iree/hal/local/task_command_buffer.c

// Copyright 2020 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//      https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "iree/hal/local/task_command_buffer.h"

#include "iree/base/debugging.h"
#include "iree/base/tracing.h"
#include "iree/hal/local/local_descriptor_set_layout.h"
#include "iree/hal/local/local_executable.h"
#include "iree/hal/local/local_executable_layout.h"
#include "iree/task/list.h"
#include "iree/task/submission.h"
#include "iree/task/task.h"

//===----------------------------------------------------------------------===//
// iree_hal_task_command_buffer_t
//===----------------------------------------------------------------------===//

// iree/task/-based command buffer.
// We track a minimal amount of state here and incrementally build out the task
// DAG that we can submit to the task system directly. There's no intermediate
// data structures and we produce the iree_task_ts directly. In the steady state
// all allocations are served from a shared per-device block pool with no
// additional allocations required during recording or execution. That means our
// command buffer here is essentially just a builder for the task system types
// and manager of the lifetime of the tasks.
typedef struct {
  iree_hal_resource_t resource;

  iree_hal_device_t* device;
  iree_task_scope_t* scope;
  iree_hal_command_buffer_mode_t mode;
  iree_hal_command_category_t allowed_categories;

  // Arena used for all allocations; references the shared device block pool.
  iree_arena_allocator_t arena;

  // One or more tasks at the root of the command buffer task DAG.
  // These tasks are all able to execute concurrently and will be the initial
  // ready task set in the submission.
  iree_task_list_t root_tasks;

  // One or more tasks at the leaves of the DAG.
  // Only once all these tasks have completed execution will the command buffer
  // be considered completed as a whole.
  //
  // An empty list indicates that root_tasks are also the leaves.
  iree_task_list_t leaf_tasks;

  // TODO(benvanik): move this out of the struct and allocate from the arena -
  // we only need this during recording and it's ~4KB of waste otherwise.
  // State tracked within the command buffer during recording only.
  struct {
    // The last global barrier that was inserted, if any.
    // The barrier is allocated and inserted into the DAG when requested but the
    // actual barrier dependency list is only allocated and set on flushes.
    // This lets us allocate the appropriately sized barrier task list from the
    // arena even though when the barrier is recorded we don't yet know what
    // other tasks we'll be emitting as we walk the command stream.
    iree_task_barrier_t* open_barrier;

    // The number of tasks in the open barrier (|open_tasks|), used to quickly
    // allocate storage for the task list without needing to walk the list.
    iree_host_size_t open_task_count;

    // All execution tasks emitted that must execute after |open_barrier|.
    iree_task_list_t open_tasks;

    // A flattened list of all available descriptor set bindings.
    // As descriptor sets are pushed/bound the bindings will be updated to
    // represent the fully-translated binding data pointer.
    // TODO(benvanik): support proper mapping semantics and track the
    // iree_hal_buffer_mapping_t and map/unmap where appropriate.
    iree_hal_executable_binding_ptr_t
        bindings[IREE_HAL_LOCAL_MAX_DESCRIPTOR_SET_COUNT *
                 IREE_HAL_LOCAL_MAX_DESCRIPTOR_BINDING_COUNT];
    iree_device_size_t
        binding_lengths[IREE_HAL_LOCAL_MAX_DESCRIPTOR_SET_COUNT *
                        IREE_HAL_LOCAL_MAX_DESCRIPTOR_BINDING_COUNT];

    // All available push constants updated each time push_constants is called.
    // Reset only with the command buffer and otherwise will maintain its values
    // during recording to allow for partial push_constants updates.
    uint32_t push_constants[IREE_HAL_LOCAL_MAX_PUSH_CONSTANT_COUNT];
  } state;
} iree_hal_task_command_buffer_t;

static const iree_hal_command_buffer_vtable_t
    iree_hal_task_command_buffer_vtable;

static iree_hal_task_command_buffer_t* iree_hal_task_command_buffer_cast(
    iree_hal_command_buffer_t* base_value) {
  IREE_HAL_ASSERT_TYPE(base_value, &iree_hal_task_command_buffer_vtable);
  return (iree_hal_task_command_buffer_t*)base_value;
}

