Cycles: changes to source code folders structure

* Split render/ into scene/ and session/. The scene/ folder now contains the
  scene and its nodes. The session/ folder contains the render session and
  associated data structures like drivers and render buffers.
* Move top level kernel headers into new folders kernel/camera/, kernel/film/,
  kernel/light/, kernel/sample/, kernel/util/
* Move integrator related kernel headers into kernel/integrator/
* Move OSL shaders from kernel/shaders/ to kernel/osl/shaders/

For patches and branches, git merge and rebase should be able to detect the
renames and move over code to the right file.

13 files changed, 4217 insertions, 0 deletions

diff --git a/intern/cycles/session/CMakeLists.txt b/intern/cycles/session/CMakeLists.txt
new file mode 100644
index 00000000000..f441def128e
--- /dev/null
+++ b/intern/cycles/session/CMakeLists.txt
@@ -0,0 +1,48 @@
+# Copyright 2011-2021 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.
+
+set(INC
+  ..
+)
+
+set(SRC
+  buffers.cpp
+  denoising.cpp
+  merge.cpp
+  session.cpp
+  tile.cpp
+)
+
+set(SRC_HEADERS
+  buffers.h
+  display_driver.h
+  denoising.h
+  merge.h
+  output_driver.h
+  session.h
+  tile.h
+)
+
+set(LIB
+  cycles_device
+  cycles_integrator
+  cycles_util
+)
+
+include_directories(${INC})
+include_directories(SYSTEM ${INC_SYS})
+
+add_definitions(${GL_DEFINITIONS})
+
+cycles_add_library(cycles_session "${LIB}" ${SRC} ${SRC_HEADERS})
diff --git a/intern/cycles/session/buffers.cpp b/intern/cycles/session/buffers.cpp
new file mode 100644
index 00000000000..439c0f826ea
--- /dev/null
+++ b/intern/cycles/session/buffers.cpp
@@ -0,0 +1,384 @@
+/*
+ * 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 <stdlib.h>
+
+#include "device/device.h"
+#include "session/buffers.h"
+
+#include "util/util_foreach.h"
+#include "util/util_hash.h"
+#include "util/util_math.h"
+#include "util/util_time.h"
+#include "util/util_types.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* --------------------------------------------------------------------
+ * Convert part information to an index of `BufferParams::pass_offset_`.
+ */
+
+static int pass_type_mode_to_index(PassType pass_type, PassMode mode)
+{
+  int index = static_cast<int>(pass_type) * 2;
+
+  if (mode == PassMode::DENOISED) {
+    ++index;
+  }
+
+  return index;
+}
+
+static int pass_to_index(const BufferPass &pass)
+{
+  return pass_type_mode_to_index(pass.type, pass.mode);
+}
+
+/* --------------------------------------------------------------------
+ * Buffer pass.
+ */
+
+NODE_DEFINE(BufferPass)
+{
+  NodeType *type = NodeType::add("buffer_pass", create);
+
+  const NodeEnum *pass_type_enum = Pass::get_type_enum();
+  const NodeEnum *pass_mode_enum = Pass::get_mode_enum();
+
+  SOCKET_ENUM(type, "Type", *pass_type_enum, PASS_COMBINED);
+  SOCKET_ENUM(mode, "Mode", *pass_mode_enum, static_cast<int>(PassMode::DENOISED));
+  SOCKET_STRING(name, "Name", ustring());
+  SOCKET_BOOLEAN(include_albedo, "Include Albedo", false);
+
+  SOCKET_INT(offset, "Offset", -1);
+
+  return type;
+}
+
+BufferPass::BufferPass() : Node(get_node_type())
+{
+}
+
+BufferPass::BufferPass(const Pass *scene_pass)
+    : Node(get_node_type()),
+      type(scene_pass->get_type()),
+      mode(scene_pass->get_mode()),
+      name(scene_pass->get_name()),
+      include_albedo(scene_pass->get_include_albedo())
+{
+}
+
+PassInfo BufferPass::get_info() const
+{
+  return Pass::get_info(type, include_albedo);
+}
+
+/* --------------------------------------------------------------------
+ * Buffer Params.
+ */
+
+NODE_DEFINE(BufferParams)
+{
+  NodeType *type = NodeType::add("buffer_params", create);
+
+  SOCKET_INT(width, "Width", 0);
+  SOCKET_INT(height, "Height", 0);
+
+  SOCKET_INT(window_x, "Window X", 0);
+  SOCKET_INT(window_y, "Window Y", 0);
+  SOCKET_INT(window_width, "Window Width", 0);
+  SOCKET_INT(window_height, "Window Height", 0);
+
+  SOCKET_INT(full_x, "Full X", 0);
+  SOCKET_INT(full_y, "Full Y", 0);
+  SOCKET_INT(full_width, "Full Width", 0);
+  SOCKET_INT(full_height, "Full Height", 0);
+
+  SOCKET_STRING(layer, "Layer", ustring());
+  SOCKET_STRING(view, "View", ustring());
+  SOCKET_INT(samples, "Samples", 0);
+  SOCKET_FLOAT(exposure, "Exposure", 1.0f);
+  SOCKET_BOOLEAN(use_approximate_shadow_catcher, "Use Approximate Shadow Catcher", false);
+  SOCKET_BOOLEAN(use_transparent_background, "Transparent Background", false);
+
+  /* Notes:
+   *  - Skip passes since they do not follow typical container socket definition.
+   *    Might look into covering those as a socket in the future.
+   *
+   *  - Skip offset, stride, and pass stride since those can be delivered from the passes and
+   *    rest of the sockets. */
+
+  return type;
+}
+
+BufferParams::BufferParams() : Node(get_node_type())
+{
+  reset_pass_offset();
+}
+
+void BufferParams::update_passes()
+{
+  update_offset_stride();
+  reset_pass_offset();
+
+  pass_stride = 0;
+  for (const BufferPass &pass : passes) {
+    if (pass.offset != PASS_UNUSED) {
+      const int index = pass_to_index(pass);
+      if (pass_offset_[index] == PASS_UNUSED) {
+        pass_offset_[index] = pass_stride;
+      }
+
+      pass_stride += pass.get_info().num_components;
+    }
+  }
+}
+
+void BufferParams::update_passes(const vector<Pass *> &scene_passes)
+{
+  passes.clear();
+
+  pass_stride = 0;
+  for (const Pass *scene_pass : scene_passes) {
+    BufferPass buffer_pass(scene_pass);
+
+    if (scene_pass->is_written()) {
+      buffer_pass.offset = pass_stride;
+      pass_stride += scene_pass->get_info().num_components;
+    }
+    else {
+      buffer_pass.offset = PASS_UNUSED;
+    }
+
+    passes.emplace_back(std::move(buffer_pass));
+  }
+
+  update_passes();
+}
+
+void BufferParams::reset_pass_offset()
+{
+  for (int i = 0; i < kNumPassOffsets; ++i) {
+    pass_offset_[i] = PASS_UNUSED;
+  }
+}
+
+int BufferParams::get_pass_offset(PassType pass_type, PassMode mode) const
+{
+  if (pass_type == PASS_NONE || pass_type == PASS_UNUSED) {
+    return PASS_UNUSED;
+  }
+
+  const int index = pass_type_mode_to_index(pass_type, mode);
+  return pass_offset_[index];
+}
+
+const BufferPass *BufferParams::find_pass(string_view name) const
+{
+  for (const BufferPass &pass : passes) {
+    if (pass.name == name) {
+      return &pass;
+    }
+  }
+
+  return nullptr;
+}
+
+const BufferPass *BufferParams::find_pass(PassType type, PassMode mode) const
+{
+  for (const BufferPass &pass : passes) {
+    if (pass.type == type && pass.mode == mode) {
+      return &pass;
+    }
+  }
+
+  return nullptr;
+}
+
+const BufferPass *BufferParams::get_actual_display_pass(PassType type, PassMode mode) const
+{
+  const BufferPass *pass = find_pass(type, mode);
+  return get_actual_display_pass(pass);
+}
+
+const BufferPass *BufferParams::get_actual_display_pass(const BufferPass *pass) const
+{
+  if (!pass) {
+    return nullptr;
+  }
+
+  if (pass->type == PASS_COMBINED) {
+    const BufferPass *shadow_catcher_matte_pass = find_pass(PASS_SHADOW_CATCHER_MATTE, pass->mode);
+    if (shadow_catcher_matte_pass) {
+      pass = shadow_catcher_matte_pass;
+    }
+  }
+
+  return pass;
+}
+
+void BufferParams::update_offset_stride()
+{
+  offset = -(full_x + full_y * width);
+  stride = width;
+}
+
+bool BufferParams::modified(const BufferParams &other) const
+{
+  if (width != other.width || height != other.height) {
+    return true;
+  }
+
+  if (full_x != other.full_x || full_y != other.full_y || full_width != other.full_width ||
+      full_height != other.full_height) {
+    return true;
+  }
+
+  if (window_x != other.window_x || window_y != other.window_y ||
+      window_width != other.window_width || window_height != other.window_height) {
+    return true;
+  }
+
+  if (offset != other.offset || stride != other.stride || pass_stride != other.pass_stride) {
+    return true;
+  }
+
+  if (layer != other.layer || view != other.view) {
+    return false;
+  }
+
+  if (exposure != other.exposure ||
+      use_approximate_shadow_catcher != other.use_approximate_shadow_catcher ||
+      use_transparent_background != other.use_transparent_background) {
+    return true;
+  }
+
+  return !(passes == other.passes);
+}
+
+/* --------------------------------------------------------------------
+ * Render Buffers.
+ */
+
+RenderBuffers::RenderBuffers(Device *device) : buffer(device, "RenderBuffers", MEM_READ_WRITE)
+{
+}
+
+RenderBuffers::~RenderBuffers()
+{
+  buffer.free();
+}
+
+void RenderBuffers::reset(const BufferParams &params_)
+{
+  DCHECK(params_.pass_stride != -1);
+
+  params = params_;
+
+  /* re-allocate buffer */
+  buffer.alloc(params.width * params.pass_stride, params.height);
+}
+
+void RenderBuffers::zero()
+{
+  buffer.zero_to_device();
+}
+
+bool RenderBuffers::copy_from_device()
+{
+  DCHECK(params.pass_stride != -1);
+
+  if (!buffer.device_pointer)
+    return false;
+
+  buffer.copy_from_device(0, params.width * params.pass_stride, params.height);
+
+  return true;
+}
+
+void RenderBuffers::copy_to_device()
+{
+  buffer.copy_to_device();
+}
+
+void render_buffers_host_copy_denoised(RenderBuffers *dst,
+                                       const BufferParams &dst_params,
+                                       const RenderBuffers *src,
+                                       const BufferParams &src_params,
+                                       const size_t src_offset)
+{
+  DCHECK_EQ(dst_params.width, src_params.width);
+  /* TODO(sergey): More sanity checks to avoid buffer overrun. */
+
+  /* Create a map of pass offsets to be copied.
+   * Assume offsets are different to allow copying passes between buffers with different set of
+   * passes. */
+
+  struct {
+    int dst_offset;
+    int src_offset;
+  } pass_offsets[PASS_NUM];
+
+  int num_passes = 0;
+
+  for (int i = 0; i < PASS_NUM; ++i) {
+    const PassType pass_type = static_cast<PassType>(i);
+
+    const int dst_pass_offset = dst_params.get_pass_offset(pass_type, PassMode::DENOISED);
+    if (dst_pass_offset == PASS_UNUSED) {
+      continue;
+    }
+
+    const int src_pass_offset = src_params.get_pass_offset(pass_type, PassMode::DENOISED);
+    if (src_pass_offset == PASS_UNUSED) {
+      continue;
+    }
+
+    pass_offsets[num_passes].dst_offset = dst_pass_offset;
+    pass_offsets[num_passes].src_offset = src_pass_offset;
+    ++num_passes;
+  }
+
+  /* Copy passes. */
+  /* TODO(sergey): Make it more reusable, allowing implement copy of noisy passes. */
+
+  const int64_t dst_width = dst_params.width;
+  const int64_t dst_height = dst_params.height;
+  const int64_t dst_pass_stride = dst_params.pass_stride;
+  const int64_t dst_num_pixels = dst_width * dst_height;
+
+  const int64_t src_pass_stride = src_params.pass_stride;
+  const int64_t src_offset_in_floats = src_offset * src_pass_stride;
+
+  const float *src_pixel = src->buffer.data() + src_offset_in_floats;
+  float *dst_pixel = dst->buffer.data();
+
+  for (int i = 0; i < dst_num_pixels;
+       ++i, src_pixel += src_pass_stride, dst_pixel += dst_pass_stride) {
+    for (int pass_offset_idx = 0; pass_offset_idx < num_passes; ++pass_offset_idx) {
+      const int dst_pass_offset = pass_offsets[pass_offset_idx].dst_offset;
+      const int src_pass_offset = pass_offsets[pass_offset_idx].src_offset;
+
+      /* TODO(sergey): Support non-RGBA passes. */
+      dst_pixel[dst_pass_offset + 0] = src_pixel[src_pass_offset + 0];
+      dst_pixel[dst_pass_offset + 1] = src_pixel[src_pass_offset + 1];
+      dst_pixel[dst_pass_offset + 2] = src_pixel[src_pass_offset + 2];
+      dst_pixel[dst_pass_offset + 3] = src_pixel[src_pass_offset + 3];
+    }
+  }
+}
+
+CCL_NAMESPACE_END
diff --git a/intern/cycles/session/buffers.h b/intern/cycles/session/buffers.h
new file mode 100644
index 00000000000..4c261430bb6
--- /dev/null
+++ b/intern/cycles/session/buffers.h
@@ -0,0 +1,199 @@
+/*
+ * 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.
