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_init(struct anv_cmd_buffer *cmd_buffer)
 117 {
 118    struct anv_cmd_state *state = &cmd_buffer->state;
 119
 120    memset(state, 0, sizeof(*state));
 121
 122    state->current_pipeline = UINT32_MAX;
 123    state->restart_index = UINT32_MAX;
 124    state->gfx.dynamic = default_dynamic_state;
 125 }
 126
 127 static void
 128 anv_cmd_pipeline_state_finish(struct anv_cmd_buffer *cmd_buffer,
 129                               struct anv_cmd_pipeline_state *pipe_state)
 130 {
 131    for (uint32_t i = 0; i < ARRAY_SIZE(pipe_state->push_descriptors); i++)
 132       vk_free(&cmd_buffer->pool->alloc, pipe_state->push_descriptors[i]);
 133 }
 134
 135 static void
 136 anv_cmd_state_finish(struct anv_cmd_buffer *cmd_buffer)
 137 {
 138    struct anv_cmd_state *state = &cmd_buffer->state;
 139
 140    anv_cmd_pipeline_state_finish(cmd_buffer, &state->gfx.base);
 141    anv_cmd_pipeline_state_finish(cmd_buffer, &state->compute.base);
 142
 143    for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++)
 144       vk_free(&cmd_buffer->pool->alloc, state->push_constants[i]);
 145
 146    vk_free(&cmd_buffer->pool->alloc, state->attachments);
 147 }
 148
 149 static void
 150 anv_cmd_state_reset(struct anv_cmd_buffer *cmd_buffer)
 151 {
 152    anv_cmd_state_finish(cmd_buffer);
 153    anv_cmd_state_init(cmd_buffer);
 154 }
 155
 156 /**
 157  * This function updates the size of the push constant buffer we need to emit.
 158  * This is called in various parts of the driver to ensure that different
 159  * pieces of push constant data get emitted as needed. However, it is important
 160  * that we never shrink the size of the buffer. For example, a compute shader
 161  * dispatch will always call this for the base group id, which has an
 162  * offset in the push constant buffer that is smaller than the offset for
 163  * storage image data. If the compute shader has storage images, we will call
 164  * this again with a larger size during binding table emission. However,
 165  * if we dispatch the compute shader again without dirtying our descriptors,
 166  * we would still call this function with a smaller size for the base group
 167  * id, and not for the images, which would incorrectly shrink the size of the
 168  * push constant data we emit with that dispatch, making us drop the image data.
 169  */
 170 VkResult
 171 anv_cmd_buffer_ensure_push_constants_size(struct anv_cmd_buffer *cmd_buffer,
 172                                           gl_shader_stage stage, uint32_t size)
 173 {
 174    struct anv_push_constants **ptr = &cmd_buffer->state.push_constants[stage];
 175
 176    if (*ptr == NULL) {
 177       *ptr = vk_alloc(&cmd_buffer->pool->alloc, size, 8,
 178                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 179       if (*ptr == NULL) {
 180          anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
 181          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 182       }
 183       (*ptr)->size = size;
 184    } else if ((*ptr)->size < size) {
 185       *ptr = vk_realloc(&cmd_buffer->pool->alloc, *ptr, size, 8,
 186                          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 187       if (*ptr == NULL) {
 188          anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
 189          return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 190       }
 191       (*ptr)->size = size;
 192    }
 193
 194    return VK_SUCCESS;
 195 }
 196
 197 static VkResult anv_create_cmd_buffer(
 198     struct anv_device *                         device,
 199     struct anv_cmd_pool *                       pool,
 200     VkCommandBufferLevel                        level,
 201     VkCommandBuffer*                            pCommandBuffer)
 202 {
 203    struct anv_cmd_buffer *cmd_buffer;
 204    VkResult result;
 205
 206    cmd_buffer = vk_alloc(&pool->alloc, sizeof(*cmd_buffer), 8,
 207                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 208    if (cmd_buffer == NULL)
 209       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 210
 211    cmd_buffer->batch.status = VK_SUCCESS;
 212
 213    cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 214    cmd_buffer->device = device;
 215    cmd_buffer->pool = pool;
 216    cmd_buffer->level = level;
 217
 218    result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer);
 219    if (result != VK_SUCCESS)
 220       goto fail;
 221
 222    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 223                          &device->surface_state_pool, 4096);
 224    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 225                          &device->dynamic_state_pool, 16384);
 226
 227    anv_cmd_state_init(cmd_buffer);
 228
 229    if (pool) {
 230       list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
 231    } else {
 232       /* Init the pool_link so we can safefly call list_del when we destroy
 233        * the command buffer
 234        */
 235       list_inithead(&cmd_buffer->pool_link);
 236    }
 237
