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    memset(&state->descriptors, 0, sizeof(state->descriptors));
 121    memset(&state->push_constants, 0, sizeof(state->push_constants));
 122    memset(state->binding_tables, 0, sizeof(state->binding_tables));
 123    memset(state->samplers, 0, sizeof(state->samplers));
 124
 125    /* 0 isn't a valid config.  This ensures that we always configure L3$. */
 126    cmd_buffer->state.current_l3_config = 0;
 127
 128    state->dirty = 0;
 129    state->vb_dirty = 0;
 130    state->pending_pipe_bits = 0;
 131    state->descriptors_dirty = 0;
 132    state->push_constants_dirty = 0;
 133    state->pipeline = NULL;
 134    state->push_constant_stages = 0;
 135    state->restart_index = UINT32_MAX;
 136    state->dynamic = default_dynamic_state;
 137    state->need_query_wa = true;
 138
 139    if (state->attachments != NULL) {
 140       anv_free(&cmd_buffer->pool->alloc, state->attachments);
 141       state->attachments = NULL;
 142    }
 143
 144    state->gen7.index_buffer = NULL;
 145 }
 146
 147 /**
 148  * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
 149  */
 150 void
 151 anv_cmd_state_setup_attachments(struct anv_cmd_buffer *cmd_buffer,
 152                                 const VkRenderPassBeginInfo *info)
 153 {
 154    struct anv_cmd_state *state = &cmd_buffer->state;
 155    ANV_FROM_HANDLE(anv_render_pass, pass, info->renderPass);
 156
 157    anv_free(&cmd_buffer->pool->alloc, state->attachments);
 158
 159    if (pass->attachment_count == 0) {
 160       state->attachments = NULL;
 161       return;
 162    }
 163
 164    state->attachments = anv_alloc(&cmd_buffer->pool->alloc,
 165                                   pass->attachment_count *
 166                                        sizeof(state->attachments[0]),
 167                                   8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 168    if (state->attachments == NULL) {
 169       /* FIXME: Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
 170       abort();
 171    }
 172
 173    for (uint32_t i = 0; i < pass->attachment_count; ++i) {
 174       struct anv_render_pass_attachment *att = &pass->attachments[i];
 175       VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
 176       VkImageAspectFlags clear_aspects = 0;
 177
 178       if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
 179          /* color attachment */
 180          if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 181             clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
 182          }
 183       } else {
 184          /* depthstencil attachment */
 185          if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
 186              att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 187             clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
 188          }
 189          if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
 190              att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 191             clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
 192          }
 193       }
 194
 195       state->attachments[i].pending_clear_aspects = clear_aspects;
 196       if (clear_aspects) {
 197          assert(info->clearValueCount > i);
 198          state->attachments[i].clear_value = info->pClearValues[i];
 199       }
 200    }
 201 }
 202
 203 static VkResult
 204 anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer,
 205                                           gl_shader_stage stage, uint32_t size)
 206 {
 207    struct anv_push_constants **ptr = &cmd_buffer->state.push_constants[stage];
 208
 209    if (*ptr == NULL) {
 210       *ptr = anv_alloc(&cmd_buffer->pool->alloc, size, 8,
 211                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 212       if (*ptr == NULL)
 213          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 214    } else if ((*ptr)->size < size) {
 215       *ptr = anv_realloc(&cmd_buffer->pool->alloc, *ptr, size, 8,
 216                          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 217       if (*ptr == NULL)
 218          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 219    }
 220    (*ptr)->size = size;
 221
 222    return VK_SUCCESS;
 223 }
 224
 225 #define anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, field) \
 226    anv_cmd_buffer_ensure_push_constants_size(cmd_buffer, stage, \
 227       (offsetof(struct anv_push_constants, field) + \
 228        sizeof(cmd_buffer->state.push_constants[0]->field)))
 229
 230 static VkResult anv_create_cmd_buffer(
 231     struct anv_device *                         device,
 232     struct anv_cmd_pool *                       pool,
 233     VkCommandBufferLevel                        level,
 234     VkCommandBuffer*                            pCommandBuffer)
 235 {
 236    struct anv_cmd_buffer *cmd_buffer;
 237    VkResult result;
 238
 239    cmd_buffer = anv_alloc(&pool->alloc, sizeof(*cmd_buffer), 8,
 240                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 241    if (cmd_buffer == NULL)
 242       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 243
 244    cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 245    cmd_buffer->device = device;
 246    cmd_buffer->pool = pool;
 247    cmd_buffer->level = level;
