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 /** \file anv_cmd_buffer.c
  33  *
  34  * This file contains all of the stuff for emitting commands into a command
  35  * buffer.  This includes implementations of most of the vkCmd*
  36  * entrypoints.  This file is concerned entirely with state emission and
  37  * not with the command buffer data structure itself.  As far as this file
  38  * is concerned, most of anv_cmd_buffer is magic.
  39  */
  40
  41 /* TODO: These are taken from GLES.  We should check the Vulkan spec */
  42 const struct anv_dynamic_state default_dynamic_state = {
  43    .viewport = {
  44       .count = 0,
  45    },
  46    .scissor = {
  47       .count = 0,
  48    },
  49    .line_width = 1.0f,
  50    .depth_bias = {
  51       .bias = 0.0f,
  52       .clamp = 0.0f,
  53       .slope = 0.0f,
  54    },
  55    .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
  56    .depth_bounds = {
  57       .min = 0.0f,
  58       .max = 1.0f,
  59    },
  60    .stencil_compare_mask = {
  61       .front = ~0u,
  62       .back = ~0u,
  63    },
  64    .stencil_write_mask = {
  65       .front = ~0u,
  66       .back = ~0u,
  67    },
  68    .stencil_reference = {
  69       .front = 0u,
  70       .back = 0u,
  71    },
  72 };
  73
  74 void
  75 anv_dynamic_state_copy(struct anv_dynamic_state *dest,
  76                        const struct anv_dynamic_state *src,
  77                        uint32_t copy_mask)
  78 {
  79    if (copy_mask & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
  80       dest->viewport.count = src->viewport.count;
  81       typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
  82                    src->viewport.count);
  83    }
  84
  85    if (copy_mask & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
  86       dest->scissor.count = src->scissor.count;
  87       typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
  88                    src->scissor.count);
  89    }
  90
  91    if (copy_mask & (1 << VK_DYNAMIC_STATE_LINE_WIDTH))
  92       dest->line_width = src->line_width;
  93
  94    if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS))
  95       dest->depth_bias = src->depth_bias;
  96
  97    if (copy_mask & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS))
  98       typed_memcpy(dest->blend_constants, src->blend_constants, 4);
  99
 100    if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS))
 101       dest->depth_bounds = src->depth_bounds;
 102
 103    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK))
 104       dest->stencil_compare_mask = src->stencil_compare_mask;
 105
 106    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK))
 107       dest->stencil_write_mask = src->stencil_write_mask;
 108
 109    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE))
 110       dest->stencil_reference = src->stencil_reference;
 111 }
 112
 113 static void
 114 anv_cmd_state_reset(struct anv_cmd_buffer *cmd_buffer)
 115 {
 116    struct anv_cmd_state *state = &cmd_buffer->state;
 117
 118    memset(&state->descriptors, 0, sizeof(state->descriptors));
 119    memset(&state->push_constants, 0, sizeof(state->push_constants));
 120    memset(state->binding_tables, 0, sizeof(state->binding_tables));
 121    memset(state->samplers, 0, sizeof(state->samplers));
 122
 123    /* 0 isn't a valid config.  This ensures that we always configure L3$. */
 124    cmd_buffer->state.current_l3_config = 0;
 125
 126    state->dirty = 0;
 127    state->vb_dirty = 0;
 128    state->descriptors_dirty = 0;
 129    state->push_constants_dirty = 0;
 130    state->pipeline = NULL;
 131    state->restart_index = UINT32_MAX;
 132    state->dynamic = default_dynamic_state;
 133    state->need_query_wa = true;
 134
 135    if (state->attachments != NULL) {
 136       anv_free(&cmd_buffer->pool->alloc, state->attachments);
 137       state->attachments = NULL;
 138    }
 139
 140    state->gen7.index_buffer = NULL;
 141 }
 142
 143 /**
 144  * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
 145  */
 146 void
 147 anv_cmd_state_setup_attachments(struct anv_cmd_buffer *cmd_buffer,
 148                                 const VkRenderPassBeginInfo *info)
 149 {
 150    struct anv_cmd_state *state = &cmd_buffer->state;
 151    ANV_FROM_HANDLE(anv_render_pass, pass, info->renderPass);
 152
 153    anv_free(&cmd_buffer->pool->alloc, state->attachments);
 154
 155    if (pass->attachment_count == 0) {
 156       state->attachments = NULL;
 157       return;
 158    }
 159
 160    state->attachments = anv_alloc(&cmd_buffer->pool->alloc,
 161                                   pass->attachment_count *
 162                                        sizeof(state->attachments[0]),
 163                                   8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 164    if (state->attachments == NULL) {
 165       /* FIXME: Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
 166       abort();
 167    }
 168
 169    for (uint32_t i = 0; i < pass->attachment_count; ++i) {
 170       struct anv_render_pass_attachment *att = &pass->attachments[i];
 171       VkImageAspectFlags clear_aspects = 0;
 172
 173       if (anv_format_is_color(att->format)) {
 174          /* color attachment */
 175          if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 176             clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
 177          }
 178       } else {
 179          /* depthstencil attachment */
 180          if (att->format->has_depth &&
 181              att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 182             clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
 183          }
 184          if (att->format->has_stencil &&
 185              att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 186             clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
 187          }
 188       }
 189
 190       state->attachments[i].pending_clear_aspects = clear_aspects;
 191       if (clear_aspects) {
 192          assert(info->clearValueCount > i);
 193          state->attachments[i].clear_value = info->pClearValues[i];
 194       }
 195    }
 196 }
 197
 198 static VkResult
 199 anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer,
 200                                           gl_shader_stage stage, uint32_t size)
 201 {
 202    struct anv_push_constants **ptr = &cmd_buffer->state.push_constants[stage];
 203
 204    if (*ptr == NULL) {