iree_status_t iree_hal_task_command_buffer_create(
    iree_hal_device_t* device, iree_task_scope_t* scope,
    iree_hal_command_buffer_mode_t mode,
    iree_hal_command_category_t command_categories,
    iree_arena_block_pool_t* block_pool,
    iree_hal_command_buffer_t** out_command_buffer) {
  IREE_ASSERT_ARGUMENT(device);
  IREE_ASSERT_ARGUMENT(out_command_buffer);
  *out_command_buffer = NULL;
  if (mode != IREE_HAL_COMMAND_BUFFER_MODE_ONE_SHOT) {
    // If we want reuse we'd need to support duplicating the task DAG after
    // recording or have some kind of copy-on-submit behavior that does so if
    // a command buffer is submitted for execution twice. Allowing for the same
    // command buffer to be enqueued multiple times would be fine so long as
    // execution doesn't overlap (`cmdbuf|cmdbuf` vs
    // `cmdbuf -> semaphore -> cmdbuf`) though we'd still need to be careful
    // that we did the enqueuing and reset of the task structures at the right
    // times. Definitely something that'll be useful in the future... but not
    // today :)
    return iree_make_status(IREE_STATUS_UNIMPLEMENTED,
                            "only one-shot command buffer usage is supported");
  }

  IREE_TRACE_ZONE_BEGIN(z0);

  iree_hal_task_command_buffer_t* command_buffer = NULL;
  iree_status_t status =
      iree_allocator_malloc(iree_hal_device_host_allocator(device),
                            sizeof(*command_buffer), (void**)&command_buffer);
  if (iree_status_is_ok(status)) {
    iree_hal_resource_initialize(&iree_hal_task_command_buffer_vtable,
                                 &command_buffer->resource);
    command_buffer->device = device;
    command_buffer->scope = scope;
    command_buffer->mode = mode;
    command_buffer->allowed_categories = command_categories;
    iree_arena_initialize(block_pool, &command_buffer->arena);
    iree_task_list_initialize(&command_buffer->root_tasks);
    iree_task_list_initialize(&command_buffer->leaf_tasks);
    memset(&command_buffer->state, 0, sizeof(command_buffer->state));
    *out_command_buffer = (iree_hal_command_buffer_t*)command_buffer;
  }

  IREE_TRACE_ZONE_END(z0);
  return status;
}

static void iree_hal_task_command_buffer_reset(
    iree_hal_task_command_buffer_t* command_buffer) {
  memset(&command_buffer->state, 0, sizeof(command_buffer->state));
  iree_task_list_discard(&command_buffer->leaf_tasks);
  iree_task_list_discard(&command_buffer->root_tasks);
  iree_arena_reset(&command_buffer->arena);
}

static void iree_hal_task_command_buffer_destroy(
    iree_hal_command_buffer_t* base_command_buffer) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);
  iree_allocator_t host_allocator =
      iree_hal_device_host_allocator(command_buffer->device);
  IREE_TRACE_ZONE_BEGIN(z0);

  iree_hal_task_command_buffer_reset(command_buffer);
  iree_arena_deinitialize(&command_buffer->arena);
  iree_allocator_free(host_allocator, command_buffer);

  IREE_TRACE_ZONE_END(z0);
}

static iree_hal_command_category_t
iree_hal_task_command_buffer_allowed_categories(
    const iree_hal_command_buffer_t* base_command_buffer) {
  return ((const iree_hal_task_command_buffer_t*)base_command_buffer)
      ->allowed_categories;
}

//===----------------------------------------------------------------------===//
// iree_hal_task_command_buffer_t recording
//===----------------------------------------------------------------------===//

static iree_status_t iree_hal_task_command_buffer_flush_tasks(
    iree_hal_task_command_buffer_t* command_buffer);

static iree_status_t iree_hal_task_command_buffer_begin(
    iree_hal_command_buffer_t* base_command_buffer) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);
  iree_hal_task_command_buffer_reset(command_buffer);
  return iree_ok_status();
}

static iree_status_t iree_hal_task_command_buffer_end(
    iree_hal_command_buffer_t* base_command_buffer) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  // Flush any open barriers.
  IREE_RETURN_IF_ERROR(
      iree_hal_task_command_buffer_flush_tasks(command_buffer));

  // Move the tasks from the leaf list (tail) to the root list (head) if this
  // was the first set of tasks recorded.
  if (iree_task_list_is_empty(&command_buffer->root_tasks) &&
      !iree_task_list_is_empty(&command_buffer->leaf_tasks)) {
    iree_task_list_move(&command_buffer->leaf_tasks,
                        &command_buffer->root_tasks);
  }