+ */
+
+#ifndef __BUFFERS_H__
+#define __BUFFERS_H__
+
+#include "device/device_memory.h"
+#include "graph/node.h"
+#include "scene/pass.h"
+
+#include "kernel/kernel_types.h"
+
+#include "util/util_half.h"
+#include "util/util_string.h"
+#include "util/util_thread.h"
+#include "util/util_types.h"
+
+CCL_NAMESPACE_BEGIN
+
+class Device;
+struct DeviceDrawParams;
+struct float4;
+
+/* NOTE: Is not a real scene node. Using Node API for ease of (de)serialization. */
+class BufferPass : public Node {
+ public:
+  NODE_DECLARE
+
+  PassType type = PASS_NONE;
+  PassMode mode = PassMode::NOISY;
+  ustring name;
+  bool include_albedo = false;
+
+  int offset = -1;
+
+  BufferPass();
+  explicit BufferPass(const Pass *scene_pass);
+
+  BufferPass(BufferPass &&other) noexcept = default;
+  BufferPass(const BufferPass &other) = default;
+
+  BufferPass &operator=(BufferPass &&other) = default;
+  BufferPass &operator=(const BufferPass &other) = default;
+
+  ~BufferPass() = default;
+
+  PassInfo get_info() const;
+
+  inline bool operator==(const BufferPass &other) const
+  {
+    return type == other.type && mode == other.mode && name == other.name &&
+           include_albedo == other.include_albedo && offset == other.offset;
+  }
+  inline bool operator!=(const BufferPass &other) const
+  {
+    return !(*this == other);
+  }
+};
+
+/* Buffer Parameters
+ * Size of render buffer and how it fits in the full image (border render). */
+
+/* NOTE: Is not a real scene node. Using Node API for ease of (de)serialization. */
+class BufferParams : public Node {
+ public:
+  NODE_DECLARE
+
+  /* Width/height of the physical buffer. */
+  int width = 0;
+  int height = 0;
+
+  /* Windows defines which part of the buffers is visible. The part outside of the window is
+   * considered an "overscan".
+   *
+   * Window X and Y are relative to the position of the buffer in the full buffer. */
+  int window_x = 0;
+  int window_y = 0;
+  int window_width = 0;
+  int window_height = 0;
+
+  /* Offset into and width/height of the full buffer. */
+  int full_x = 0;
+  int full_y = 0;
+  int full_width = 0;
+  int full_height = 0;
+
+  /* Runtime fields, only valid after `update_passes()` or `update_offset_stride()`. */
+  int offset = -1, stride = -1;
+
+  /* Runtime fields, only valid after `update_passes()`. */
+  int pass_stride = -1;
+
+  /* Properties which are used for accessing buffer pixels outside of scene graph. */
+  vector<BufferPass> passes;
+  ustring layer;
+  ustring view;
+  int samples = 0;
+  float exposure = 1.0f;
+  bool use_approximate_shadow_catcher = false;
+  bool use_transparent_background = false;
+
+  BufferParams();
+
+  BufferParams(BufferParams &&other) noexcept = default;
+  BufferParams(const BufferParams &other) = default;
+
+  BufferParams &operator=(BufferParams &&other) = default;
+  BufferParams &operator=(const BufferParams &other) = default;
+
+  ~BufferParams() = default;
+
+  /* Pre-calculate all fields which depends on the passes.
+   *
+   * When the scene passes are given, the buffer passes will be created from them and stored in
+   * this params, and then params are updated for those passes.
+   * The `update_passes()` without parameters updates offsets and strides which are stored outside
+   * of the passes. */
+  void update_passes();
+  void update_passes(const vector<Pass *> &scene_passes);
+
+  /* Returns PASS_UNUSED if there is no such pass in the buffer. */
+  int get_pass_offset(PassType type, PassMode mode = PassMode::NOISY) const;
+
+  /* Returns nullptr if pass with given name does not exist. */
+  const BufferPass *find_pass(string_view name) const;
+  const BufferPass *find_pass(PassType type, PassMode mode = PassMode::NOISY) const;
+
+  /* Get display pass from its name.
+   * Will do special logic to replace combined pass with shadow catcher matte. */
+  const BufferPass *get_actual_display_pass(PassType type, PassMode mode = PassMode::NOISY) const;
+  const BufferPass *get_actual_display_pass(const BufferPass *pass) const;
+
+  void update_offset_stride();
+
+  bool modified(const BufferParams &other) const;
+
+ protected:
+  void reset_pass_offset();
+
+  /* Multiplied by 2 to be able to store noisy and denoised pass types. */
+  static constexpr int kNumPassOffsets = PASS_NUM * 2;
+
+  /* Indexed by an index derived from pass type and mode, indicates offset of the corresponding
+   * pass in the buffer.
+   * If there are multiple passes with same type and mode contains lowest offset of all of them. */
+  int pass_offset_[kNumPassOffsets];
+};
+
+/* Render Buffers */
+
+class RenderBuffers {
+ public:
+  /* buffer parameters */
+  BufferParams params;
+
+  /* float buffer */
+  device_vector<float> buffer;
+
+  explicit RenderBuffers(Device *device);
+  ~RenderBuffers();
+
+  void reset(const BufferParams &params);
+  void zero();
+
+  bool copy_from_device();
+  void copy_to_device();
+};
+
+/* Copy denoised passes form source to destination.
+ *
+ * Buffer parameters are provided explicitly, allowing to copy pixels between render buffers which
+ * content corresponds to a render result at a non-unit resolution divider.
+ *
+ * `src_offset` allows to offset source pixel index which is used when a fraction of the source
+ * buffer is to be copied.
+ *
+ * Copy happens of the number of pixels in the destination. */
+void render_buffers_host_copy_denoised(RenderBuffers *dst,
+                                       const BufferParams &dst_params,
+                                       const RenderBuffers *src,
+                                       const BufferParams &src_params,
+                                       const size_t src_offset = 0);
+
+CCL_NAMESPACE_END
+
+#endif /* __BUFFERS_H__ */
diff --git a/intern/cycles/session/denoising.cpp b/intern/cycles/session/denoising.cpp
new file mode 100644
index 00000000000..21df068092a
--- /dev/null
+++ b/intern/cycles/session/denoising.cpp
@@ -0,0 +1,934 @@
+/*
+ * Copyright 2011-2018 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 "session/denoising.h"
+
+#if 0
+
+#  include "kernel/filter/filter_defines.h"
+
+#  include "util/util_foreach.h"
+#  include "util/util_map.h"
+#  include "util/util_system.h"
+#  include "util/util_task.h"
+#  include "util/util_time.h"
+
+#  include <OpenImageIO/filesystem.h>
+
+CCL_NAMESPACE_BEGIN
+
+/* Utility Functions */
+
+static void print_progress(int num, int total, int frame, int num_frames)
+{
+  const char *label = "Denoise Frame ";
+  int cols = system_console_width();
+
+  cols -= strlen(label);
+
+  int len = 1;
+  for (int x = total; x > 9; x /= 10) {
+    len++;
+  }
+
+  int bars = cols - 2 * len - 6;
+
+  printf("\r%s", label);
+
+  if (num_frames > 1) {
+    int frame_len = 1;
+    for (int x = num_frames - 1; x > 9; x /= 10) {
+      frame_len++;
+    }
+    bars -= frame_len + 2;
+    printf("%*d ", frame_len, frame);
+  }
+
+  int v = int(float(num) * bars / total);
+  printf("[");
+  for (int i = 0; i < v; i++) {
+    printf("=");
+  }
+  if (v < bars) {
+    printf(">");
+  }
+  for (int i = v + 1; i < bars; i++) {
+    printf(" ");
+  }
+  printf(string_printf("] %%%dd / %d", len, total).c_str(), num);
+  fflush(stdout);
+}
+
+/* Splits in at its last dot, setting suffix to the part after the dot and in to the part before
+ * it. Returns whether a dot was found. */
+static bool split_last_dot(string &in, string &suffix)
+{
+  size_t pos = in.rfind(".");
+  if (pos == string::npos) {
+    return false;
+  }
+  suffix = in.substr(pos + 1);
+  in = in.substr(0, pos);
+  return true;
+}
+
+/* Separate channel names as generated by Blender.
+ * If views is true:
+ *   Inputs are expected in the form RenderLayer.Pass.View.Channel, sets renderlayer to
+ *   "RenderLayer.View" Otherwise: Inputs are expected in the form RenderLayer.Pass.Channel */
+static bool parse_channel_name(
+    string name, string &renderlayer, string &pass, string &channel, bool multiview_channels)
+{
+  if (!split_last_dot(name, channel)) {
+    return false;
+  }
+  string view;
+  if (multiview_channels && !split_last_dot(name, view)) {
+    return false;
+  }
+  if (!split_last_dot(name, pass)) {
+    return false;
+  }
+  renderlayer = name;
+
+  if (multiview_channels) {
+    renderlayer += "." + view;
+  }
+
+  return true;
+}
+
+/* Channel Mapping */
+
+struct ChannelMapping {
+  int channel;
+  string name;
+};
+
+static void fill_mapping(vector<ChannelMapping> &map, int pos, string name, string channels)
+{
+  for (const char *chan = channels.c_str(); *chan; chan++) {
+    map.push_back({pos++, name + "." + *chan});
+  }
+}
+
+static const int INPUT_NUM_CHANNELS = 15;
+static const int INPUT_DENOISING_DEPTH = 0;
+static const int INPUT_DENOISING_NORMAL = 1;
+static const int INPUT_DENOISING_SHADOWING = 4;
+static const int INPUT_DENOISING_ALBEDO = 5;
+static const int INPUT_NOISY_IMAGE = 8;
+static const int INPUT_DENOISING_VARIANCE = 11;
+static const int INPUT_DENOISING_INTENSITY = 14;
+static vector<ChannelMapping> input_channels()
+{
+  vector<ChannelMapping> map;
+  fill_mapping(map, INPUT_DENOISING_DEPTH, "Denoising Depth", "Z");
+  fill_mapping(map, INPUT_DENOISING_NORMAL, "Denoising Normal", "XYZ");
+  fill_mapping(map, INPUT_DENOISING_SHADOWING, "Denoising Shadowing", "X");
+  fill_mapping(map, INPUT_DENOISING_ALBEDO, "Denoising Albedo", "RGB");
+  fill_mapping(map, INPUT_NOISY_IMAGE, "Noisy Image", "RGB");
+  fill_mapping(map, INPUT_DENOISING_VARIANCE, "Denoising Variance", "RGB");
+  fill_mapping(map, INPUT_DENOISING_INTENSITY, "Denoising Intensity", "X");
+  return map;
+}
+
+static const int OUTPUT_NUM_CHANNELS = 3;
+static vector<ChannelMapping> output_channels()
+{
+  vector<ChannelMapping> map;
+  fill_mapping(map, 0, "Combined", "RGB");
+  return map;
+}
+
+/* Renderlayer Handling */
+
+bool DenoiseImageLayer::detect_denoising_channels()
+{
+  /* Map device input to image channels. */
+  input_to_image_channel.clear();
+  input_to_image_channel.resize(INPUT_NUM_CHANNELS, -1);
+
+  foreach (const ChannelMapping &mapping, input_channels()) {
+    vector<string>::iterator i = find(channels.begin(), channels.end(), mapping.name);
+    if (i == channels.end()) {
+      return false;
+    }
+
+    size_t input_channel = mapping.channel;
+    size_t layer_channel = i - channels.begin();
+    input_to_image_channel[input_channel] = layer_to_image_channel[layer_channel];
+  }
+
+  /* Map device output to image channels. */
+  output_to_image_channel.clear();
+  output_to_image_channel.resize(OUTPUT_NUM_CHANNELS, -1);
+
+  foreach (const ChannelMapping &mapping, output_channels()) {
+    vector<string>::iterator i = find(channels.begin(), channels.end(), mapping.name);
+    if (i == channels.end()) {
+      return false;
+    }
+
+    size_t output_channel = mapping.channel;
+    size_t layer_channel = i - channels.begin();
+    output_to_image_channel[output_channel] = layer_to_image_channel[layer_channel];
+  }
+
+  /* Check that all buffer channels are correctly set. */
+  for (int i = 0; i < INPUT_NUM_CHANNELS; i++) {
+    assert(input_to_image_channel[i] >= 0);
+  }
+  for (int i = 0; i < OUTPUT_NUM_CHANNELS; i++) {
+    assert(output_to_image_channel[i] >= 0);
+  }
+
+  return true;
+}
+
+bool DenoiseImageLayer::match_channels(int neighbor,
+                                       const std::vector<string> &channelnames,
+                                       const std::vector<string> &neighbor_channelnames)
+{
+  neighbor_input_to_image_channel.resize(neighbor + 1);
+  vector<int> &mapping = neighbor_input_to_image_channel[neighbor];
+
+  assert(mapping.size() == 0);
+  mapping.resize(input_to_image_channel.size(), -1);
+
+  for (int i = 0; i < input_to_image_channel.size(); i++) {
+    const string &channel = channelnames[input_to_image_channel[i]];
+    std::vector<string>::const_iterator frame_channel = find(
+        neighbor_channelnames.begin(), neighbor_channelnames.end(), channel);
+
+    if (frame_channel == neighbor_channelnames.end()) {
+      return false;
+    }
+
+    mapping[i] = frame_channel - neighbor_channelnames.begin();
+  }
+
+  return true;
+}
+
+/* Denoise Task */
+
+DenoiseTask::DenoiseTask(Device *device,
+                         DenoiserPipeline *denoiser,
+                         int frame,
+                         const vector<int> &neighbor_frames)
+    : denoiser(denoiser),
+      device(device),
+      frame(frame),
+      neighbor_frames(neighbor_frames),
+      current_layer(0),
+      input_pixels(device, "filter input buffer", MEM_READ_ONLY),
+      num_tiles(0)
+{
+  image.samples = denoiser->samples_override;
+}
+
+DenoiseTask::~DenoiseTask()
+{
+  free();
+}
+
+/* Device callbacks */
+
+bool DenoiseTask::acquire_tile(Device *device, Device *tile_device, RenderTile &tile)
+{
+  thread_scoped_lock tile_lock(tiles_mutex);
+
+  if (tiles.empty()) {
+    return false;
+  }
+
+  tile = tiles.front();
+  tiles.pop_front();
+
+  device->map_tile(tile_device, tile);
+
+  print_progress(num_tiles - tiles.size(), num_tiles, frame, denoiser->num_frames);
+
+  return true;
+}
+
+/* Mapping tiles is required for regular rendering since each tile has its separate memory
+ * which may be allocated on a different device.
+ * For standalone denoising, there is a single memory that is present on all devices, so the only
+ * thing that needs to be done here is to specify the surrounding tile geometry.