 205       *ptr = anv_alloc(&cmd_buffer->pool->alloc, size, 8,
 206                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 207       if (*ptr == NULL)
 208          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 209    } else if ((*ptr)->size < size) {
 210       *ptr = anv_realloc(&cmd_buffer->pool->alloc, *ptr, size, 8,
 211                          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 212       if (*ptr == NULL)
 213          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 214    }
 215    (*ptr)->size = size;
 216
 217    return VK_SUCCESS;
 218 }
 219
 220 #define anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, field) \
 221    anv_cmd_buffer_ensure_push_constants_size(cmd_buffer, stage, \
 222       (offsetof(struct anv_push_constants, field) + \
 223        sizeof(cmd_buffer->state.push_constants[0]->field)))
 224
 225 static VkResult anv_create_cmd_buffer(
 226     struct anv_device *                         device,
 227     struct anv_cmd_pool *                       pool,
 228     VkCommandBufferLevel                        level,
 229     VkCommandBuffer*                            pCommandBuffer)
 230 {
 231    struct anv_cmd_buffer *cmd_buffer;
 232    VkResult result;
 233
 234    cmd_buffer = anv_alloc(&pool->alloc, sizeof(*cmd_buffer), 8,
 235                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 236    if (cmd_buffer == NULL)
 237       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 238
 239    cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 240    cmd_buffer->device = device;
 241    cmd_buffer->pool = pool;
 242    cmd_buffer->level = level;
 243    cmd_buffer->state.attachments = NULL;
 244
 245    result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer);
 246    if (result != VK_SUCCESS)
 247       goto fail;
 248
 249    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 250                          &device->surface_state_block_pool);
 251    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 252                          &device->dynamic_state_block_pool);
 253
 254    if (pool) {
 255       list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
 256    } else {
 257       /* Init the pool_link so we can safefly call list_del when we destroy
 258        * the command buffer
 259        */
 260       list_inithead(&cmd_buffer->pool_link);
 261    }
 262
 263    *pCommandBuffer = anv_cmd_buffer_to_handle(cmd_buffer);
 264
 265    return VK_SUCCESS;
 266
 267  fail:
 268    anv_free(&cmd_buffer->pool->alloc, cmd_buffer);
 269
 270    return result;
 271 }
 272
 273 VkResult anv_AllocateCommandBuffers(
 274     VkDevice                                    _device,
 275     const VkCommandBufferAllocateInfo*          pAllocateInfo,
 276     VkCommandBuffer*                            pCommandBuffers)
 277 {
 278    ANV_FROM_HANDLE(anv_device, device, _device);
 279    ANV_FROM_HANDLE(anv_cmd_pool, pool, pAllocateInfo->commandPool);
 280
 281    VkResult result = VK_SUCCESS;
 282    uint32_t i;
 283
 284    for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
 285       result = anv_create_cmd_buffer(device, pool, pAllocateInfo->level,
 286                                      &pCommandBuffers[i]);
 287       if (result != VK_SUCCESS)
 288          break;
 289    }
 290
 291    if (result != VK_SUCCESS)
 292       anv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
 293                              i, pCommandBuffers);
 294
 295    return result;
 296 }
 297
 298 static void
 299 anv_cmd_buffer_destroy(struct anv_cmd_buffer *cmd_buffer)
 300 {
 301    list_del(&cmd_buffer->pool_link);
 302
 303    anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer);
 304
 305    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 306    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 307
 308    anv_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
 309    anv_free(&cmd_buffer->pool->alloc, cmd_buffer);
 310 }
 311
 312 void anv_FreeCommandBuffers(
 313     VkDevice                                    device,
 314     VkCommandPool                               commandPool,
 315     uint32_t                                    commandBufferCount,
 316     const VkCommandBuffer*                      pCommandBuffers)
 317 {
 318    for (uint32_t i = 0; i < commandBufferCount; i++) {
 319       ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
 320
 321       anv_cmd_buffer_destroy(cmd_buffer);
 322    }
 323 }
 324
 325 VkResult anv_ResetCommandBuffer(
 326     VkCommandBuffer                             commandBuffer,
 327     VkCommandBufferResetFlags                   flags)
 328 {
 329    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 330
 331    cmd_buffer->usage_flags = 0;
 332    cmd_buffer->state.current_pipeline = UINT32_MAX;
 333    anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer);
 334    anv_cmd_state_reset(cmd_buffer);
 335
 336    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 337    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 338                          &cmd_buffer->device->surface_state_block_pool);
 339
 340    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 341    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 342                          &cmd_buffer->device->dynamic_state_block_pool);
 343
 344    return VK_SUCCESS;
 345 }
 346
 347 void
 348 anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer)
 349 {
 350    switch (cmd_buffer->device->info.gen) {
 351    case 7:
 352       if (cmd_buffer->device->info.is_haswell)
 353          return gen7_cmd_buffer_emit_state_base_address(cmd_buffer);
 354       else
 355          return gen7_cmd_buffer_emit_state_base_address(cmd_buffer);
 356    case 8:
 357       return gen8_cmd_buffer_emit_state_base_address(cmd_buffer);
 358    case 9:
 359       return gen9_cmd_buffer_emit_state_base_address(cmd_buffer);
 360    default:
 361       unreachable("unsupported gen\n");
 362    }
 363 }
 364
 365 VkResult anv_BeginCommandBuffer(
 366     VkCommandBuffer                             commandBuffer,
 367     const VkCommandBufferBeginInfo*             pBeginInfo)
 368 {
 369    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 370
 371    /* If this is the first vkBeginCommandBuffer, we must *initialize* the
 372     * command buffer's state. Otherwise, we must *reset* its state. In both
 373     * cases we reset it.