 248    cmd_buffer->state.attachments = NULL;
 249
 250    result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer);
 251    if (result != VK_SUCCESS)
 252       goto fail;
 253
 254    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 255                          &device->surface_state_block_pool);
 256    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 257                          &device->dynamic_state_block_pool);
 258
 259    if (pool) {
 260       list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
 261    } else {
 262       /* Init the pool_link so we can safefly call list_del when we destroy
 263        * the command buffer
 264        */
 265       list_inithead(&cmd_buffer->pool_link);
 266    }
 267
 268    *pCommandBuffer = anv_cmd_buffer_to_handle(cmd_buffer);
 269
 270    return VK_SUCCESS;
 271
 272  fail:
 273    anv_free(&cmd_buffer->pool->alloc, cmd_buffer);
 274
 275    return result;
 276 }
 277
 278 VkResult anv_AllocateCommandBuffers(
 279     VkDevice                                    _device,
 280     const VkCommandBufferAllocateInfo*          pAllocateInfo,
 281     VkCommandBuffer*                            pCommandBuffers)
 282 {
 283    ANV_FROM_HANDLE(anv_device, device, _device);
 284    ANV_FROM_HANDLE(anv_cmd_pool, pool, pAllocateInfo->commandPool);
 285
 286    VkResult result = VK_SUCCESS;
 287    uint32_t i;
 288
 289    for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
 290       result = anv_create_cmd_buffer(device, pool, pAllocateInfo->level,
 291                                      &pCommandBuffers[i]);
 292       if (result != VK_SUCCESS)
 293          break;
 294    }
 295
 296    if (result != VK_SUCCESS)
 297       anv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
 298                              i, pCommandBuffers);
 299
 300    return result;
 301 }
 302
 303 static void
 304 anv_cmd_buffer_destroy(struct anv_cmd_buffer *cmd_buffer)
 305 {
 306    list_del(&cmd_buffer->pool_link);
 307
 308    anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer);
 309
 310    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 311    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 312
 313    anv_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
 314    anv_free(&cmd_buffer->pool->alloc, cmd_buffer);
 315 }
 316
 317 void anv_FreeCommandBuffers(
 318     VkDevice                                    device,
 319     VkCommandPool                               commandPool,
 320     uint32_t                                    commandBufferCount,
 321     const VkCommandBuffer*                      pCommandBuffers)
 322 {
 323    for (uint32_t i = 0; i < commandBufferCount; i++) {
 324       ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
 325
 326       anv_cmd_buffer_destroy(cmd_buffer);
 327    }
 328 }
 329
 330 VkResult anv_ResetCommandBuffer(
 331     VkCommandBuffer                             commandBuffer,
 332     VkCommandBufferResetFlags                   flags)
 333 {
 334    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 335
 336    cmd_buffer->usage_flags = 0;
 337    cmd_buffer->state.current_pipeline = UINT32_MAX;
 338    anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer);
 339    anv_cmd_state_reset(cmd_buffer);
 340
 341    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 342    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 343                          &cmd_buffer->device->surface_state_block_pool);
 344
 345    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 346    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 347                          &cmd_buffer->device->dynamic_state_block_pool);
 348
 349    return VK_SUCCESS;
 350 }
 351
 352 void
 353 anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer)
 354 {
 355    switch (cmd_buffer->device->info.gen) {
 356    case 7:
 357       if (cmd_buffer->device->info.is_haswell)
 358          return gen7_cmd_buffer_emit_state_base_address(cmd_buffer);
 359       else
 360          return gen7_cmd_buffer_emit_state_base_address(cmd_buffer);
 361    case 8:
 362       return gen8_cmd_buffer_emit_state_base_address(cmd_buffer);
 363    case 9:
 364       return gen9_cmd_buffer_emit_state_base_address(cmd_buffer);
 365    default:
 366       unreachable("unsupported gen\n");
 367    }
 368 }
 369
 370 VkResult anv_BeginCommandBuffer(
 371     VkCommandBuffer                             commandBuffer,
 372     const VkCommandBufferBeginInfo*             pBeginInfo)
 373 {
 374    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 375
 376    /* If this is the first vkBeginCommandBuffer, we must *initialize* the
 377     * command buffer's state. Otherwise, we must *reset* its state. In both
 378     * cases we reset it.
 379     *
 380     * From the Vulkan 1.0 spec:
 381     *
 382     *    If a command buffer is in the executable state and the command buffer
 383     *    was allocated from a command pool with the
 384     *    VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then
 385     *    vkBeginCommandBuffer implicitly resets the command buffer, behaving
 386     *    as if vkResetCommandBuffer had been called with
 387     *    VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts
 388     *    the command buffer in the recording state.