  return iree_ok_status();
}

// Flushes all open tasks to the previous barrier and prepares for more
// recording. The root tasks are also populated here when required as this is
// the one place where we can see both halves of the most recent synchronization
// event: those tasks recorded prior (if any) and the task that marks the set of
// tasks that will be recorded after (if any).
static iree_status_t iree_hal_task_command_buffer_flush_tasks(
    iree_hal_task_command_buffer_t* command_buffer) {
  iree_task_barrier_t* open_barrier = command_buffer->state.open_barrier;
  if (open_barrier != NULL) {
    // There is an open barrier we need to fixup the fork out to all of the open
    // tasks that were recorded after it.
    iree_task_t* task_head =
        iree_task_list_front(&command_buffer->state.open_tasks);
    iree_host_size_t dependent_task_count =
        command_buffer->state.open_task_count;
    if (dependent_task_count == 1) {
      // Special-case: only one open task so we can avoid the additional barrier
      // overhead by reusing the completion task.
      iree_task_set_completion_task(&open_barrier->header, task_head);
    } else if (dependent_task_count > 1) {
      // Allocate the list of tasks we'll stash back on the previous barrier.
      // Since we couldn't know at the time how many tasks would end up in the
      // barrier we had to defer it until now.
      iree_task_t** dependent_tasks = NULL;
      IREE_RETURN_IF_ERROR(iree_arena_allocate(
          &command_buffer->arena, dependent_task_count * sizeof(iree_task_t*),
          (void**)&dependent_tasks));
      iree_task_t* task = task_head;
      for (iree_host_size_t i = 0; i < dependent_task_count; ++i) {
        dependent_tasks[i] = task;
        task = task->next_task;
      }
      iree_task_barrier_set_dependent_tasks(open_barrier, dependent_task_count,
                                            dependent_tasks);
    }
  }
  command_buffer->state.open_barrier = NULL;

  // Move the open tasks to the tail as they represent the first half of the
  // *next* barrier that will be inserted.
  if (command_buffer->state.open_task_count > 0) {
    iree_task_list_move(&command_buffer->state.open_tasks,
                        &command_buffer->leaf_tasks);
    command_buffer->state.open_task_count = 0;
  }

  return iree_ok_status();
}

// Emits a global barrier, splitting execution into all prior recorded tasks
// and all subsequent recorded tasks. This is currently the critical piece that
// limits our concurrency: changing to fine-grained barriers (via barrier
// buffers or events) will allow more work to overlap at the cost of more brain
// to build out the proper task graph.
static iree_status_t iree_hal_task_command_buffer_emit_global_barrier(
    iree_hal_task_command_buffer_t* command_buffer) {
  // Flush open tasks to the previous barrier. This resets our state such that
  // we can assign the new open barrier and start recording tasks for it.
  // Previous tasks will be moved into the leaf_tasks list.
  IREE_RETURN_IF_ERROR(
      iree_hal_task_command_buffer_flush_tasks(command_buffer));

  // Allocate the new open barrier.
  // As we are recording forward we can't yet assign the dependent tasks (the
  // second half of the synchronization domain) and instead are just inserting
  // it so we can setup the join from previous tasks (the first half of the
  // synchronization domain).
  iree_task_barrier_t* barrier = NULL;
  IREE_RETURN_IF_ERROR(iree_arena_allocate(&command_buffer->arena,
                                           sizeof(*barrier), (void**)&barrier));
  iree_task_barrier_initialize_empty(command_buffer->scope, barrier);

  // If there were previous tasks then join them to the barrier.
  for (iree_task_t* task = iree_task_list_front(&command_buffer->leaf_tasks);
       task != NULL; task = task->next_task) {
    iree_task_set_completion_task(task, &barrier->header);
  }

  // Move the tasks from the leaf list (tail) to the root list (head) if this
  // was the first set of tasks recorded.
  if (iree_task_list_is_empty(&command_buffer->root_tasks) &&
      !iree_task_list_is_empty(&command_buffer->leaf_tasks)) {
    iree_task_list_move(&command_buffer->leaf_tasks,
                        &command_buffer->root_tasks);
  }

  // Reset the tail of the command buffer to the barrier. This leaves us in a
  // consistent state if the recording ends immediate after this (the barrier
  // will be the last task).
  iree_task_list_initialize(&command_buffer->leaf_tasks);
  iree_task_list_push_back(&command_buffer->leaf_tasks, &barrier->header);

