src/intel/vulkan/anv_cmd_buffer.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25 #include <stdbool.h>
  26 #include <string.h>
  27 #include <unistd.h>
  28 #include <fcntl.h>
  29
  30 #include "anv_private.h"
  31
  32 #include "vk_format_info.h"
  33
  34 /** \file anv_cmd_buffer.c
  35  *
  36  * This file contains all of the stuff for emitting commands into a command
  37  * buffer.  This includes implementations of most of the vkCmd*
  38  * entrypoints.  This file is concerned entirely with state emission and
  39  * not with the command buffer data structure itself.  As far as this file
  40  * is concerned, most of anv_cmd_buffer is magic.
  41  */
  42
  43 /* TODO: These are taken from GLES.  We should check the Vulkan spec */
  44 const struct anv_dynamic_state default_dynamic_state = {
  45    .viewport = {
  46       .count = 0,
  47    },
  48    .scissor = {
  49       .count = 0,
  50    },
  51    .line_width = 1.0f,
  52    .depth_bias = {
  53       .bias = 0.0f,
  54       .clamp = 0.0f,
  55       .slope = 0.0f,
  56    },
  57    .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
  58    .depth_bounds = {
  59       .min = 0.0f,
  60       .max = 1.0f,
  61    },
  62    .stencil_compare_mask = {
  63       .front = ~0u,
  64       .back = ~0u,
  65    },
  66    .stencil_write_mask = {
  67       .front = ~0u,
  68       .back = ~0u,
  69    },
  70    .stencil_reference = {
  71       .front = 0u,
  72       .back = 0u,
  73    },
  74 };
  75
  76 void
  77 anv_dynamic_state_copy(struct anv_dynamic_state *dest,
  78                        const struct anv_dynamic_state *src,
  79                        uint32_t copy_mask)
  80 {
  81    if (copy_mask & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
  82       dest->viewport.count = src->viewport.count;
  83       typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
  84                    src->viewport.count);
  85    }
  86
  87    if (copy_mask & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
  88       dest->scissor.count = src->scissor.count;
  89       typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
  90                    src->scissor.count);
  91    }
  92
  93    if (copy_mask & (1 << VK_DYNAMIC_STATE_LINE_WIDTH))
  94       dest->line_width = src->line_width;
  95
  96    if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS))
  97       dest->depth_bias = src->depth_bias;
  98
  99    if (copy_mask & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS))
 100       typed_memcpy(dest->blend_constants, src->blend_constants, 4);
 101
 102    if (copy_mask & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS))
 103       dest->depth_bounds = src->depth_bounds;
 104
 105    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK))
 106       dest->stencil_compare_mask = src->stencil_compare_mask;
 107
 108    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK))
 109       dest->stencil_write_mask = src->stencil_write_mask;
 110
 111    if (copy_mask & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE))
 112       dest->stencil_reference = src->stencil_reference;
 113 }
 114
 115 static void
 116 anv_cmd_state_reset(struct anv_cmd_buffer *cmd_buffer)
 117 {
 118    struct anv_cmd_state *state = &cmd_buffer->state;
 119
 120    memset(&state->descriptors, 0, sizeof(state->descriptors));
 121    memset(&state->push_constants, 0, sizeof(state->push_constants));
 122    memset(state->binding_tables, 0, sizeof(state->binding_tables));
 123    memset(state->samplers, 0, sizeof(state->samplers));
 124
 125    /* 0 isn't a valid config.  This ensures that we always configure L3$. */
 126    cmd_buffer->state.current_l3_config = 0;
 127
 128    state->dirty = 0;
 129    state->vb_dirty = 0;
 130    state->pending_pipe_bits = 0;
 131    state->descriptors_dirty = 0;
 132    state->push_constants_dirty = 0;
 133    state->pipeline = NULL;
 134    state->push_constant_stages = 0;
 135    state->restart_index = UINT32_MAX;
 136    state->dynamic = default_dynamic_state;
 137    state->need_query_wa = true;
 138
 139    if (state->attachments != NULL) {
 140       anv_free(&cmd_buffer->pool->alloc, state->attachments);
 141       state->attachments = NULL;
 142    }
 143
 144    state->gen7.index_buffer = NULL;
 145 }
 146
 147 /**
 148  * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
 149  */
 150 void
 151 anv_cmd_state_setup_attachments(struct anv_cmd_buffer *cmd_buffer,
 152                                 const VkRenderPassBeginInfo *info)
 153 {
 154    struct anv_cmd_state *state = &cmd_buffer->state;
 155    ANV_FROM_HANDLE(anv_render_pass, pass, info->renderPass);
 156
 157    anv_free(&cmd_buffer->pool->alloc, state->attachments);
 158
 159    if (pass->attachment_count == 0) {
 160       state->attachments = NULL;
 161       return;
 162    }
 163
 164    state->attachments = anv_alloc(&cmd_buffer->pool->alloc,
 165                                   pass->attachment_count *
 166                                        sizeof(state->attachments[0]),
 167                                   8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 168    if (state->attachments == NULL) {
 169       /* FIXME: Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
 170       abort();
 171    }
 172
 173    for (uint32_t i = 0; i < pass->attachment_count; ++i) {
 174       struct anv_render_pass_attachment *att = &pass->attachments[i];
 175       VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
 176       VkImageAspectFlags clear_aspects = 0;
 177
 178       if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
 179          /* color attachment */
 180          if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 181             clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
 182          }
 183       } else {
 184          /* depthstencil attachment */
 185          if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
 186              att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 187             clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
