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    for (uint32_t i = 0; i < descriptorSetCount; i++) {
 624       ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
 625       set_layout = layout->set[firstSet + i].layout;
 626
 627       if (cmd_buffer->state.descriptors[firstSet + i] != set) {
 628          cmd_buffer->state.descriptors[firstSet + i] = set;
 629          cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 630       }
 631
 632       if (set_layout->dynamic_offset_count > 0) {
 633          anv_foreach_stage(s, set_layout->shader_stages) {
 634             anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, s, dynamic);
 635
 636             struct anv_push_constants *push =
 637                cmd_buffer->state.push_constants[s];
 638
 639             unsigned d = layout->set[firstSet + i].dynamic_offset_start;
 640             const uint32_t *offsets = pDynamicOffsets;
 641             struct anv_descriptor *desc = set->descriptors;
 642
 643             for (unsigned b = 0; b < set_layout->binding_count; b++) {
 644                if (set_layout->binding[b].dynamic_offset_index < 0)
 645                   continue;
 646
 647                unsigned array_size = set_layout->binding[b].array_size;
 648                for (unsigned j = 0; j < array_size; j++) {
 649                   uint32_t range = 0;
 650                   if (desc->buffer_view)
 651                      range = desc->buffer_view->range;
 652                   push->dynamic[d].offset = *(offsets++);
 653                   push->dynamic[d].range = range;
 654                   desc++;
 655                   d++;
 656                }
 657             }
 658          }
 659          cmd_buffer->state.push_constants_dirty |= set_layout->shader_stages;
 660       }
 661    }
 662 }
 663
 664 void anv_CmdBindVertexBuffers(
 665     VkCommandBuffer                             commandBuffer,
 666     uint32_t                                    firstBinding,
 667     uint32_t                                    bindingCount,
 668     const VkBuffer*                             pBuffers,
 669     const VkDeviceSize*                         pOffsets)
 670 {
 671    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 672    struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 673
 674    /* We have to defer setting up vertex buffer since we need the buffer
 675     * stride from the pipeline. */
 676
 677    assert(firstBinding + bindingCount < MAX_VBS);
 678    for (uint32_t i = 0; i < bindingCount; i++) {
 679       vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
 680       vb[firstBinding + i].offset = pOffsets[i];
 681       cmd_buffer->state.vb_dirty |= 1 << (firstBinding + i);
 682    }
 683 }
 684
 685 static void
 686 add_surface_state_reloc(struct anv_cmd_buffer *cmd_buffer,
 687                         struct anv_state state, struct anv_bo *bo, uint32_t offset)
 688 {
 689    /* The address goes in SURFACE_STATE dword 1 for gens < 8 and dwords 8 and
 690     * 9 for gen8+.  We only write the first dword for gen8+ here and rely on
 691     * the initial state to set the high bits to 0. */
 692
 693    const uint32_t dword = cmd_buffer->device->info.gen < 8 ? 1 : 8;
 694
 695    anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
 696                       state.offset + dword * 4, bo, offset);
 697 }
 698
 699 enum isl_format
 700 anv_isl_format_for_descriptor_type(VkDescriptorType type)
 701 {
 702    switch (type) {
 703    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 704    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 705       return ISL_FORMAT_R32G32B32A32_FLOAT;
 706
 707    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 708    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 709       return ISL_FORMAT_RAW;
 710
 711    default:
 712       unreachable("Invalid descriptor type");
 713    }
 714 }
 715
 716 static struct anv_state
 717 anv_cmd_buffer_alloc_null_surface_state(struct anv_cmd_buffer *cmd_buffer,
 718                                         struct anv_framebuffer *fb)
 719 {
 720    switch (cmd_buffer->device->info.gen) {
 721    case 7:
 722       if (cmd_buffer->device->info.is_haswell) {
 723          return gen75_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 724       } else {
 725          return gen7_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 726       }
 727    case 8:
 728       return gen8_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 729    case 9:
 730       return gen9_cmd_buffer_alloc_null_surface_state(cmd_buffer, fb);
 731    default:
 732       unreachable("Invalid hardware generation");
 733    }
 734 }
 735
 736 VkResult
 737 anv_cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
 738                                   gl_shader_stage stage,
 739                                   struct anv_state *bt_state)
 740 {
 741    struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 742    struct anv_subpass *subpass = cmd_buffer->state.subpass;
 743    struct anv_pipeline_bind_map *map;
 744    uint32_t bias, state_offset;
 745
 746    switch (stage) {
 747    case  MESA_SHADER_COMPUTE:
 748       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 749       bias = 1;
 750       break;
 751    default:
