blender/intern/cycles/device/device_multi.cpp

/*
 * Copyright 2011-2013 Blender Foundation
 *
 * 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
 *
 * http://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 <sstream>
#include <stdlib.h>

#include "bvh/bvh_multi.h"

#include "device/device.h"
#include "device/device_intern.h"
#include "device/device_network.h"

#include "render/buffers.h"
#include "render/geometry.h"

#include "util/util_foreach.h"
#include "util/util_list.h"
#include "util/util_logging.h"
#include "util/util_map.h"
#include "util/util_time.h"

CCL_NAMESPACE_BEGIN

class MultiDevice : public Device {
 public:
  struct SubDevice {
    Stats stats;
    Device *device;
    map<device_ptr, device_ptr> ptr_map;
    int peer_island_index = -1;
  };

  list<SubDevice> devices, denoising_devices;
  device_ptr unique_key;
  vector<vector<SubDevice *>> peer_islands;
  bool use_denoising;
  bool matching_rendering_and_denoising_devices;

  MultiDevice(DeviceInfo &info, Stats &stats, Profiler &profiler, bool background_)
      : Device(info, stats, profiler, background_),
        unique_key(1),
        use_denoising(!info.denoising_devices.empty())
  {
    foreach (DeviceInfo &subinfo, info.multi_devices) {
      /* Always add CPU devices at the back since GPU devices can change
       * host memory pointers, which CPU uses as device pointer. */
      SubDevice *sub;
      if (subinfo.type == DEVICE_CPU) {
        devices.emplace_back();
        sub = &devices.back();
      }
      else {
        devices.emplace_front();
        sub = &devices.front();
      }

      /* The pointer to 'sub->stats' will stay valid even after new devices
       * are added, since 'devices' is a linked list. */
      sub->device = Device::create(subinfo, sub->stats, profiler, background);
    }

    foreach (DeviceInfo &subinfo, info.denoising_devices) {
      denoising_devices.emplace_front();
      SubDevice *sub = &denoising_devices.front();

      sub->device = Device::create(subinfo, sub->stats, profiler, background);
    }

    /* Build a list of peer islands for the available render devices */
    foreach (SubDevice &sub, devices) {
      /* First ensure that every device is in at least once peer island */
      if (sub.peer_island_index < 0) {
        peer_islands.emplace_back();
        sub.peer_island_index = (int)peer_islands.size() - 1;
        peer_islands[sub.peer_island_index].push_back(&sub);
      }

      if (!info.has_peer_memory) {
        continue;
      }

      /* Second check peer access between devices and fill up the islands accordingly */
      foreach (SubDevice &peer_sub, devices) {
        if (peer_sub.peer_island_index < 0 &&
            peer_sub.device->info.type == sub.device->info.type &&
            peer_sub.device->check_peer_access(sub.device)) {
          peer_sub.peer_island_index = sub.peer_island_index;
          peer_islands[sub.peer_island_index].push_back(&peer_sub);
        }
      }
    }

    /* Try to re-use memory when denoising and render devices use the same physical devices
     * (e.g. OptiX denoising and CUDA rendering device pointing to the same GPU).
     * Ordering has to match as well, so that 'DeviceTask::split' behaves consistent. */
    matching_rendering_and_denoising_devices = denoising_devices.empty() ||
                                               (devices.size() == denoising_devices.size());
    if (matching_rendering_and_denoising_devices) {
      for (list<SubDevice>::iterator device_it = devices.begin(),
                                     denoising_device_it = denoising_devices.begin();
           device_it != devices.end() && denoising_device_it != denoising_devices.end();
           ++device_it, ++denoising_device_it) {
        const DeviceInfo &info = device_it->device->info;
        const DeviceInfo &denoising_info = denoising_device_it->device->info;
        if ((info.type != DEVICE_CUDA && info.type != DEVICE_OPTIX) ||
            (denoising_info.type != DEVICE_CUDA && denoising_info.type != DEVICE_OPTIX) ||
            info.num != denoising_info.num) {
          matching_rendering_and_denoising_devices = false;
          break;
        }
      }
    }