 238    *pCommandBuffer = anv_cmd_buffer_to_handle(cmd_buffer);
 239
 240    return VK_SUCCESS;
 241
 242  fail:
 243    vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
 244
 245    return result;
 246 }
 247
 248 VkResult anv_AllocateCommandBuffers(
 249     VkDevice                                    _device,
 250     const VkCommandBufferAllocateInfo*          pAllocateInfo,
 251     VkCommandBuffer*                            pCommandBuffers)
 252 {
 253    ANV_FROM_HANDLE(anv_device, device, _device);
 254    ANV_FROM_HANDLE(anv_cmd_pool, pool, pAllocateInfo->commandPool);
 255
 256    VkResult result = VK_SUCCESS;
 257    uint32_t i;
 258
 259    for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
 260       result = anv_create_cmd_buffer(device, pool, pAllocateInfo->level,
 261                                      &pCommandBuffers[i]);
 262       if (result != VK_SUCCESS)
 263          break;
 264    }
 265
 266    if (result != VK_SUCCESS) {
 267       anv_FreeCommandBuffers(_device, pAllocateInfo->commandPool,
 268                              i, pCommandBuffers);
 269       for (i = 0; i < pAllocateInfo->commandBufferCount; i++)
 270          pCommandBuffers[i] = VK_NULL_HANDLE;
 271    }
 272
 273    return result;
 274 }
 275
 276 static void
 277 anv_cmd_buffer_destroy(struct anv_cmd_buffer *cmd_buffer)
 278 {
 279    list_del(&cmd_buffer->pool_link);
 280
 281    anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer);
 282
 283    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 284    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 285
 286    anv_cmd_state_finish(cmd_buffer);
 287
 288    vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
 289 }
 290
 291 void anv_FreeCommandBuffers(
 292     VkDevice                                    device,
 293     VkCommandPool                               commandPool,
 294     uint32_t                                    commandBufferCount,
 295     const VkCommandBuffer*                      pCommandBuffers)
 296 {
 297    for (uint32_t i = 0; i < commandBufferCount; i++) {
 298       ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
 299
 300       if (!cmd_buffer)
 301          continue;
 302
 303       anv_cmd_buffer_destroy(cmd_buffer);
 304    }
 305 }
 306
 307 VkResult
 308 anv_cmd_buffer_reset(struct anv_cmd_buffer *cmd_buffer)
 309 {
 310    cmd_buffer->usage_flags = 0;
 311    anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer);
 312    anv_cmd_state_reset(cmd_buffer);
 313
 314    anv_state_stream_finish(&cmd_buffer->surface_state_stream);
 315    anv_state_stream_init(&cmd_buffer->surface_state_stream,
 316                          &cmd_buffer->device->surface_state_pool, 4096);
 317
 318    anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
 319    anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
 320                          &cmd_buffer->device->dynamic_state_pool, 16384);
 321    return VK_SUCCESS;
 322 }
 323
 324 VkResult anv_ResetCommandBuffer(
 325     VkCommandBuffer                             commandBuffer,
 326     VkCommandBufferResetFlags                   flags)
 327 {
 328    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 329    return anv_cmd_buffer_reset(cmd_buffer);
 330 }
 331
 332 #define anv_genX_call(devinfo, func, ...)          \
 333    switch ((devinfo)->gen) {                       \
 334    case 7:                                         \
 335       if ((devinfo)->is_haswell) {                 \
 336          gen75_##func(__VA_ARGS__);                \
 337       } else {                                     \
 338          gen7_##func(__VA_ARGS__);                 \
 339       }                                            \
 340       break;                                       \
 341    case 8:                                         \
 342       gen8_##func(__VA_ARGS__);                    \
 343       break;                                       \
 344    case 9:                                         \
 345       gen9_##func(__VA_ARGS__);                    \
 346       break;                                       \
 347    case 10:                                        \
 348       gen10_##func(__VA_ARGS__);                   \
 349       break;                                       \
 350    case 11:                                        \
 351       gen11_##func(__VA_ARGS__);                   \
 352       break;                                       \
 353    default:                                        \
 354       assert(!"Unknown hardware generation");      \
 355    }
 356
 357 void
 358 anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer)
 359 {
 360    anv_genX_call(&cmd_buffer->device->info,
 361                  cmd_buffer_emit_state_base_address,
 362                  cmd_buffer);
 363 }
 364
 365 void
 366 anv_cmd_buffer_mark_image_written(struct anv_cmd_buffer *cmd_buffer,
 367                                   const struct anv_image *image,
 368                                   VkImageAspectFlagBits aspect,
 369                                   enum isl_aux_usage aux_usage,