 752       map = &cmd_buffer->state.pipeline->bindings[stage];
 753       bias = 0;
 754       break;
 755    }
 756
 757    if (bias + map->surface_count == 0) {
 758       *bt_state = (struct anv_state) { 0, };
 759       return VK_SUCCESS;
 760    }
 761
 762    *bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer,
 763                                                   bias + map->surface_count,
 764                                                   &state_offset);
 765    uint32_t *bt_map = bt_state->map;
 766
 767    if (bt_state->map == NULL)
 768       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 769
 770    if (stage == MESA_SHADER_COMPUTE &&
 771        get_cs_prog_data(cmd_buffer->state.compute_pipeline)->uses_num_work_groups) {
 772       struct anv_bo *bo = cmd_buffer->state.num_workgroups_bo;
 773       uint32_t bo_offset = cmd_buffer->state.num_workgroups_offset;
 774
 775       struct anv_state surface_state;
 776       surface_state =
 777          anv_cmd_buffer_alloc_surface_state(cmd_buffer);
 778
 779       const enum isl_format format =
 780          anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
 781       anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
 782                                     format, bo_offset, 12, 1);
 783
 784       bt_map[0] = surface_state.offset + state_offset;
 785       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 786    }
 787
 788    if (map->surface_count == 0)
 789       goto out;
 790
 791    if (map->image_count > 0) {
 792       VkResult result =
 793          anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images);
 794       if (result != VK_SUCCESS)
 795          return result;
 796
 797       cmd_buffer->state.push_constants_dirty |= 1 << stage;
 798    }
 799
 800    uint32_t image = 0;
 801    for (uint32_t s = 0; s < map->surface_count; s++) {
 802       struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];
 803
 804       struct anv_state surface_state;
 805       struct anv_bo *bo;
 806       uint32_t bo_offset;
 807
 808       if (binding->set == ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS) {
 809          /* Color attachment binding */
 810          assert(stage == MESA_SHADER_FRAGMENT);
 811          assert(binding->binding == 0);
 812          if (binding->index < subpass->color_count) {
 813             const struct anv_image_view *iview =
 814                fb->attachments[subpass->color_attachments[binding->index]];
 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       uint32_t offset = set->layout->binding[binding->binding].descriptor_index;
 834       struct anv_descriptor *desc = &set->descriptors[offset + binding->index];
 835
 836       switch (desc->type) {
 837       case VK_DESCRIPTOR_TYPE_SAMPLER:
 838          /* Nothing for us to do here */
 839          continue;
 840
 841       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 842       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 843       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 844          surface_state = desc->image_view->sampler_surface_state;
 845          assert(surface_state.alloc_size);
 846          bo = desc->image_view->bo;
 847          bo_offset = desc->image_view->offset;
 848          break;
 849
 850       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
 851          surface_state = desc->image_view->storage_surface_state;
 852          assert(surface_state.alloc_size);
 853          bo = desc->image_view->bo;
 854          bo_offset = desc->image_view->offset;
 855
 856          struct brw_image_param *image_param =
 857             &cmd_buffer->state.push_constants[stage]->images[image++];
 858
 859          *image_param = desc->image_view->storage_image_param;
 860          image_param->surface_idx = bias + s;
 861          break;
 862       }
 863
 864       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 865       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 866       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 867       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 868       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
 869          surface_state = desc->buffer_view->surface_state;
 870          assert(surface_state.alloc_size);
 871          bo = desc->buffer_view->bo;
 872          bo_offset = desc->buffer_view->offset;
 873          break;
 874
 875       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
 876          surface_state = desc->buffer_view->storage_surface_state;
 877          assert(surface_state.alloc_size);
 878          bo = desc->buffer_view->bo;
 879          bo_offset = desc->buffer_view->offset;
 880
 881          struct brw_image_param *image_param =
 882             &cmd_buffer->state.push_constants[stage]->images[image++];
 883
 884          *image_param = desc->buffer_view->storage_image_param;
 885          image_param->surface_idx = bias + s;
 886          break;
 887
 888       default:
 889          assert(!"Invalid descriptor type");
 890          continue;
 891       }
 892
 893       bt_map[bias + s] = surface_state.offset + state_offset;
 894       add_surface_state_reloc(cmd_buffer, surface_state, bo, bo_offset);
 895    }
 896    assert(image == map->image_count);