#ifdef WITH_NETWORK
    /* try to add network devices */
    ServerDiscovery discovery(true);
    time_sleep(1.0);

    vector<string> servers = discovery.get_server_list();

    foreach (string &server, servers) {
      Device *device = device_network_create(info, stats, profiler, server.c_str());
      if (device)
        devices.push_back(SubDevice(device));
    }
#endif
  }

  ~MultiDevice()
  {
    foreach (SubDevice &sub, devices)
      delete sub.device;
    foreach (SubDevice &sub, denoising_devices)
      delete sub.device;
  }

  const string &error_message() override
  {
    error_msg.clear();

    foreach (SubDevice &sub, devices)
      error_msg += sub.device->error_message();
    foreach (SubDevice &sub, denoising_devices)
      error_msg += sub.device->error_message();

    return error_msg;
  }

  virtual bool show_samples() const override
  {
    if (devices.size() > 1) {
      return false;
    }
    return devices.front().device->show_samples();
  }

  virtual BVHLayoutMask get_bvh_layout_mask() const override
  {
    BVHLayoutMask bvh_layout_mask = BVH_LAYOUT_ALL;
    BVHLayoutMask bvh_layout_mask_all = BVH_LAYOUT_NONE;
    foreach (const SubDevice &sub_device, devices) {
      BVHLayoutMask device_bvh_layout_mask = sub_device.device->get_bvh_layout_mask();
      bvh_layout_mask &= device_bvh_layout_mask;
      bvh_layout_mask_all |= device_bvh_layout_mask;
    }

    /* With multiple OptiX devices, every device needs its own acceleration structure */
    if (bvh_layout_mask == BVH_LAYOUT_OPTIX) {
      return BVH_LAYOUT_MULTI_OPTIX;
    }

    /* When devices do not share a common BVH layout, fall back to creating one for each */
    const BVHLayoutMask BVH_LAYOUT_OPTIX_EMBREE = (BVH_LAYOUT_OPTIX | BVH_LAYOUT_EMBREE);
    if ((bvh_layout_mask_all & BVH_LAYOUT_OPTIX_EMBREE) == BVH_LAYOUT_OPTIX_EMBREE) {
      return BVH_LAYOUT_MULTI_OPTIX_EMBREE;
    }

    return bvh_layout_mask;
  }

  bool load_kernels(const DeviceRequestedFeatures &requested_features) override
  {
    foreach (SubDevice &sub, devices)
      if (!sub.device->load_kernels(requested_features))
        return false;

    use_denoising = requested_features.use_denoising;
    if (requested_features.use_denoising) {
      /* Only need denoising feature, everything else is unused. */
      DeviceRequestedFeatures denoising_features;
      denoising_features.use_denoising = true;
      foreach (SubDevice &sub, denoising_devices)
        if (!sub.device->load_kernels(denoising_features))
          return false;
    }

    return true;
  }

  bool wait_for_availability(const DeviceRequestedFeatures &requested_features) override
  {
    foreach (SubDevice &sub, devices)
      if (!sub.device->wait_for_availability(requested_features))
        return false;

    if (requested_features.use_denoising) {
      foreach (SubDevice &sub, denoising_devices)
        if (!sub.device->wait_for_availability(requested_features))
          return false;
    }

    return true;
  }

  DeviceKernelStatus get_active_kernel_switch_state() override
  {
    DeviceKernelStatus result = DEVICE_KERNEL_USING_FEATURE_KERNEL;

    foreach (SubDevice &sub, devices) {
      DeviceKernelStatus subresult = sub.device->get_active_kernel_switch_state();
      switch (subresult) {
        case DEVICE_KERNEL_FEATURE_KERNEL_INVALID:
        case DEVICE_KERNEL_FEATURE_KERNEL_AVAILABLE:
          return subresult;

        case DEVICE_KERNEL_USING_FEATURE_KERNEL:
        case DEVICE_KERNEL_UNKNOWN:
          break;
      }
    }

    return result;
  }

  void build_bvh(BVH *bvh, Progress &progress, bool refit) override
  {
    /* Try to build and share a single acceleration structure, if possible */
    if (bvh->params.bvh_layout == BVH_LAYOUT_BVH2 || bvh->params.bvh_layout == BVH_LAYOUT_EMBREE) {
      devices.back().device->build_bvh(bvh, progress, refit);
      return;
    }

    assert(bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX ||
           bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX_EMBREE);