 374     *
 375     * From the Vulkan 1.0 spec:
 376     *
 377     *    If a command buffer is in the executable state and the command buffer
 378     *    was allocated from a command pool with the
 379     *    VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then
 380     *    vkBeginCommandBuffer implicitly resets the command buffer, behaving
 381     *    as if vkResetCommandBuffer had been called with
 382     *    VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts
 383     *    the command buffer in the recording state.
 384     */
 385    anv_ResetCommandBuffer(commandBuffer, /*flags*/ 0);
 386
 387    cmd_buffer->usage_flags = pBeginInfo->flags;
 388
 389    assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY ||
 390           !(cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT));
 391
 392    anv_cmd_buffer_emit_state_base_address(cmd_buffer);
 393
 394    if (cmd_buffer->usage_flags &
 395        VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
 396       cmd_buffer->state.framebuffer =
 397          anv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);
 398       cmd_buffer->state.pass =
 399          anv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
 400
 401       struct anv_subpass *subpass =
 402          &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
 403
 404       anv_cmd_buffer_set_subpass(cmd_buffer, subpass);
 405    }
 406
 407    return VK_SUCCESS;
 408 }
 409
 410 VkResult anv_EndCommandBuffer(
 411     VkCommandBuffer                             commandBuffer)
 412 {
 413    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 414    struct anv_device *device = cmd_buffer->device;
 415
 416    anv_cmd_buffer_end_batch_buffer(cmd_buffer);
 417
 418    if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
 419       /* The algorithm used to compute the validate list is not threadsafe as
 420        * it uses the bo->index field.  We have to lock the device around it.
 421        * Fortunately, the chances for contention here are probably very low.
 422        */
 423       pthread_mutex_lock(&device->mutex);
 424       anv_cmd_buffer_prepare_execbuf(cmd_buffer);
 425       pthread_mutex_unlock(&device->mutex);
 426    }
 427
 428    return VK_SUCCESS;
 429 }
 430
 431 void anv_CmdBindPipeline(
 432     VkCommandBuffer                             commandBuffer,
 433     VkPipelineBindPoint                         pipelineBindPoint,
 434     VkPipeline                                  _pipeline)
 435 {
 436    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 437    ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 438
 439    switch (pipelineBindPoint) {
 440    case VK_PIPELINE_BIND_POINT_COMPUTE:
 441       cmd_buffer->state.compute_pipeline = pipeline;
 442       cmd_buffer->state.compute_dirty |= ANV_CMD_DIRTY_PIPELINE;
 443       cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 444       cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 445       break;
 446
 447    case VK_PIPELINE_BIND_POINT_GRAPHICS:
 448       cmd_buffer->state.pipeline = pipeline;
 449       cmd_buffer->state.vb_dirty |= pipeline->vb_used;
 450       cmd_buffer->state.dirty |= ANV_CMD_DIRTY_PIPELINE;
 451       cmd_buffer->state.push_constants_dirty |= pipeline->active_stages;
 452       cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;
 453
 454       /* Apply the dynamic state from the pipeline */
 455       cmd_buffer->state.dirty |= pipeline->dynamic_state_mask;
 456       anv_dynamic_state_copy(&cmd_buffer->state.dynamic,
 457                              &pipeline->dynamic_state,
 458                              pipeline->dynamic_state_mask);
 459       break;
 460
 461    default:
 462       assert(!"invalid bind point");
 463       break;
 464    }
 465 }
 466
 467 void anv_CmdSetViewport(
 468     VkCommandBuffer                             commandBuffer,
 469     uint32_t                                    firstViewport,
 470     uint32_t                                    viewportCount,
 471     const VkViewport*                           pViewports)
 472 {
 473    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 474
 475    const uint32_t total_count = firstViewport + viewportCount;
 476    if (cmd_buffer->state.dynamic.viewport.count < total_count)
 477       cmd_buffer->state.dynamic.viewport.count = total_count;
 478
 479    memcpy(cmd_buffer->state.dynamic.viewport.viewports + firstViewport,
 480           pViewports, viewportCount * sizeof(*pViewports));
 481
 482    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT;
 483 }
 484
 485 void anv_CmdSetScissor(
 486     VkCommandBuffer                             commandBuffer,
 487     uint32_t                                    firstScissor,
 488     uint32_t                                    scissorCount,
 489     const VkRect2D*                             pScissors)
 490 {
 491    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 492
 493    const uint32_t total_count = firstScissor + scissorCount;
 494    if (cmd_buffer->state.dynamic.scissor.count < total_count)
 495       cmd_buffer->state.dynamic.scissor.count = total_count;
 496
 497    memcpy(cmd_buffer->state.dynamic.scissor.scissors + firstScissor,
 498           pScissors, scissorCount * sizeof(*pScissors));
 499
 500    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR;
 501 }
 502
 503 void anv_CmdSetLineWidth(
 504     VkCommandBuffer                             commandBuffer,
 505     float                                       lineWidth)
 506 {
 507    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 508
 509    cmd_buffer->state.dynamic.line_width = lineWidth;
 510    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