 188          }
 189          if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
 190              att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
 191             clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
 192          }
 193       }
 194
 195       state->attachments[i].pending_clear_aspects = clear_aspects;
 196       if (clear_aspects) {
 197          assert(info->clearValueCount > i);
 198          state->attachments[i].clear_value = info->pClearValues[i];
 199       }
 200    }
 201 }
 202
 203 static VkResult
 204 anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer,
 205                                           gl_shader_stage stage, uint32_t size)
 206 {
 207    struct anv_push_constants **ptr = &cmd_buffer->state.push_constants[stage];
 208
 209    if (*ptr == NULL) {
 210       *ptr = anv_alloc(&cmd_buffer->pool->alloc, size, 8,
 211                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 212       if (*ptr == NULL)
 213          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 214    } else if ((*ptr)->size < size) {
 215       *ptr = anv_realloc(&cmd_buffer->pool->alloc, *ptr, size, 8,
 216                          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 217       if (*ptr == NULL)
 218          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 219    }
 220    (*ptr)->size = size;
 221
 222    return VK_SUCCESS;
 223 }
 224
 225 #define anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, field) \
 226    anv_cmd_buffer_ensure_push_constants_size(cmd_buffer, stage, \
 227       (offsetof(struct anv_push_constants, field) + \
 228        sizeof(cmd_buffer->state.push_constants[0]->field)))
 229
 230 static VkResult anv_create_cmd_buffer(
 231     struct anv_device *                         device,
 232     struct anv_cmd_pool *                       pool,
 233     VkCommandBufferLevel                        level,
 234     VkCommandBuffer*                            pCommandBuffer)
 235 {
 236    struct anv_cmd_buffer *cmd_buffer;
 237    VkResult result;
 238
 239    cmd_buffer = anv_alloc(&pool->alloc, sizeof(*cmd_buffer), 8,
 240                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 241    if (cmd_buffer == NULL)
 242       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 243
 244    cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 245    cmd_buffer->device = device;
 246    cmd_buffer->pool = pool;
 247    cmd_buffer->level = level;
 248    cmd_buffer->state.attachments = NULL;
 249
 250    result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer);
 251    if (result != VK_SUCCESS)
 252       goto fail;
 253
 254    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 255                          &device->surface_state_block_pool);
 256    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 257                          &device->dynamic_state_block_pool);
 258
 259    if (pool) {
 260       list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
 261    } else {
 262       /* Init the pool_link so we can safefly call list_del when we destroy
 263        * the command buffer
 264        */
 265       list_inithead(&cmd_buffer->pool_link);
 266    }
 267
 268    *pCommandBuffer = anv_cmd_buffer_to_handle(cmd_buffer);
 269
 270    return VK_SUCCESS;
 271
 272  fail:
 273    anv_free(&cmd_buffer->pool->alloc, cmd_buffer);
 274
 275    return result;
 276 }
 277
 278 VkResult anv_AllocateCommandBuffers(
 279     VkDevice                                    _device,
 280     const VkCommandBufferAllocateInfo*          pAllocateInfo,
 281     VkCommandBuffer*                            pCommandBuffers)
 282 {
 283    ANV_FROM_HANDLE(anv_device, device, _device);
 284    ANV_FROM_HANDLE(anv_cmd_pool, pool, pAllocateInfo->commandPool);
 285
 286    VkResult result = VK_SUCCESS;
 287    uint32_t i;
 288
 289    for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
 290       result = anv_create_cmd_buffer(device, pool, pAllocateInfo->level,
 291                                      &pCommandBuffers[i]);
 292       if (result != VK_SUCCESS)
 293          break;
 294    }
 295
 296    if (result != VK_SUCCESS)
 297       anv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
 298                              i, pCommandBuffers);
 299
 300    return result;
 301 }
 302
 303 static void
 304 anv_cmd_buffer_destroy(struct anv_cmd_buffer *cmd_buffer)
 305 {
 306    list_del(&cmd_buffer->pool_link);
 307
 308    anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer);
 309
 310    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 311    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 312
 313    anv_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
 314    anv_free(&cmd_buffer->pool->alloc, cmd_buffer);
 315 }
 316
 317 void anv_FreeCommandBuffers(
 318     VkDevice                                    device,
 319     VkCommandPool                               commandPool,
 320     uint32_t                                    commandBufferCount,
 321     const VkCommandBuffer*                      pCommandBuffers)
 322 {
 323    for (uint32_t i = 0; i < commandBufferCount; i++) {
 324       ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
 325
 326       anv_cmd_buffer_destroy(cmd_buffer);
 327    }
 328 }
 329
 330 static VkResult
 331 anv_cmd_buffer_reset(struct anv_cmd_buffer *cmd_buffer)
 332 {
 333    cmd_buffer->usage_flags = 0;
 334    cmd_buffer->state.current_pipeline = UINT32_MAX;
 335    anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer);
 336    anv_cmd_state_reset(cmd_buffer);
 337
 338    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 339    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 340                          &cmd_buffer->device->surface_state_block_pool);
 341
 342    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 343    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 344                          &cmd_buffer->device->dynamic_state_block_pool);
 345    return VK_SUCCESS;