    BVHMulti *const bvh_multi = static_cast<BVHMulti *>(bvh);
    bvh_multi->sub_bvhs.resize(devices.size());

    vector<BVHMulti *> geom_bvhs;
    geom_bvhs.reserve(bvh->geometry.size());
    foreach (Geometry *geom, bvh->geometry) {
      geom_bvhs.push_back(static_cast<BVHMulti *>(geom->bvh));
    }

    /* Broadcast acceleration structure build to all render devices */
    size_t i = 0;
    foreach (SubDevice &sub, devices) {
      /* Change geometry BVH pointers to the sub BVH */
      for (size_t k = 0; k < bvh->geometry.size(); ++k) {
        bvh->geometry[k]->bvh = geom_bvhs[k]->sub_bvhs[i];
      }

      if (!bvh_multi->sub_bvhs[i]) {
        BVHParams params = bvh->params;
        if (bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX)
          params.bvh_layout = BVH_LAYOUT_OPTIX;
        else if (bvh->params.bvh_layout == BVH_LAYOUT_MULTI_OPTIX_EMBREE)
          params.bvh_layout = sub.device->info.type == DEVICE_OPTIX ? BVH_LAYOUT_OPTIX :
                                                                      BVH_LAYOUT_EMBREE;

        /* Skip building a bottom level acceleration structure for non-instanced geometry on Embree
         * (since they are put into the top level directly, see bvh_embree.cpp) */
        if (!params.top_level && params.bvh_layout == BVH_LAYOUT_EMBREE &&
            !bvh->geometry[0]->is_instanced()) {
          i++;
          continue;
        }

        bvh_multi->sub_bvhs[i] = BVH::create(params, bvh->geometry, bvh->objects, sub.device);
      }

      sub.device->build_bvh(bvh_multi->sub_bvhs[i], progress, refit);
      i++;
    }

    /* Change geometry BVH pointers back to the multi BVH. */
    for (size_t k = 0; k < bvh->geometry.size(); ++k) {
      bvh->geometry[k]->bvh = geom_bvhs[k];
    }
  }

  virtual void *osl_memory() override
  {
    if (devices.size() > 1) {
      return NULL;
    }
    return devices.front().device->osl_memory();
  }

  bool is_resident(device_ptr key, Device *sub_device) override
  {
    foreach (SubDevice &sub, devices) {
      if (sub.device == sub_device) {
        return find_matching_mem_device(key, sub)->device == sub_device;
      }
    }
    return false;
  }

  SubDevice *find_matching_mem_device(device_ptr key, SubDevice &sub)
  {
    assert(key != 0 && (sub.peer_island_index >= 0 || sub.ptr_map.find(key) != sub.ptr_map.end()));

    /* Get the memory owner of this key (first try current device, then peer devices) */
    SubDevice *owner_sub = &sub;
    if (owner_sub->ptr_map.find(key) == owner_sub->ptr_map.end()) {
      foreach (SubDevice *island_sub, peer_islands[sub.peer_island_index]) {
        if (island_sub != owner_sub &&
            island_sub->ptr_map.find(key) != island_sub->ptr_map.end()) {
          owner_sub = island_sub;
        }
      }
    }
    return owner_sub;
  }

  SubDevice *find_suitable_mem_device(device_ptr key, const vector<SubDevice *> &island)
  {
    assert(!island.empty());