 511 }
 512
 513 void anv_CmdSetDepthBias(
 514     VkCommandBuffer                             commandBuffer,
 515     float                                       depthBiasConstantFactor,
 516     float                                       depthBiasClamp,
 517     float                                       depthBiasSlopeFactor)
 518 {
 519    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 520
 521    cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
 522    cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
 523    cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
 524
 525    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
 526 }
 527
 528 void anv_CmdSetBlendConstants(
 529     VkCommandBuffer                             commandBuffer,
 530     const float                                 blendConstants[4])
 531 {
 532    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 533
 534    memcpy(cmd_buffer->state.dynamic.blend_constants,
 535           blendConstants, sizeof(float) * 4);
 536
 537    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
 538 }
 539
 540 void anv_CmdSetDepthBounds(
 541     VkCommandBuffer                             commandBuffer,
 542     float                                       minDepthBounds,
 543     float                                       maxDepthBounds)
 544 {
 545    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 546
 547    cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
 548    cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
 549
 550    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
 551 }
 552
 553 void anv_CmdSetStencilCompareMask(
 554     VkCommandBuffer                             commandBuffer,
 555     VkStencilFaceFlags                          faceMask,
 556     uint32_t                                    compareMask)
 557 {
 558    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 559
 560    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 561       cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
 562    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 563       cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
 564
 565    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
 566 }
 567
 568 void anv_CmdSetStencilWriteMask(
 569     VkCommandBuffer                             commandBuffer,
 570     VkStencilFaceFlags                          faceMask,
 571     uint32_t                                    writeMask)
 572 {
 573    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 574
 575    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 576       cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
 577    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 578       cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
 579
 580    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
 581 }
 582
 583 void anv_CmdSetStencilReference(
 584     VkCommandBuffer                             commandBuffer,
 585     VkStencilFaceFlags                          faceMask,
 586     uint32_t                                    reference)
 587 {
 588    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 589
 590    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 591       cmd_buffer->state.dynamic.stencil_reference.front = reference;
 592    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 593       cmd_buffer->state.dynamic.stencil_reference.back = reference;
 594
 595    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
 596 }
 597
 598 void anv_CmdBindDescriptorSets(
 599     VkCommandBuffer                             commandBuffer,
 600     VkPipelineBindPoint                         pipelineBindPoint,
 601     VkPipelineLayout                            _layout,
 602     uint32_t                                    firstSet,
 603     uint32_t                                    descriptorSetCount,
 604     const VkDescriptorSet*                      pDescriptorSets,
 605     uint32_t                                    dynamicOffsetCount,
 606     const uint32_t*                             pDynamicOffsets)
 607 {
 608    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 609    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 610    struct anv_descriptor_set_layout *set_layout;
 611
 612    assert(firstSet + descriptorSetCount < MAX_SETS);
 613
 614    uint32_t dynamic_slot = 0;
 615    for (uint32_t i = 0; i < descriptorSetCount; i++) {
 616       ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
 617       set_layout = layout->set[firstSet + i].layout;
 618
 619       if (cmd_buffer->state.descriptors[firstSet + i] != set) {
 620          cmd_buffer->state.descriptors[firstSet + i] = set;
 621          cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 622       }
 623
 624       if (set_layout->dynamic_offset_count > 0) {
 625          anv_foreach_stage(s, set_layout->shader_stages) {
 626             anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, s, dynamic);
 627
 628             struct anv_push_constants *push =
 629                cmd_buffer->state.push_constants[s];
 630
 631             unsigned d = layout->set[firstSet + i].dynamic_offset_start;
 632             const uint32_t *offsets = pDynamicOffsets + dynamic_slot;
 633             struct anv_descriptor *desc = set->descriptors;
 634
 635             for (unsigned b = 0; b < set_layout->binding_count; b++) {
 636                if (set_layout->binding[b].dynamic_offset_index < 0)
 637                   continue;
 638
 639                unsigned array_size = set_layout->binding[b].array_size;
 640                for (unsigned j = 0; j < array_size; j++) {