 389     */
 390    anv_ResetCommandBuffer(commandBuffer, /*flags*/ 0);
 391
 392    cmd_buffer->usage_flags = pBeginInfo->flags;
 393
 394    assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY ||
 395           !(cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT));
 396
 397    anv_cmd_buffer_emit_state_base_address(cmd_buffer);
 398
 399    if (cmd_buffer->usage_flags &
 400        VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
 401       cmd_buffer->state.framebuffer =
 402          anv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);
 403       cmd_buffer->state.pass =
 404          anv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
 405
 406       struct anv_subpass *subpass =
 407          &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
 408
 409       anv_cmd_buffer_set_subpass(cmd_buffer, subpass);
 410    }
 411
 412    return VK_SUCCESS;
 413 }
 414
 415 VkResult anv_EndCommandBuffer(
 416     VkCommandBuffer                             commandBuffer)
 417 {
 418    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 419    struct anv_device *device = cmd_buffer->device;
 420
 421    anv_cmd_buffer_end_batch_buffer(cmd_buffer);
 422
 423    if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
 424       /* The algorithm used to compute the validate list is not threadsafe as
 425        * it uses the bo->index field.  We have to lock the device around it.
 426        * Fortunately, the chances for contention here are probably very low.
 427        */
 428       pthread_mutex_lock(&device->mutex);
 429       anv_cmd_buffer_prepare_execbuf(cmd_buffer);
 430       pthread_mutex_unlock(&device->mutex);
 431    }
 432
 433    return VK_SUCCESS;
 434 }
 435
 436 void anv_CmdBindPipeline(
 437     VkCommandBuffer                             commandBuffer,
 438     VkPipelineBindPoint                         pipelineBindPoint,
 439     VkPipeline                                  _pipeline)
 440 {
 441    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 442    ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 443
 444    switch (pipelineBindPoint) {
 445    case VK_PIPELINE_BIND_POINT_COMPUTE:
 446       cmd_buffer->state.compute_pipeline = pipeline;
 447       cmd_buffer->state.compute_dirty |= ANV_CMD_DIRTY_PIPELINE;
 448       cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 449       cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 450       break;
 451
 452    case VK_PIPELINE_BIND_POINT_GRAPHICS:
 453       cmd_buffer->state.pipeline = pipeline;
 454       cmd_buffer->state.vb_dirty |= pipeline->vb_used;
 455       cmd_buffer->state.dirty |= ANV_CMD_DIRTY_PIPELINE;
 456       cmd_buffer->state.push_constants_dirty |= pipeline->active_stages;
 457       cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;
 458
 459       /* Apply the dynamic state from the pipeline */
 460       cmd_buffer->state.dirty |= pipeline->dynamic_state_mask;
 461       anv_dynamic_state_copy(&cmd_buffer->state.dynamic,
 462                              &pipeline->dynamic_state,
 463                              pipeline->dynamic_state_mask);
 464       break;
 465
 466    default:
 467       assert(!"invalid bind point");
 468       break;
 469    }
 470 }
 471
 472 void anv_CmdSetViewport(
 473     VkCommandBuffer                             commandBuffer,
 474     uint32_t                                    firstViewport,
 475     uint32_t                                    viewportCount,
 476     const VkViewport*                           pViewports)
 477 {
 478    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 479
 480    const uint32_t total_count = firstViewport + viewportCount;
 481    if (cmd_buffer->state.dynamic.viewport.count < total_count)
 482       cmd_buffer->state.dynamic.viewport.count = total_count;
 483
 484    memcpy(cmd_buffer->state.dynamic.viewport.viewports + firstViewport,
 485           pViewports, viewportCount * sizeof(*pViewports));
 486
 487    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT;
 488 }
 489
 490 void anv_CmdSetScissor(
 491     VkCommandBuffer                             commandBuffer,
 492     uint32_t                                    firstScissor,
 493     uint32_t                                    scissorCount,
 494     const VkRect2D*                             pScissors)
 495 {
 496    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 497
 498    const uint32_t total_count = firstScissor + scissorCount;
 499    if (cmd_buffer->state.dynamic.scissor.count < total_count)
 500       cmd_buffer->state.dynamic.scissor.count = total_count;
 501
 502    memcpy(cmd_buffer->state.dynamic.scissor.scissors + firstScissor,
 503           pScissors, scissorCount * sizeof(*pScissors));
 504
 505    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR;
 506 }
 507
 508 void anv_CmdSetLineWidth(
 509     VkCommandBuffer                             commandBuffer,
 510     float                                       lineWidth)
 511 {
 512    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 513
 514    cmd_buffer->state.dynamic.line_width = lineWidth;
 515    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
 516 }
 517
 518 void anv_CmdSetDepthBias(
 519     VkCommandBuffer                             commandBuffer,
 520     float                                       depthBiasConstantFactor,
 521     float                                       depthBiasClamp,
 522     float                                       depthBiasSlopeFactor)
 523 {
 524    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 525