  // NOTE: all new tasks emitted will be executed after this barrier.
  command_buffer->state.open_barrier = barrier;
  command_buffer->state.open_task_count = 0;

  return iree_ok_status();
}

// Emits a the given execution |task| into the current open synchronization
// scope (after state.open_barrier and before the next barrier).
static iree_status_t iree_hal_task_command_buffer_emit_execution_task(
    iree_hal_task_command_buffer_t* command_buffer, iree_task_t* task) {
  if (command_buffer->state.open_barrier == NULL) {
    // If there is no open barrier then we are at the head and going right into
    // the task DAG.
    iree_task_list_push_back(&command_buffer->leaf_tasks, task);
  } else {
    // Append to the open task list that will be flushed to the open barrier.
    iree_task_list_push_back(&command_buffer->state.open_tasks, task);
    ++command_buffer->state.open_task_count;
  }
  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_task_command_buffer_t execution
//===----------------------------------------------------------------------===//

iree_status_t iree_hal_task_command_buffer_issue(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_task_queue_state_t* queue_state, iree_task_t* retire_task,
    iree_arena_allocator_t* arena, iree_task_submission_t* pending_submission) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  // If the command buffer is empty (valid!) then we are a no-op.
  bool has_root_tasks = !iree_task_list_is_empty(&command_buffer->root_tasks);
  if (!has_root_tasks) {
    return iree_ok_status();
  }

  bool has_leaf_tasks = !iree_task_list_is_empty(&command_buffer->leaf_tasks);
  if (has_leaf_tasks) {
    // Chain the retire task onto the leaf tasks as their completion indicates
    // that all commands have completed.
    for (iree_task_t* task = command_buffer->leaf_tasks.head; task != NULL;
         task = task->next_task) {
      iree_task_set_completion_task(task, retire_task);
    }
  } else {
    // If we have no leaf tasks it means that this is a single layer DAG and
    // after the root tasks complete the entire command buffer has completed.
    for (iree_task_t* task = command_buffer->root_tasks.head; task != NULL;
         task = task->next_task) {
      iree_task_set_completion_task(task, retire_task);
    }
  }

  // Enqueue all root tasks that are ready to run immediately.
  // After this all of the command buffer tasks are owned by the submission and
  // we need to ensure the command buffer doesn't try to discard them.
  iree_task_submission_enqueue_list(pending_submission,
                                    &command_buffer->root_tasks);
  iree_task_list_initialize(&command_buffer->leaf_tasks);

  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_execution_barrier
//===----------------------------------------------------------------------===//

static iree_status_t iree_hal_task_command_buffer_execution_barrier(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_execution_stage_t source_stage_mask,
    iree_hal_execution_stage_t target_stage_mask,
    iree_host_size_t memory_barrier_count,
    const iree_hal_memory_barrier_t* memory_barriers,
    iree_host_size_t buffer_barrier_count,
    const iree_hal_buffer_barrier_t* buffer_barriers) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  // TODO(benvanik): actual DAG construction. Right now we are just doing simple
  // global barriers each time and forcing a join-fork point.
  return iree_hal_task_command_buffer_emit_global_barrier(command_buffer);
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_signal_event
//===----------------------------------------------------------------------===//

static iree_status_t iree_hal_task_command_buffer_signal_event(
    iree_hal_command_buffer_t* base_command_buffer, iree_hal_event_t* event,
    iree_hal_execution_stage_t source_stage_mask) {
  // TODO(#4518): implement events. For now we just insert global barriers.
  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_reset_event
//===----------------------------------------------------------------------===//

static iree_status_t iree_hal_task_command_buffer_reset_event(
    iree_hal_command_buffer_t* base_command_buffer, iree_hal_event_t* event,
    iree_hal_execution_stage_t source_stage_mask) {
  // TODO(#4518): implement events. For now we just insert global barriers.
  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_wait_events
//===----------------------------------------------------------------------===//

static iree_status_t iree_hal_task_command_buffer_wait_events(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_host_size_t event_count, const iree_hal_event_t** events,
    iree_hal_execution_stage_t source_stage_mask,
    iree_hal_execution_stage_t target_stage_mask,
    iree_host_size_t memory_barrier_count,
    const iree_hal_memory_barrier_t* memory_barriers,
    iree_host_size_t buffer_barrier_count,
    const iree_hal_buffer_barrier_t* buffer_barriers) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);
  // TODO(#4518): implement events. For now we just insert global barriers.
  return iree_hal_task_command_buffer_emit_global_barrier(command_buffer);
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_discard_buffer
//===----------------------------------------------------------------------===//

static iree_status_t iree_hal_task_command_buffer_discard_buffer(
    iree_hal_command_buffer_t* base_command_buffer, iree_hal_buffer_t* buffer) {
  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_fill_buffer
//===----------------------------------------------------------------------===//
// NOTE: for large fills we could dispatch this as tiles for parallelism.
// We'd want to do some measurement for when it's worth it; filling a 200KB
// buffer: maybe not, filling a 200MB buffer: yeah.