 346 }
 347
 348 VkResult anv_ResetCommandBuffer(
 349     VkCommandBuffer                             commandBuffer,
 350     VkCommandBufferResetFlags                   flags)
 351 {
 352    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 353    return anv_cmd_buffer_reset(cmd_buffer);
 354 }
 355
 356 void
 357 anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer)
 358 {
 359    switch (cmd_buffer->device->info.gen) {
 360    case 7:
 361       if (cmd_buffer->device->info.is_haswell)
 362          return gen7_cmd_buffer_emit_state_base_address(cmd_buffer);
 363       else
 364          return gen7_cmd_buffer_emit_state_base_address(cmd_buffer);
 365    case 8:
 366       return gen8_cmd_buffer_emit_state_base_address(cmd_buffer);
 367    case 9:
 368       return gen9_cmd_buffer_emit_state_base_address(cmd_buffer);
 369    default:
 370       unreachable("unsupported gen\n");
 371    }
 372 }
 373
 374 VkResult anv_BeginCommandBuffer(
 375     VkCommandBuffer                             commandBuffer,
 376     const VkCommandBufferBeginInfo*             pBeginInfo)
 377 {
 378    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 379
 380    /* If this is the first vkBeginCommandBuffer, we must *initialize* the
 381     * command buffer's state. Otherwise, we must *reset* its state. In both
 382     * cases we reset it.
 383     *
 384     * From the Vulkan 1.0 spec:
 385     *
 386     *    If a command buffer is in the executable state and the command buffer
 387     *    was allocated from a command pool with the
 388     *    VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then
 389     *    vkBeginCommandBuffer implicitly resets the command buffer, behaving
 390     *    as if vkResetCommandBuffer had been called with
 391     *    VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts
 392     *    the command buffer in the recording state.
 393     */
 394    anv_cmd_buffer_reset(cmd_buffer);
 395
 396    cmd_buffer->usage_flags = pBeginInfo->flags;
 397
 398    assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY ||
 399           !(cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT));
 400
 401    anv_cmd_buffer_emit_state_base_address(cmd_buffer);
 402
 403    if (cmd_buffer->usage_flags &
 404        VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
 405       cmd_buffer->state.framebuffer =
 406          anv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);
 407       cmd_buffer->state.pass =
 408          anv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
 409
 410       struct anv_subpass *subpass =
 411          &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
 412
 413       anv_cmd_buffer_set_subpass(cmd_buffer, subpass);
 414    }
 415
 416    return VK_SUCCESS;
 417 }
 418
 419 VkResult anv_EndCommandBuffer(
 420     VkCommandBuffer                             commandBuffer)
 421 {
 422    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 423    struct anv_device *device = cmd_buffer->device;
 424
 425    anv_cmd_buffer_end_batch_buffer(cmd_buffer);
 426
 427    if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
 428       /* The algorithm used to compute the validate list is not threadsafe as
 429        * it uses the bo->index field.  We have to lock the device around it.
 430        * Fortunately, the chances for contention here are probably very low.
 431        */
 432       pthread_mutex_lock(&device->mutex);
 433       anv_cmd_buffer_prepare_execbuf(cmd_buffer);
 434       pthread_mutex_unlock(&device->mutex);
 435    }
 436
 437    return VK_SUCCESS;
 438 }
 439
 440 void anv_CmdBindPipeline(
 441     VkCommandBuffer                             commandBuffer,
 442     VkPipelineBindPoint                         pipelineBindPoint,
 443     VkPipeline                                  _pipeline)
 444 {
 445    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 446    ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 447
 448    switch (pipelineBindPoint) {
 449    case VK_PIPELINE_BIND_POINT_COMPUTE:
 450       cmd_buffer->state.compute_pipeline = pipeline;
 451       cmd_buffer->state.compute_dirty |= ANV_CMD_DIRTY_PIPELINE;
 452       cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 453       cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 454       break;
 455
 456    case VK_PIPELINE_BIND_POINT_GRAPHICS:
 457       cmd_buffer->state.pipeline = pipeline;
 458       cmd_buffer->state.vb_dirty |= pipeline->vb_used;
 459       cmd_buffer->state.dirty |= ANV_CMD_DIRTY_PIPELINE;
 460       cmd_buffer->state.push_constants_dirty |= pipeline->active_stages;
 461       cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;
 462
 463       /* Apply the dynamic state from the pipeline */
 464       cmd_buffer->state.dirty |= pipeline->dynamic_state_mask;
 465       anv_dynamic_state_copy(&cmd_buffer->state.dynamic,
 466                              &pipeline->dynamic_state,
 467                              pipeline->dynamic_state_mask);
 468       break;
 469
 470    default:
 471       assert(!"invalid bind point");
 472       break;
 473    }
 474 }
 475
 476 void anv_CmdSetViewport(
 477     VkCommandBuffer                             commandBuffer,
 478     uint32_t                                    firstViewport,
 479     uint32_t                                    viewportCount,
 480     const VkViewport*                           pViewports)
 481 {
 482    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 483
 484    const uint32_t total_count = firstViewport + viewportCount;
 485    if (cmd_buffer->state.dynamic.viewport.count < total_count)
 486       cmd_buffer->state.dynamic.viewport.count = total_count;
 487
 488    memcpy(cmd_buffer->state.dynamic.viewport.viewports + firstViewport,
 489           pViewports, viewportCount * sizeof(*pViewports));
 490
 491    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT;
 492 }
 493
 494 void anv_CmdSetScissor(
 495     VkCommandBuffer                             commandBuffer,