 370                                   uint32_t level,
 371                                   uint32_t base_layer,
 372                                   uint32_t layer_count)
 373 {
 374    anv_genX_call(&cmd_buffer->device->info,
 375                  cmd_buffer_mark_image_written,
 376                  cmd_buffer, image, aspect, aux_usage,
 377                  level, base_layer, layer_count);
 378 }
 379
 380 void anv_CmdBindPipeline(
 381     VkCommandBuffer                             commandBuffer,
 382     VkPipelineBindPoint                         pipelineBindPoint,
 383     VkPipeline                                  _pipeline)
 384 {
 385    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 386    ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 387
 388    switch (pipelineBindPoint) {
 389    case VK_PIPELINE_BIND_POINT_COMPUTE:
 390       cmd_buffer->state.compute.base.pipeline = pipeline;
 391       cmd_buffer->state.compute.pipeline_dirty = true;
 392       cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 393       cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 394       break;
 395
 396    case VK_PIPELINE_BIND_POINT_GRAPHICS:
 397       cmd_buffer->state.gfx.base.pipeline = pipeline;
 398       cmd_buffer->state.gfx.vb_dirty |= pipeline->vb_used;
 399       cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_PIPELINE;
 400       cmd_buffer->state.push_constants_dirty |= pipeline->active_stages;
 401       cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;
 402
 403       /* Apply the dynamic state from the pipeline */
 404       cmd_buffer->state.gfx.dirty |= pipeline->dynamic_state_mask;
 405       anv_dynamic_state_copy(&cmd_buffer->state.gfx.dynamic,
 406                              &pipeline->dynamic_state,
 407                              pipeline->dynamic_state_mask);
 408       break;
 409
 410    default:
 411       assert(!"invalid bind point");
 412       break;
 413    }
 414 }
 415
 416 void anv_CmdSetViewport(
 417     VkCommandBuffer                             commandBuffer,
 418     uint32_t                                    firstViewport,
 419     uint32_t                                    viewportCount,
 420     const VkViewport*                           pViewports)
 421 {
 422    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 423
 424    const uint32_t total_count = firstViewport + viewportCount;
 425    if (cmd_buffer->state.gfx.dynamic.viewport.count < total_count)
 426       cmd_buffer->state.gfx.dynamic.viewport.count = total_count;
 427
 428    memcpy(cmd_buffer->state.gfx.dynamic.viewport.viewports + firstViewport,
 429           pViewports, viewportCount * sizeof(*pViewports));
 430
 431    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_VIEWPORT;
 432 }
 433
 434 void anv_CmdSetScissor(
 435     VkCommandBuffer                             commandBuffer,
 436     uint32_t                                    firstScissor,
 437     uint32_t                                    scissorCount,
 438     const VkRect2D*                             pScissors)
 439 {
 440    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 441
 442    const uint32_t total_count = firstScissor + scissorCount;
 443    if (cmd_buffer->state.gfx.dynamic.scissor.count < total_count)
 444       cmd_buffer->state.gfx.dynamic.scissor.count = total_count;
 445
 446    memcpy(cmd_buffer->state.gfx.dynamic.scissor.scissors + firstScissor,
 447           pScissors, scissorCount * sizeof(*pScissors));
 448
 449    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_SCISSOR;
 450 }
 451
 452 void anv_CmdSetLineWidth(
 453     VkCommandBuffer                             commandBuffer,
 454     float                                       lineWidth)
 455 {
 456    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 457
 458    cmd_buffer->state.gfx.dynamic.line_width = lineWidth;
 459    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
 460 }
 461
 462 void anv_CmdSetDepthBias(
 463     VkCommandBuffer                             commandBuffer,
 464     float                                       depthBiasConstantFactor,
 465     float                                       depthBiasClamp,
 466     float                                       depthBiasSlopeFactor)
 467 {
 468    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 469
 470    cmd_buffer->state.gfx.dynamic.depth_bias.bias = depthBiasConstantFactor;
 471    cmd_buffer->state.gfx.dynamic.depth_bias.clamp = depthBiasClamp;
 472    cmd_buffer->state.gfx.dynamic.depth_bias.slope = depthBiasSlopeFactor;
 473
 474    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS;
 475 }
 476
 477 void anv_CmdSetBlendConstants(
 478     VkCommandBuffer                             commandBuffer,
 479     const float                                 blendConstants[4])
 480 {
 481    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 482
 483    memcpy(cmd_buffer->state.gfx.dynamic.blend_constants,
 484           blendConstants, sizeof(float) * 4);
 485
 486    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS;