 897
 898  out:
 899    if (!cmd_buffer->device->info.has_llc)
 900       anv_state_clflush(*bt_state);
 901
 902    return VK_SUCCESS;
 903 }
 904
 905 VkResult
 906 anv_cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer,
 907                              gl_shader_stage stage, struct anv_state *state)
 908 {
 909    struct anv_pipeline_bind_map *map;
 910
 911    if (stage == MESA_SHADER_COMPUTE)
 912       map = &cmd_buffer->state.compute_pipeline->bindings[stage];
 913    else
 914       map = &cmd_buffer->state.pipeline->bindings[stage];
 915
 916    if (map->sampler_count == 0) {
 917       *state = (struct anv_state) { 0, };
 918       return VK_SUCCESS;
 919    }
 920
 921    uint32_t size = map->sampler_count * 16;
 922    *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32);
 923
 924    if (state->map == NULL)
 925       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 926
 927    for (uint32_t s = 0; s < map->sampler_count; s++) {
 928       struct anv_pipeline_binding *binding = &map->sampler_to_descriptor[s];
 929       struct anv_descriptor_set *set =
 930          cmd_buffer->state.descriptors[binding->set];
 931       uint32_t offset = set->layout->binding[binding->binding].descriptor_index;
 932       struct anv_descriptor *desc = &set->descriptors[offset + binding->index];
 933
 934       if (desc->type != VK_DESCRIPTOR_TYPE_SAMPLER &&
 935           desc->type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
 936          continue;
 937
 938       struct anv_sampler *sampler = desc->sampler;
 939
 940       /* This can happen if we have an unfilled slot since TYPE_SAMPLER
 941        * happens to be zero.
 942        */
 943       if (sampler == NULL)
 944          continue;
 945
 946       memcpy(state->map + (s * 16),
 947              sampler->state, sizeof(sampler->state));
 948    }
 949
 950    if (!cmd_buffer->device->info.has_llc)
 951       anv_state_clflush(*state);
 952
 953    return VK_SUCCESS;
 954 }
 955
 956 uint32_t
 957 anv_cmd_buffer_flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer)
 958 {
 959    VkShaderStageFlags dirty = cmd_buffer->state.descriptors_dirty &
 960                               cmd_buffer->state.pipeline->active_stages;
 961
 962    VkResult result = VK_SUCCESS;
 963    anv_foreach_stage(s, dirty) {
 964       result = anv_cmd_buffer_emit_samplers(cmd_buffer, s,
 965                                             &cmd_buffer->state.samplers[s]);
 966       if (result != VK_SUCCESS)
 967          break;
 968       result = anv_cmd_buffer_emit_binding_table(cmd_buffer, s,
 969                                                  &cmd_buffer->state.binding_tables[s]);
 970       if (result != VK_SUCCESS)
 971          break;
 972    }
 973
 974    if (result != VK_SUCCESS) {
 975       assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY);
 976
 977       result = anv_cmd_buffer_new_binding_table_block(cmd_buffer);
 978       assert(result == VK_SUCCESS);
 979
 980       /* Re-emit state base addresses so we get the new surface state base
 981        * address before we start emitting binding tables etc.
 982        */
 983       anv_cmd_buffer_emit_state_base_address(cmd_buffer);
 984
 985       /* Re-emit all active binding tables */
 986       dirty |= cmd_buffer->state.pipeline->active_stages;
 987       anv_foreach_stage(s, dirty) {
 988          result = anv_cmd_buffer_emit_samplers(cmd_buffer, s,
 989                                                &cmd_buffer->state.samplers[s]);
 990          if (result != VK_SUCCESS)
 991             return result;
 992          result = anv_cmd_buffer_emit_binding_table(cmd_buffer, s,
 993                                                     &cmd_buffer->state.binding_tables[s]);
 994          if (result != VK_SUCCESS)
 995             return result;
 996       }
 997    }
 998
 999    cmd_buffer->state.descriptors_dirty &= ~dirty;
1000
1001    return dirty;
1002 }
1003
1004 struct anv_state
1005 anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
1006                             const void *data, uint32_t size, uint32_t alignment)
1007 {
1008    struct anv_state state;
1009
1010    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment);
1011    memcpy(state.map, data, size);
1012
1013    if (!cmd_buffer->device->info.has_llc)
1014       anv_state_clflush(state);
1015
1016    VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size));
1017
1018    return state;
1019 }
1020
1021 struct anv_state
1022 anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
1023                              uint32_t *a, uint32_t *b,
1024                              uint32_t dwords, uint32_t alignment)
1025 {
1026    struct anv_state state;
1027    uint32_t *p;
1028
1029    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1030                                               dwords * 4, alignment);
1031    p = state.map;
1032    for (uint32_t i = 0; i < dwords; i++)
1033       p[i] = a[i] | b[i];
1034
1035    if (!cmd_buffer->device->info.has_llc)
1036       anv_state_clflush(state);
1037
1038    VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
1039
1040    return state;
1041 }
1042
1043 /**
1044  * @brief Setup the command buffer for recording commands inside the given
1045  * subpass.