 641                   uint32_t range = 0;
 642                   if (desc->buffer_view)
 643                      range = desc->buffer_view->range;
 644                   push->dynamic[d].offset = *(offsets++);
 645                   push->dynamic[d].range = range;
 646                   desc++;
 647                   d++;
 648                }
 649             }
 650          }
 651          cmd_buffer->state.push_constants_dirty |= set_layout->shader_stages;
 652       }
 653    }
 654 }
 655
 656 void anv_CmdBindVertexBuffers(
 657     VkCommandBuffer                             commandBuffer,
 658     uint32_t                                    firstBinding,
 659     uint32_t                                    bindingCount,
 660     const VkBuffer*                             pBuffers,
 661     const VkDeviceSize*                         pOffsets)
 662 {
 663    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 664    struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 665
 666    /* We have to defer setting up vertex buffer since we need the buffer
 667     * stride from the pipeline. */
 668
 669    assert(firstBinding + bindingCount < MAX_VBS);
 670    for (uint32_t i = 0; i < bindingCount; i++) {
 671       vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
 672       vb[firstBinding + i].offset = pOffsets[i];
 673       cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i);
 674    }
 675 }
 676
 677 static void
 678 add_surface_state_reloc(struct anv_cmd_buffer *cmd_buffer,
 679                         struct anv_state state, struct anv_bo *bo, uint32_t offset)
 680 {
 681    /* The address goes in SURFACE_STATE dword 1 for gens < 8 and dwords 8 and
 682     * 9 for gen8+.  We only write the first dword for gen8+ here and rely on
 683     * the initial state to set the high bits to 0. */
 684
 685    const uint32_t dword = cmd_buffer->device->info.gen < 8 ? 1 : 8;
 686
 687    anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
 688                       state.offset + dword * 4, bo, offset);
 689 }
 690
 691 const struct anv_format *
 692 anv_format_for_descriptor_type(VkDescriptorType type)
 693 {
 694    switch (type) {
 695    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 696    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 697       return anv_format_for_vk_format(VK_FORMAT_R32G32B32A32_SFLOAT);
 698
 699    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 700    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 701       return anv_format_for_vk_format(VK_FORMAT_UNDEFINED);
 702
 703    default:
 704       unreachable("Invalid descriptor type");
 705    }
 706 }
 707
 708 static struct anv_state
 709 anv_cmd_buffer_alloc_null_surface_state(struct anv_cmd_buffer *cmd_buffer,
 710                                         struct anv_framebuffer *fb)
 711 {
 712    switch (cmd_buffer->device->info.gen) {
 713    case 7:
 714       if (cmd_buffer->device->info.is_haswell) {
 715          return gen75_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 716       } else {
 717          return gen7_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 718       }
 719    case 8:
 720       return gen8_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 721    case 9:
 722       return gen9_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 723    default:
 724       unreachable("Invalid hardware generation");
 725    }
 726 }
 727
 728 VkResult
 729 anv_cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
 730                                   gl_shader_stage stage,
 731                                   struct anv_state *bt_state)
 732 {
 733    struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 734    struct anv_subpass *subpass = cmd_buffer->state.subpass;
 735    struct anv_pipeline_bind_map *map;
 736    uint32_t bias, state_offset;
 737
 738    switch (stage) {
 739    case  MESA_SHADER_COMPUTE:
 740       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 741       bias = 1;
 742       break;
 743    default:
 744       map = &cmd_buffer->state.pipeline->bindings[stage];
 745       bias = 0;
 746       break;
 747    }
 748
 749    if (bias + map->surface_count == 0) {
 750       *bt_state = (struct anv_state) { 0, };
 751       return VK_SUCCESS;
 752    }
 753
 754    *bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer,
 755                                                   bias + map->surface_count,
 756                                                   &state_offset);
 757    uint32_t *bt_map = bt_state->map;
 758
 759    if (bt_state->map == NULL)
 760       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 761
 762    if (stage == MESA_SHADER_COMPUTE &&
 763        get_cs_prog_data(cmd_buffer->state.compute_pipeline)->uses_num_work_groups) {
 764       struct anv_bo *bo = cmd_buffer->state.num_workgroups_bo;
 765       uint32_t bo_offset = cmd_buffer->state.num_workgroups_offset;
 766
 767       struct anv_state surface_state;
 768       surface_state =
 769          anv_cmd_buffer_alloc_surface_state(cmd_buffer);
 770
 771       const struct anv_format *format =
 772          anv_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
 773       anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
 774                                     format->isl_format, bo_offset, 12, 1);
 775
 776       bt_map[0] = surface_state.offset + state_offset;
 777       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 778    }
 779
 780    if (map->surface_count == 0)
 781       goto out;
 782
 783    if (map->image_count > 0) {
 784       VkResult result =
 785          anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images);
 786       if (result != VK_SUCCESS)