+ *
+ * However, since there is only one large memory, the denoised result has to be written to
+ * a different buffer to avoid having to copy an entire horizontal slice of the image. */
+void DenoiseTask::map_neighboring_tiles(RenderTileNeighbors &neighbors, Device *tile_device)
+{
+  RenderTile &center_tile = neighbors.tiles[RenderTileNeighbors::CENTER];
+  RenderTile &target_tile = neighbors.target;
+
+  /* Fill tile information. */
+  for (int i = 0; i < RenderTileNeighbors::SIZE; i++) {
+    if (i == RenderTileNeighbors::CENTER) {
+      continue;
+    }
+
+    RenderTile &tile = neighbors.tiles[i];
+    int dx = (i % 3) - 1;
+    int dy = (i / 3) - 1;
+    tile.x = clamp(center_tile.x + dx * denoiser->tile_size.x, 0, image.width);
+    tile.w = clamp(center_tile.x + (dx + 1) * denoiser->tile_size.x, 0, image.width) - tile.x;
+    tile.y = clamp(center_tile.y + dy * denoiser->tile_size.y, 0, image.height);
+    tile.h = clamp(center_tile.y + (dy + 1) * denoiser->tile_size.y, 0, image.height) - tile.y;
+
+    tile.buffer = center_tile.buffer;
+    tile.offset = center_tile.offset;
+    tile.stride = image.width;
+  }
+
+  /* Allocate output buffer. */
+  device_vector<float> *output_mem = new device_vector<float>(
+      tile_device, "denoising_output", MEM_READ_WRITE);
+  output_mem->alloc(OUTPUT_NUM_CHANNELS * center_tile.w * center_tile.h);
+
+  /* Fill output buffer with noisy image, assumed by kernel_filter_finalize
+   * when skipping denoising of some pixels. */
+  float *result = output_mem->data();
+  float *in = &image.pixels[image.num_channels * (center_tile.y * image.width + center_tile.x)];
+
+  const DenoiseImageLayer &layer = image.layers[current_layer];
+  const int *input_to_image_channel = layer.input_to_image_channel.data();
+
+  for (int y = 0; y < center_tile.h; y++) {
+    for (int x = 0; x < center_tile.w; x++, result += OUTPUT_NUM_CHANNELS) {
+      for (int i = 0; i < OUTPUT_NUM_CHANNELS; i++) {
+        result[i] = in[image.num_channels * x + input_to_image_channel[INPUT_NOISY_IMAGE + i]];
+      }
+    }
+    in += image.num_channels * image.width;
+  }
+
+  output_mem->copy_to_device();
+
+  /* Fill output tile info. */
+  target_tile = center_tile;
+  target_tile.buffer = output_mem->device_pointer;
+  target_tile.stride = target_tile.w;
+  target_tile.offset -= target_tile.x + target_tile.y * target_tile.stride;
+
+  thread_scoped_lock output_lock(output_mutex);
+  assert(output_pixels.count(center_tile.tile_index) == 0);
+  output_pixels[target_tile.tile_index] = output_mem;
+}
+
+void DenoiseTask::unmap_neighboring_tiles(RenderTileNeighbors &neighbors)
+{
+  RenderTile &center_tile = neighbors.tiles[RenderTileNeighbors::CENTER];
+  RenderTile &target_tile = neighbors.target;
+
+  thread_scoped_lock output_lock(output_mutex);
+  assert(output_pixels.count(center_tile.tile_index) == 1);
+  device_vector<float> *output_mem = output_pixels[target_tile.tile_index];
+  output_pixels.erase(center_tile.tile_index);
+  output_lock.unlock();
+
+  /* Copy denoised pixels from device. */
+  output_mem->copy_from_device(0, OUTPUT_NUM_CHANNELS * target_tile.w, target_tile.h);
+
+  float *result = output_mem->data();
+  float *out = &image.pixels[image.num_channels * (target_tile.y * image.width + target_tile.x)];
+
+  const DenoiseImageLayer &layer = image.layers[current_layer];
+  const int *output_to_image_channel = layer.output_to_image_channel.data();
+
+  for (int y = 0; y < target_tile.h; y++) {
+    for (int x = 0; x < target_tile.w; x++, result += OUTPUT_NUM_CHANNELS) {
+      for (int i = 0; i < OUTPUT_NUM_CHANNELS; i++) {
+        out[image.num_channels * x + output_to_image_channel[i]] = result[i];
+      }
+    }
+    out += image.num_channels * image.width;
+  }
+
+  /* Free device buffer. */
+  output_mem->free();
+  delete output_mem;
+}
+
+void DenoiseTask::release_tile()
+{
+}
+
+bool DenoiseTask::get_cancel()
+{
+  return false;
+}
+
+void DenoiseTask::create_task(DeviceTask &task)
+{
+  /* Callback functions. */
+  task.acquire_tile = function_bind(&DenoiseTask::acquire_tile, this, device, _1, _2);
+  task.map_neighbor_tiles = function_bind(&DenoiseTask::map_neighboring_tiles, this, _1, _2);
+  task.unmap_neighbor_tiles = function_bind(&DenoiseTask::unmap_neighboring_tiles, this, _1);
+  task.release_tile = function_bind(&DenoiseTask::release_tile, this);
+  task.get_cancel = function_bind(&DenoiseTask::get_cancel, this);
+
+  /* Denoising parameters. */
+  task.denoising = denoiser->params;
+  task.denoising.type = DENOISER_NLM;
+  task.denoising.use = true;
+  task.denoising_from_render = false;
+
+  task.denoising_frames.resize(neighbor_frames.size());
+  for (int i = 0; i < neighbor_frames.size(); i++) {
+    task.denoising_frames[i] = neighbor_frames[i] - frame;
+  }
+
+  /* Buffer parameters. */
+  task.pass_stride = INPUT_NUM_CHANNELS;
+  task.target_pass_stride = OUTPUT_NUM_CHANNELS;
+  task.pass_denoising_data = 0;
+  task.pass_denoising_clean = -1;
+  task.frame_stride = image.width * image.height * INPUT_NUM_CHANNELS;
+
+  /* Create tiles. */
+  thread_scoped_lock tile_lock(tiles_mutex);
+  thread_scoped_lock output_lock(output_mutex);
+
+  tiles.clear();
+  assert(output_pixels.empty());
+  output_pixels.clear();
+
+  int tiles_x = divide_up(image.width, denoiser->tile_size.x);
+  int tiles_y = divide_up(image.height, denoiser->tile_size.y);
+
+  for (int ty = 0; ty < tiles_y; ty++) {
+    for (int tx = 0; tx < tiles_x; tx++) {
+      RenderTile tile;
+      tile.x = tx * denoiser->tile_size.x;
+      tile.y = ty * denoiser->tile_size.y;
+      tile.w = min(image.width - tile.x, denoiser->tile_size.x);
+      tile.h = min(image.height - tile.y, denoiser->tile_size.y);
+      tile.start_sample = 0;
+      tile.num_samples = image.layers[current_layer].samples;
+      tile.sample = 0;
+      tile.offset = 0;
+      tile.stride = image.width;
+      tile.tile_index = ty * tiles_x + tx;
+      tile.task = RenderTile::DENOISE;
+      tile.buffers = NULL;
+      tile.buffer = input_pixels.device_pointer;
+      tiles.push_back(tile);
+    }
+  }
+
+  num_tiles = tiles.size();
+}
+
+/* Denoiser Operations */
+
+bool DenoiseTask::load_input_pixels(int layer)
+{
+  int w = image.width;
+  int h = image.height;
+  int num_pixels = image.width * image.height;
+  int frame_stride = num_pixels * INPUT_NUM_CHANNELS;
+
+  /* Load center image */
+  DenoiseImageLayer &image_layer = image.layers[layer];
+
+  float *buffer_data = input_pixels.data();
+  image.read_pixels(image_layer, buffer_data);
+  buffer_data += frame_stride;
+
+  /* Load neighbor images */
+  for (int i = 0; i < image.in_neighbors.size(); i++) {
+    if (!image.read_neighbor_pixels(i, image_layer, buffer_data)) {
+      error = "Failed to read neighbor frame pixels";
+      return false;
+    }
+    buffer_data += frame_stride;
+  }
+
+  /* Preprocess */
+  buffer_data = input_pixels.data();
+  for (int neighbor = 0; neighbor < image.in_neighbors.size() + 1; neighbor++) {
+    /* Clamp */
+    if (denoiser->params.clamp_input) {
+      for (int i = 0; i < num_pixels * INPUT_NUM_CHANNELS; i++) {
+        buffer_data[i] = clamp(buffer_data[i], -1e8f, 1e8f);
+      }
+    }
+
+    /* Box blur */
+    int r = 5 * denoiser->params.radius;
+    float *data = buffer_data + 14;
+    array<float> temp(num_pixels);
+
+    for (int y = 0; y < h; y++) {
+      for (int x = 0; x < w; x++) {
+        int n = 0;
+        float sum = 0.0f;
+        for (int dx = max(x - r, 0); dx < min(x + r + 1, w); dx++, n++) {
+          sum += data[INPUT_NUM_CHANNELS * (y * w + dx)];
+        }
+        temp[y * w + x] = sum / n;
+      }
+    }
+
+    for (int y = 0; y < h; y++) {
+      for (int x = 0; x < w; x++) {
+        int n = 0;
+        float sum = 0.0f;
+
+        for (int dy = max(y - r, 0); dy < min(y + r + 1, h); dy++, n++) {
+          sum += temp[dy * w + x];
+        }
+
+        data[INPUT_NUM_CHANNELS * (y * w + x)] = sum / n;
+      }
+    }
+
+    /* Highlight compression */
+    data = buffer_data + 8;
+    for (int y = 0; y < h; y++) {
+      for (int x = 0; x < w; x++) {
+        int idx = INPUT_NUM_CHANNELS * (y * w + x);
+        float3 color = make_float3(data[idx], data[idx + 1], data[idx + 2]);
+        color = color_highlight_compress(color, NULL);
+        data[idx] = color.x;
+        data[idx + 1] = color.y;
+        data[idx + 2] = color.z;
+      }
+    }
+
+    buffer_data += frame_stride;
+  }
+
+  /* Copy to device */
+  input_pixels.copy_to_device();
+
+  return true;
+}
+
+/* Task stages */
+
+bool DenoiseTask::load()
+{
+  string center_filepath = denoiser->input[frame];
+  if (!image.load(center_filepath, error)) {
+    return false;
+  }
+
+  if (!image.load_neighbors(denoiser->input, neighbor_frames, error)) {
+    return false;
+  }
+
+  if (image.layers.empty()) {
+    error = "No image layers found to denoise in " + center_filepath;
+    return false;
+  }
+
+  /* Allocate device buffer. */
+  int num_frames = image.in_neighbors.size() + 1;
+  input_pixels.alloc(image.width * INPUT_NUM_CHANNELS, image.height * num_frames);
+  input_pixels.zero_to_device();
+
+  /* Read pixels for first layer. */
+  current_layer = 0;
+  if (!load_input_pixels(current_layer)) {
+    return false;
+  }
+
+  return true;
+}
+
+bool DenoiseTask::exec()
+{
+  for (current_layer = 0; current_layer < image.layers.size(); current_layer++) {
+    /* Read pixels for secondary layers, first was already loaded. */
+    if (current_layer > 0) {
+      if (!load_input_pixels(current_layer)) {
+        return false;
+      }
+    }
+
+    /* Run task on device. */
+    DeviceTask task(DeviceTask::RENDER);
+    create_task(task);
+    device->task_add(task);
+    device->task_wait();
+
+    printf("\n");
+  }
+
+  return true;
+}
+
+bool DenoiseTask::save()
+{
+  bool ok = image.save_output(denoiser->output[frame], error);
+  free();
+  return ok;
+}
+
+void DenoiseTask::free()
+{
+  image.free();
+  input_pixels.free();
+  assert(output_pixels.empty());
+}
+
+/* Denoise Image Storage */
+
+DenoiseImage::DenoiseImage()
+{
+  width = 0;
+  height = 0;
+  num_channels = 0;
+  samples = 0;
+}
+
+DenoiseImage::~DenoiseImage()
+{
+  free();
+}
+
+void DenoiseImage::close_input()
+{
+  in_neighbors.clear();
+}
+
+void DenoiseImage::free()
+{
+  close_input();
+  pixels.clear();
+}
+
+bool DenoiseImage::parse_channels(const ImageSpec &in_spec, string &error)
+{
+  const std::vector<string> &channels = in_spec.channelnames;
+  const ParamValue *multiview = in_spec.find_attribute("multiView");
+  const bool multiview_channels = (multiview && multiview->type().basetype == TypeDesc::STRING &&
+                                   multiview->type().arraylen >= 2);
+
+  layers.clear();
+
+  /* Loop over all the channels in the file, parse their name and sort them
+   * by RenderLayer.
+   * Channels that can't be parsed are directly passed through to the output. */
+  map<string, DenoiseImageLayer> file_layers;
+  for (int i = 0; i < channels.size(); i++) {
+    string layer, pass, channel;
+    if (parse_channel_name(channels[i], layer, pass, channel, multiview_channels)) {
+      file_layers[layer].channels.push_back(pass + "." + channel);
+      file_layers[layer].layer_to_image_channel.push_back(i);
+    }
+  }
+
+  /* Loop over all detected RenderLayers, check whether they contain a full set of input channels.
+   * Any channels that won't be processed internally are also passed through. */
+  for (map<string, DenoiseImageLayer>::iterator i = file_layers.begin(); i != file_layers.end();
+       ++i) {
+    const string &name = i->first;
+    DenoiseImageLayer &layer = i->second;
+
+    /* Check for full pass set. */
+    if (!layer.detect_denoising_channels()) {
+      continue;
+    }
+
+    layer.name = name;
+    layer.samples = samples;
+
+    /* If the sample value isn't set yet, check if there is a layer-specific one in the input file.