1046  *
1047  * This does not record all commands needed for starting the subpass.
1048  * Starting the subpass may require additional commands.
1049  *
1050  * Note that vkCmdBeginRenderPass, vkCmdNextSubpass, and vkBeginCommandBuffer
1051  * with VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT, all setup the
1052  * command buffer for recording commands for some subpass.  But only the first
1053  * two, vkCmdBeginRenderPass and vkCmdNextSubpass, can start a subpass.
1054  */
1055 void
1056 anv_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
1057                            struct anv_subpass *subpass)
1058 {
1059    switch (cmd_buffer->device->info.gen) {
1060    case 7:
1061       if (cmd_buffer->device->info.is_haswell) {
1062          gen75_cmd_buffer_set_subpass(cmd_buffer, subpass);
1063       } else {
1064          gen7_cmd_buffer_set_subpass(cmd_buffer, subpass);
1065       }
1066       break;
1067    case 8:
1068       gen8_cmd_buffer_set_subpass(cmd_buffer, subpass);
1069       break;
1070    case 9:
1071       gen9_cmd_buffer_set_subpass(cmd_buffer, subpass);
1072       break;
1073    default:
1074       unreachable("unsupported gen\n");
1075    }
1076 }
1077
1078 struct anv_state
1079 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
1080                               gl_shader_stage stage)
1081 {
1082    struct anv_push_constants *data =
1083       cmd_buffer->state.push_constants[stage];
1084    const struct brw_stage_prog_data *prog_data =
1085       cmd_buffer->state.pipeline->prog_data[stage];
1086
1087    /* If we don't actually have any push constants, bail. */
1088    if (data == NULL || prog_data == NULL || prog_data->nr_params == 0)
1089       return (struct anv_state) { .offset = 0 };
1090
1091    struct anv_state state =
1092       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1093                                          prog_data->nr_params * sizeof(float),
1094                                          32 /* bottom 5 bits MBZ */);
1095
1096    /* Walk through the param array and fill the buffer with data */
1097    uint32_t *u32_map = state.map;
1098    for (unsigned i = 0; i < prog_data->nr_params; i++) {
1099       uint32_t offset = (uintptr_t)prog_data->param[i];
1100       u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
1101    }
1102
1103    if (!cmd_buffer->device->info.has_llc)
1104       anv_state_clflush(state);
1105
1106    return state;
1107 }
1108
1109 struct anv_state
1110 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
1111 {
1112    struct anv_push_constants *data =
1113       cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
1114    struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
1115    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
1116    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
1117
1118    /* If we don't actually have any push constants, bail. */
1119    if (cs_prog_data->push.total.size == 0)
1120       return (struct anv_state) { .offset = 0 };
1121
1122    const unsigned push_constant_alignment =
1123       cmd_buffer->device->info.gen < 8 ? 32 : 64;
1124    const unsigned aligned_total_push_constants_size =
1125       ALIGN(cs_prog_data->push.total.size, push_constant_alignment);
1126    struct anv_state state =
1127       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
1128                                          aligned_total_push_constants_size,
1129                                          push_constant_alignment);
1130
1131    /* Walk through the param array and fill the buffer with data */
1132    uint32_t *u32_map = state.map;
1133
1134    if (cs_prog_data->push.cross_thread.size > 0) {
1135       assert(cs_prog_data->thread_local_id_index < 0 ||
1136              cs_prog_data->thread_local_id_index >=
1137                 cs_prog_data->push.cross_thread.dwords);
1138       for (unsigned i = 0;
1139            i < cs_prog_data->push.cross_thread.dwords;
1140            i++) {
1141          uint32_t offset = (uintptr_t)prog_data->param[i];
1142          u32_map[i] = *(uint32_t *)((uint8_t *)data + offset);
1143       }
1144    }
1145
1146    if (cs_prog_data->push.per_thread.size > 0) {
1147       for (unsigned t = 0; t < cs_prog_data->threads; t++) {
1148          unsigned dst =
1149             8 * (cs_prog_data->push.per_thread.regs * t +
1150                  cs_prog_data->push.cross_thread.regs);
1151          unsigned src = cs_prog_data->push.cross_thread.dwords;
1152          for ( ; src < prog_data->nr_params; src++, dst++) {
1153             if (src != cs_prog_data->thread_local_id_index) {