 487 }
 488
 489 void anv_CmdSetDepthBounds(
 490     VkCommandBuffer                             commandBuffer,
 491     float                                       minDepthBounds,
 492     float                                       maxDepthBounds)
 493 {
 494    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 495
 496    cmd_buffer->state.gfx.dynamic.depth_bounds.min = minDepthBounds;
 497    cmd_buffer->state.gfx.dynamic.depth_bounds.max = maxDepthBounds;
 498
 499    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS;
 500 }
 501
 502 void anv_CmdSetStencilCompareMask(
 503     VkCommandBuffer                             commandBuffer,
 504     VkStencilFaceFlags                          faceMask,
 505     uint32_t                                    compareMask)
 506 {
 507    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 508
 509    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 510       cmd_buffer->state.gfx.dynamic.stencil_compare_mask.front = compareMask;
 511    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 512       cmd_buffer->state.gfx.dynamic.stencil_compare_mask.back = compareMask;
 513
 514    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
 515 }
 516
 517 void anv_CmdSetStencilWriteMask(
 518     VkCommandBuffer                             commandBuffer,
 519     VkStencilFaceFlags                          faceMask,
 520     uint32_t                                    writeMask)
 521 {
 522    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 523
 524    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 525       cmd_buffer->state.gfx.dynamic.stencil_write_mask.front = writeMask;
 526    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 527       cmd_buffer->state.gfx.dynamic.stencil_write_mask.back = writeMask;
 528
 529    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
 530 }
 531
 532 void anv_CmdSetStencilReference(
 533     VkCommandBuffer                             commandBuffer,
 534     VkStencilFaceFlags                          faceMask,
 535     uint32_t                                    reference)
 536 {
 537    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 538
 539    if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
 540       cmd_buffer->state.gfx.dynamic.stencil_reference.front = reference;
 541    if (faceMask & VK_STENCIL_FACE_BACK_BIT)
 542       cmd_buffer->state.gfx.dynamic.stencil_reference.back = reference;
 543
 544    cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
 545 }
 546
 547 static void
 548 anv_cmd_buffer_bind_descriptor_set(struct anv_cmd_buffer *cmd_buffer,
 549                                    VkPipelineBindPoint bind_point,
 550                                    struct anv_pipeline_layout *layout,
 551                                    uint32_t set_index,
 552                                    struct anv_descriptor_set *set,
 553                                    uint32_t *dynamic_offset_count,
 554                                    const uint32_t **dynamic_offsets)
 555 {
 556    struct anv_descriptor_set_layout *set_layout =
 557       layout->set[set_index].layout;
 558
 559    struct anv_cmd_pipeline_state *pipe_state;
 560    if (bind_point == VK_PIPELINE_BIND_POINT_COMPUTE) {
 561       pipe_state = &cmd_buffer->state.compute.base;
 562    } else {
 563       assert(bind_point == VK_PIPELINE_BIND_POINT_GRAPHICS);
 564       pipe_state = &cmd_buffer->state.gfx.base;
 565    }
 566    pipe_state->descriptors[set_index] = set;
 567
 568    if (dynamic_offsets) {
 569       if (set_layout->dynamic_offset_count > 0) {
 570          uint32_t dynamic_offset_start =
 571             layout->set[set_index].dynamic_offset_start;
 572
 573          /* Assert that everything is in range */
 574          assert(set_layout->dynamic_offset_count <= *dynamic_offset_count);
 575          assert(dynamic_offset_start + set_layout->dynamic_offset_count <=
 576                 ARRAY_SIZE(pipe_state->dynamic_offsets));
 577
 578          typed_memcpy(&pipe_state->dynamic_offsets[dynamic_offset_start],
 579                       *dynamic_offsets, set_layout->dynamic_offset_count);
 580
 581          *dynamic_offsets += set_layout->dynamic_offset_count;
 582          *dynamic_offset_count -= set_layout->dynamic_offset_count;
 583       }
 584    }
 585
 586    if (bind_point == VK_PIPELINE_BIND_POINT_COMPUTE) {
 587       cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
 588    } else {
 589       assert(bind_point == VK_PIPELINE_BIND_POINT_GRAPHICS);
 590       cmd_buffer->state.descriptors_dirty |=
 591          set_layout->shader_stages & VK_SHADER_STAGE_ALL_GRAPHICS;
 592    }
 593
 594    /* Pipeline layout objects are required to live at least while any command
 595     * buffers that use them are in recording state. We need to grab a reference
 596     * to the pipeline layout being bound here so we can compute correct dynamic
 597     * offsets for VK_DESCRIPTOR_TYPE_*_DYNAMIC in dynamic_offset_for_binding()
 598     * when we record draw commands that come after this.