typedef struct {
  iree_task_call_t task;
  iree_hal_buffer_t* target_buffer;
  iree_device_size_t target_offset;
  iree_device_size_t length;
  uint32_t pattern_length;
  uint8_t pattern[8];
} iree_hal_cmd_fill_buffer_t;

static iree_status_t iree_hal_cmd_fill_buffer(
    uintptr_t user_context, iree_task_t* task,
    iree_task_submission_t* pending_submission) {
  const iree_hal_cmd_fill_buffer_t* cmd =
      (const iree_hal_cmd_fill_buffer_t*)user_context;
  IREE_TRACE_ZONE_BEGIN(z0);
  iree_status_t status =
      iree_hal_buffer_fill(cmd->target_buffer, cmd->target_offset, cmd->length,
                           cmd->pattern, cmd->pattern_length);
  IREE_TRACE_ZONE_END(z0);
  return status;
}

static iree_status_t iree_hal_task_command_buffer_fill_buffer(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_buffer_t* target_buffer, iree_device_size_t target_offset,
    iree_device_size_t length, const void* pattern,
    iree_host_size_t pattern_length) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  iree_hal_cmd_fill_buffer_t* cmd = NULL;
  IREE_RETURN_IF_ERROR(
      iree_arena_allocate(&command_buffer->arena, sizeof(*cmd), (void**)&cmd));

  iree_task_call_initialize(
      command_buffer->scope,
      iree_task_make_call_closure(iree_hal_cmd_fill_buffer, (uintptr_t)cmd),
      &cmd->task);
  cmd->target_buffer = target_buffer;
  cmd->target_offset = target_offset;
  cmd->length = length;
  memcpy(cmd->pattern, pattern, pattern_length);
  cmd->pattern_length = pattern_length;

  return iree_hal_task_command_buffer_emit_execution_task(command_buffer,
                                                          &cmd->task.header);
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_update_buffer
//===----------------------------------------------------------------------===//

typedef struct {
  iree_task_call_t task;
  iree_hal_buffer_t* target_buffer;
  iree_device_size_t target_offset;
  iree_device_size_t length;
  uint8_t source_buffer[];
} iree_hal_cmd_update_buffer_t;

static iree_status_t iree_hal_cmd_update_buffer(
    uintptr_t user_context, iree_task_t* task,
    iree_task_submission_t* pending_submission) {
  const iree_hal_cmd_update_buffer_t* cmd =
      (const iree_hal_cmd_update_buffer_t*)user_context;
  IREE_TRACE_ZONE_BEGIN(z0);
  iree_status_t status = iree_hal_buffer_write_data(
      cmd->target_buffer, cmd->target_offset, cmd->source_buffer, cmd->length);
  IREE_TRACE_ZONE_END(z0);
  return status;
}

static iree_status_t iree_hal_task_command_buffer_update_buffer(
    iree_hal_command_buffer_t* base_command_buffer, const void* source_buffer,
    iree_host_size_t source_offset, iree_hal_buffer_t* target_buffer,
    iree_device_size_t target_offset, iree_device_size_t length) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  iree_host_size_t total_cmd_size =
      sizeof(iree_hal_cmd_update_buffer_t) + length;

  iree_hal_cmd_update_buffer_t* cmd = NULL;
  IREE_RETURN_IF_ERROR(iree_arena_allocate(&command_buffer->arena,
                                           total_cmd_size, (void**)&cmd));

  iree_task_call_initialize(
      command_buffer->scope,
      iree_task_make_call_closure(iree_hal_cmd_update_buffer, (uintptr_t)cmd),
      &cmd->task);
  cmd->target_buffer = (iree_hal_buffer_t*)target_buffer;
  cmd->target_offset = target_offset;
  cmd->length = length;

  memcpy(cmd->source_buffer, (const uint8_t*)source_buffer + source_offset,
         cmd->length);

  return iree_hal_task_command_buffer_emit_execution_task(command_buffer,
                                                          &cmd->task.header);
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_copy_buffer
//===----------------------------------------------------------------------===//
// NOTE: for large copies we could dispatch this as tiles for parallelism.
// We'd want to do some measurement for when it's worth it; copying a 200KB
// buffer: maybe not, copying a 200MB buffer: yeah.