 787          return result;
 788
 789       cmd_buffer->state.push_constants_dirty |= 1 << stage;
 790    }
 791
 792    uint32_t image = 0;
 793    for (uint32_t s = 0; s < map->surface_count; s++) {
 794       struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];
 795
 796       struct anv_state surface_state;
 797       struct anv_bo *bo;
 798       uint32_t bo_offset;
 799
 800       if (binding->set == ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS) {
 801          /* Color attachment binding */
 802          assert(stage == MESA_SHADER_FRAGMENT);
 803          if (binding->offset < subpass->color_count) {
 804             const struct anv_image_view *iview =
 805                fb->attachments[subpass->color_attachments[binding->offset]];
 806
 807             assert(iview->color_rt_surface_state.alloc_size);
 808             surface_state = iview->color_rt_surface_state;
 809             add_surface_state_reloc(cmd_buffer, iview->color_rt_surface_state,
 810                                     iview->bo, iview->offset);
 811          } else {
 812             /* Null render target */
 813             struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 814             surface_state =
 815                anv_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 816          }
 817
 818          bt_map[bias + s] = surface_state.offset + state_offset;
 819          continue;
 820       }
 821
 822       struct anv_descriptor_set *set =
 823          cmd_buffer->state.descriptors[binding->set];
 824       struct anv_descriptor *desc = &set->descriptors[binding->offset];
 825
 826       switch (desc->type) {
 827       case VK_DESCRIPTOR_TYPE_SAMPLER:
 828          /* Nothing for us to do here */
 829          continue;
 830
 831       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 832       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 833       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 834          surface_state = desc->image_view->sampler_surface_state;
 835          assert(surface_state.alloc_size);
 836          bo = desc->image_view->bo;
 837          bo_offset = desc->image_view->offset;
 838          break;
 839
 840       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
 841          surface_state = desc->image_view->storage_surface_state;
 842          assert(surface_state.alloc_size);
 843          bo = desc->image_view->bo;
 844          bo_offset = desc->image_view->offset;
 845
 846          struct brw_image_param *image_param =
 847             &cmd_buffer->state.push_constants[stage]->images[image++];
 848
 849          *image_param = desc->image_view->storage_image_param;
 850          image_param->surface_idx = bias + s;
 851          break;
 852       }
 853
 854       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 855       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 856       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 857       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 858       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
 859          surface_state = desc->buffer_view->surface_state;
 860          assert(surface_state.alloc_size);
 861          bo = desc->buffer_view->bo;
 862          bo_offset = desc->buffer_view->offset;
 863          break;
 864
 865       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
 866          surface_state = desc->buffer_view->storage_surface_state;
 867          assert(surface_state.alloc_size);
 868          bo = desc->buffer_view->bo;
 869          bo_offset = desc->buffer_view->offset;
 870
 871          struct brw_image_param *image_param =
 872             &cmd_buffer->state.push_constants[stage]->images[image++];
 873
 874          *image_param = desc->buffer_view->storage_image_param;
 875          image_param->surface_idx = bias + s;
 876          break;
 877
 878       default:
 879          assert(!"Invalid descriptor type");
 880          continue;
 881       }
 882
 883       bt_map[bias + s] = surface_state.offset + state_offset;
 884       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 885    }
 886    assert(image == map->image_count);
 887
 888  out:
 889    if (!cmd_buffer->device->info.has_llc)
 890       anv_state_clflush(*bt_state);
 891
 892    return VK_SUCCESS;
 893 }
 894
 895 VkResult
 896 anv_cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer,
 897                              gl_shader_stage stage, struct anv_state *state)
 898 {
 899    struct anv_pipeline_bind_map *map;
 900
 901    if (stage == MESA_SHADER_COMPUTE)
 902       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 903    else
 904       map = &cmd_buffer->state.pipeline->bindings[stage];
 905
 906    if (map->sampler_count == 0) {
 907       *state = (struct anv_state) { 0, };
 908       return VK_SUCCESS;
 909    }
 910
 911    uint32_t size = map->sampler_count * 16;
 912    *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32);
 913
 914    if (state->map == NULL)
 915       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 916
 917    for (uint32_t s = 0; s < map->sampler_count; s++) {
 918       struct anv_pipeline_binding *binding = &map->sampler_to_descriptor[s];
 919       struct anv_descriptor_set *set =
 920          cmd_buffer->state.descriptors[binding->set];
 921       struct anv_descriptor *desc = &set->descriptors[binding->offset];
 922
 923       if (desc->type != VK_DESCRIPTOR_TYPE_SAMPLER &&
 924           desc->type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
 925          continue;
 926
 927       struct anv_sampler *sampler = desc->sampler;
 928
 929       /* This can happen if we have an unfilled slot since TYPE_SAMPLER
 930        * happens to be zero.
