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