typedef struct {
  iree_task_call_t task;
  iree_hal_buffer_t* source_buffer;
  iree_device_size_t source_offset;
  iree_hal_buffer_t* target_buffer;
  iree_device_size_t target_offset;
  iree_device_size_t length;
} iree_hal_cmd_copy_buffer_t;

static iree_status_t iree_hal_cmd_copy_buffer(
    uintptr_t user_context, iree_task_t* task,
    iree_task_submission_t* pending_submission) {
  const iree_hal_cmd_copy_buffer_t* cmd =
      (const iree_hal_cmd_copy_buffer_t*)user_context;
  IREE_TRACE_ZONE_BEGIN(z0);
  iree_status_t status = iree_hal_buffer_copy_data(
      cmd->source_buffer, cmd->source_offset, cmd->target_buffer,
      cmd->target_offset, cmd->length);
  IREE_TRACE_ZONE_END(z0);
  return status;
}

static iree_status_t iree_hal_task_command_buffer_copy_buffer(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_buffer_t* source_buffer, iree_device_size_t source_offset,
    iree_hal_buffer_t* target_buffer, iree_device_size_t target_offset,
    iree_device_size_t length) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  iree_hal_cmd_copy_buffer_t* cmd = NULL;
  IREE_RETURN_IF_ERROR(
      iree_arena_allocate(&command_buffer->arena, sizeof(*cmd), (void**)&cmd));

  iree_task_call_initialize(
      command_buffer->scope,
      iree_task_make_call_closure(iree_hal_cmd_copy_buffer, (uintptr_t)cmd),
      &cmd->task);
  cmd->source_buffer = (iree_hal_buffer_t*)source_buffer;
  cmd->source_offset = source_offset;
  cmd->target_buffer = (iree_hal_buffer_t*)target_buffer;
  cmd->target_offset = target_offset;
  cmd->length = length;

  return iree_hal_task_command_buffer_emit_execution_task(command_buffer,
                                                          &cmd->task.header);
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_push_constants
//===----------------------------------------------------------------------===//
// NOTE: command buffer state change only; enqueues no tasks.

static iree_status_t iree_hal_task_command_buffer_push_constants(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_executable_layout_t* executable_layout, iree_host_size_t offset,
    const void* values, iree_host_size_t values_length) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  if (IREE_UNLIKELY(offset + values_length >=
                    sizeof(command_buffer->state.push_constants))) {
    return iree_make_status(IREE_STATUS_INVALID_ARGUMENT,
                            "push constant range %zu (length=%zu) out of range",
                            offset, values_length);
  }

  memcpy((uint8_t*)&command_buffer->state.push_constants + offset, values,
         values_length);

  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_push_descriptor_set
//===----------------------------------------------------------------------===//
// NOTE: command buffer state change only; enqueues no tasks.

static iree_status_t iree_hal_task_command_buffer_push_descriptor_set(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_executable_layout_t* executable_layout, uint32_t set,
    iree_host_size_t binding_count,
    const iree_hal_descriptor_set_binding_t* bindings) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  if (IREE_UNLIKELY(set >= IREE_HAL_LOCAL_MAX_DESCRIPTOR_SET_COUNT)) {
    return iree_make_status(IREE_STATUS_INVALID_ARGUMENT,
                            "set %u out of bounds", set);
  }

  iree_hal_local_executable_layout_t* local_executable_layout =
      iree_hal_local_executable_layout_cast(executable_layout);
  iree_hal_local_descriptor_set_layout_t* local_set_layout =
      iree_hal_local_descriptor_set_layout_cast(
          local_executable_layout->set_layouts[set]);

  iree_host_size_t binding_base =
      set * IREE_HAL_LOCAL_MAX_DESCRIPTOR_BINDING_COUNT;
  for (iree_host_size_t i = 0; i < binding_count; ++i) {
    if (IREE_UNLIKELY(bindings[i].binding >=
                      IREE_HAL_LOCAL_MAX_DESCRIPTOR_BINDING_COUNT)) {
      return iree_make_status(IREE_STATUS_INVALID_ARGUMENT,
                              "buffer binding index out of bounds");
    }
    iree_host_size_t binding_ordinal = binding_base + bindings[i].binding;

    // TODO(benvanik): track mapping so we can properly map/unmap/flush/etc.
    iree_hal_buffer_mapping_t buffer_mapping;
    IREE_RETURN_IF_ERROR(iree_hal_buffer_map_range(
        bindings[i].buffer, local_set_layout->bindings[binding_ordinal].access,
        bindings[i].offset, bindings[i].length, &buffer_mapping));
    command_buffer->state.bindings[binding_ordinal] =
        buffer_mapping.contents.data;
    command_buffer->state.binding_lengths[binding_ordinal] =
        buffer_mapping.contents.data_length;
  }

  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_bind_descriptor_set
//===----------------------------------------------------------------------===//
// NOTE: command buffer state change only; enqueues no tasks.