    /* Get the memory owner of this key or the device with the lowest memory usage when new */
    SubDevice *owner_sub = island.front();
    foreach (SubDevice *island_sub, island) {
      if (key ? (island_sub->ptr_map.find(key) != island_sub->ptr_map.end()) :
                (island_sub->device->stats.mem_used < owner_sub->device->stats.mem_used)) {
        owner_sub = island_sub;
      }
    }
    return owner_sub;
  }

  inline device_ptr find_matching_mem(device_ptr key, SubDevice &sub)
  {
    return find_matching_mem_device(key, sub)->ptr_map[key];
  }

  void mem_alloc(device_memory &mem) override
  {
    device_ptr key = unique_key++;

    if (mem.type == MEM_PIXELS) {
      /* Always allocate pixels memory on all devices
       * This is necessary to ensure PBOs are registered everywhere, which FILM_CONVERT uses */
      foreach (SubDevice &sub, devices) {
        mem.device = sub.device;
        mem.device_pointer = 0;
        mem.device_size = 0;

        sub.device->mem_alloc(mem);
        sub.ptr_map[key] = mem.device_pointer;
      }
    }
    else {
      assert(mem.type == MEM_READ_ONLY || mem.type == MEM_READ_WRITE ||
             mem.type == MEM_DEVICE_ONLY);
      /* The remaining memory types can be distributed across devices */
      foreach (const vector<SubDevice *> &island, peer_islands) {
        SubDevice *owner_sub = find_suitable_mem_device(key, island);
        mem.device = owner_sub->device;
        mem.device_pointer = 0;
        mem.device_size = 0;

        owner_sub->device->mem_alloc(mem);
        owner_sub->ptr_map[key] = mem.device_pointer;
      }
    }

    mem.device = this;
    mem.device_pointer = key;
    stats.mem_alloc(mem.device_size);
  }

  void mem_copy_to(device_memory &mem) override
  {
    device_ptr existing_key = mem.device_pointer;
    device_ptr key = (existing_key) ? existing_key : unique_key++;
    size_t existing_size = mem.device_size;

    /* The tile buffers are allocated on each device (see below), so copy to all of them */
    if (strcmp(mem.name, "RenderBuffers") == 0 && use_denoising) {
      foreach (SubDevice &sub, devices) {
        mem.device = sub.device;
        mem.device_pointer = (existing_key) ? sub.ptr_map[existing_key] : 0;
        mem.device_size = existing_size;

        sub.device->mem_copy_to(mem);
        sub.ptr_map[key] = mem.device_pointer;
      }
    }
    else {
      foreach (const vector<SubDevice *> &island, peer_islands) {
        SubDevice *owner_sub = find_suitable_mem_device(existing_key, island);
        mem.device = owner_sub->device;
        mem.device_pointer = (existing_key) ? owner_sub->ptr_map[existing_key] : 0;
        mem.device_size = existing_size;

        owner_sub->device->mem_copy_to(mem);
        owner_sub->ptr_map[key] = mem.device_pointer;

        if (mem.type == MEM_GLOBAL || mem.type == MEM_TEXTURE) {
          /* Need to create texture objects and update pointer in kernel globals on all devices */
          foreach (SubDevice *island_sub, island) {
            if (island_sub != owner_sub) {
              island_sub->device->mem_copy_to(mem);
            }
          }
        }
      }
    }

    mem.device = this;
    mem.device_pointer = key;
    stats.mem_alloc(mem.device_size - existing_size);
  }

  void mem_copy_from(device_memory &mem, int y, int w, int h, int elem) override
  {
    device_ptr key = mem.device_pointer;
    int i = 0, sub_h = h / devices.size();

    foreach (SubDevice &sub, devices) {
      int sy = y + i * sub_h;
      int sh = (i == (int)devices.size() - 1) ? h - sub_h * i : sub_h;