 496     uint32_t                                    firstScissor,
 497     uint32_t                                    scissorCount,
 498     const VkRect2D*                             pScissors)
 499 {
 500    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 501
 502    const uint32_t total_count = firstScissor + scissorCount;
 503    if (cmd_buffer->state.dynamic.scissor.count < total_count)
 504       cmd_buffer->state.dynamic.scissor.count = total_count;
 505
 506    memcpy(cmd_buffer->state.dynamic.scissor.scissors + firstScissor,
 507           pScissors, scissorCount * sizeof(*pScissors));
 508
 509    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR;
 510 }
 511
 512 void anv_CmdSetLineWidth(
 513     VkCommandBuffer                             commandBuffer,
 514     float                                       lineWidth)
 515 {
 516    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 517
 518    cmd_buffer->state.dynamic.line_width = lineWidth;
 519    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
 520 }
 521
 522 void anv_CmdSetDepthBias(
 523     VkCommandBuffer                             commandBuffer,
 524     float                                       depthBiasConstantFactor,
 525     float                                       depthBiasClamp,
 526     float                                       depthBiasSlopeFactor)
 527 {
 528    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 529
 530    cmd_buffer->state.dynamic.depth_bias.bias = depthBiasConstantFactor;
 531    cmd_buffer->state.dynamic.depth_bias.clamp = depthBiasClamp;
 532    cmd_buffer->state.dynamic.depth_bias.slope = depthBiasSlopeFactor;
 533
 534    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
 535 }
 536
 537 void anv_CmdSetBlendConstants(
 538     VkCommandBuffer                             commandBuffer,
 539     const float                                 blendConstants[4])
 540 {
 541    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 542
 543    memcpy(cmd_buffer->state.dynamic.blend_constants,
 544           blendConstants, sizeof(float) * 4);
 545
 546    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
 547 }
 548
 549 void anv_CmdSetDepthBounds(
 550     VkCommandBuffer                             commandBuffer,
 551     float                                       minDepthBounds,
 552     float                                       maxDepthBounds)
 553 {
 554    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 555
 556    cmd_buffer->state.dynamic.depth_bounds.min = minDepthBounds;
 557    cmd_buffer->state.dynamic.depth_bounds.max = maxDepthBounds;
 558
 559    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
 560 }
 561
 562 void anv_CmdSetStencilCompareMask(
 563     VkCommandBuffer                             commandBuffer,
 564     VkStencilFaceFlags                          faceMask,
 565     uint32_t                                    compareMask)
 566 {
 567    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 568
 569    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 570       cmd_buffer->state.dynamic.stencil_compare_mask.front = compareMask;
 571    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 572       cmd_buffer->state.dynamic.stencil_compare_mask.back = compareMask;
 573
 574    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
 575 }
 576
 577 void anv_CmdSetStencilWriteMask(
 578     VkCommandBuffer                             commandBuffer,
 579     VkStencilFaceFlags                          faceMask,
 580     uint32_t                                    writeMask)
 581 {
 582    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 583
 584    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 585       cmd_buffer->state.dynamic.stencil_write_mask.front = writeMask;
 586    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 587       cmd_buffer->state.dynamic.stencil_write_mask.back = writeMask;
 588
 589    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
 590 }
 591
 592 void anv_CmdSetStencilReference(
 593     VkCommandBuffer                             commandBuffer,
 594     VkStencilFaceFlags                          faceMask,
 595     uint32_t                                    reference)
 596 {
 597    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 598
 599    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 600       cmd_buffer->state.dynamic.stencil_reference.front = reference;
 601    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 602       cmd_buffer->state.dynamic.stencil_reference.back = reference;
 603
 604    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
 605 }
 606
 607 void anv_CmdBindDescriptorSets(
 608     VkCommandBuffer                             commandBuffer,
 609     VkPipelineBindPoint                         pipelineBindPoint,
 610     VkPipelineLayout                            _layout,
 611     uint32_t                                    firstSet,
 612     uint32_t                                    descriptorSetCount,
 613     const VkDescriptorSet*                      pDescriptorSets,
 614     uint32_t                                    dynamicOffsetCount,
 615     const uint32_t*                             pDynamicOffsets)
 616 {
 617    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 618    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 619    struct anv_descriptor_set_layout *set_layout;
 620
 621    assert(firstSet + descriptorSetCount < MAX_SETS);
 622
 623    uint32_t dynamic_slot = 0;
 624    for (uint32_t i = 0; i < descriptorSetCount; i++) {
 625       ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
 626       set_layout = layout->set[firstSet + i].layout;
 627
 628       if (cmd_buffer->state.descriptors[firstSet + i] != set) {
 629          cmd_buffer->state.descriptors[firstSet + i] = set;
 630          cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 631       }
 632
 633       if (set_layout->dynamic_offset_count > 0) {
 634          anv_foreach_stage(s, set_layout->shader_stages) {