 526    cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
 527    cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
 528    cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
 529
 530    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
 531 }
 532
 533 void anv_CmdSetBlendConstants(
 534     VkCommandBuffer                             commandBuffer,
 535     const float                                 blendConstants[4])
 536 {
 537    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 538
 539    memcpy(cmd_buffer->state.dynamic.blend_constants,
 540           blendConstants, sizeof(float) * 4);
 541
 542    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
 543 }
 544
 545 void anv_CmdSetDepthBounds(
 546     VkCommandBuffer                             commandBuffer,
 547     float                                       minDepthBounds,
 548     float                                       maxDepthBounds)
 549 {
 550    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 551
 552    cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
 553    cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
 554
 555    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
 556 }
 557
 558 void anv_CmdSetStencilCompareMask(
 559     VkCommandBuffer                             commandBuffer,
 560     VkStencilFaceFlags                          faceMask,
 561     uint32_t                                    compareMask)
 562 {
 563    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 564
 565    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 566       cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
 567    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 568       cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
 569
 570    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
 571 }
 572
 573 void anv_CmdSetStencilWriteMask(
 574     VkCommandBuffer                             commandBuffer,
 575     VkStencilFaceFlags                          faceMask,
 576     uint32_t                                    writeMask)
 577 {
 578    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 579
 580    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 581       cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
 582    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 583       cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
 584
 585    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
 586 }
 587
 588 void anv_CmdSetStencilReference(
 589     VkCommandBuffer                             commandBuffer,
 590     VkStencilFaceFlags                          faceMask,
 591     uint32_t                                    reference)
 592 {
 593    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 594
 595    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 596       cmd_buffer->state.dynamic.stencil_reference.front = reference;
 597    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 598       cmd_buffer->state.dynamic.stencil_reference.back = reference;
 599
 600    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
 601 }
 602
 603 void anv_CmdBindDescriptorSets(
 604     VkCommandBuffer                             commandBuffer,
 605     VkPipelineBindPoint                         pipelineBindPoint,
 606     VkPipelineLayout                            _layout,
 607     uint32_t                                    firstSet,
 608     uint32_t                                    descriptorSetCount,
 609     const VkDescriptorSet*                      pDescriptorSets,
 610     uint32_t                                    dynamicOffsetCount,
 611     const uint32_t*                             pDynamicOffsets)
 612 {
 613    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 614    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 615    struct anv_descriptor_set_layout *set_layout;
 616
 617    assert(firstSet + descriptorSetCount < MAX_SETS);
 618
 619    uint32_t dynamic_slot = 0;
 620    for (uint32_t i = 0; i < descriptorSetCount; i++) {
 621       ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
 622       set_layout = layout->set[firstSet + i].layout;
 623
 624       if (cmd_buffer->state.descriptors[firstSet + i] != set) {
 625          cmd_buffer->state.descriptors[firstSet + i] = set;
 626          cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 627       }
 628
 629       if (set_layout->dynamic_offset_count > 0) {
 630          anv_foreach_stage(s, set_layout->shader_stages) {
 631             anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, s, dynamic);
 632
 633             struct anv_push_constants *push =
 634                cmd_buffer->state.push_constants[s];
 635
 636             unsigned d = layout->set[firstSet + i].dynamic_offset_start;
 637             const uint32_t *offsets = pDynamicOffsets + dynamic_slot;
 638             struct anv_descriptor *desc = set->descriptors;
 639
 640             for (unsigned b = 0; b < set_layout->binding_count; b++) {
 641                if (set_layout->binding[b].dynamic_offset_index < 0)
 642                   continue;
 643
 644                unsigned array_size = set_layout->binding[b].array_size;
 645                for (unsigned j = 0; j < array_size; j++) {
 646                   uint32_t range = 0;
 647                   if (desc->buffer_view)
 648                      range = desc->buffer_view->range;
 649                   push->dynamic[d].offset = *(offsets++);
 650                   push->dynamic[d].range = range;