 599     */
 600    pipe_state->layout = layout;
 601 }
 602
 603 void anv_CmdBindDescriptorSets(
 604     VkCommandBuffer                             commandBuffer,
 605     VkPipelineBindPoint                         pipelineBindPoint,
 606     VkPipelineLayout                            _layout,
 607     uint32_t                                    firstSet,
 608     uint32_t                                    descriptorSetCount,
 609     const VkDescriptorSet*                      pDescriptorSets,
 610     uint32_t                                    dynamicOffsetCount,
 611     const uint32_t*                             pDynamicOffsets)
 612 {
 613    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 614    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 615
 616    assert(firstSet + descriptorSetCount < MAX_SETS);
 617
 618    for (uint32_t i = 0; i < descriptorSetCount; i++) {
 619       ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
 620       anv_cmd_buffer_bind_descriptor_set(cmd_buffer, pipelineBindPoint,
 621                                          layout, firstSet + i, set,
 622                                          &dynamicOffsetCount,
 623                                          &pDynamicOffsets);
 624    }
 625 }
 626
 627 void anv_CmdBindVertexBuffers(
 628     VkCommandBuffer                             commandBuffer,
 629     uint32_t                                    firstBinding,
 630     uint32_t                                    bindingCount,
 631     const VkBuffer*                             pBuffers,
 632     const VkDeviceSize*                         pOffsets)
 633 {
 634    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 635    struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 636
 637    /* We have to defer setting up vertex buffer since we need the buffer
 638     * stride from the pipeline. */
 639
 640    assert(firstBinding + bindingCount <= MAX_VBS);
 641    for (uint32_t i = 0; i < bindingCount; i++) {
 642       vb[firstBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
 643       vb[firstBinding + i].offset = pOffsets[i];
 644       cmd_buffer->state.gfx.vb_dirty |= 1 << (firstBinding + i);
 645    }
 646 }
 647
 648 enum isl_format
 649 anv_isl_format_for_descriptor_type(VkDescriptorType type)
 650 {
 651    switch (type) {
 652    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 653    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 654       return ISL_FORMAT_R32G32B32A32_FLOAT;
 655
 656    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 657    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 658       return ISL_FORMAT_RAW;
 659
 660    default:
 661       unreachable("Invalid descriptor type");
 662    }
 663 }
 664
 665 struct anv_state
 666 anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
 667                             const void *data, uint32_t size, uint32_t alignment)
 668 {
 669    struct anv_state state;
 670
 671    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, alignment);
 672    memcpy(state.map, data, size);
 673
 674    anv_state_flush(cmd_buffer->device, state);
 675
 676    VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, size));
 677
 678    return state;
 679 }
 680
 681 struct anv_state
 682 anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
 683                              uint32_t *a, uint32_t *b,
 684                              uint32_t dwords, uint32_t alignment)
 685 {
 686    struct anv_state state;
 687    uint32_t *p;
 688
 689    state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 690                                               dwords * 4, alignment);
 691    p = state.map;
 692    for (uint32_t i = 0; i < dwords; i++)
 693       p[i] = a[i] | b[i];
 694
 695    anv_state_flush(cmd_buffer->device, state);
 696
 697    VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
 698
 699    return state;
 700 }
 701
 702 static uint32_t
 703 anv_push_constant_value(struct anv_push_constants *data, uint32_t param)
 704 {
 705    if (BRW_PARAM_IS_BUILTIN(param)) {
 706       switch (param) {
 707       case BRW_PARAM_BUILTIN_ZERO:
 708          return 0;
 709       case BRW_PARAM_BUILTIN_BASE_WORK_GROUP_ID_X:
 710          return data->base_work_group_id[0];
 711       case BRW_PARAM_BUILTIN_BASE_WORK_GROUP_ID_Y:
 712          return data->base_work_group_id[1];
 713       case BRW_PARAM_BUILTIN_BASE_WORK_GROUP_ID_Z:
 714          return data->base_work_group_id[2];
 715       default:
 716          unreachable("Invalid param builtin");
 717       }
 718    } else {
 719       uint32_t offset = ANV_PARAM_PUSH_OFFSET(param);
 720       assert(offset % sizeof(uint32_t) == 0);
 721       if (offset < data->size)
 722          return *(uint32_t *)((uint8_t *)data + offset);
 723       else
 724          return 0;
 725    }
 726 }
 727
 728 struct anv_state
 729 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
 730                               gl_shader_stage stage)