+     */
+    if (layer.samples < 1) {
+      string sample_string = in_spec.get_string_attribute("cycles." + name + ".samples", "");
+      if (sample_string != "") {
+        if (!sscanf(sample_string.c_str(), "%d", &layer.samples)) {
+          error = "Failed to parse samples metadata: " + sample_string;
+          return false;
+        }
+      }
+    }
+
+    if (layer.samples < 1) {
+      error = string_printf(
+          "No sample number specified in the file for layer %s or on the command line",
+          name.c_str());
+      return false;
+    }
+
+    layers.push_back(layer);
+  }
+
+  return true;
+}
+
+void DenoiseImage::read_pixels(const DenoiseImageLayer &layer, float *input_pixels)
+{
+  /* Pixels from center file have already been loaded into pixels.
+   * We copy a subset into the device input buffer with channels reshuffled. */
+  const int *input_to_image_channel = layer.input_to_image_channel.data();
+
+  for (int i = 0; i < width * height; i++) {
+    for (int j = 0; j < INPUT_NUM_CHANNELS; j++) {
+      int image_channel = input_to_image_channel[j];
+      input_pixels[i * INPUT_NUM_CHANNELS + j] =
+          pixels[((size_t)i) * num_channels + image_channel];
+    }
+  }
+}
+
+bool DenoiseImage::read_neighbor_pixels(int neighbor,
+                                        const DenoiseImageLayer &layer,
+                                        float *input_pixels)
+{
+  /* Load pixels from neighboring frames, and copy them into device buffer
+   * with channels reshuffled. */
+  size_t num_pixels = (size_t)width * (size_t)height;
+  array<float> neighbor_pixels(num_pixels * num_channels);
+  if (!in_neighbors[neighbor]->read_image(TypeDesc::FLOAT, neighbor_pixels.data())) {
+    return false;
+  }
+
+  const int *input_to_image_channel = layer.neighbor_input_to_image_channel[neighbor].data();
+
+  for (int i = 0; i < width * height; i++) {
+    for (int j = 0; j < INPUT_NUM_CHANNELS; j++) {
+      int image_channel = input_to_image_channel[j];
+      input_pixels[i * INPUT_NUM_CHANNELS + j] =
+          neighbor_pixels[((size_t)i) * num_channels + image_channel];
+    }
+  }
+
+  return true;
+}
+
+bool DenoiseImage::load(const string &in_filepath, string &error)
+{
+  if (!Filesystem::is_regular(in_filepath)) {
+    error = "Couldn't find file: " + in_filepath;
+    return false;
+  }
+
+  unique_ptr<ImageInput> in(ImageInput::open(in_filepath));
+  if (!in) {
+    error = "Couldn't open file: " + in_filepath;
+    return false;
+  }
+
+  in_spec = in->spec();
+  width = in_spec.width;
+  height = in_spec.height;
+  num_channels = in_spec.nchannels;
+
+  if (!parse_channels(in_spec, error)) {
+    return false;
+  }
+
+  if (layers.size() == 0) {
+    error = "Could not find a render layer containing denoising info";
+    return false;
+  }
+
+  size_t num_pixels = (size_t)width * (size_t)height;
+  pixels.resize(num_pixels * num_channels);
+
+  /* Read all channels into buffer. Reading all channels at once is faster
+   * than individually due to interleaved EXR channel storage. */
+  if (!in->read_image(TypeDesc::FLOAT, pixels.data())) {
+    error = "Failed to read image: " + in_filepath;
+    return false;
+  }
+
+  return true;
+}
+
+bool DenoiseImage::load_neighbors(const vector<string> &filepaths,
+                                  const vector<int> &frames,
+                                  string &error)
+{
+  if (frames.size() > DENOISE_MAX_FRAMES - 1) {
+    error = string_printf("Maximum number of neighbors (%d) exceeded\n", DENOISE_MAX_FRAMES - 1);
+    return false;
+  }
+
+  for (int neighbor = 0; neighbor < frames.size(); neighbor++) {
+    int frame = frames[neighbor];
+    const string &filepath = filepaths[frame];
+
+    if (!Filesystem::is_regular(filepath)) {
+      error = "Couldn't find neighbor frame: " + filepath;
+      return false;
+    }
+
+    unique_ptr<ImageInput> in_neighbor(ImageInput::open(filepath));
+    if (!in_neighbor) {
+      error = "Couldn't open neighbor frame: " + filepath;
+      return false;
+    }
+
+    const ImageSpec &neighbor_spec = in_neighbor->spec();
+    if (neighbor_spec.width != width || neighbor_spec.height != height) {
+      error = "Neighbor frame has different dimensions: " + filepath;
+      return false;
+    }
+
+    foreach (DenoiseImageLayer &layer, layers) {
+      if (!layer.match_channels(neighbor, in_spec.channelnames, neighbor_spec.channelnames)) {
+        error = "Neighbor frame misses denoising data passes: " + filepath;
+        return false;
+      }
+    }
+
+    in_neighbors.push_back(std::move(in_neighbor));
+  }
+
+  return true;
+}
+
+bool DenoiseImage::save_output(const string &out_filepath, string &error)
+{
+  /* Save image with identical dimensions, channels and metadata. */
+  ImageSpec out_spec = in_spec;
+
+  /* Ensure that the output frame contains sample information even if the input didn't. */
+  for (int i = 0; i < layers.size(); i++) {
+    string name = "cycles." + layers[i].name + ".samples";
+    if (!out_spec.find_attribute(name, TypeDesc::STRING)) {
+      out_spec.attribute(name, TypeDesc::STRING, string_printf("%d", layers[i].samples));
+    }
+  }
+
+  /* We don't need input anymore at this point, and will possibly
+   * overwrite the same file. */
+  close_input();
+
+  /* Write to temporary file path, so we denoise images in place and don't
+   * risk destroying files when something goes wrong in file saving. */
+  string extension = OIIO::Filesystem::extension(out_filepath);
+  string unique_name = ".denoise-tmp-" + OIIO::Filesystem::unique_path();
+  string tmp_filepath = out_filepath + unique_name + extension;
+  unique_ptr<ImageOutput> out(ImageOutput::create(tmp_filepath));
+
+  if (!out) {
+    error = "Failed to open temporary file " + tmp_filepath + " for writing";
+    return false;
+  }
+
+  /* Open temporary file and write image buffers. */
+  if (!out->open(tmp_filepath, out_spec)) {
+    error = "Failed to open file " + tmp_filepath + " for writing: " + out->geterror();
+    return false;
+  }
+
+  bool ok = true;
+  if (!out->write_image(TypeDesc::FLOAT, pixels.data())) {
+    error = "Failed to write to file " + tmp_filepath + ": " + out->geterror();
+    ok = false;
+  }
+
+  if (!out->close()) {
+    error = "Failed to save to file " + tmp_filepath + ": " + out->geterror();
+    ok = false;
+  }
+
+  out.reset();
+
+  /* Copy temporary file to output filepath. */
+  string rename_error;
+  if (ok && !OIIO::Filesystem::rename(tmp_filepath, out_filepath, rename_error)) {
+    error = "Failed to move denoised image to " + out_filepath + ": " + rename_error;
+    ok = false;
+  }
+
+  if (!ok) {
+    OIIO::Filesystem::remove(tmp_filepath);
+  }
+
+  return ok;
+}
+
+/* File pattern handling and outer loop over frames */
+
+DenoiserPipeline::DenoiserPipeline(DeviceInfo &device_info)
+{
+  samples_override = 0;
+  tile_size = make_int2(64, 64);
+
+  num_frames = 0;
+
+  /* Initialize task scheduler. */
+  TaskScheduler::init();
+
+  /* Initialize device. */
+  device = Device::create(device_info, stats, profiler, true);
+
+  device->load_kernels(KERNEL_FEATURE_DENOISING);
+}
+
+DenoiserPipeline::~DenoiserPipeline()
+{
+  delete device;
+  TaskScheduler::exit();
+}
+
+bool DenoiserPipeline::run()
+{
+  assert(input.size() == output.size());
+
+  num_frames = output.size();
+
+  for (int frame = 0; frame < num_frames; frame++) {
+    /* Skip empty output paths. */
+    if (output[frame].empty()) {
+      continue;
+    }
+
+    /* Determine neighbor frame numbers that should be used for filtering. */
+    vector<int> neighbor_frames;
+    for (int f = frame - params.neighbor_frames; f <= frame + params.neighbor_frames; f++) {
+      if (f >= 0 && f < num_frames && f != frame) {
+        neighbor_frames.push_back(f);
+      }
+    }
+
+    /* Execute task. */
+    DenoiseTask task(device, this, frame, neighbor_frames);
+    if (!task.load()) {
+      error = task.error;
+      return false;
+    }
+
+    if (!task.exec()) {
+      error = task.error;
+      return false;
+    }
+
+    if (!task.save()) {
+      error = task.error;
+      return false;
+    }
+
+    task.free();
+  }
+
+  return true;
+}
+
+CCL_NAMESPACE_END
+
+#endif
diff --git a/intern/cycles/session/denoising.h b/intern/cycles/session/denoising.h
new file mode 100644
index 00000000000..097cc570d06
--- /dev/null
+++ b/intern/cycles/session/denoising.h
@@ -0,0 +1,216 @@
+/*
+ * Copyright 2011-2018 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.
+ */
+
+#ifndef __DENOISING_H__
+#define __DENOISING_H__
+
+#if 0
+
+/* TODO(sergey): Make it explicit and clear when something is a denoiser, its pipeline or
+ * parameters. Currently it is an annoying mixture of terms used interchangeably. */
+
+#  include "device/device.h"
+
+#  include "render/buffers.h"
+
+#  include "util/util_string.h"
+#  include "util/util_unique_ptr.h"
+#  include "util/util_vector.h"
+
+#  include <OpenImageIO/imageio.h>
+
+OIIO_NAMESPACE_USING
+
+CCL_NAMESPACE_BEGIN
+
+/* Denoiser pipeline */
+
+class DenoiserPipeline {
+ public:
+  DenoiserPipeline(DeviceInfo &device_info);
+  ~DenoiserPipeline();
+
+  bool run();
+
+  /* Error message after running, in case of failure. */
+  string error;
+
+  /* Sequential list of frame filepaths to denoise. */
+  vector<string> input;
+  /* Sequential list of frame filepaths to write result to. Empty entries
+   * are skipped, so only a subset of the sequence can be denoised while
+   * taking into account all input frames. */
+  vector<string> output;
+
+  /* Sample number override, takes precedence over values from input frames. */
+  int samples_override;
+  /* Tile size for processing on device. */
+  int2 tile_size;
+
+  /* Equivalent to the settings in the regular denoiser. */
+  DenoiseParams params;
+
+ protected:
+  friend class DenoiseTask;
+
+  Stats stats;
+  Profiler profiler;
+  Device *device;
+
+  int num_frames;
+};
+
+/* Denoise Image Layer */
+
+struct DenoiseImageLayer {
+  string name;
+  /* All channels belonging to this DenoiseImageLayer. */
+  vector<string> channels;
+  /* Layer to image channel mapping. */
+  vector<int> layer_to_image_channel;
+
+  /* Sample amount that was used for rendering this layer. */
+  int samples;
+
+  /* Device input channel will be copied from image channel input_to_image_channel[i]. */
+  vector<int> input_to_image_channel;
+
+  /* input_to_image_channel of the secondary frames, if any are used. */
+  vector<vector<int>> neighbor_input_to_image_channel;
+
+  /* Write i-th channel of the processing output to output_to_image_channel[i]-th channel of the
+   * file. */
+  vector<int> output_to_image_channel;
+
+  /* Detect whether this layer contains a full set of channels and set up the offsets accordingly.
+   */
+  bool detect_denoising_channels();
+
+  /* Map the channels of a secondary frame to the channels that are required for processing,
+   * fill neighbor_input_to_image_channel if all are present or return false if a channel are
+   * missing. */
+  bool match_channels(int neighbor,
+                      const std::vector<string> &channelnames,
+                      const std::vector<string> &neighbor_channelnames);
+};
+
+/* Denoise Image Data */
+
+class DenoiseImage {
+ public:
+  DenoiseImage();
+  ~DenoiseImage();
+
+  /* Dimensions */
+  int width, height, num_channels;
+
+  /* Samples */
+  int samples;
+
+  /* Pixel buffer with interleaved channels. */
+  array<float> pixels;
+
+  /* Image file handles */
+  ImageSpec in_spec;
+  vector<unique_ptr<ImageInput>> in_neighbors;
+
+  /* Render layers */
+  vector<DenoiseImageLayer> layers;
+
+  void free();
+
+  /* Open the input image, parse its channels, open the output image and allocate the output
+   * buffer. */
+  bool load(const string &in_filepath, string &error);
+
+  /* Load neighboring frames. */
+  bool load_neighbors(const vector<string> &filepaths, const vector<int> &frames, string &error);
+
+  /* Load subset of pixels from file buffer into input buffer, as needed for denoising
+   * on the device. Channels are reshuffled following the provided mapping. */
+  void read_pixels(const DenoiseImageLayer &layer, float *input_pixels);
+  bool read_neighbor_pixels(int neighbor, const DenoiseImageLayer &layer, float *input_pixels);
+
+  bool save_output(const string &out_filepath, string &error);
+
+ protected:
+  /* Parse input file channels, separate them into DenoiseImageLayers,
+   * detect DenoiseImageLayers with full channel sets,
+   * fill layers and set up the output channels and passthrough map. */
+  bool parse_channels(const ImageSpec &in_spec, string &error);
+
+  void close_input();
+};
+
+/* Denoise Task */
+
+class DenoiseTask {
+ public:
+  DenoiseTask(Device *device,
+              DenoiserPipeline *denoiser,
+              int frame,
+              const vector<int> &neighbor_frames);
+  ~DenoiseTask();
+
+  /* Task stages */
+  bool load();
+  bool exec();
+  bool save();
+  void free();
+
+  string error;
+
+ protected:
+  /* Denoiser parameters and device */
+  DenoiserPipeline *denoiser;
+  Device *device;
+
+  /* Frame number to be denoised */
+  int frame;
+  vector<int> neighbor_frames;
+
+  /* Image file data */
+  DenoiseImage image;
+  int current_layer;
+
+  /* Device input buffer */
+  device_vector<float> input_pixels;
+
+  /* Tiles */
+  thread_mutex tiles_mutex;
+  list<RenderTile> tiles;
+  int num_tiles;
+
+  thread_mutex output_mutex;
+  map<int, device_vector<float> *> output_pixels;
+
+  /* Task handling */
+  bool load_input_pixels(int layer);
+  void create_task(DeviceTask &task);
+
+  /* Device task callbacks */
+  bool acquire_tile(Device *device, Device *tile_device, RenderTile &tile);
+  void map_neighboring_tiles(RenderTileNeighbors &neighbors, Device *tile_device);
+  void unmap_neighboring_tiles(RenderTileNeighbors &neighbors);
+  void release_tile();
+  bool get_cancel();
+};
+
+CCL_NAMESPACE_END
+
+#endif
+
+#endif /* __DENOISING_H__ */
diff --git a/intern/cycles/session/display_driver.h b/intern/cycles/session/display_driver.h
new file mode 100644
index 00000000000..85f305034d7
--- /dev/null
+++ b/intern/cycles/session/display_driver.h
@@ -0,0 +1,131 @@
+/*
+ * Copyright 2021 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.