 651                   desc++;
 652                   d++;
 653                }
 654             }
 655          }
 656          cmd_buffer->state.push_constants_dirty |= set_layout->shader_stages;
 657       }
 658    }
 659 }
 660
 661 void anv_CmdBindVertexBuffers(
 662     VkCommandBuffer                             commandBuffer,
 663     uint32_t                                    firstBinding,
 664     uint32_t                                    bindingCount,
 665     const VkBuffer*                             pBuffers,
 666     const VkDeviceSize*                         pOffsets)
 667 {
 668    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 669    struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 670
 671    /* We have to defer setting up vertex buffer since we need the buffer
 672     * stride from the pipeline. */
 673
 674    assert(firstBinding + bindingCount < MAX_VBS);
 675    for (uint32_t i = 0; i < bindingCount; i++) {
 676       vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
 677       vb[firstBinding + i].offset = pOffsets[i];
 678       cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i);
 679    }
 680 }
 681
 682 static void
 683 add_surface_state_reloc(struct anv_cmd_buffer *cmd_buffer,
 684                         struct anv_state state, struct anv_bo *bo, uint32_t offset)
 685 {
 686    /* The address goes in SURFACE_STATE dword 1 for gens < 8 and dwords 8 and
 687     * 9 for gen8+.  We only write the first dword for gen8+ here and rely on
 688     * the initial state to set the high bits to 0. */
 689
 690    const uint32_t dword = cmd_buffer->device->info.gen < 8 ? 1 : 8;
 691
 692    anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
 693                       state.offset + dword * 4, bo, offset);
 694 }
 695
 696 enum isl_format
 697 anv_isl_format_for_descriptor_type(VkDescriptorType type)
 698 {
 699    switch (type) {
 700    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 701    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 702       return ISL_FORMAT_R32G32B32A32_FLOAT;
 703
 704    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 705    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 706       return ISL_FORMAT_RAW;
 707
 708    default:
 709       unreachable("Invalid descriptor type");
 710    }
 711 }
 712
 713 static struct anv_state
 714 anv_cmd_buffer_alloc_null_surface_state(struct anv_cmd_buffer *cmd_buffer,
 715                                         struct anv_framebuffer *fb)
 716 {
 717    switch (cmd_buffer->device->info.gen) {
 718    case 7:
 719       if (cmd_buffer->device->info.is_haswell) {
 720          return gen75_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 721       } else {
 722          return gen7_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 723       }
 724    case 8:
 725       return gen8_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 726    case 9:
 727       return gen9_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 728    default:
 729       unreachable("Invalid hardware generation");
 730    }
 731 }
 732
 733 VkResult
 734 anv_cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
 735                                   gl_shader_stage stage,
 736                                   struct anv_state *bt_state)
 737 {
 738    struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 739    struct anv_subpass *subpass = cmd_buffer->state.subpass;
 740    struct anv_pipeline_bind_map *map;
 741    uint32_t bias, state_offset;
 742
 743    switch (stage) {
 744    case  MESA_SHADER_COMPUTE:
 745       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 746       bias = 1;
 747       break;
 748    default:
 749       map = &cmd_buffer->state.pipeline->bindings[stage];
 750       bias = 0;
 751       break;
 752    }
 753
 754    if (bias + map->surface_count == 0) {
 755       *bt_state = (struct anv_state) { 0, };
 756       return VK_SUCCESS;
 757    }
 758
 759    *bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer,
 760                                                   bias + map->surface_count,
 761                                                   &state_offset);
 762    uint32_t *bt_map = bt_state->map;
 763
 764    if (bt_state->map == NULL)
 765       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 766
 767    if (stage == MESA_SHADER_COMPUTE &&
 768        get_cs_prog_data(cmd_buffer->state.compute_pipeline)->uses_num_work_groups) {
 769       struct anv_bo *bo = cmd_buffer->state.num_workgroups_bo;
 770       uint32_t bo_offset = cmd_buffer->state.num_workgroups_offset;
 771
 772       struct anv_state surface_state;
 773       surface_state =
 774          anv_cmd_buffer_alloc_surface_state(cmd_buffer);
 775
 776       const enum isl_format format =
 777          anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
 778       anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
 779                                     format, bo_offset, 12, 1);
 780
 781       bt_map[0] = surface_state.offset + state_offset;
 782       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 783    }
 784
 785    if (map->surface_count == 0)
 786       goto out;
 787
 788    if (map->image_count > 0) {
 789       VkResult result =
 790          anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images);
 791       if (result != VK_SUCCESS)
 792          return result;
 793
 794       cmd_buffer->state.push_constants_dirty |= 1 << stage;
 795    }
 796
 797    uint32_t image = 0;