 635             anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, s, dynamic);
 636
 637             struct anv_push_constants *push =
 638                cmd_buffer->state.push_constants[s];
 639
 640             unsigned d = layout->set[firstSet + i].dynamic_offset_start;
 641             const uint32_t *offsets = pDynamicOffsets + dynamic_slot;
 642             struct anv_descriptor *desc = set->descriptors;
 643
 644             for (unsigned b = 0; b < set_layout->binding_count; b++) {
 645                if (set_layout->binding[b].dynamic_offset_index < 0)
 646                   continue;
 647
 648                unsigned array_size = set_layout->binding[b].array_size;
 649                for (unsigned j = 0; j < array_size; j++) {
 650                   uint32_t range = 0;
 651                   if (desc->buffer_view)
 652                      range = desc->buffer_view->range;
 653                   push->dynamic[d].offset = *(offsets++);
 654                   push->dynamic[d].range = range;
 655                   desc++;
 656                   d++;
 657                }
 658             }
 659          }
 660          cmd_buffer->state.push_constants_dirty |= set_layout->shader_stages;
 661       }
 662    }
 663 }
 664
 665 void anv_CmdBindVertexBuffers(
 666     VkCommandBuffer                             commandBuffer,
 667     uint32_t                                    firstBinding,
 668     uint32_t                                    bindingCount,
 669     const VkBuffer*                             pBuffers,
 670     const VkDeviceSize*                         pOffsets)
 671 {
 672    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 673    struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 674
 675    /* We have to defer setting up vertex buffer since we need the buffer
 676     * stride from the pipeline. */
 677
 678    assert(firstBinding + bindingCount < MAX_VBS);
 679    for (uint32_t i = 0; i < bindingCount; i++) {
 680       vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
 681       vb[firstBinding + i].offset = pOffsets[i];
 682       cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i);
 683    }
 684 }
 685
 686 static void
 687 add_surface_state_reloc(struct anv_cmd_buffer *cmd_buffer,
 688                         struct anv_state state, struct anv_bo *bo, uint32_t offset)
 689 {
 690    /* The address goes in SURFACE_STATE dword 1 for gens < 8 and dwords 8 and
 691     * 9 for gen8+.  We only write the first dword for gen8+ here and rely on
 692     * the initial state to set the high bits to 0. */
 693
 694    const uint32_t dword = cmd_buffer->device->info.gen < 8 ? 1 : 8;
 695
 696    anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
 697                       state.offset + dword * 4, bo, offset);
 698 }
 699
 700 enum isl_format
 701 anv_isl_format_for_descriptor_type(VkDescriptorType type)
 702 {
 703    switch (type) {
 704    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 705    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 706       return ISL_FORMAT_R32G32B32A32_FLOAT;
 707
 708    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 709    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 710       return ISL_FORMAT_RAW;
 711
 712    default:
 713       unreachable("Invalid descriptor type");
 714    }
 715 }
 716
 717 static struct anv_state
 718 anv_cmd_buffer_alloc_null_surface_state(struct anv_cmd_buffer *cmd_buffer,
 719                                         struct anv_framebuffer *fb)
 720 {
 721    switch (cmd_buffer->device->info.gen) {
 722    case 7:
 723       if (cmd_buffer->device->info.is_haswell) {
 724          return gen75_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 725       } else {
 726          return gen7_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 727       }
 728    case 8:
 729       return gen8_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 730    case 9:
 731       return gen9_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 732    default:
 733       unreachable("Invalid hardware generation");
 734    }
 735 }
 736
 737 VkResult
 738 anv_cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
 739                                   gl_shader_stage stage,
 740                                   struct anv_state *bt_state)
 741 {
 742    struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 743    struct anv_subpass *subpass = cmd_buffer->state.subpass;
 744    struct anv_pipeline_bind_map *map;
 745    uint32_t bias, state_offset;
 746
 747    switch (stage) {
 748    case  MESA_SHADER_COMPUTE:
 749       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 750       bias = 1;
 751       break;
 752    default:
 753       map = &cmd_buffer->state.pipeline->bindings[stage];
 754       bias = 0;
 755       break;
 756    }
 757
 758    if (bias + map->surface_count == 0) {
 759       *bt_state = (struct anv_state) { 0, };
 760       return VK_SUCCESS;
 761    }
 762
 763    *bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer,
 764                                                   bias + map->surface_count,
 765                                                   &state_offset);
 766    uint32_t *bt_map = bt_state->map;
 767
 768    if (bt_state->map == NULL)
 769       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 770
 771    if (stage == MESA_SHADER_COMPUTE &&
 772        get_cs_prog_data(cmd_buffer->state.compute_pipeline)->uses_num_work_groups) {
 773       struct anv_bo *bo = cmd_buffer->state.num_workgroups_bo;
 774       uint32_t bo_offset = cmd_buffer->state.num_workgroups_offset;
 775
 776       struct anv_state surface_state;
 777       surface_state =
 778          anv_cmd_buffer_alloc_surface_state(cmd_buffer);
 779
 780       const enum isl_format format =
 781          anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
 782       anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
 783                                     format, bo_offset, 12, 1);
 784
 785       bt_map[0] = surface_state.offset + state_offset;