      SubDevice *owner_sub = find_matching_mem_device(key, sub);
      mem.device = owner_sub->device;
      mem.device_pointer = owner_sub->ptr_map[key];

      owner_sub->device->mem_copy_from(mem, sy, w, sh, elem);
      i++;
    }

    mem.device = this;
    mem.device_pointer = key;
  }

  void mem_zero(device_memory &mem) override
  {
    device_ptr existing_key = mem.device_pointer;
    device_ptr key = (existing_key) ? existing_key : unique_key++;
    size_t existing_size = mem.device_size;

    /* This is a hack to only allocate the tile buffers on denoising devices
     * Similarly the tile buffers also need to be allocated separately on all devices so any
     * overlap rendered for denoising does not interfere with each other */
    if (strcmp(mem.name, "RenderBuffers") == 0 && use_denoising) {
      vector<device_ptr> device_pointers;
      device_pointers.reserve(devices.size());

      foreach (SubDevice &sub, devices) {
        mem.device = sub.device;
        mem.device_pointer = (existing_key) ? sub.ptr_map[existing_key] : 0;
        mem.device_size = existing_size;

        sub.device->mem_zero(mem);
        sub.ptr_map[key] = mem.device_pointer;

        device_pointers.push_back(mem.device_pointer);
      }
      foreach (SubDevice &sub, denoising_devices) {
        if (matching_rendering_and_denoising_devices) {
          sub.ptr_map[key] = device_pointers.front();
          device_pointers.erase(device_pointers.begin());
        }
        else {
          mem.device = sub.device;
          mem.device_pointer = (existing_key) ? sub.ptr_map[existing_key] : 0;
          mem.device_size = existing_size;

          sub.device->mem_zero(mem);
          sub.ptr_map[key] = mem.device_pointer;
        }
      }
    }
    else {
      foreach (const vector<SubDevice *> &island, peer_islands) {
        SubDevice *owner_sub = find_suitable_mem_device(existing_key, island);
        mem.device = owner_sub->device;
        mem.device_pointer = (existing_key) ? owner_sub->ptr_map[existing_key] : 0;
        mem.device_size = existing_size;

        owner_sub->device->mem_zero(mem);
        owner_sub->ptr_map[key] = mem.device_pointer;
      }
    }

    mem.device = this;
    mem.device_pointer = key;
    stats.mem_alloc(mem.device_size - existing_size);
  }

  void mem_free(device_memory &mem) override
  {
    device_ptr key = mem.device_pointer;
    size_t existing_size = mem.device_size;

    /* Free memory that was allocated for all devices (see above) on each device */
    if (mem.type == MEM_PIXELS || (strcmp(mem.name, "RenderBuffers") == 0 && use_denoising)) {
      foreach (SubDevice &sub, devices) {
        mem.device = sub.device;
        mem.device_pointer = sub.ptr_map[key];
        mem.device_size = existing_size;

        sub.device->mem_free(mem);
        sub.ptr_map.erase(sub.ptr_map.find(key));
      }
      foreach (SubDevice &sub, denoising_devices) {
        if (matching_rendering_and_denoising_devices) {
          sub.ptr_map.erase(key);
        }
        else {
          mem.device = sub.device;
          mem.device_pointer = sub.ptr_map[key];
          mem.device_size = existing_size;

          sub.device->mem_free(mem);
          sub.ptr_map.erase(sub.ptr_map.find(key));
        }
      }
    }
    else {
      foreach (const vector<SubDevice *> &island, peer_islands) {
        SubDevice *owner_sub = find_matching_mem_device(key, *island.front());
        mem.device = owner_sub->device;
        mem.device_pointer = owner_sub->ptr_map[key];
        mem.device_size = existing_size;

        owner_sub->device->mem_free(mem);
        owner_sub->ptr_map.erase(owner_sub->ptr_map.find(key));

        if (mem.type == MEM_TEXTURE) {
          /* Free texture objects on all devices */
          foreach (SubDevice *island_sub, island) {
            if (island_sub != owner_sub) {
              island_sub->device->mem_free(mem);
            }
          }
        }
      }
    }