 731 {
 732    struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
 733
 734    /* If we don't have this stage, bail. */
 735    if (!anv_pipeline_has_stage(pipeline, stage))
 736       return (struct anv_state) { .offset = 0 };
 737
 738    struct anv_push_constants *data =
 739       cmd_buffer->state.push_constants[stage];
 740    const struct brw_stage_prog_data *prog_data =
 741       pipeline->shaders[stage]->prog_data;
 742
 743    /* If we don't actually have any push constants, bail. */
 744    if (data == NULL || prog_data == NULL || prog_data->nr_params == 0)
 745       return (struct anv_state) { .offset = 0 };
 746
 747    struct anv_state state =
 748       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 749                                          prog_data->nr_params * sizeof(float),
 750                                          32 /* bottom 5 bits MBZ */);
 751
 752    /* Walk through the param array and fill the buffer with data */
 753    uint32_t *u32_map = state.map;
 754    for (unsigned i = 0; i < prog_data->nr_params; i++)
 755       u32_map[i] = anv_push_constant_value(data, prog_data->param[i]);
 756
 757    anv_state_flush(cmd_buffer->device, state);
 758
 759    return state;
 760 }
 761
 762 struct anv_state
 763 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer)
 764 {
 765    struct anv_push_constants *data =
 766       cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
 767    struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
 768    const struct brw_cs_prog_data *cs_prog_data = get_cs_prog_data(pipeline);
 769    const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
 770
 771    /* If we don't actually have any push constants, bail. */
 772    if (cs_prog_data->push.total.size == 0)
 773       return (struct anv_state) { .offset = 0 };
 774
 775    const unsigned push_constant_alignment =
 776       cmd_buffer->device->info.gen < 8 ? 32 : 64;
 777    const unsigned aligned_total_push_constants_size =
 778       ALIGN(cs_prog_data->push.total.size, push_constant_alignment);
 779    struct anv_state state =
 780       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
 781                                          aligned_total_push_constants_size,
 782                                          push_constant_alignment);
 783
 784    /* Walk through the param array and fill the buffer with data */
 785    uint32_t *u32_map = state.map;
 786
 787    if (cs_prog_data->push.cross_thread.size > 0) {
 788       for (unsigned i = 0;
 789            i < cs_prog_data->push.cross_thread.dwords;
 790            i++) {
 791          assert(prog_data->param[i] != BRW_PARAM_BUILTIN_SUBGROUP_ID);
 792          u32_map[i] = anv_push_constant_value(data, prog_data->param[i]);
 793       }
 794    }
 795
 796    if (cs_prog_data->push.per_thread.size > 0) {
 797       for (unsigned t = 0; t < cs_prog_data->threads; t++) {
 798          unsigned dst =
 799             8 * (cs_prog_data->push.per_thread.regs * t +
 800                  cs_prog_data->push.cross_thread.regs);
 801          unsigned src = cs_prog_data->push.cross_thread.dwords;
 802          for ( ; src < prog_data->nr_params; src++, dst++) {
 803             if (prog_data->param[src] == BRW_PARAM_BUILTIN_SUBGROUP_ID) {
 804                u32_map[dst] = t;
 805             } else {
 806                u32_map[dst] =
 807                   anv_push_constant_value(data, prog_data->param[src]);
 808             }
 809          }
 810       }
 811    }
 812
 813    anv_state_flush(cmd_buffer->device, state);
 814
 815    return state;
 816 }
 817
 818 void anv_CmdPushConstants(
 819     VkCommandBuffer                             commandBuffer,
 820     VkPipelineLayout                            layout,
 821     VkShaderStageFlags                          stageFlags,
 822     uint32_t                                    offset,
 823     uint32_t                                    size,
 824     const void*                                 pValues)
 825 {
 826    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 827
 828    anv_foreach_stage(stage, stageFlags) {
 829       VkResult result =
 830          anv_cmd_buffer_ensure_push_constant_field(cmd_buffer,
 831                                                    stage, client_data);
 832       if (result != VK_SUCCESS)
 833          return;
 834
 835       memcpy(cmd_buffer->state.push_constants[stage]->client_data + offset,
 836              pValues, size);
 837    }
 838
 839    cmd_buffer->state.push_constants_dirty |= stageFlags;
 840 }
 841
 842 VkResult anv_CreateCommandPool(
 843     VkDevice                                    _device,
 844     const VkCommandPoolCreateInfo*              pCreateInfo,
 845     const VkAllocationCallbacks*                pAllocator,
 846     VkCommandPool*                              pCmdPool)
 847 {
 848    ANV_FROM_HANDLE(anv_device, device, _device);
 849    struct anv_cmd_pool *pool;
 850