+ */
+
+#pragma once
+
+#include "util/util_half.h"
+#include "util/util_types.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* Display driver for efficient interactive display of renders.
+ *
+ * Host applications implement this interface for viewport rendering. For best performance, we
+ * recommend:
+ * - Allocating a texture on the GPU to be interactively updated
+ * - Using the graphics interop mechanism to avoid CPU-GPU copying overhead
+ * - Using a dedicated or thread-safe graphics API context for updates, to avoid
+ *   blocking the host application.
+ */
+class DisplayDriver {
+ public:
+  DisplayDriver() = default;
+  virtual ~DisplayDriver() = default;
+
+  /* Render buffer parameters. */
+  struct Params {
+   public:
+    /* Render resolution, ignoring progressive resolution changes.
+     * The texture buffer should be allocated with this size. */
+    int2 size = make_int2(0, 0);
+
+    /* For border rendering, the full resolution of the render, and the offset within that larger
+     * render. */
+    int2 full_size = make_int2(0, 0);
+    int2 full_offset = make_int2(0, 0);
+
+    bool modified(const Params &other) const
+    {
+      return !(full_offset == other.full_offset && full_size == other.full_size &&
+               size == other.size);
+    }
+  };
+
+  /* Update the render from the rendering thread.
+   *
+   * Cycles periodically updates the render to be displayed. For multithreaded updates with
+   * potentially multiple rendering devices, it will call these methods as follows.
+   *
+   * if (driver.update_begin(params, width, height)) {
+   *     parallel_for_each(rendering_device) {
+   *         buffer = driver.map_texture_buffer();
+   *         if (buffer) {
+   *             fill(buffer);
+   *             driver.unmap_texture_buffer();
+   *         }
+   *     }
+   *     driver.update_end();
+   * }
+   *
+   * The parameters may dynamically change due to camera changes in the scene, and resources should
+   * be re-allocated accordingly.
+   *
+   * The width and height passed to update_begin() are the effective render resolution taking into
+   * account progressive resolution changes, which may be equal to or smaller than the params.size.
+   * For efficiency, changes in this resolution should be handled without re-allocating resources,
+   * but rather by using a subset of the full resolution buffer. */
+  virtual bool update_begin(const Params &params, int width, int height) = 0;
+  virtual void update_end() = 0;
+
+  virtual half4 *map_texture_buffer() = 0;
+  virtual void unmap_texture_buffer() = 0;
+
+  /* Optionally return a handle to a native graphics API texture buffer. If supported,
+   * the rendering device may write directly to this buffer instead of calling
+   * map_texture_buffer() and unmap_texture_buffer(). */
+  class GraphicsInterop {
+   public:
+    /* Dimensions of the buffer, in pixels. */
+    int buffer_width = 0;
+    int buffer_height = 0;
+
+    /* OpenGL pixel buffer object. */
+    int opengl_pbo_id = 0;
+
+    /* Clear the entire buffer before doing partial write to it. */
+    bool need_clear = false;
+  };
+
+  virtual GraphicsInterop graphics_interop_get()
+  {
+    return GraphicsInterop();
+  }
+
+  /* (De)activate graphics context required for editing or deleting the graphics interop
+   * object.
+   *
+   * For example, destruction of the CUDA object associated with an OpenGL requires the
+   * OpenGL context to be active. */
+  virtual void graphics_interop_activate(){};
+  virtual void graphics_interop_deactivate(){};
+
+  /* Clear the display buffer by filling it with zeros. */
+  virtual void clear() = 0;
+
+  /* Draw the render using the native graphics API.
+   *
+   * Note that this may be called in parallel to updates. The implementation is responsible for
+   * mutex locking or other mechanisms to avoid conflicts.
+   *
+   * The parameters may have changed since the last update. The implementation is responsible for
+   * deciding to skip or adjust render display for such changes.
+   *
+   * Host application drawing the render buffer should use Session.draw(), which will
+   * call this method. */
+  virtual void draw(const Params &params) = 0;
+};
+
+CCL_NAMESPACE_END
diff --git a/intern/cycles/session/merge.cpp b/intern/cycles/session/merge.cpp
new file mode 100644
index 00000000000..97e9c75d5f7
--- /dev/null
+++ b/intern/cycles/session/merge.cpp
@@ -0,0 +1,516 @@
+/*
+ * Copyright 2011-2019 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 "session/merge.h"
+
+#include "util/util_array.h"
+#include "util/util_map.h"
+#include "util/util_system.h"
+#include "util/util_time.h"
+#include "util/util_unique_ptr.h"
+
+#include <OpenImageIO/filesystem.h>
+#include <OpenImageIO/imageio.h>
+
+OIIO_NAMESPACE_USING
+
+CCL_NAMESPACE_BEGIN
+
+/* Merge Image Layer */
+
+enum MergeChannelOp {
+  MERGE_CHANNEL_NOP,
+  MERGE_CHANNEL_COPY,
+  MERGE_CHANNEL_SUM,
+  MERGE_CHANNEL_AVERAGE
+};
+
+struct MergeImagePass {
+  /* Full channel name. */
+  string channel_name;
+  /* Channel format in the file. */
+  TypeDesc format;
+  /* Type of operation to perform when merging. */
+  MergeChannelOp op;
+  /* Offset of layer channels in input image. */
+  int offset;
+  /* Offset of layer channels in merged image. */
+  int merge_offset;
+};
+
+struct MergeImageLayer {
+  /* Layer name. */
+  string name;
+  /* Passes. */
+  vector<MergeImagePass> passes;
+  /* Sample amount that was used for rendering this layer. */
+  int samples;
+};
+
+/* Merge Image */
+
+struct MergeImage {
+  /* OIIO file handle. */
+  unique_ptr<ImageInput> in;
+  /* Image file path. */
+  string filepath;
+  /* Render layers. */
+  vector<MergeImageLayer> layers;
+};
+
+/* Channel Parsing */
+
+static MergeChannelOp parse_channel_operation(const string &pass_name)
+{
+  if (pass_name == "Depth" || pass_name == "IndexMA" || pass_name == "IndexOB" ||
+      string_startswith(pass_name, "Crypto")) {
+    return MERGE_CHANNEL_COPY;
+  }
+  else if (string_startswith(pass_name, "Debug BVH") ||
+           string_startswith(pass_name, "Debug Ray") ||
+           string_startswith(pass_name, "Debug Render Time")) {
+    return MERGE_CHANNEL_SUM;
+  }
+  else {
+    return MERGE_CHANNEL_AVERAGE;
+  }
+}
+
+/* Splits in at its last dot, setting suffix to the part after the dot and
+ * into the part before it. Returns whether a dot was found. */
+static bool split_last_dot(string &in, string &suffix)
+{
+  size_t pos = in.rfind(".");
+  if (pos == string::npos) {
+    return false;
+  }
+  suffix = in.substr(pos + 1);
+  in = in.substr(0, pos);
+  return true;
+}
+
+/* Separate channel names as generated by Blender.
+ * Multiview format: RenderLayer.Pass.View.Channel
+ * Otherwise: RenderLayer.Pass.Channel */
+static bool parse_channel_name(
+    string name, string &renderlayer, string &pass, string &channel, bool multiview_channels)
+{
+  if (!split_last_dot(name, channel)) {
+    return false;
+  }
+  string view;
+  if (multiview_channels && !split_last_dot(name, view)) {
+    return false;
+  }
+  if (!split_last_dot(name, pass)) {
+    return false;
+  }
+  renderlayer = name;
+
+  if (multiview_channels) {
+    renderlayer += "." + view;
+  }
+
+  return true;
+}
+
+static bool parse_channels(const ImageSpec &in_spec,
+                           vector<MergeImageLayer> &layers,
+                           string &error)
+{
+  const ParamValue *multiview = in_spec.find_attribute("multiView");
+  const bool multiview_channels = (multiview && multiview->type().basetype == TypeDesc::STRING &&
+                                   multiview->type().arraylen >= 2);
+
+  layers.clear();
+
+  /* Loop over all the channels in the file, parse their name and sort them
+   * by RenderLayer.
+   * Channels that can't be parsed are directly passed through to the output. */
+  map<string, MergeImageLayer> file_layers;
+  for (int i = 0; i < in_spec.nchannels; i++) {
+    MergeImagePass pass;
+    pass.channel_name = in_spec.channelnames[i];
+    pass.format = (in_spec.channelformats.size() > 0) ? in_spec.channelformats[i] : in_spec.format;
+    pass.offset = i;
+    pass.merge_offset = i;
+
+    string layername, passname, channelname;
+    if (parse_channel_name(
+            pass.channel_name, layername, passname, channelname, multiview_channels)) {
+      /* Channel part of a render layer. */
+      pass.op = parse_channel_operation(passname);
+    }
+    else {
+      /* Other channels are added in unnamed layer. */
+      layername = "";
+      pass.op = parse_channel_operation(pass.channel_name);
+    }
+
+    file_layers[layername].passes.push_back(pass);
+  }
+
+  /* Loop over all detected render-layers, check whether they contain a full set of input channels.
+   * Any channels that won't be processed internally are also passed through. */
+  for (auto &i : file_layers) {
+    const string &name = i.first;
+    MergeImageLayer &layer = i.second;
+
+    layer.name = name;
+    layer.samples = 0;
+
+    /* Determine number of samples from metadata. */
+    if (layer.name == "") {
+      layer.samples = 1;
+    }
+    else if (layer.samples < 1) {
+      string sample_string = in_spec.get_string_attribute("cycles." + name + ".samples", "");
+      if (sample_string != "") {
+        if (!sscanf(sample_string.c_str(), "%d", &layer.samples)) {
+          error = "Failed to parse samples metadata: " + sample_string;
+          return false;
+        }
+      }
+    }
+
+    if (layer.samples < 1) {
+      error = string_printf(
+          "No sample number specified in the file for layer %s or on the command line",
+          name.c_str());
+      return false;
+    }
+
+    layers.push_back(layer);
+  }
+
+  return true;
+}
+
+static bool open_images(const vector<string> &filepaths, vector<MergeImage> &images, string &error)
+{
+  for (const string &filepath : filepaths) {
+    unique_ptr<ImageInput> in(ImageInput::open(filepath));
+    if (!in) {
+      error = "Couldn't open file: " + filepath;
+      return false;
+    }
+
+    MergeImage image;
+    image.in = std::move(in);
+    image.filepath = filepath;
+    if (!parse_channels(image.in->spec(), image.layers, error)) {
+      return false;
+    }
+
+    if (image.layers.size() == 0) {
+      error = "Could not find a render layer for merging";
+      return false;
+    }
+
+    if (image.in->spec().deep) {
+      error = "Merging deep images not supported.";
+      return false;
+    }
+
+    if (images.size() > 0) {
+      const ImageSpec &base_spec = images[0].in->spec();
+      const ImageSpec &spec = image.in->spec();
+
+      if (base_spec.width != spec.width || base_spec.height != spec.height ||
+          base_spec.depth != spec.depth || base_spec.format != spec.format ||
+          base_spec.deep != spec.deep) {
+        error = "Images do not have matching size and data layout.";
+        return false;
+      }
+    }
+
+    images.push_back(std::move(image));
+  }
+
+  return true;
+}
+
+static void merge_render_time(ImageSpec &spec,
+                              const vector<MergeImage> &images,
+                              const string &name,
+                              const bool average)
+{
+  double time = 0.0;
+
+  for (const MergeImage &image : images) {
+    string time_str = image.in->spec().get_string_attribute(name, "");
+    time += time_human_readable_to_seconds(time_str);
+  }
+
+  if (average) {
+    time /= images.size();
+  }
+
+  spec.attribute(name, TypeDesc::STRING, time_human_readable_from_seconds(time));
+}
+
+static void merge_layer_render_time(ImageSpec &spec,
+                                    const vector<MergeImage> &images,
+                                    const string &layer_name,
+                                    const string &time_name,
+                                    const bool average)
+{
+  string name = "cycles." + layer_name + "." + time_name;
+  double time = 0.0;
+
+  for (const MergeImage &image : images) {
+    string time_str = image.in->spec().get_string_attribute(name, "");
+    time += time_human_readable_to_seconds(time_str);
+  }
+
+  if (average) {
+    time /= images.size();
+  }
+
+  spec.attribute(name, TypeDesc::STRING, time_human_readable_from_seconds(time));
+}
+
+static void merge_channels_metadata(vector<MergeImage> &images,
+                                    ImageSpec &out_spec,
+                                    vector<int> &channel_total_samples)
+{
+  /* Based on first image. */
+  out_spec = images[0].in->spec();
+
+  /* Merge channels and compute offsets. */
+  out_spec.nchannels = 0;
+  out_spec.channelformats.clear();
+  out_spec.channelnames.clear();
+
+  for (MergeImage &image : images) {
+    for (MergeImageLayer &layer : image.layers) {
+      for (MergeImagePass &pass : layer.passes) {
+        /* Test if matching channel already exists in merged image. */
+        bool found = false;
+
+        for (size_t i = 0; i < out_spec.nchannels; i++) {
+          if (pass.channel_name == out_spec.channelnames[i]) {
+            pass.merge_offset = i;
+            channel_total_samples[i] += layer.samples;
+            /* First image wins for channels that can't be averaged or summed. */
+            if (pass.op == MERGE_CHANNEL_COPY) {
+              pass.op = MERGE_CHANNEL_NOP;
+            }
+            found = true;
+            break;
+          }
+        }
+
+        if (!found) {
+          /* Add new channel. */
+          pass.merge_offset = out_spec.nchannels;
+          channel_total_samples.push_back(layer.samples);
+
+          out_spec.channelnames.push_back(pass.channel_name);
+          out_spec.channelformats.push_back(pass.format);
+          out_spec.nchannels++;
+        }
+      }
+    }
+  }
+
+  /* Merge metadata. */
+  merge_render_time(out_spec, images, "RenderTime", false);
+
+  map<string, int> layer_num_samples;
+  for (MergeImage &image : images) {
+    for (MergeImageLayer &layer : image.layers) {
+      if (layer.name != "") {
+        layer_num_samples[layer.name] += layer.samples;
+      }
+    }
+  }
+
+  for (const auto &i : layer_num_samples) {
+    string name = "cycles." + i.first + ".samples";
+    out_spec.attribute(name, TypeDesc::STRING, string_printf("%d", i.second));
+
+    merge_layer_render_time(out_spec, images, i.first, "total_time", false);
+    merge_layer_render_time(out_spec, images, i.first, "render_time", false);