 931        */
 932       if (sampler == NULL)
 933          continue;
 934
 935       memcpy(state->map + (s * 16),
 936              sampler->state, sizeof(sampler->state));
 937    }
 938
 939    if (!cmd_buffer->device->info.has_llc)
 940       anv_state_clflush(*state);
 941
 942    return VK_SUCCESS;
 943 }
 944
 945 struct anv_state
 946 anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
 947                             const void *data, uint32_t size, uint32_t alignment)
 948 {
 949    struct anv_state state;
 950
 951    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment);
 952    memcpy(state.map, data, size);
 953
 954    if (!cmd_buffer->device->info.has_llc)
 955       anv_state_clflush(state);
 956
 957    VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size));
 958
 959    return state;
 960 }
 961
 962 struct anv_state
 963 anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
 964                              uint32_t *a, uint32_t *b,
 965                              uint32_t dwords, uint32_t alignment)
 966 {
 967    struct anv_state state;
 968    uint32_t *p;
 969
 970    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 971                                               dwords * 4, alignment);
 972    p = state.map;
 973    for (uint32_t i = 0; i < dwords; i++)
 974       p[i] = a[i] | b[i];
 975
 976    if (!cmd_buffer->device->info.has_llc)
 977       anv_state_clflush(state);
 978
 979    VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
 980
 981    return state;
 982 }
 983
 984 /**
 985  * @brief Setup the command buffer for recording commands inside the given
 986  * subpass.
 987  *
 988  * This does not record all commands needed for starting the subpass.
 989  * Starting the subpass may require additional commands.
 990  *
 991  * Note that vkCmdBeginRenderPass, vkCmdNextSubpass, and vkBeginCommandBuffer
 992  * with VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT, all setup the
 993  * command buffer for recording commands for some subpass.  But only the first
 994  * two, vkCmdBeginRenderPass and vkCmdNextSubpass, can start a subpass.
 995  */
 996 void
 997 anv_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
 998                            struct anv_subpass *subpass)
 999 {
1000    switch (cmd_buffer->device->info.gen) {
1001    case 7:
1002       if (cmd_buffer->device->info.is_haswell) {
1003          gen75_cmd_buffer_set_subpass(cmd_buffer, subpass);
1004       } else {
1005          gen7_cmd_buffer_set_subpass(cmd_buffer, subpass);
1006       }
1007       break;
1008    case 8:
1009       gen8_cmd_buffer_set_subpass(cmd_buffer, subpass);
1010       break;
1011    case 9:
1012       gen9_cmd_buffer_set_subpass(cmd_buffer, subpass);
1013       break;
1014    default:
1015       unreachable("unsupported gen\n");
1016    }
1017 }
1018
1019 struct anv_state
1020 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
1021                               gl_shader_stage stage)
1022 {
1023    struct anv_push_constants *data =
1024       cmd_buffer->state.push_constants[stage];
1025    const struct brw_stage_prog_data *prog_data =
1026       cmd_buffer->state.pipeline->prog_data[stage];
1027
1028    /* If we don't actually have any push constants, bail. */
1029    if (data == NULL || prog_data->nr_params == 0)
1030       return (struct anv_state) { .offset = 0 };
1031
1032    struct anv_state state =
1033       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1034                                          prog_data->nr_params * sizeof(float),
1035                                          32 /* bottom 5 bits MBZ */);
1036
1037    /* Walk through the param array and fill the buffer with data */
1038    uint32_t *u32_map = state.map;
1039    for (unsigned i = 0; i < prog_data->nr_params; i++) {
1040       uint32_t offset = (uintptr_t)prog_data->param[i];
1041       u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
1042    }
1043
1044    if (!cmd_buffer->device->info.has_llc)
1045       anv_state_clflush(state);
1046
1047    return state;
1048 }
1049
1050 struct anv_state
1051 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
1052 {
1053    struct anv_push_constants *data =
1054       cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
1055    struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
1056    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
1057    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
1058
1059    const unsigned local_id_dwords = cs_prog_data->local_invocation_id_regs * 8;
1060    const unsigned push_constant_data_size =
1061       (local_id_dwords + prog_data->nr_params) * 4;
1062    const unsigned reg_aligned_constant_size = ALIGN(push_constant_data_size, 32);
1063    const unsigned param_aligned_count =
1064       reg_aligned_constant_size / sizeof(uint32_t);
1065
1066    /* If we don't actually have any push constants, bail. */
1067    if (reg_aligned_constant_size == 0)
1068       return (struct anv_state) { .offset = 0 };
1069
1070    const unsigned threads = pipeline->cs_thread_width_max;
1071    const unsigned total_push_constants_size =
1072       reg_aligned_constant_size * threads;
1073    const unsigned push_constant_alignment =
1074       cmd_buffer->device->info.gen < 8 ? 32 : 64;
1075    const unsigned aligned_total_push_constants_size =
1076       ALIGN(total_push_constants_size, push_constant_alignment);
1077    struct anv_state state =
1078       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1079                                          aligned_total_push_constants_size,