 851    pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
 852                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 853    if (pool == NULL)
 854       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 855
 856    if (pAllocator)
 857       pool->alloc = *pAllocator;
 858    else
 859       pool->alloc = device->alloc;
 860
 861    list_inithead(&pool->cmd_buffers);
 862
 863    *pCmdPool = anv_cmd_pool_to_handle(pool);
 864
 865    return VK_SUCCESS;
 866 }
 867
 868 void anv_DestroyCommandPool(
 869     VkDevice                                    _device,
 870     VkCommandPool                               commandPool,
 871     const VkAllocationCallbacks*                pAllocator)
 872 {
 873    ANV_FROM_HANDLE(anv_device, device, _device);
 874    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
 875
 876    if (!pool)
 877       return;
 878
 879    list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
 880                             &pool->cmd_buffers, pool_link) {
 881       anv_cmd_buffer_destroy(cmd_buffer);
 882    }
 883
 884    vk_free2(&device->alloc, pAllocator, pool);
 885 }
 886
 887 VkResult anv_ResetCommandPool(
 888     VkDevice                                    device,
 889     VkCommandPool                               commandPool,
 890     VkCommandPoolResetFlags                     flags)
 891 {
 892    ANV_FROM_HANDLE(anv_cmd_pool, pool, commandPool);
 893
 894    list_for_each_entry(struct anv_cmd_buffer, cmd_buffer,
 895                        &pool->cmd_buffers, pool_link) {
 896       anv_cmd_buffer_reset(cmd_buffer);
 897    }
 898
 899    return VK_SUCCESS;
 900 }
 901
 902 void anv_TrimCommandPool(
 903     VkDevice                                    device,
 904     VkCommandPool                               commandPool,
 905     VkCommandPoolTrimFlags                      flags)
 906 {
 907    /* Nothing for us to do here.  Our pools stay pretty tidy. */
 908 }
 909
 910 /**
 911  * Return NULL if the current subpass has no depthstencil attachment.
 912  */
 913 const struct anv_image_view *
 914 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer)
 915 {
 916    const struct anv_subpass *subpass = cmd_buffer->state.subpass;
 917    const struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
 918
 919    if (subpass->depth_stencil_attachment == NULL)
 920       return NULL;
 921
 922    const struct anv_image_view *iview =
 923       fb->attachments[subpass->depth_stencil_attachment->attachment];
 924
 925    assert(iview->aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT |
 926                                 VK_IMAGE_ASPECT_STENCIL_BIT));
 927
 928    return iview;
 929 }
 930
 931 static struct anv_push_descriptor_set *
 932 anv_cmd_buffer_get_push_descriptor_set(struct anv_cmd_buffer *cmd_buffer,
 933                                        VkPipelineBindPoint bind_point,
 934                                        uint32_t set)
 935 {
 936    struct anv_cmd_pipeline_state *pipe_state;
 937    if (bind_point == VK_PIPELINE_BIND_POINT_COMPUTE) {
 938       pipe_state = &cmd_buffer->state.compute.base;
 939    } else {
 940       assert(bind_point == VK_PIPELINE_BIND_POINT_GRAPHICS);
 941       pipe_state = &cmd_buffer->state.gfx.base;
 942    }
 943
 944    struct anv_push_descriptor_set **push_set =
 945       &pipe_state->push_descriptors[set];
 946
 947    if (*push_set == NULL) {
 948       *push_set = vk_alloc(&cmd_buffer->pool->alloc,
 949                            sizeof(struct anv_push_descriptor_set), 8,
 950                            VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 951       if (*push_set == NULL) {
 952          anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
 953          return NULL;
 954       }
 955    }
 956
 957    return *push_set;
 958 }
 959
 960 void anv_CmdPushDescriptorSetKHR(
 961     VkCommandBuffer commandBuffer,
 962     VkPipelineBindPoint pipelineBindPoint,
 963     VkPipelineLayout _layout,
 964     uint32_t _set,
 965     uint32_t descriptorWriteCount,
 966     const VkWriteDescriptorSet* pDescriptorWrites)
 967 {
 968    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 969    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
 970
 971    assert(_set < MAX_SETS);
 972
 973    struct anv_descriptor_set_layout *set_layout = layout->set[_set].layout;
 974
 975    struct anv_push_descriptor_set *push_set =
 976       anv_cmd_buffer_get_push_descriptor_set(cmd_buffer,
 977                                              pipelineBindPoint, _set);
 978    if (!push_set)
 979       return;
 980
 981    struct anv_descriptor_set *set = &push_set->set;
 982
 983    set->layout = set_layout;
 984    set->size = anv_descriptor_set_layout_size(set_layout);
 985    set->buffer_count = set_layout->buffer_count;
 986    set->buffer_views = push_set->buffer_views;
 987
 988    /* Go through the user supplied descriptors. */
 989    for (uint32_t i = 0; i < descriptorWriteCount; i++) {