 786       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 787    }
 788
 789    if (map->surface_count == 0)
 790       goto out;
 791
 792    if (map->image_count > 0) {
 793       VkResult result =
 794          anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images);
 795       if (result != VK_SUCCESS)
 796          return result;
 797
 798       cmd_buffer->state.push_constants_dirty |= 1 << stage;
 799    }
 800
 801    uint32_t image = 0;
 802    for (uint32_t s = 0; s < map->surface_count; s++) {
 803       struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];
 804
 805       struct anv_state surface_state;
 806       struct anv_bo *bo;
 807       uint32_t bo_offset;
 808
 809       if (binding->set == ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS) {
 810          /* Color attachment binding */
 811          assert(stage == MESA_SHADER_FRAGMENT);
 812          if (binding->offset < subpass->color_count) {
 813             const struct anv_image_view *iview =
 814                fb->attachments[subpass->color_attachments[binding->offset]];
 815
 816             assert(iview->color_rt_surface_state.alloc_size);
 817             surface_state = iview->color_rt_surface_state;
 818             add_surface_state_reloc(cmd_buffer, iview->color_rt_surface_state,
 819                                     iview->bo, iview->offset);
 820          } else {
 821             /* Null render target */
 822             struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 823             surface_state =
 824                anv_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 825          }
 826
 827          bt_map[bias + s] = surface_state.offset + state_offset;
 828          continue;
 829       }
 830
 831       struct anv_descriptor_set *set =
 832          cmd_buffer->state.descriptors[binding->set];
 833       struct anv_descriptor *desc = &set->descriptors[binding->offset];
 834
 835       switch (desc->type) {
 836       case VK_DESCRIPTOR_TYPE_SAMPLER:
 837          /* Nothing for us to do here */
 838          continue;
 839
 840       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 841       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 842       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 843          surface_state = desc->image_view->sampler_surface_state;
 844          assert(surface_state.alloc_size);
 845          bo = desc->image_view->bo;
 846          bo_offset = desc->image_view->offset;
 847          break;
 848
 849       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
 850          surface_state = desc->image_view->storage_surface_state;
 851          assert(surface_state.alloc_size);
 852          bo = desc->image_view->bo;
 853          bo_offset = desc->image_view->offset;
 854
 855          struct brw_image_param *image_param =
 856             &cmd_buffer->state.push_constants[stage]->images[image++];
 857
 858          *image_param = desc->image_view->storage_image_param;
 859          image_param->surface_idx = bias + s;
 860          break;
 861       }
 862
 863       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 864       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 865       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 866       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 867       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
 868          surface_state = desc->buffer_view->surface_state;
 869          assert(surface_state.alloc_size);
 870          bo = desc->buffer_view->bo;
 871          bo_offset = desc->buffer_view->offset;
 872          break;
 873
 874       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
 875          surface_state = desc->buffer_view->storage_surface_state;
 876          assert(surface_state.alloc_size);
 877          bo = desc->buffer_view->bo;
 878          bo_offset = desc->buffer_view->offset;
 879
 880          struct brw_image_param *image_param =
 881             &cmd_buffer->state.push_constants[stage]->images[image++];
 882
 883          *image_param = desc->buffer_view->storage_image_param;
 884          image_param->surface_idx = bias + s;
 885          break;
 886
 887       default:
 888          assert(!"Invalid descriptor type");
 889          continue;
 890       }
 891
 892       bt_map[bias + s] = surface_state.offset + state_offset;
 893       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 894    }
 895    assert(image == map->image_count);
 896
 897  out:
 898    if (!cmd_buffer->device->info.has_llc)
 899       anv_state_clflush(*bt_state);
 900
 901    return VK_SUCCESS;
 902 }
 903
 904 VkResult
 905 anv_cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer,
 906                              gl_shader_stage stage, struct anv_state *state)
 907 {
 908    struct anv_pipeline_bind_map *map;
 909
 910    if (stage == MESA_SHADER_COMPUTE)
 911       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 912    else
 913       map = &cmd_buffer->state.pipeline->bindings[stage];
 914
 915    if (map->sampler_count == 0) {
 916       *state = (struct anv_state) { 0, };
 917       return VK_SUCCESS;
 918    }
 919
 920    uint32_t size = map->sampler_count * 16;
 921    *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32);
 922
 923    if (state->map == NULL)
 924       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 925
 926    for (uint32_t s = 0; s < map->sampler_count; s++) {
 927       struct anv_pipeline_binding *binding = &map->sampler_to_descriptor[s];
 928       struct anv_descriptor_set *set =
 929          cmd_buffer->state.descriptors[binding->set];
 930       struct anv_descriptor *desc = &set->descriptors[binding->offset];
 931
 932       if (desc->type != VK_DESCRIPTOR_TYPE_SAMPLER &&
 933           desc->type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
 934          continue;
 935
 936       struct anv_sampler *sampler = desc->sampler;
 937
 938       /* This can happen if we have an unfilled slot since TYPE_SAMPLER
 939        * happens to be zero.