1154                uint32_t offset = (uintptr_t)prog_data->param[src];
1155                u32_map[dst] = *(uint32_t *)((uint8_t *)data + offset);
1156             } else {
1157                u32_map[dst] = t * cs_prog_data->simd_size;
1158             }
1159          }
1160       }
1161    }
1162
1163    if (!cmd_buffer->device->info.has_llc)
1164       anv_state_clflush(state);
1165
1166    return state;
1167 }
1168
1169 void anv_CmdPushConstants(
1170     VkCommandBuffer                             commandBuffer,
1171     VkPipelineLayout                            layout,
1172     VkShaderStageFlags                          stageFlags,
1173     uint32_t                                    offset,
1174     uint32_t                                    size,
1175     const void*                                 pValues)
1176 {
1177    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1178
1179    anv_foreach_stage(stage, stageFlags) {
1180       anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, client_data);
1181
1182       memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
1183              pValues, size);
1184    }
1185
1186    cmd_buffer->state.push_constants_dirty |= stageFlags;
1187 }
1188
1189 void anv_CmdExecuteCommands(
1190     VkCommandBuffer                             commandBuffer,
1191     uint32_t                                    commandBufferCount,
1192     const VkCommandBuffer*                      pCmdBuffers)
1193 {
1194    ANV_FROM_HANDLE(anv_cmd_buffer, primary, commandBuffer);
1195
1196    assert(primary->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1197
1198    for (uint32_t i = 0; i < commandBufferCount; i++) {
1199       ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);
1200
1201       assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
1202
1203       anv_cmd_buffer_add_secondary(primary, secondary);
1204    }
1205 }
1206
1207 VkResult anv_CreateCommandPool(
1208     VkDevice                                    _device,
1209     const VkCommandPoolCreateInfo*              pCreateInfo,
1210     const VkAllocationCallbacks*                pAllocator,
1211     VkCommandPool*                              pCmdPool)
1212 {
1213    ANV_FROM_HANDLE(anv_device, device, _device);
1214    struct anv_cmd_pool *pool;
1215
1216    pool = anv_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
1217                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1218    if (pool == NULL)
1219       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1220
1221    if (pAllocator)
1222       pool->alloc = *pAllocator;
1223    else
1224       pool->alloc = device->alloc;
1225
1226    list_inithead(&pool->cmd_buffers);
1227
1228    *pCmdPool = anv_cmd_pool_to_handle(pool);
1229
1230    return VK_SUCCESS;
1231 }
1232
1233 void anv_DestroyCommandPool(
1234     VkDevice                                    _device,
1235     VkCommandPool                               commandPool,
1236     const VkAllocationCallbacks*                pAllocator)
1237 {
1238    ANV_FROM_HANDLE(anv_device, device, _device);
1239    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1240
1241    list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
1242                             &pool->cmd_buffers, pool_link) {
1243       anv_cmd_buffer_destroy(cmd_buffer);
1244    }
1245
1246    anv_free2(&device->alloc, pAllocator, pool);
1247 }
1248
1249 VkResult anv_ResetCommandPool(
1250     VkDevice                                    device,
1251     VkCommandPool                               commandPool,
1252     VkCommandPoolResetFlags                     flags)
1253 {
1254    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
1255
1256    list_for_each_entry(struct anv_cmd_buffer, cmd_buffer,
1257                        &pool->cmd_buffers, pool_link) {
1258       anv_cmd_buffer_reset(cmd_buffer);
1259    }
1260
1261    return VK_SUCCESS;
1262 }
1263
1264 /**
1265  * Return NULL if the current subpass has no depthstencil attachment.
1266  */
1267 const struct anv_image_view *
1268 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
1269 {
1270    const struct anv_subpass *subpass = cmd_buffer->state.subpass;
1271    const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
1272
1273    if (subpass->depth_stencil_attachment == VK_ATTACHMENT_UNUSED)
1274       return NULL;
1275
1276    const struct anv_image_view *iview =
1277       fb->attachments[subpass->depth_stencil_attachment];
1278
1279    assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
1280                                 VK_IMAGE_ASPECT_STENCIL_BIT));
1281
1282    return iview;
1283 }