1080                                          push_constant_alignment);
1081
1082    /* Walk through the param array and fill the buffer with data */
1083    uint32_t *u32_map = state.map;
1084
1085    brw_cs_fill_local_id_payload(cs_prog_data, u32_map, threads,
1086                                 reg_aligned_constant_size);
1087
1088    /* Setup uniform data for the first thread */
1089    for (unsigned i = 0; i < prog_data->nr_params; i++) {
1090       uint32_t offset = (uintptr_t)prog_data->param[i];
1091       u32_map[local_id_dwords + i] = *(uint32_t *)((uint8_t *)data + offset);
1092    }
1093
1094    /* Copy uniform data from the first thread to every other thread */
1095    const size_t uniform_data_size = prog_data->nr_params * sizeof(uint32_t);
1096    for (unsigned t = 1; t < threads; t++) {
1097       memcpy(&u32_map[t * param_aligned_count + local_id_dwords],
1098              &u32_map[local_id_dwords],
1099              uniform_data_size);
1100    }
1101
1102    if (!cmd_buffer->device->info.has_llc)
1103       anv_state_clflush(state);
1104
1105    return state;
1106 }
1107
1108 void anv_CmdPushConstants(
1109     VkCommandBuffer                             commandBuffer,
1110     VkPipelineLayout                            layout,
1111     VkShaderStageFlags                          stageFlags,
1112     uint32_t                                    offset,
1113     uint32_t                                    size,
1114     const void*                                 pValues)
1115 {
1116    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1117
1118    anv_foreach_stage(stage, stageFlags) {
1119       anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, client_data);
1120
1121       memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
1122              pValues, size);
1123    }
1124
1125    cmd_buffer->state.push_constants_dirty |= stageFlags;
1126 }
1127
1128 void anv_CmdExecuteCommands(
1129     VkCommandBuffer                             commandBuffer,
1130     uint32_t                                    commandBufferCount,
1131     const VkCommandBuffer*                      pCmdBuffers)
1132 {
1133    ANV_FROM_HANDLE(anv_cmd_buffer, primary, commandBuffer);
1134
1135    assert(primary->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1136
1137    for (uint32_t i = 0; i < commandBufferCount; i++) {
1138       ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);
1139
1140       assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
1141
1142       anv_cmd_buffer_add_secondary(primary, secondary);
1143    }
1144 }
1145
1146 VkResult anv_CreateCommandPool(
1147     VkDevice                                    _device,
1148     const VkCommandPoolCreateInfo*              pCreateInfo,
1149     const VkAllocationCallbacks*                pAllocator,
1150     VkCommandPool*                              pCmdPool)
1151 {
1152    ANV_FROM_HANDLE(anv_device, device, _device);
1153    struct anv_cmd_pool *pool;
1154
1155    pool = anv_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
1156                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1157    if (pool == NULL)
1158       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1159
1160    if (pAllocator)
1161       pool->alloc = *pAllocator;
1162    else
1163       pool->alloc = device->alloc;
1164
1165    list_inithead(&pool->cmd_buffers);
1166
1167    *pCmdPool = anv_cmd_pool_to_handle(pool);
1168
1169    return VK_SUCCESS;
1170 }
1171
1172 void anv_DestroyCommandPool(
1173     VkDevice                                    _device,
1174     VkCommandPool                               commandPool,
1175     const VkAllocationCallbacks*                pAllocator)
1176 {
1177    ANV_FROM_HANDLE(anv_device, device, _device);
1178    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1179
1180    anv_ResetCommandPool(_device, commandPool, 0);
1181
1182    anv_free2(&device->alloc, pAllocator, pool);
1183 }
1184
1185 VkResult anv_ResetCommandPool(
1186     VkDevice                                    device,
1187     VkCommandPool                               commandPool,
1188     VkCommandPoolResetFlags                     flags)
1189 {
1190    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1191
1192    /* FIXME: vkResetCommandPool must not destroy its command buffers. The
1193     * Vulkan 1.0 spec requires that it only reset them:
1194     *
1195     *    Resetting a command pool recycles all of the resources from all of
1196     *    the command buffers allocated from the command pool back to the
1197     *    command pool. All command buffers that have been allocated from the
1198     *    command pool are put in the initial state.
1199     */
1200    list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
1201                             &pool->cmd_buffers, pool_link) {
1202       anv_cmd_buffer_destroy(cmd_buffer);
1203    }
1204
1205    return VK_SUCCESS;
1206 }
1207
1208 /**
1209  * Return NULL if the current subpass has no depthstencil attachment.
1210  */
1211 const struct anv_image_view *
1212 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
1213 {
1214    const struct anv_subpass *subpass = cmd_buffer->state.subpass;
1215    const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
1216
1217    if (subpass->depth_stencil_attachment == VK_ATTACHMENT_UNUSED)
1218       return NULL;
1219
1220    const struct anv_image_view *iview =
1221       fb->attachments[subpass->depth_stencil_attachment];
1222
1223    assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
1224                                 VK_IMAGE_ASPECT_STENCIL_BIT));
1225
1226    return iview;
1227 }