static iree_status_t iree_hal_task_command_buffer_bind_descriptor_set(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_executable_layout_t* executable_layout, uint32_t set,
    iree_hal_descriptor_set_t* descriptor_set,
    iree_host_size_t dynamic_offset_count,
    const iree_device_size_t* dynamic_offsets) {
  return iree_make_status(IREE_STATUS_UNIMPLEMENTED,
                          "descriptor set binding not yet implemented");
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_dispatch
//===----------------------------------------------------------------------===//

typedef struct {
  iree_task_dispatch_t task;
  iree_hal_local_executable_t* executable;
  iree_host_size_t ordinal;
  iree_hal_executable_binding_ptr_t* IREE_RESTRICT bindings;
  iree_device_size_t* IREE_RESTRICT binding_lengths;
  uint32_t* IREE_RESTRICT push_constants;
} iree_hal_cmd_dispatch_t;

static iree_status_t iree_hal_cmd_dispatch_tile(
    uintptr_t user_context, const iree_task_tile_context_t* tile_context,
    iree_task_submission_t* pending_submission) {
  const iree_hal_cmd_dispatch_t* cmd =
      (const iree_hal_cmd_dispatch_t*)user_context;
  IREE_TRACE_ZONE_BEGIN(z0);

  iree_hal_executable_dispatch_state_v0_t state;
  // TODO(benvanik): wire up device state (imports, etc) and cache on the
  // command buffer for reuse across all tiles.

  iree_hal_local_executable_call_t call = {
      .state = &state,
      .push_constants = cmd->push_constants,
      .bindings = cmd->bindings,
      .binding_lengths = cmd->binding_lengths,
  };
  memcpy(call.workgroup_id.value, tile_context->workgroup_xyz,
         sizeof(iree_hal_vec3_t));
  memcpy(call.workgroup_size.value, tile_context->workgroup_size,
         sizeof(iree_hal_vec3_t));
  memcpy(call.workgroup_count.value, tile_context->workgroup_count,
         sizeof(iree_hal_vec3_t));
  iree_status_t status = iree_hal_local_executable_issue_call(
      cmd->executable, cmd->ordinal, &call);

  IREE_TRACE_ZONE_END(z0);
  return status;
}

static iree_status_t iree_hal_task_command_buffer_build_dispatch(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_executable_t* executable, int32_t entry_point,
    uint32_t workgroup_x, uint32_t workgroup_y, uint32_t workgroup_z,
    iree_hal_cmd_dispatch_t** out_cmd) {
  iree_hal_task_command_buffer_t* command_buffer =
      iree_hal_task_command_buffer_cast(base_command_buffer);

  iree_hal_local_executable_t* local_executable =
      iree_hal_local_executable_cast(executable);
  iree_host_size_t push_constant_count =
      local_executable->layout->push_constants;
  iree_hal_local_binding_mask_t used_binding_mask =
      local_executable->layout->used_bindings;
  iree_host_size_t used_binding_count =
      iree_math_count_ones_u64(used_binding_mask);

  iree_hal_cmd_dispatch_t* cmd = NULL;
  iree_host_size_t total_cmd_size =
      sizeof(*cmd) + push_constant_count * sizeof(uint32_t) +
      used_binding_count * sizeof(iree_hal_executable_binding_ptr_t) +
      used_binding_count * sizeof(iree_device_size_t);
  IREE_RETURN_IF_ERROR(iree_arena_allocate(&command_buffer->arena,
                                           total_cmd_size, (void**)&cmd));

  cmd->executable = local_executable;
  cmd->ordinal = entry_point;

  uint32_t workgroup_count[3] = {workgroup_x, workgroup_y, workgroup_z};
  // TODO(benvanik): expose on API or keep fixed on executable.
  uint32_t workgroup_size[3] = {1, 1, 1};
  iree_task_dispatch_initialize(command_buffer->scope,
                                iree_task_make_dispatch_closure(
                                    iree_hal_cmd_dispatch_tile, (uintptr_t)cmd),
                                workgroup_size, workgroup_count, &cmd->task);

  // Copy only the push constant range used by the executable.
  uint8_t* cmd_ptr = (uint8_t*)cmd + sizeof(*cmd);
  cmd->push_constants = (uint32_t*)cmd_ptr;
  memcpy(cmd->push_constants, command_buffer->state.push_constants,
         push_constant_count * sizeof(*cmd->push_constants));
  cmd_ptr += push_constant_count * sizeof(*cmd->push_constants);