 940        */
 941       if (sampler == NULL)
 942          continue;
 943
 944       memcpy(state->map + (s * 16),
 945              sampler->state, sizeof(sampler->state));
 946    }
 947
 948    if (!cmd_buffer->device->info.has_llc)
 949       anv_state_clflush(*state);
 950
 951    return VK_SUCCESS;
 952 }
 953
 954 struct anv_state
 955 anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
 956                             const void *data, uint32_t size, uint32_t alignment)
 957 {
 958    struct anv_state state;
 959
 960    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment);
 961    memcpy(state.map, data, size);
 962
 963    if (!cmd_buffer->device->info.has_llc)
 964       anv_state_clflush(state);
 965
 966    VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size));
 967
 968    return state;
 969 }
 970
 971 struct anv_state
 972 anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
 973                              uint32_t *a, uint32_t *b,
 974                              uint32_t dwords, uint32_t alignment)
 975 {
 976    struct anv_state state;
 977    uint32_t *p;
 978
 979    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 980                                               dwords * 4, alignment);
 981    p = state.map;
 982    for (uint32_t i = 0; i < dwords; i++)
 983       p[i] = a[i] | b[i];
 984
 985    if (!cmd_buffer->device->info.has_llc)
 986       anv_state_clflush(state);
 987
 988    VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
 989
 990    return state;
 991 }
 992
 993 /**
 994  * @brief Setup the command buffer for recording commands inside the given
 995  * subpass.
 996  *
 997  * This does not record all commands needed for starting the subpass.
 998  * Starting the subpass may require additional commands.
 999  *
1000  * Note that vkCmdBeginRenderPass, vkCmdNextSubpass, and vkBeginCommandBuffer
1001  * with VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT, all setup the
1002  * command buffer for recording commands for some subpass.  But only the first
1003  * two, vkCmdBeginRenderPass and vkCmdNextSubpass, can start a subpass.
1004  */
1005 void
1006 anv_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
1007                            struct anv_subpass *subpass)
1008 {
1009    switch (cmd_buffer->device->info.gen) {
1010    case 7:
1011       if (cmd_buffer->device->info.is_haswell) {
1012          gen75_cmd_buffer_set_subpass(cmd_buffer, subpass);
1013       } else {
1014          gen7_cmd_buffer_set_subpass(cmd_buffer, subpass);
1015       }
1016       break;
1017    case 8:
1018       gen8_cmd_buffer_set_subpass(cmd_buffer, subpass);
1019       break;
1020    case 9:
1021       gen9_cmd_buffer_set_subpass(cmd_buffer, subpass);
1022       break;
1023    default:
1024       unreachable("unsupported gen\n");
1025    }
1026 }
1027
1028 struct anv_state
1029 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
1030                               gl_shader_stage stage)
1031 {
1032    struct anv_push_constants *data =
1033       cmd_buffer->state.push_constants[stage];
1034    const struct brw_stage_prog_data *prog_data =
1035       cmd_buffer->state.pipeline->prog_data[stage];
1036
1037    /* If we don't actually have any push constants, bail. */
1038    if (data == NULL || prog_data->nr_params == 0)
1039       return (struct anv_state) { .offset = 0 };
1040
1041    struct anv_state state =
1042       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1043                                          prog_data->nr_params * sizeof(float),
1044                                          32 /* bottom 5 bits MBZ */);
1045
1046    /* Walk through the param array and fill the buffer with data */
1047    uint32_t *u32_map = state.map;
1048    for (unsigned i = 0; i < prog_data->nr_params; i++) {
1049       uint32_t offset = (uintptr_t)prog_data->param[i];
1050       u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
1051    }
1052
1053    if (!cmd_buffer->device->info.has_llc)
1054       anv_state_clflush(state);
1055
1056    return state;
1057 }
1058
1059 struct anv_state
1060 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
1061 {
1062    struct anv_push_constants *data =
1063       cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
1064    struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
1065    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
1066    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
1067
1068    /* If we don't actually have any push constants, bail. */
1069    if (cs_prog_data->push.total.size == 0)
1070       return (struct anv_state) { .offset = 0 };
1071
1072    const unsigned push_constant_alignment =
1073       cmd_buffer->device->info.gen < 8 ? 32 : 64;
1074    const unsigned aligned_total_push_constants_size =
1075       ALIGN(cs_prog_data->push.total.size, push_constant_alignment);
1076    struct anv_state state =
1077       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1078                                          aligned_total_push_constants_size,
1079                                          push_constant_alignment);
1080
1081    /* Walk through the param array and fill the buffer with data */
1082    uint32_t *u32_map = state.map;
1083
1084    if (cs_prog_data->push.cross_thread.size > 0) {
1085       assert(cs_prog_data->thread_local_id_index < 0 ||
1086              cs_prog_data->thread_local_id_index >=
1087                 cs_prog_data->push.cross_thread.dwords);
1088       for (unsigned i = 0;
1089            i < cs_prog_data->push.cross_thread.dwords;