 990       const VkWriteDescriptorSet *write = &pDescriptorWrites[i];
 991
 992       switch (write->descriptorType) {
 993       case VK_DESCRIPTOR_TYPE_SAMPLER:
 994       case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
 995       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
 996       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
 997       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
 998          for (uint32_t j = 0; j < write->descriptorCount; j++) {
 999             anv_descriptor_set_write_image_view(set, &cmd_buffer->device->info,
1000                                                 write->pImageInfo + j,
1001                                                 write->descriptorType,
1002                                                 write->dstBinding,
1003                                                 write->dstArrayElement + j);
1004          }
1005          break;
1006
1007       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
1008       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
1009          for (uint32_t j = 0; j < write->descriptorCount; j++) {
1010             ANV_FROM_HANDLE(anv_buffer_view, bview,
1011                             write->pTexelBufferView[j]);
1012
1013             anv_descriptor_set_write_buffer_view(set,
1014                                                  write->descriptorType,
1015                                                  bview,
1016                                                  write->dstBinding,
1017                                                  write->dstArrayElement + j);
1018          }
1019          break;
1020
1021       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
1022       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
1023       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
1024       case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
1025          for (uint32_t j = 0; j < write->descriptorCount; j++) {
1026             assert(write->pBufferInfo[j].buffer);
1027             ANV_FROM_HANDLE(anv_buffer, buffer, write->pBufferInfo[j].buffer);
1028             assert(buffer);
1029
1030             anv_descriptor_set_write_buffer(set,
1031                                             cmd_buffer->device,
1032                                             &cmd_buffer->surface_state_stream,
1033                                             write->descriptorType,
1034                                             buffer,
1035                                             write->dstBinding,
1036                                             write->dstArrayElement + j,
1037                                             write->pBufferInfo[j].offset,
1038                                             write->pBufferInfo[j].range);
1039          }
1040          break;
1041
1042       default:
1043          break;
1044       }
1045    }
1046
1047    anv_cmd_buffer_bind_descriptor_set(cmd_buffer, pipelineBindPoint,
1048                                       layout, _set, set, NULL, NULL);
1049 }
1050
1051 void anv_CmdPushDescriptorSetWithTemplateKHR(
1052     VkCommandBuffer                             commandBuffer,
1053     VkDescriptorUpdateTemplate                  descriptorUpdateTemplate,
1054     VkPipelineLayout                            _layout,
1055     uint32_t                                    _set,
1056     const void*                                 pData)
1057 {
1058    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1059    ANV_FROM_HANDLE(anv_descriptor_update_template, template,
1060                    descriptorUpdateTemplate);
1061    ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
1062
1063    assert(_set < MAX_PUSH_DESCRIPTORS);
1064
1065    struct anv_descriptor_set_layout *set_layout = layout->set[_set].layout;
1066
1067    struct anv_push_descriptor_set *push_set =
1068       anv_cmd_buffer_get_push_descriptor_set(cmd_buffer,
1069                                              template->bind_point, _set);
1070    if (!push_set)
1071       return;
1072
1073    struct anv_descriptor_set *set = &push_set->set;
1074
1075    set->layout = set_layout;
1076    set->size = anv_descriptor_set_layout_size(set_layout);
1077    set->buffer_count = set_layout->buffer_count;
1078    set->buffer_views = push_set->buffer_views;
1079
1080    anv_descriptor_set_write_template(set,
1081                                      cmd_buffer->device,
1082                                      &cmd_buffer->surface_state_stream,
1083                                      template,
1084                                      pData);
1085
1086    anv_cmd_buffer_bind_descriptor_set(cmd_buffer, template->bind_point,
1087                                       layout, _set, set, NULL, NULL);
1088 }
1089
1090 void anv_CmdSetDeviceMask(
1091     VkCommandBuffer                             commandBuffer,
1092     uint32_t                                    deviceMask)
1093 {
1094    /* No-op */
1095 }