  // Produce the dense binding list based on the declared bindings used.
  // This allows us to change the descriptor sets and bindings counts supported
  // in the HAL independent of any executable as each executable just gets the
  // flat dense list and doesn't care about our descriptor set stuff.
  //
  // Note that we are just directly setting the binding data pointers here with
  // no ownership/retaining/etc - it's part of the HAL contract that buffers are
  // kept valid for the duration they may be in use.
  cmd->bindings = (iree_hal_executable_binding_ptr_t*)cmd_ptr;
  cmd_ptr += used_binding_count * sizeof(*cmd->bindings);
  cmd->binding_lengths = (iree_device_size_t*)cmd_ptr;
  cmd_ptr += used_binding_count * sizeof(*cmd->binding_lengths);
  iree_host_size_t binding_base = 0;
  for (iree_host_size_t i = 0; i < used_binding_count; ++i) {
    int mask_offset = iree_math_count_trailing_zeros_u64(used_binding_mask);
    int binding_ordinal = binding_base + mask_offset;
    binding_base += mask_offset + 1;
    used_binding_mask = iree_shr(used_binding_mask, mask_offset + 1);
    cmd->bindings[i] = command_buffer->state.bindings[binding_ordinal];
    cmd->binding_lengths[i] =
        command_buffer->state.binding_lengths[binding_ordinal];
    if (!cmd->bindings[i]) {
      return iree_make_status(IREE_STATUS_FAILED_PRECONDITION,
                              "(flat) binding %d is NULL", binding_ordinal);
    }
  }

  *out_cmd = cmd;
  return iree_hal_task_command_buffer_emit_execution_task(command_buffer,
                                                          &cmd->task.header);
}

static iree_status_t iree_hal_task_command_buffer_dispatch(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_executable_t* executable, int32_t entry_point,
    uint32_t workgroup_x, uint32_t workgroup_y, uint32_t workgroup_z) {
  iree_hal_cmd_dispatch_t* cmd = NULL;
  return iree_hal_task_command_buffer_build_dispatch(
      base_command_buffer, executable, entry_point, workgroup_x, workgroup_y,
      workgroup_z, &cmd);
}

static iree_status_t iree_hal_task_command_buffer_dispatch_indirect(
    iree_hal_command_buffer_t* base_command_buffer,
    iree_hal_executable_t* executable, int32_t entry_point,
    iree_hal_buffer_t* workgroups_buffer,
    iree_device_size_t workgroups_offset) {
  // TODO(benvanik): track mapping so we can properly map/unmap/flush/etc.
  iree_hal_buffer_mapping_t buffer_mapping;
  IREE_RETURN_IF_ERROR(iree_hal_buffer_map_range(
      workgroups_buffer, IREE_HAL_MEMORY_ACCESS_READ, workgroups_offset,
      3 * sizeof(uint32_t), &buffer_mapping));

  iree_hal_cmd_dispatch_t* cmd = NULL;
  IREE_RETURN_IF_ERROR(iree_hal_task_command_buffer_build_dispatch(
      base_command_buffer, executable, entry_point, 0, 0, 0, &cmd));
  cmd->task.workgroup_count.ptr = (const uint32_t*)buffer_mapping.contents.data;
  cmd->task.header.flags |= IREE_TASK_FLAG_DISPATCH_INDIRECT;
  return iree_ok_status();
}

//===----------------------------------------------------------------------===//
// iree_hal_command_buffer_vtable_t
//===----------------------------------------------------------------------===//

static const iree_hal_command_buffer_vtable_t
    iree_hal_task_command_buffer_vtable = {
        .destroy = iree_hal_task_command_buffer_destroy,
        .allowed_categories = iree_hal_task_command_buffer_allowed_categories,
        .begin = iree_hal_task_command_buffer_begin,
        .end = iree_hal_task_command_buffer_end,
        .execution_barrier = iree_hal_task_command_buffer_execution_barrier,
        .signal_event = iree_hal_task_command_buffer_signal_event,
        .reset_event = iree_hal_task_command_buffer_reset_event,
        .wait_events = iree_hal_task_command_buffer_wait_events,
        .discard_buffer = iree_hal_task_command_buffer_discard_buffer,
        .fill_buffer = iree_hal_task_command_buffer_fill_buffer,
        .update_buffer = iree_hal_task_command_buffer_update_buffer,
        .copy_buffer = iree_hal_task_command_buffer_copy_buffer,
        .push_constants = iree_hal_task_command_buffer_push_constants,
        .push_descriptor_set = iree_hal_task_command_buffer_push_descriptor_set,
        .bind_descriptor_set = iree_hal_task_command_buffer_bind_descriptor_set,
        .dispatch = iree_hal_task_command_buffer_dispatch,
        .dispatch_indirect = iree_hal_task_command_buffer_dispatch_indirect,
};