+    merge_layer_render_time(out_spec, images, i.first, "synchronization_time", true);
+  }
+}
+
+static void alloc_pixels(const ImageSpec &spec, array<float> &pixels)
+{
+  const size_t width = spec.width;
+  const size_t height = spec.height;
+  const size_t num_channels = spec.nchannels;
+
+  const size_t num_pixels = (size_t)width * (size_t)height;
+  pixels.resize(num_pixels * num_channels);
+}
+
+static bool merge_pixels(const vector<MergeImage> &images,
+                         const ImageSpec &out_spec,
+                         const vector<int> &channel_total_samples,
+                         array<float> &out_pixels,
+                         string &error)
+{
+  alloc_pixels(out_spec, out_pixels);
+  memset(out_pixels.data(), 0, out_pixels.size() * sizeof(float));
+
+  for (const MergeImage &image : images) {
+    /* Read all channels into buffer. Reading all channels at once is
+     * faster than individually due to interleaved EXR channel storage. */
+    array<float> pixels;
+    alloc_pixels(image.in->spec(), pixels);
+
+    if (!image.in->read_image(TypeDesc::FLOAT, pixels.data())) {
+      error = "Failed to read image: " + image.filepath;
+      return false;
+    }
+
+    for (size_t li = 0; li < image.layers.size(); li++) {
+      const MergeImageLayer &layer = image.layers[li];
+
+      const size_t stride = image.in->spec().nchannels;
+      const size_t out_stride = out_spec.nchannels;
+      const size_t num_pixels = pixels.size();
+
+      for (const MergeImagePass &pass : layer.passes) {
+        size_t offset = pass.offset;
+        size_t out_offset = pass.merge_offset;
+
+        switch (pass.op) {
+          case MERGE_CHANNEL_NOP:
+            break;
+          case MERGE_CHANNEL_COPY:
+            for (; offset < num_pixels; offset += stride, out_offset += out_stride) {
+              out_pixels[out_offset] = pixels[offset];
+            }
+            break;
+          case MERGE_CHANNEL_SUM:
+            for (; offset < num_pixels; offset += stride, out_offset += out_stride) {
+              out_pixels[out_offset] += pixels[offset];
+            }
+            break;
+          case MERGE_CHANNEL_AVERAGE:
+            /* Weights based on sample metadata. Per channel since not
+             * all files are guaranteed to have the same channels. */
+            const int total_samples = channel_total_samples[out_offset];
+            const float t = (float)layer.samples / (float)total_samples;
+
+            for (; offset < num_pixels; offset += stride, out_offset += out_stride) {
+              out_pixels[out_offset] += t * pixels[offset];
+            }
+            break;
+        }
+      }
+    }
+  }
+
+  return true;
+}
+
+static bool save_output(const string &filepath,
+                        const ImageSpec &spec,
+                        const array<float> &pixels,
+                        string &error)
+{
+  /* Write to temporary file path, so we merge images in place and don't
+   * risk destroying files when something goes wrong in file saving. */
+  string extension = OIIO::Filesystem::extension(filepath);
+  string unique_name = ".merge-tmp-" + OIIO::Filesystem::unique_path();
+  string tmp_filepath = filepath + unique_name + extension;
+  unique_ptr<ImageOutput> out(ImageOutput::create(tmp_filepath));
+
+  if (!out) {
+    error = "Failed to open temporary file " + tmp_filepath + " for writing";
+    return false;
+  }
+
+  /* Open temporary file and write image buffers. */
+  if (!out->open(tmp_filepath, spec)) {
+    error = "Failed to open file " + tmp_filepath + " for writing: " + out->geterror();
+    return false;
+  }
+
+  bool ok = true;
+  if (!out->write_image(TypeDesc::FLOAT, pixels.data())) {
+    error = "Failed to write to file " + tmp_filepath + ": " + out->geterror();
+    ok = false;
+  }
+
+  if (!out->close()) {
+    error = "Failed to save to file " + tmp_filepath + ": " + out->geterror();
+    ok = false;
+  }
+
+  out.reset();
+
+  /* Copy temporary file to output filepath. */
+  string rename_error;
+  if (ok && !OIIO::Filesystem::rename(tmp_filepath, filepath, rename_error)) {
+    error = "Failed to move merged image to " + filepath + ": " + rename_error;
+    ok = false;
+  }
+
+  if (!ok) {
+    OIIO::Filesystem::remove(tmp_filepath);
+  }
+
+  return ok;
+}
+
+/* Image Merger */
+
+ImageMerger::ImageMerger()
+{
+}
+
+bool ImageMerger::run()
+{
+  if (input.empty()) {
+    error = "No input file paths specified.";
+    return false;
+  }
+  if (output.empty()) {
+    error = "No output file path specified.";
+    return false;
+  }
+
+  /* Open images and verify they have matching layout. */
+  vector<MergeImage> images;
+  if (!open_images(input, images, error)) {
+    return false;
+  }
+
+  /* Merge metadata and setup channels and offsets. */
+  ImageSpec out_spec;
+  vector<int> channel_total_samples;
+  merge_channels_metadata(images, out_spec, channel_total_samples);
+
+  /* Merge pixels. */
+  array<float> out_pixels;
+  if (!merge_pixels(images, out_spec, channel_total_samples, out_pixels, error)) {
+    return false;
+  }
+
+  /* We don't need input anymore at this point, and will possibly
+   * overwrite the same file. */
+  images.clear();
+
+  /* Save output file. */
+  return save_output(output, out_spec, out_pixels, error);
+}
+
+CCL_NAMESPACE_END
diff --git a/intern/cycles/session/merge.h b/intern/cycles/session/merge.h
new file mode 100644
index 00000000000..87e5d2d4723
--- /dev/null
+++ b/intern/cycles/session/merge.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright 2011-2019 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.
+ */
+
+#ifndef __MERGE_H__
+#define __MERGE_H__
+
+#include "util/util_string.h"
+#include "util/util_vector.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* Merge OpenEXR multilayer renders. */
+
+class ImageMerger {
+ public:
+  ImageMerger();
+  bool run();
+
+  /* Error message after running, in case of failure. */
+  string error;
+
+  /* List of image filepaths to merge. */
+  vector<string> input;
+  /* Output filepath. */
+  string output;
+};
+
+CCL_NAMESPACE_END
+
+#endif /* __MERGE_H__ */
diff --git a/intern/cycles/session/output_driver.h b/intern/cycles/session/output_driver.h
new file mode 100644
index 00000000000..b7e980d71d4
--- /dev/null
+++ b/intern/cycles/session/output_driver.h
@@ -0,0 +1,82 @@
+/*
+ * Copyright 2021 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.
+ */
+
+#pragma once
+
+#include "util/util_math.h"
+#include "util/util_string.h"
+#include "util/util_types.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* Output driver for reading render buffers.
+ *
+ * Host applications implement this interface for outputting render buffers for offline rendering.
+ * Drivers can be used to copy the buffers into the host application or write them directly to
+ * disk. This interface may also be used for interactive display, however the DisplayDriver is more
+ * efficient for that purpose.
+ */
+class OutputDriver {
+ public:
+  OutputDriver() = default;
+  virtual ~OutputDriver() = default;
+
+  class Tile {
+   public:
+    Tile(const int2 offset,
+         const int2 size,
+         const int2 full_size,
+         const string_view layer,
+         const string_view view)
+        : offset(offset), size(size), full_size(full_size), layer(layer), view(view)
+    {
+    }
+    virtual ~Tile() = default;
+
+    const int2 offset;
+    const int2 size;
+    const int2 full_size;
+    const string layer;
+    const string view;
+
+    virtual bool get_pass_pixels(const string_view pass_name,
+                                 const int num_channels,
+                                 float *pixels) const = 0;
+    virtual bool set_pass_pixels(const string_view pass_name,
+                                 const int num_channels,
+                                 const float *pixels) const = 0;
+  };
+
+  /* Write tile once it has finished rendering. */
+  virtual void write_render_tile(const Tile &tile) = 0;
+
+  /* Update tile while rendering is in progress. Return true if any update
+   * was performed. */
+  virtual bool update_render_tile(const Tile & /* tile */)
+  {
+    return false;
+  }
+
+  /* For baking, read render pass PASS_BAKE_PRIMITIVE and PASS_BAKE_DIFFERENTIAL
+   * to determine which shading points to use for baking at each pixel. Return
+   * true if any data was read. */
+  virtual bool read_render_tile(const Tile & /* tile */)
+  {
+    return false;
+  }
+};
+
+CCL_NAMESPACE_END
diff --git a/intern/cycles/session/session.cpp b/intern/cycles/session/session.cpp
new file mode 100644
index 00000000000..f8fc892f127
--- /dev/null
+++ b/intern/cycles/session/session.cpp
@@ -0,0 +1,624 @@
+/*
+ * 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 <limits.h>
+#include <string.h>
+
+#include "device/cpu/device.h"
+#include "device/device.h"
+#include "integrator/pass_accessor_cpu.h"
+#include "integrator/path_trace.h"
+#include "scene/background.h"
+#include "scene/bake.h"
+#include "scene/camera.h"
+#include "scene/integrator.h"
+#include "scene/light.h"
+#include "scene/mesh.h"
+#include "scene/object.h"
+#include "scene/scene.h"
+#include "scene/shader_graph.h"
+#include "session/buffers.h"
+#include "session/display_driver.h"
+#include "session/output_driver.h"
+#include "session/session.h"
+
+#include "util/util_foreach.h"
+#include "util/util_function.h"
+#include "util/util_logging.h"
+#include "util/util_math.h"
+#include "util/util_task.h"
+#include "util/util_time.h"
+
+CCL_NAMESPACE_BEGIN
+
+Session::Session(const SessionParams &params_, const SceneParams &scene_params)
+    : params(params_), render_scheduler_(tile_manager_, params)
+{
+  TaskScheduler::init(params.threads);
+
+  session_thread_ = nullptr;
+
+  delayed_reset_.do_reset = false;
+
+  pause_ = false;
+  cancel_ = false;
+  new_work_added_ = false;
+
+  device = Device::create(params.device, stats, profiler);
+
+  scene = new Scene(scene_params, device);
+
+  /* Configure path tracer. */
+  path_trace_ = make_unique<PathTrace>(
+      device, scene->film, &scene->dscene, render_scheduler_, tile_manager_);
+  path_trace_->set_progress(&progress);
+  path_trace_->progress_update_cb = [&]() { update_status_time(); };
+
+  tile_manager_.full_buffer_written_cb = [&](string_view filename) {
+    if (!full_buffer_written_cb) {
+      return;
+    }
+    full_buffer_written_cb(filename);
+  };
+}
+
+Session::~Session()
+{
+  cancel();
+
+  /* Make sure path tracer is destroyed before the device. This is needed because destruction might
+   * need to access device for device memory free. */
+  /* TODO(sergey): Convert device to be unique_ptr, and rely on C++ to destruct objects in the
+   * pre-defined order. */
+  path_trace_.reset();
+
+  delete scene;
+  delete device;
+
+  TaskScheduler::exit();
+}
+
+void Session::start()
+{
+  if (!session_thread_) {
+    session_thread_ = new thread(function_bind(&Session::run, this));
+  }
+}
+
+void Session::cancel(bool quick)
+{
+  if (quick && path_trace_) {
+    path_trace_->cancel();
+  }
+
+  if (session_thread_) {
+    /* wait for session thread to end */
+    progress.set_cancel("Exiting");
+
+    {
+      thread_scoped_lock pause_lock(pause_mutex_);
+      pause_ = false;
+      cancel_ = true;
+    }
+    pause_cond_.notify_all();
+
+    wait();
+  }
+}
+
+bool Session::ready_to_reset()
+{
+  return path_trace_->ready_to_reset();
+}
+
+void Session::run_main_render_loop()
+{
+  path_trace_->clear_display();
+
+  while (true) {
+    RenderWork render_work = run_update_for_next_iteration();
+
+    if (!render_work) {
+      if (VLOG_IS_ON(2)) {
+        double total_time, render_time;
+        progress.get_time(total_time, render_time);
+        VLOG(2) << "Rendering in main loop is done in " << render_time << " seconds.";
+        VLOG(2) << path_trace_->full_report();
+      }
+
+      if (params.background) {
+        /* if no work left and in background mode, we can stop immediately. */
+        progress.set_status("Finished");
+        break;
+      }
+    }
+
+    const bool did_cancel = progress.get_cancel();
+    if (did_cancel) {
+      render_scheduler_.render_work_reschedule_on_cancel(render_work);
+      if (!render_work) {
+        break;
+      }
+    }
+    else if (run_wait_for_work(render_work)) {
+      continue;
+    }
+
+    /* Stop rendering if error happened during scene update or other step of preparing scene
+     * for render. */
+    if (device->have_error()) {
+      progress.set_error(device->error_message());
+      break;
+    }
+
+    {
+      /* buffers mutex is locked entirely while rendering each
+       * sample, and released/reacquired on each iteration to allow
+       * reset and draw in between */
+      thread_scoped_lock buffers_lock(buffers_mutex_);
+
+      /* update status and timing */
+      update_status_time();
+
+      /* render */
+      path_trace_->render(render_work);
+
+      /* update status and timing */
+      update_status_time();
+
+      /* Stop rendering if error happened during path tracing. */
+      if (device->have_error()) {
+        progress.set_error(device->error_message());
+        break;
+      }
+    }
+
+    progress.set_update();
+
+    if (did_cancel) {
+      break;
+    }
+  }
+}
+
+void Session::run()
+{
+  if (params.use_profiling && (params.device.type == DEVICE_CPU)) {
+    profiler.start();
+  }
+
+  /* session thread loop */
+  progress.set_status("Waiting for render to start");
+
+  /* run */
+  if (!progress.get_cancel()) {
+    /* reset number of rendered samples */
+    progress.reset_sample();
+
+    run_main_render_loop();
+  }
+
+  profiler.stop();
+
+  /* progress update */
+  if (progress.get_cancel())
+    progress.set_status(progress.get_cancel_message());
+  else
+    progress.set_update();
+}
+
+RenderWork Session::run_update_for_next_iteration()
+{
+  RenderWork render_work;
+
+  thread_scoped_lock scene_lock(scene->mutex);
+  thread_scoped_lock reset_lock(delayed_reset_.mutex);
+
+  bool have_tiles = true;
+  bool switched_to_new_tile = false;
+
+  const bool did_reset = delayed_reset_.do_reset;
+  if (delayed_reset_.do_reset) {
+    thread_scoped_lock buffers_lock(buffers_mutex_);
+    do_delayed_reset();
+
+    /* After reset make sure the tile manager is at the first big tile. */
+    have_tiles = tile_manager_.next();
+    switched_to_new_tile = true;
+  }
+
+  /* Update number of samples in the integrator.
