descriptor.buffer_views;
 946
 947    anv_descriptor_set_write_template(set,
 948                                      cmd_buffer->device,
 949                                      &cmd_buffer->surface_state_stream,
 950                                      template,
 951                                      pData);
 952
 953    cmd_buffer->state.descriptors[_set] = set;
 954    cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 955 }