 798    for (uint32_t s = 0; s < map->surface_count; s++) {
 799       struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];
 800
 801       struct anv_state surface_state;
 802       struct anv_bo *bo;
 803       uint32_t bo_offset;
 804
 805       if (binding->set == ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS) {
 806          /* Color attachment binding */
 807          assert(stage == MESA_SHADER_FRAGMENT);
 808          if (binding->offset < subpass->color_count) {
 809             const struct anv_image_view *iview =
 810                fb->attachments[subpass->color_attachments[binding->offset]];
 811
 812             assert(iview->color_rt_surface_state.alloc_size);
 813             surface_state = iview->color_rt_surface_state;
 814             add_surface_state_reloc(cmd_buffer, iview->color_rt_surface_state,
 815                                     iview->bo, iview->offset);
 816          } else {
 817             /* Null render target */
 818             struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 819             surface_state =
 820                anv_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 821          }
 822
 823          bt_map[bias + s] = surface_state.offset + state_offset;
 824          continue;
 825       }
 826
 827       struct anv_descriptor_set *set =
 828          cmd_buffer->state.descriptors[binding->set];
 829       struct anv_descriptor *desc = &set->descriptors[binding->offset];
 830
 831       switch (desc->type) {
 832       case VK_DESCRIPTOR_TYPE_SAMPLER:
 833          /* Nothing for us to do here */
 834          continue;
 835
 836       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 837       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 838       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 839          surface_state = desc->image_view->sampler_surface_state;
 840          assert(surface_state.alloc_size);
 841          bo = desc->image_view->bo;
 842          bo_offset = desc->image_view->offset;
 843          break;
 844
 845       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
 846          surface_state = desc->image_view->storage_surface_state;
 847          assert(surface_state.alloc_size);
 848          bo = desc->image_view->bo;
 849          bo_offset = desc->image_view->offset;
 850
 851          struct brw_image_param *image_param =
 852             &cmd_buffer->state.push_constants[stage]->images[image++];
 853
 854          *image_param = desc->image_view->storage_image_param;
 855          image_param->surface_idx = bias + s;
 856          break;
 857       }
 858
 859       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 860       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 861       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 862       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 863       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
 864          surface_state = desc->buffer_view->surface_state;
 865          assert(surface_state.alloc_size);
 866          bo = desc->buffer_view->bo;
 867          bo_offset = desc->buffer_view->offset;
 868          break;
 869
 870       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
 871          surface_state = desc->buffer_view->storage_surface_state;
 872          assert(surface_state.alloc_size);
 873          bo = desc->buffer_view->bo;
 874          bo_offset = desc->buffer_view->offset;
 875
 876          struct brw_image_param *image_param =
 877             &cmd_buffer->state.push_constants[stage]->images[image++];
 878
 879          *image_param = desc->buffer_view->storage_image_param;
 880          image_param->surface_idx = bias + s;
 881          break;
 882
 883       default:
 884          assert(!"Invalid descriptor type");
 885          continue;
 886       }
 887
 888       bt_map[bias + s] = surface_state.offset + state_offset;
 889       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 890    }
 891    assert(image == map->image_count);
 892
 893  out:
 894    if (!cmd_buffer->device->info.has_llc)
 895       anv_state_clflush(*bt_state);
 896
 897    return VK_SUCCESS;
 898 }
 899
 900 VkResult
 901 anv_cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer,
 902                              gl_shader_stage stage, struct anv_state *state)
 903 {
 904    struct anv_pipeline_bind_map *map;
 905
 906    if (stage == MESA_SHADER_COMPUTE)
 907       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 908    else
 909       map = &cmd_buffer->state.pipeline->bindings[stage];
 910
 911    if (map->sampler_count == 0) {
 912       *state = (struct anv_state) { 0, };
 913       return VK_SUCCESS;
 914    }
 915
 916    uint32_t size = map->sampler_count * 16;
 917    *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32);
 918
 919    if (state->map == NULL)
 920       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 921
 922    for (uint32_t s = 0; s < map->sampler_count; s++) {
 923       struct anv_pipeline_binding *binding = &map->sampler_to_descriptor[s];
 924       struct anv_descriptor_set *set =
 925          cmd_buffer->state.descriptors[binding->set];
 926       struct anv_descriptor *desc = &set->descriptors[binding->offset];
 927
 928       if (desc->type != VK_DESCRIPTOR_TYPE_SAMPLER &&
 929           desc->type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
 930          continue;
 931
 932       struct anv_sampler *sampler = desc->sampler;
 933
 934       /* This can happen if we have an unfilled slot since TYPE_SAMPLER
 935        * happens to be zero.