+   * Ideally this would need to happen once in `Session::set_samples()`, but the issue there is
+   * the initial configuration when Session is created where the `set_samples()` is not used.
+   *
+   * NOTE: Unless reset was requested only allow increasing number of samples. */
+  if (did_reset || scene->integrator->get_aa_samples() < params.samples) {
+    scene->integrator->set_aa_samples(params.samples);
+  }
+
+  /* Update denoiser settings. */
+  {
+    const DenoiseParams denoise_params = scene->integrator->get_denoise_params();
+    path_trace_->set_denoiser_params(denoise_params);
+  }
+
+  /* Update adaptive sampling. */
+  {
+    const AdaptiveSampling adaptive_sampling = scene->integrator->get_adaptive_sampling();
+    path_trace_->set_adaptive_sampling(adaptive_sampling);
+  }
+
+  render_scheduler_.set_num_samples(params.samples);
+  render_scheduler_.set_time_limit(params.time_limit);
+
+  while (have_tiles) {
+    render_work = render_scheduler_.get_render_work();
+    if (render_work) {
+      break;
+    }
+
+    progress.add_finished_tile(false);
+
+    have_tiles = tile_manager_.next();
+    if (have_tiles) {
+      render_scheduler_.reset_for_next_tile();
+      switched_to_new_tile = true;
+    }
+  }
+
+  if (render_work) {
+    scoped_timer update_timer;
+
+    if (switched_to_new_tile) {
+      BufferParams tile_params = buffer_params_;
+
+      const Tile &tile = tile_manager_.get_current_tile();
+
+      tile_params.width = tile.width;
+      tile_params.height = tile.height;
+
+      tile_params.window_x = tile.window_x;
+      tile_params.window_y = tile.window_y;
+      tile_params.window_width = tile.window_width;
+      tile_params.window_height = tile.window_height;
+
+      tile_params.full_x = tile.x + buffer_params_.full_x;
+      tile_params.full_y = tile.y + buffer_params_.full_y;
+      tile_params.full_width = buffer_params_.full_width;
+      tile_params.full_height = buffer_params_.full_height;
+
+      tile_params.update_offset_stride();
+
+      path_trace_->reset(buffer_params_, tile_params);
+    }
+
+    const int resolution = render_work.resolution_divider;
+    const int width = max(1, buffer_params_.full_width / resolution);
+    const int height = max(1, buffer_params_.full_height / resolution);
+
+    if (update_scene(width, height)) {
+      profiler.reset(scene->shaders.size(), scene->objects.size());
+    }
+    progress.add_skip_time(update_timer, params.background);
+  }
+
+  return render_work;
+}
+
+bool Session::run_wait_for_work(const RenderWork &render_work)
+{
+  /* In an offline rendering there is no pause, and no tiles will mean the job is fully done. */
+  if (params.background) {
+    return false;
+  }
+
+  thread_scoped_lock pause_lock(pause_mutex_);
+
+  if (!pause_ && render_work) {
+    /* Rendering is not paused and there is work to be done. No need to wait for anything. */
+    return false;
+  }
+
+  const bool no_work = !render_work;
+  update_status_time(pause_, no_work);
+
+  /* Only leave the loop when rendering is not paused. But even if the current render is un-paused
+   * but there is nothing to render keep waiting until new work is added. */
+  while (!cancel_) {
+    scoped_timer pause_timer;
+
+    if (!pause_ && (render_work || new_work_added_ || delayed_reset_.do_reset)) {
+      break;
+    }
+
+    /* Wait for either pause state changed, or extra samples added to render. */
+    pause_cond_.wait(pause_lock);
+
+    if (pause_) {
+      progress.add_skip_time(pause_timer, params.background);
+    }
+
+    update_status_time(pause_, no_work);
+    progress.set_update();
+  }
+
+  new_work_added_ = false;
+
+  return no_work;
+}
+
+void Session::draw()
+{
+  path_trace_->draw();
+}
+
+int2 Session::get_effective_tile_size() const
+{
+  /* No support yet for baking with tiles. */
+  if (!params.use_auto_tile || scene->bake_manager->get_baking()) {
+    return make_int2(buffer_params_.width, buffer_params_.height);
+  }
+
+  /* TODO(sergey): Take available memory into account, and if there is enough memory do not tile
+   * and prefer optimal performance. */
+  const int tile_size = tile_manager_.compute_render_tile_size(params.tile_size);
+  return make_int2(tile_size, tile_size);
+}
+
+void Session::do_delayed_reset()
+{
+  if (!delayed_reset_.do_reset) {
+    return;
+  }
+  delayed_reset_.do_reset = false;
+
+  params = delayed_reset_.session_params;
+  buffer_params_ = delayed_reset_.buffer_params;
+
+  /* Store parameters used for buffers access outside of scene graph.  */
+  buffer_params_.samples = params.samples;
+  buffer_params_.exposure = scene->film->get_exposure();
+  buffer_params_.use_approximate_shadow_catcher =
+      scene->film->get_use_approximate_shadow_catcher();
+  buffer_params_.use_transparent_background = scene->background->get_transparent();
+
+  /* Tile and work scheduling. */
+  tile_manager_.reset_scheduling(buffer_params_, get_effective_tile_size());
+  render_scheduler_.reset(buffer_params_, params.samples);
+
+  /* Passes. */
+  /* When multiple tiles are used SAMPLE_COUNT pass is used to keep track of possible partial
+   * tile results. It is safe to use generic update function here which checks for changes since
+   * changes in tile settings re-creates session, which ensures film is fully updated on tile
+   * changes. */
+  scene->film->update_passes(scene, tile_manager_.has_multiple_tiles());
+
+  /* Update for new state of scene and passes. */
+  buffer_params_.update_passes(scene->passes);
+  tile_manager_.update(buffer_params_, scene);
+
+  /* Progress. */
+  progress.reset_sample();
+  progress.set_total_pixel_samples(static_cast<uint64_t>(buffer_params_.width) *
+                                   buffer_params_.height * params.samples);
+
+  if (!params.background) {
+    progress.set_start_time();
+  }
+  progress.set_render_start_time();
+}
+
+void Session::reset(const SessionParams &session_params, const BufferParams &buffer_params)
+{
+  {
+    thread_scoped_lock reset_lock(delayed_reset_.mutex);
+    thread_scoped_lock pause_lock(pause_mutex_);
+
+    delayed_reset_.do_reset = true;
+    delayed_reset_.session_params = session_params;
+    delayed_reset_.buffer_params = buffer_params;
+
+    path_trace_->cancel();
+  }
+
+  pause_cond_.notify_all();
+}
+
+void Session::set_samples(int samples)
+{
+  if (samples == params.samples) {
+    return;
+  }
+
+  params.samples = samples;
+
+  {
+    thread_scoped_lock pause_lock(pause_mutex_);
+    new_work_added_ = true;
+  }
+
+  pause_cond_.notify_all();
+}
+
+void Session::set_time_limit(double time_limit)
+{
+  if (time_limit == params.time_limit) {
+    return;
+  }
+
+  params.time_limit = time_limit;
+
+  {
+    thread_scoped_lock pause_lock(pause_mutex_);
+    new_work_added_ = true;
+  }
+
+  pause_cond_.notify_all();
+}
+
+void Session::set_pause(bool pause)
+{
+  bool notify = false;
+
+  {
+    thread_scoped_lock pause_lock(pause_mutex_);
+
+    if (pause != pause_) {
+      pause_ = pause;
+      notify = true;
+    }
+  }
+
+  if (session_thread_) {
+    if (notify) {
+      pause_cond_.notify_all();
+    }
+  }
+  else if (pause_) {
+    update_status_time(pause_);
+  }
+}
+
+void Session::set_output_driver(unique_ptr<OutputDriver> driver)
+{
+  path_trace_->set_output_driver(move(driver));
+}
+
+void Session::set_display_driver(unique_ptr<DisplayDriver> driver)
+{
+  path_trace_->set_display_driver(move(driver));
+}
+
+double Session::get_estimated_remaining_time() const
+{
+  const float completed = progress.get_progress();
+  if (completed == 0.0f) {
+    return 0.0;
+  }
+
+  double total_time, render_time;
+  progress.get_time(total_time, render_time);
+  double remaining = (1.0 - (double)completed) * (render_time / (double)completed);
+
+  const double time_limit = render_scheduler_.get_time_limit();
+  if (time_limit != 0.0) {
+    remaining = min(remaining, max(time_limit - render_time, 0.0));
+  }
+
+  return remaining;
+}
+
+void Session::wait()
+{
+  if (session_thread_) {
+    session_thread_->join();
+    delete session_thread_;
+  }
+
+  session_thread_ = nullptr;
+}
+
+bool Session::update_scene(int width, int height)
+{
+  /* Update camera if dimensions changed for progressive render. the camera
+   * knows nothing about progressive or cropped rendering, it just gets the
+   * image dimensions passed in. */
+  Camera *cam = scene->camera;
+  cam->set_screen_size(width, height);
+
+  const bool scene_update_result = scene->update(progress);
+
+  path_trace_->load_kernels();
+  path_trace_->alloc_work_memory();
+
+  return scene_update_result;
+}
+
+static string status_append(const string &status, const string &suffix)
+{
+  string prefix = status;
+  if (!prefix.empty()) {
+    prefix += ", ";
+  }
+  return prefix + suffix;
+}
+
+void Session::update_status_time(bool show_pause, bool show_done)
+{
+  string status, substatus;
+
+  const int current_tile = progress.get_rendered_tiles();
+  const int num_tiles = tile_manager_.get_num_tiles();
+
+  const int current_sample = progress.get_current_sample();
+  const int num_samples = render_scheduler_.get_num_samples();
+
+  /* TIle. */
+  if (tile_manager_.has_multiple_tiles()) {
+    substatus = status_append(substatus,
+                              string_printf("Rendered %d/%d Tiles", current_tile, num_tiles));
+  }
+
+  /* Sample. */
+  if (num_samples == Integrator::MAX_SAMPLES) {
+    substatus = status_append(substatus, string_printf("Sample %d", current_sample));
+  }
+  else {
+    substatus = status_append(substatus,
+                              string_printf("Sample %d/%d", current_sample, num_samples));
+  }
+
+  /* TODO(sergey): Denoising status from the path trace. */
+
+  if (show_pause) {
+    status = "Rendering Paused";
+  }
+  else if (show_done) {
+    status = "Rendering Done";
+    progress.set_end_time(); /* Save end time so that further calls to get_time are accurate. */
+  }
+  else {
+    status = substatus;
+    substatus.clear();
+  }
+
+  progress.set_status(status, substatus);
+}
+
+void Session::device_free()
+{
+  scene->device_free();
+  path_trace_->device_free();
+}
+
+void Session::collect_statistics(RenderStats *render_stats)
+{
+  scene->collect_statistics(render_stats);
+  if (params.use_profiling && (params.device.type == DEVICE_CPU)) {
+    render_stats->collect_profiling(scene, profiler);
+  }
+}
+
+/* --------------------------------------------------------------------
+ * Full-frame on-disk storage.
+ */
+
+void Session::process_full_buffer_from_disk(string_view filename)
+{
+  path_trace_->process_full_buffer_from_disk(filename);
+}
+
+CCL_NAMESPACE_END
diff --git a/intern/cycles/session/session.h b/intern/cycles/session/session.h
new file mode 100644
index 00000000000..5aa6df79ef1
--- /dev/null
+++ b/intern/cycles/session/session.h
@@ -0,0 +1,229 @@
+/*
+ * 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.
+ */
+
+#ifndef __SESSION_H__
+#define __SESSION_H__
+
+#include "device/device.h"
+#include "integrator/render_scheduler.h"
+#include "scene/shader.h"
+#include "scene/stats.h"
+#include "session/buffers.h"
+#include "session/tile.h"
+
+#include "util/util_progress.h"
+#include "util/util_stats.h"
+#include "util/util_thread.h"
+#include "util/util_unique_ptr.h"
+#include "util/util_vector.h"
+
+CCL_NAMESPACE_BEGIN
+
+class BufferParams;
+class Device;
+class DeviceScene;
+class DisplayDriver;
+class OutputDriver;
+class PathTrace;
+class Progress;
+class RenderBuffers;
+class Scene;
+class SceneParams;
+
+/* Session Parameters */
+
+class SessionParams {
+ public:
+  DeviceInfo device;
+
+  bool headless;
+  bool background;
+
+  bool experimental;
+  int samples;
+  int pixel_size;
+  int threads;
+
+  /* Limit in seconds for how long path tracing is allowed to happen.