    mem.device = this;
    mem.device_pointer = 0;
    mem.device_size = 0;
    stats.mem_free(existing_size);
  }

  void const_copy_to(const char *name, void *host, size_t size) override
  {
    foreach (SubDevice &sub, devices)
      sub.device->const_copy_to(name, host, size);
  }

  void draw_pixels(device_memory &rgba,
                   int y,
                   int w,
                   int h,
                   int width,
                   int height,
                   int dx,
                   int dy,
                   int dw,
                   int dh,
                   bool transparent,
                   const DeviceDrawParams &draw_params) override
  {
    assert(rgba.type == MEM_PIXELS);

    device_ptr key = rgba.device_pointer;
    int i = 0, sub_h = h / devices.size();
    int sub_height = height / devices.size();

    foreach (SubDevice &sub, devices) {
      int sy = y + i * sub_h;
      int sh = (i == (int)devices.size() - 1) ? h - sub_h * i : sub_h;
      int sheight = (i == (int)devices.size() - 1) ? height - sub_height * i : sub_height;
      int sdy = dy + i * sub_height;
      /* adjust math for w/width */

      rgba.device_pointer = sub.ptr_map[key];
      sub.device->draw_pixels(
          rgba, sy, w, sh, width, sheight, dx, sdy, dw, dh, transparent, draw_params);
      i++;
    }

    rgba.device_pointer = key;
  }

  void map_tile(Device *sub_device, RenderTile &tile) override
  {
    if (!tile.buffer) {
      return;
    }

    foreach (SubDevice &sub, devices) {
      if (sub.device == sub_device) {
        tile.buffer = find_matching_mem(tile.buffer, sub);
        return;
      }
    }

    foreach (SubDevice &sub, denoising_devices) {
      if (sub.device == sub_device) {
        tile.buffer = sub.ptr_map[tile.buffer];
        return;
      }
    }
  }

  int device_number(Device *sub_device) override
  {
    int i = 0;

    foreach (SubDevice &sub, devices) {
      if (sub.device == sub_device)
        return i;
      i++;
    }

    foreach (SubDevice &sub, denoising_devices) {
      if (sub.device == sub_device)
        return i;
      i++;
    }

    return -1;
  }

  void map_neighbor_tiles(Device *sub_device, RenderTileNeighbors &neighbors) override
  {
    for (int i = 0; i < RenderTileNeighbors::SIZE; i++) {
      RenderTile &tile = neighbors.tiles[i];

      if (!tile.buffers) {
        continue;
      }

      device_vector<float> &mem = tile.buffers->buffer;
      tile.buffer = mem.device_pointer;

      if (mem.device == this && matching_rendering_and_denoising_devices) {
        /* Skip unnecessary copies in viewport mode (buffer covers the
         * whole image), but still need to fix up the tile device pointer. */
        map_tile(sub_device, tile);
        continue;
      }

      /* If the tile was rendered on another device, copy its memory to
       * to the current device now, for the duration of the denoising task.
       * Note that this temporarily modifies the RenderBuffers and calls
       * the device, so this function is not thread safe. */
      if (mem.device != sub_device) {
        /* Only copy from device to host once. This is faster, but
         * also required for the case where a CPU thread is denoising
         * a tile rendered on the GPU. In that case we have to avoid
         * overwriting the buffer being de-noised by the CPU thread. */
        if (!tile.buffers->map_neighbor_copied) {
          tile.buffers->map_neighbor_copied = true;
          mem.copy_from_device();
        }

        if (mem.device == this) {
          /* Can re-use memory if tile is already allocated on the sub device. */
          map_tile(sub_device, tile);
          mem.swap_device(sub_device, mem.device_size, tile.buffer);
        }
        else {
          mem.swap_device(sub_device, 0, 0);
        }

        mem.copy_to_device();

        tile.buffer = mem.device_pointer;
        tile.device_size = mem.device_size;

        mem.restore_device();
      }
    }
  }

  void unmap_neighbor_tiles(Device *sub_device, RenderTileNeighbors &neighbors) override
  {
    RenderTile &target_tile = neighbors.target;
    device_vector<float> &mem = target_tile.buffers->buffer;

    if (mem.device == this && matching_rendering_and_denoising_devices) {
      return;
    }