1090            i++) {
1091          uint32_t offset = (uintptr_t)prog_data->param[i];
1092          u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
1093       }
1094    }
1095
1096    if (cs_prog_data->push.per_thread.size > 0) {
1097       brw_cs_fill_local_id_payload(cs_prog_data, u32_map, cs_prog_data->threads,
1098                                    cs_prog_data->push.per_thread.size);
1099       for (unsigned t = 0; t < cs_prog_data->threads; t++) {
1100          unsigned dst =
1101             8 * (cs_prog_data->push.per_thread.regs * t +
1102                  cs_prog_data->push.cross_thread.regs +
1103                  cs_prog_data->local_invocation_id_regs);
1104          unsigned src = cs_prog_data->push.cross_thread.dwords;
1105          for ( ; src < prog_data->nr_params; src++, dst++) {
1106             if (src != cs_prog_data->thread_local_id_index) {
1107                uint32_t offset = (uintptr_t)prog_data->param[src];
1108                u32_map[dst] = *(uint32_t *)((uint8_t *)data + offset);
1109             } else {
1110                u32_map[dst] = t * cs_prog_data->simd_size;
1111             }
1112          }
1113       }
1114    }
1115
1116    if (!cmd_buffer->device->info.has_llc)
1117       anv_state_clflush(state);
1118
1119    return state;
1120 }
1121
1122 void anv_CmdPushConstants(
1123     VkCommandBuffer                             commandBuffer,
1124     VkPipelineLayout                            layout,
1125     VkShaderStageFlags                          stageFlags,
1126     uint32_t                                    offset,
1127     uint32_t                                    size,
1128     const void*                                 pValues)
1129 {
1130    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1131
1132    anv_foreach_stage(stage, stageFlags) {
1133       anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, client_data);
1134
1135       memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
1136              pValues, size);
1137    }
1138
1139    cmd_buffer->state.push_constants_dirty |= stageFlags;
1140 }
1141
1142 void anv_CmdExecuteCommands(
1143     VkCommandBuffer                             commandBuffer,
1144     uint32_t                                    commandBufferCount,
1145     const VkCommandBuffer*                      pCmdBuffers)
1146 {
1147    ANV_FROM_HANDLE(anv_cmd_buffer, primary, commandBuffer);
1148
1149    assert(primary->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1150
1151    for (uint32_t i = 0; i < commandBufferCount; i++) {
1152       ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);
1153
1154       assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
1155
1156       anv_cmd_buffer_add_secondary(primary, secondary);
1157    }
1158 }
1159
1160 VkResult anv_CreateCommandPool(
1161     VkDevice                                    _device,
1162     const VkCommandPoolCreateInfo*              pCreateInfo,
1163     const VkAllocationCallbacks*                pAllocator,
1164     VkCommandPool*                              pCmdPool)
1165 {
1166    ANV_FROM_HANDLE(anv_device, device, _device);
1167    struct anv_cmd_pool *pool;
1168
1169    pool = anv_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
1170                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1171    if (pool == NULL)
1172       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1173
1174    if (pAllocator)
1175       pool->alloc = *pAllocator;
1176    else
1177       pool->alloc = device->alloc;
1178
1179    list_inithead(&pool->cmd_buffers);
1180
1181    *pCmdPool = anv_cmd_pool_to_handle(pool);
1182
1183    return VK_SUCCESS;
1184 }
1185
1186 void anv_DestroyCommandPool(
1187     VkDevice                                    _device,
1188     VkCommandPool                               commandPool,
1189     const VkAllocationCallbacks*                pAllocator)
1190 {
1191    ANV_FROM_HANDLE(anv_device, device, _device);
1192    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1193
1194    list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
1195                             &pool->cmd_buffers, pool_link) {
1196       anv_cmd_buffer_destroy(cmd_buffer);
1197    }
1198
1199    anv_free2(&device->alloc, pAllocator, pool);
1200 }
1201
1202 VkResult anv_ResetCommandPool(
1203     VkDevice                                    device,
1204     VkCommandPool                               commandPool,
1205     VkCommandPoolResetFlags                     flags)
1206 {
1207    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1208
1209    list_for_each_entry(struct anv_cmd_buffer, cmd_buffer,
1210                        &pool->cmd_buffers, pool_link) {
1211       anv_cmd_buffer_reset(cmd_buffer);
1212    }
1213
1214    return VK_SUCCESS;
1215 }
1216
1217 /**
1218  * Return NULL if the current subpass has no depthstencil attachment.
1219  */
1220 const struct anv_image_view *
1221 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
1222 {
1223    const struct anv_subpass *subpass = cmd_buffer->state.subpass;
1224    const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
1225
1226    if (subpass->depth_stencil_attachment == VK_ATTACHMENT_UNUSED)
1227       return NULL;
1228
1229    const struct anv_image_view *iview =
1230       fb->attachments[subpass->depth_stencil_attachment];
1231
1232    assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
1233                                 VK_IMAGE_ASPECT_STENCIL_BIT));
1234
1235    return iview;
1236 }