 936        */
 937       if (sampler == NULL)
 938          continue;
 939
 940       memcpy(state->map + (s * 16),
 941              sampler->state, sizeof(sampler->state));
 942    }
 943
 944    if (!cmd_buffer->device->info.has_llc)
 945       anv_state_clflush(*state);
 946
 947    return VK_SUCCESS;
 948 }
 949
 950 struct anv_state
 951 anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
 952                             const void *data, uint32_t size, uint32_t alignment)
 953 {
 954    struct anv_state state;
 955
 956    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment);
 957    memcpy(state.map, data, size);
 958
 959    if (!cmd_buffer->device->info.has_llc)
 960       anv_state_clflush(state);
 961
 962    VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size));
 963
 964    return state;
 965 }
 966
 967 struct anv_state
 968 anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
 969                              uint32_t *a, uint32_t *b,
 970                              uint32_t dwords, uint32_t alignment)
 971 {
 972    struct anv_state state;
 973    uint32_t *p;
 974
 975    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 976                                               dwords * 4, alignment);
 977    p = state.map;
 978    for (uint32_t i = 0; i < dwords; i++)
 979       p[i] = a[i] | b[i];
 980
 981    if (!cmd_buffer->device->info.has_llc)
 982       anv_state_clflush(state);
 983
 984    VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
 985
 986    return state;
 987 }
 988
 989 /**
 990  * @brief Setup the command buffer for recording commands inside the given
 991  * subpass.
 992  *
 993  * This does not record all commands needed for starting the subpass.
 994  * Starting the subpass may require additional commands.
 995  *
 996  * Note that vkCmdBeginRenderPass, vkCmdNextSubpass, and vkBeginCommandBuffer
 997  * with VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT, all setup the
 998  * command buffer for recording commands for some subpass.  But only the first
 999  * two, vkCmdBeginRenderPass and vkCmdNextSubpass, can start a subpass.
1000  */
1001 void
1002 anv_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
1003                            struct anv_subpass *subpass)
1004 {
1005    switch (cmd_buffer->device->info.gen) {
1006    case 7:
1007       if (cmd_buffer->device->info.is_haswell) {
1008          gen75_cmd_buffer_set_subpass(cmd_buffer, subpass);
1009       } else {
1010          gen7_cmd_buffer_set_subpass(cmd_buffer, subpass);
1011       }
1012       break;
1013    case 8:
1014       gen8_cmd_buffer_set_subpass(cmd_buffer, subpass);
1015       break;
1016    case 9:
1017       gen9_cmd_buffer_set_subpass(cmd_buffer, subpass);
1018       break;
1019    default:
1020       unreachable("unsupported gen\n");
1021    }
1022 }
1023
1024 struct anv_state
1025 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
1026                               gl_shader_stage stage)
1027 {
1028    struct anv_push_constants *data =
1029       cmd_buffer->state.push_constants[stage];
1030    const struct brw_stage_prog_data *prog_data =
1031       cmd_buffer->state.pipeline->prog_data[stage];
1032
1033    /* If we don't actually have any push constants, bail. */
1034    if (data == NULL || prog_data->nr_params == 0)
1035       return (struct anv_state) { .offset = 0 };
1036
1037    struct anv_state state =
1038       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1039                                          prog_data->nr_params * sizeof(float),
1040                                          32 /* bottom 5 bits MBZ */);
1041
1042    /* Walk through the param array and fill the buffer with data */
1043    uint32_t *u32_map = state.map;
1044    for (unsigned i = 0; i < prog_data->nr_params; i++) {
1045       uint32_t offset = (uintptr_t)prog_data->param[i];
1046       u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
1047    }
1048
1049    if (!cmd_buffer->device->info.has_llc)
1050       anv_state_clflush(state);
1051
1052    return state;
1053 }
1054
1055 struct anv_state
1056 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
1057 {
1058    struct anv_push_constants *data =
1059       cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
1060    struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
1061    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
1062    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
1063
1064    const unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
1065    const unsigned push_constant_data_size =
1066       (local_id_dwords + prog_data->nr_params) * 4;
1067    const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
1068    const unsigned param_aligned_count =
1069       reg_aligned_constant_size / sizeof(uint32_t);
1070
1071    /* If we don't actually have any push constants, bail. */
1072    if (reg_aligned_constant_size == 0)
1073       return (struct anv_state) { .offset = 0 };
1074
1075    const unsigned threads = pipeline->cs_thread_width_max;
1076    const unsigned total_push_constants_size =
1077       reg_aligned_constant_size * threads;
1078    const unsigned push_constant_alignment =
1079       cmd_buffer->device->info.gen < 8 ? 32 : 64;
1080    const unsigned aligned_total_push_constants_size =
1081       ALIGN(total_push_constants_size, push_constant_alignment);
1082    struct anv_state state =
1083       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1084                                          aligned_total_push_constants_size,
1085                                          push_constant_alignment);