    /* Copy denoised result back to the host. */
    mem.swap_device(sub_device, target_tile.device_size, target_tile.buffer);
    mem.copy_from_device();
    mem.restore_device();

    /* Copy denoised result to the original device. */
    mem.copy_to_device();

    for (int i = 0; i < RenderTileNeighbors::SIZE; i++) {
      RenderTile &tile = neighbors.tiles[i];
      if (!tile.buffers) {
        continue;
      }

      device_vector<float> &mem = tile.buffers->buffer;

      if (mem.device != sub_device && mem.device != this) {
        /* Free up memory again if it was allocated for the copy above. */
        mem.swap_device(sub_device, tile.device_size, tile.buffer);
        sub_device->mem_free(mem);
        mem.restore_device();
      }
    }
  }

  int get_split_task_count(DeviceTask &task) override
  {
    int total_tasks = 0;
    list<DeviceTask> tasks;
    task.split(tasks, devices.size());
    foreach (SubDevice &sub, devices) {
      if (!tasks.empty()) {
        DeviceTask subtask = tasks.front();
        tasks.pop_front();

        total_tasks += sub.device->get_split_task_count(subtask);
      }
    }
    return total_tasks;
  }

  void task_add(DeviceTask &task) override
  {
    list<SubDevice> task_devices = devices;
    if (!denoising_devices.empty()) {
      if (task.type == DeviceTask::DENOISE_BUFFER) {
        /* Denoising tasks should be redirected to the denoising devices entirely. */
        task_devices = denoising_devices;
      }
      else if (task.type == DeviceTask::RENDER && (task.tile_types & RenderTile::DENOISE)) {
        const uint tile_types = task.tile_types;
        /* For normal rendering tasks only redirect the denoising part to the denoising devices.
         * Do not need to split the task here, since they all run through 'acquire_tile'. */
        task.tile_types = RenderTile::DENOISE;
        foreach (SubDevice &sub, denoising_devices) {
          sub.device->task_add(task);
        }
        /* Rendering itself should still be executed on the rendering devices. */
        task.tile_types = tile_types ^ RenderTile::DENOISE;
      }
    }

    list<DeviceTask> tasks;
    task.split(tasks, task_devices.size());

    foreach (SubDevice &sub, task_devices) {
      if (!tasks.empty()) {
        DeviceTask subtask = tasks.front();
        tasks.pop_front();

        if (task.buffer)
          subtask.buffer = find_matching_mem(task.buffer, sub);
        if (task.rgba_byte)
          subtask.rgba_byte = sub.ptr_map[task.rgba_byte];
        if (task.rgba_half)
          subtask.rgba_half = sub.ptr_map[task.rgba_half];
        if (task.shader_input)
          subtask.shader_input = find_matching_mem(task.shader_input, sub);
        if (task.shader_output)
          subtask.shader_output = find_matching_mem(task.shader_output, sub);

        sub.device->task_add(subtask);

        if (task.buffers && task.buffers->buffer.device == this) {
          /* Synchronize access to RenderBuffers, since 'map_neighbor_tiles' is not thread-safe. */
          sub.device->task_wait();
        }
      }
    }
  }

  void task_wait() override
  {
    foreach (SubDevice &sub, devices)
      sub.device->task_wait();
    foreach (SubDevice &sub, denoising_devices)
      sub.device->task_wait();
  }

  void task_cancel() override
  {
    foreach (SubDevice &sub, devices)
      sub.device->task_cancel();
    foreach (SubDevice &sub, denoising_devices)
      sub.device->task_cancel();
  }
};

Device *device_multi_create(DeviceInfo &info, Stats &stats, Profiler &profiler, bool background)
{
  return new MultiDevice(info, stats, profiler, background);
}

CCL_NAMESPACE_END