1086
1087    /* Walk through the param array and fill the buffer with data */
1088    uint32_t *u32_map = state.map;
1089
1090    brw_cs_fill_local_id_payload(cs_prog_data, u32_map, threads,
1091                                 reg_aligned_constant_size);
1092
1093    /* Setup uniform data for the first thread */
1094    for (unsigned i = 0; i < prog_data->nr_params; i++) {
1095       uint32_t offset = (uintptr_t)prog_data->param[i];
1096       u32_map[local_id_dwords + i] = *(uint32_t *)((uint8_t *)data + offset);
1097    }
1098
1099    /* Copy uniform data from the first thread to every other thread */
1100    const size_t uniform_data_size = prog_data->nr_params * sizeof(uint32_t);
1101    for (unsigned t = 1; t < threads; t++) {
1102       memcpy(&u32_map[t * param_aligned_count + local_id_dwords],
1103              &u32_map[local_id_dwords],
1104              uniform_data_size);
1105    }
1106
1107    if (!cmd_buffer->device->info.has_llc)
1108       anv_state_clflush(state);
1109
1110    return state;
1111 }
1112
1113 void anv_CmdPushConstants(
1114     VkCommandBuffer                             commandBuffer,
1115     VkPipelineLayout                            layout,
1116     VkShaderStageFlags                          stageFlags,
1117     uint32_t                                    offset,
1118     uint32_t                                    size,
1119     const void*                                 pValues)
1120 {
1121    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1122
1123    anv_foreach_stage(stage, stageFlags) {
1124       anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, client_data);
1125
1126       memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
1127              pValues, size);
1128    }
1129
1130    cmd_buffer->state.push_constants_dirty |= stageFlags;
1131 }
1132
1133 void anv_CmdExecuteCommands(
1134     VkCommandBuffer                             commandBuffer,
1135     uint32_t                                    commandBufferCount,
1136     const VkCommandBuffer*                      pCmdBuffers)
1137 {
1138    ANV_FROM_HANDLE(anv_cmd_buffer, primary, commandBuffer);
1139
1140    assert(primary->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1141
1142    for (uint32_t i = 0; i < commandBufferCount; i++) {
1143       ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);
1144
1145       assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
1146
1147       anv_cmd_buffer_add_secondary(primary, secondary);
1148    }
1149 }
1150
1151 VkResult anv_CreateCommandPool(
1152     VkDevice                                    _device,
1153     const VkCommandPoolCreateInfo*              pCreateInfo,
1154     const VkAllocationCallbacks*                pAllocator,
1155     VkCommandPool*                              pCmdPool)
1156 {
1157    ANV_FROM_HANDLE(anv_device, device, _device);
1158    struct anv_cmd_pool *pool;
1159
1160    pool = anv_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
1161                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1162    if (pool == NULL)
1163       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1164
1165    if (pAllocator)
1166       pool->alloc = *pAllocator;
1167    else
1168       pool->alloc = device->alloc;
1169
1170    list_inithead(&pool->cmd_buffers);
1171
1172    *pCmdPool = anv_cmd_pool_to_handle(pool);
1173
1174    return VK_SUCCESS;
1175 }
1176
1177 void anv_DestroyCommandPool(
1178     VkDevice                                    _device,
1179     VkCommandPool                               commandPool,
1180     const VkAllocationCallbacks*                pAllocator)
1181 {
1182    ANV_FROM_HANDLE(anv_device, device, _device);
1183    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1184
1185    anv_ResetCommandPool(_device, commandPool, 0);
1186
1187    anv_free2(&device->alloc, pAllocator, pool);
1188 }
1189
1190 VkResult anv_ResetCommandPool(
1191     VkDevice                                    device,
1192     VkCommandPool                               commandPool,
1193     VkCommandPoolResetFlags                     flags)
1194 {
1195    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1196
1197    /* FIXME: vkResetCommandPool must not destroy its command buffers. The
1198     * Vulkan 1.0 spec requires that it only reset them:
1199     *
1200     *    Resetting a command pool recycles all of the resources from all of
1201     *    the command buffers allocated from the command pool back to the
1202     *    command pool. All command buffers that have been allocated from the
1203     *    command pool are put in the initial state.
1204     */
1205    list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
1206                             &pool->cmd_buffers, pool_link) {
1207       anv_cmd_buffer_destroy(cmd_buffer);
1208    }
1209
1210    return VK_SUCCESS;
1211 }
1212
1213 /**
1214  * Return NULL if the current subpass has no depthstencil attachment.
1215  */
1216 const struct anv_image_view *
1217 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
1218 {
1219    const struct anv_subpass *subpass = cmd_buffer->state.subpass;
1220    const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
1221
1222    if (subpass->depth_stencil_attachment == VK_ATTACHMENT_UNUSED)
1223       return NULL;
1224
1225    const struct anv_image_view *iview =
1226       fb->attachments[subpass->depth_stencil_attachment];
1227
1228    assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
1229                                 VK_IMAGE_ASPECT_STENCIL_BIT));
1230
1231    return iview;
1232 }