+   * Zero means no limit is applied. */
+  double time_limit;
+
+  bool use_profiling;
+
+  bool use_auto_tile;
+  int tile_size;
+
+  ShadingSystem shadingsystem;
+
+  SessionParams()
+  {
+    headless = false;
+    background = false;
+
+    experimental = false;
+    samples = 1024;
+    pixel_size = 1;
+    threads = 0;
+    time_limit = 0.0;
+
+    use_profiling = false;
+
+    use_auto_tile = true;
+    tile_size = 2048;
+
+    shadingsystem = SHADINGSYSTEM_SVM;
+  }
+
+  bool modified(const SessionParams &params) const
+  {
+    /* Modified means we have to recreate the session, any parameter changes
+     * that can be handled by an existing Session are omitted. */
+    return !(device == params.device && headless == params.headless &&
+             background == params.background && experimental == params.experimental &&
+             pixel_size == params.pixel_size && threads == params.threads &&
+             use_profiling == params.use_profiling && shadingsystem == params.shadingsystem &&
+             use_auto_tile == params.use_auto_tile && tile_size == params.tile_size);
+  }
+};
+
+/* Session
+ *
+ * This is the class that contains the session thread, running the render
+ * control loop and dispatching tasks. */
+
+class Session {
+ public:
+  Device *device;
+  Scene *scene;
+  Progress progress;
+  SessionParams params;
+  Stats stats;
+  Profiler profiler;
+
+  /* Callback is invoked by tile manager whenever on-dist tiles storage file is closed after
+   * writing. Allows an engine integration to keep track of those files without worry about
+   * transferring the information when it needs to re-create session during rendering. */
+  function<void(string_view)> full_buffer_written_cb;
+
+  explicit Session(const SessionParams &params, const SceneParams &scene_params);
+  ~Session();
+
+  void start();
+
+  /* When quick cancel is requested path tracing is cancels as soon as possible, without waiting
+   * for the buffer to be uniformly sampled. */
+  void cancel(bool quick = false);
+
+  void draw();
+  void wait();
+
+  bool ready_to_reset();
+  void reset(const SessionParams &session_params, const BufferParams &buffer_params);
+
+  void set_pause(bool pause);
+
+  void set_samples(int samples);
+  void set_time_limit(double time_limit);
+
+  void set_output_driver(unique_ptr<OutputDriver> driver);
+  void set_display_driver(unique_ptr<DisplayDriver> driver);
+
+  double get_estimated_remaining_time() const;
+
+  void device_free();
+
+  /* Returns the rendering progress or 0 if no progress can be determined
+   * (for example, when rendering with unlimited samples). */
+  float get_progress();
+
+  void collect_statistics(RenderStats *stats);
+
+  /* --------------------------------------------------------------------
+   * Full-frame on-disk storage.
+   */
+
+  /* Read given full-frame file from disk, perform needed processing and write it to the software
+   * via the write callback. */
+  void process_full_buffer_from_disk(string_view filename);
+
+ protected:
+  struct DelayedReset {
+    thread_mutex mutex;
+    bool do_reset;
+    SessionParams session_params;
+    BufferParams buffer_params;
+  } delayed_reset_;
+
+  void run();
+
+  /* Update for the new iteration of the main loop in run implementation (run_cpu and run_gpu).
+   *
+   * Will take care of the following things:
+   *  - Delayed reset
+   *  - Scene update
+   *  - Tile manager advance
+   *  - Render scheduler work request
+   *
+   * The updates are done in a proper order with proper locking around them, which guarantees
+   * that the device side of scene and render buffers are always in a consistent state.
+   *
+   * Returns render work which is to be rendered next. */
+  RenderWork run_update_for_next_iteration();
+
+  /* Wait for rendering to be unpaused, or for new tiles for render to arrive.
+   * Returns true if new main render loop iteration is required after this function call.
+   *
+   * The `render_work` is the work which was scheduled by the render scheduler right before
+   * checking the pause. */
+  bool run_wait_for_work(const RenderWork &render_work);
+
+  void run_main_render_loop();
+
+  bool update_scene(int width, int height);
+
+  void update_status_time(bool show_pause = false, bool show_done = false);
+
+  void do_delayed_reset();
+
+  int2 get_effective_tile_size() const;
+
+  thread *session_thread_;
+
+  bool pause_ = false;
+  bool cancel_ = false;
+  bool new_work_added_ = false;
+
+  thread_condition_variable pause_cond_;
+  thread_mutex pause_mutex_;
+  thread_mutex tile_mutex_;
+  thread_mutex buffers_mutex_;
+
+  TileManager tile_manager_;
+  BufferParams buffer_params_;
+
+  /* Render scheduler is used to get work to be rendered with the current big tile. */
+  RenderScheduler render_scheduler_;
+
+  /* Path tracer object.
+   *
+   * Is a single full-frame path tracer for interactive viewport rendering.
+   * A path tracer for the current big-tile for an offline rendering. */
+  unique_ptr<PathTrace> path_trace_;
+};
+
+CCL_NAMESPACE_END
+
+#endif /* __SESSION_H__ */
diff --git a/intern/cycles/session/tile.cpp b/intern/cycles/session/tile.cpp
new file mode 100644
index 00000000000..59332530596
--- /dev/null
+++ b/intern/cycles/session/tile.cpp
@@ -0,0 +1,629 @@
+/*
+ * 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 "session/tile.h"
+
+#include <atomic>
+
+#include "graph/node.h"
+#include "scene/background.h"
+#include "scene/film.h"
+#include "scene/integrator.h"
+#include "scene/scene.h"
+#include "util/util_algorithm.h"
+#include "util/util_foreach.h"
+#include "util/util_logging.h"
+#include "util/util_path.h"
+#include "util/util_string.h"
+#include "util/util_system.h"
+#include "util/util_types.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* --------------------------------------------------------------------
+ * Internal functions.
+ */
+
+static const char *ATTR_PASSES_COUNT = "cycles.passes.count";
+static const char *ATTR_PASS_SOCKET_PREFIX_FORMAT = "cycles.passes.%d.";
+static const char *ATTR_BUFFER_SOCKET_PREFIX = "cycles.buffer.";
+static const char *ATTR_DENOISE_SOCKET_PREFIX = "cycles.denoise.";
+
+/* Global counter of ToleManager object instances. */
+static std::atomic<uint64_t> g_instance_index = 0;
+
+/* Construct names of EXR channels which will ensure order of all channels to match exact offsets
+ * in render buffers corresponding to the given passes.
+ *
+ * Returns `std` datatypes so that it can be assigned directly to the OIIO's `ImageSpec`. */
+static std::vector<std::string> exr_channel_names_for_passes(const BufferParams &buffer_params)
+{
+  static const char *component_suffixes[] = {"R", "G", "B", "A"};
+
+  int pass_index = 0;
+  int num_channels = 0;
+  std::vector<std::string> channel_names;
+  for (const BufferPass &pass : buffer_params.passes) {
+    if (pass.offset == PASS_UNUSED) {
+      continue;
+    }
+
+    const PassInfo pass_info = pass.get_info();
+    num_channels += pass_info.num_components;
+
+    /* EXR canonically expects first part of channel names to be sorted alphabetically, which is
+     * not guaranteed to be the case with passes names. Assign a prefix based on the pass index
+     * with a fixed width to ensure ordering. This makes it possible to dump existing render
+     * buffers memory to disk and read it back without doing extra mapping. */
+    const string prefix = string_printf("%08d", pass_index);
+
+    const string channel_name_prefix = prefix + string(pass.name) + ".";
+
+    for (int i = 0; i < pass_info.num_components; ++i) {
+      channel_names.push_back(channel_name_prefix + component_suffixes[i]);
+    }
+
+    ++pass_index;
+  }
+
+  return channel_names;
+}
+
+inline string node_socket_attribute_name(const SocketType &socket, const string &attr_name_prefix)
+{
+  return attr_name_prefix + string(socket.name);
+}
+
+template<typename ValidateValueFunc, typename GetValueFunc>
+static bool node_socket_generic_to_image_spec_atttributes(
+    ImageSpec *image_spec,
+    const Node *node,
+    const SocketType &socket,
+    const string &attr_name_prefix,
+    const ValidateValueFunc &validate_value_func,
+    const GetValueFunc &get_value_func)
+{
+  if (!validate_value_func(node, socket)) {
+    return false;
+  }
+
+  image_spec->attribute(node_socket_attribute_name(socket, attr_name_prefix),
+                        get_value_func(node, socket));
+
+  return true;
+}
+
+static bool node_socket_to_image_spec_atttributes(ImageSpec *image_spec,
+                                                  const Node *node,
+                                                  const SocketType &socket,
+                                                  const string &attr_name_prefix)
+{
+  const string attr_name = node_socket_attribute_name(socket, attr_name_prefix);
+
+  switch (socket.type) {
+    case SocketType::ENUM: {
+      const ustring value = node->get_string(socket);
+
+      /* Validate that the node is consistent with the node type definition. */
+      const NodeEnum &enum_values = *socket.enum_values;
+      if (!enum_values.exists(value)) {
+        LOG(DFATAL) << "Node enum contains invalid value " << value;
+        return false;
+      }
+
+      image_spec->attribute(attr_name, value);
+
+      return true;
+    }
+
+    case SocketType::STRING:
+      image_spec->attribute(attr_name, node->get_string(socket));
+      return true;
+
+    case SocketType::INT:
+      image_spec->attribute(attr_name, node->get_int(socket));
+      return true;
+
+    case SocketType::FLOAT:
+      image_spec->attribute(attr_name, node->get_float(socket));
+      return true;
+
+    case SocketType::BOOLEAN:
+      image_spec->attribute(attr_name, node->get_bool(socket));
+      return true;
+
+    default:
+      LOG(DFATAL) << "Unhandled socket type " << socket.type << ", should never happen.";
+      return false;
+  }
+}
+
+static bool node_socket_from_image_spec_atttributes(Node *node,
+                                                    const SocketType &socket,
+                                                    const ImageSpec &image_spec,
+                                                    const string &attr_name_prefix)
+{
+  const string attr_name = node_socket_attribute_name(socket, attr_name_prefix);
+
+  switch (socket.type) {
+    case SocketType::ENUM: {
+      /* TODO(sergey): Avoid construction of `ustring` by using `string_view` in the Node API. */
+      const ustring value(image_spec.get_string_attribute(attr_name, ""));
+
+      /* Validate that the node is consistent with the node type definition. */
+      const NodeEnum &enum_values = *socket.enum_values;
+      if (!enum_values.exists(value)) {
+        LOG(ERROR) << "Invalid enumerator value " << value;
+        return false;
+      }
+
+      node->set(socket, enum_values[value]);
+
+      return true;
+    }
+
+    case SocketType::STRING:
+      /* TODO(sergey): Avoid construction of `ustring` by using `string_view` in the Node API. */
+      node->set(socket, ustring(image_spec.get_string_attribute(attr_name, "")));
+      return true;
+
+    case SocketType::INT:
+      node->set(socket, image_spec.get_int_attribute(attr_name, 0));
+      return true;
+
+    case SocketType::FLOAT:
+      node->set(socket, image_spec.get_float_attribute(attr_name, 0));
+      return true;
+
+    case SocketType::BOOLEAN:
+      node->set(socket, static_cast<bool>(image_spec.get_int_attribute(attr_name, 0)));
+      return true;
+
+    default:
+      LOG(DFATAL) << "Unhandled socket type " << socket.type << ", should never happen.";
+      return false;
+  }
+}
+
+static bool node_to_image_spec_atttributes(ImageSpec *image_spec,
+                                           const Node *node,
+                                           const string &attr_name_prefix)
+{
+  for (const SocketType &socket : node->type->inputs) {
+    if (!node_socket_to_image_spec_atttributes(image_spec, node, socket, attr_name_prefix)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+static bool node_from_image_spec_atttributes(Node *node,
+                                             const ImageSpec &image_spec,
+                                             const string &attr_name_prefix)
+{
+  for (const SocketType &socket : node->type->inputs) {
+    if (!node_socket_from_image_spec_atttributes(node, socket, image_spec, attr_name_prefix)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+static bool buffer_params_to_image_spec_atttributes(ImageSpec *image_spec,
+                                                    const BufferParams &buffer_params)
+{
+  if (!node_to_image_spec_atttributes(image_spec, &buffer_params, ATTR_BUFFER_SOCKET_PREFIX)) {
+    return false;
+  }
+
+  /* Passes storage is not covered by the node socket. so "expand" the loop manually. */
+
+  const int num_passes = buffer_params.passes.size();
+  image_spec->attribute(ATTR_PASSES_COUNT, num_passes);
+
+  for (int pass_index = 0; pass_index < num_passes; ++pass_index) {
+    const string attr_name_prefix = string_printf(ATTR_PASS_SOCKET_PREFIX_FORMAT, pass_index);
+
+    const BufferPass *pass = &buffer_params.passes[pass_index];
+    if (!node_to_image_spec_atttributes(image_spec, pass, attr_name_prefix)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+static bool buffer_params_from_image_spec_atttributes(BufferParams *buffer_params,
+                                                      const ImageSpec &image_spec)
+{
+  if (!node_from_image_spec_atttributes(buffer_params, image_spec, ATTR_BUFFER_SOCKET_PREFIX)) {
+    return false;
+  }
+
+  /* Passes storage is not covered by the node socket. so "expand" the loop manually. */
+
+  const int num_passes = image_spec.get_int_attribute(ATTR_PASSES_COUNT, 0);
+  if (num_passes == 0) {
+    LOG(ERROR) << "Missing passes count attribute.";
+    return false;
+  }
+
+  for (int pass_index = 0; pass_index < num_passes; ++pass_index) {
+    const string attr_name_prefix = string_printf(ATTR_PASS_SOCKET_PREFIX_FORMAT, pass_index);
+
+    BufferPass pass;
+
+    if (!node_from_image_spec_atttributes(&pass, image_spec, attr_name_prefix)) {
+      return false;
+    }
+
+    buffer_params->passes.emplace_back(std::move(pass));
+  }
+
+  buffer_params->update_passes();
+
+  return true;
+}
+
+/* Configure image specification for the given buffer parameters and passes.
+ *
+ * Image channels will be strictly ordered to match content of corresponding buffer, and the
+ * metadata will be set so that the render buffers and passes can be reconstructed from it.
+ *
+ * If the tile size different from (0, 0) the image specification will be configured to use the
+ * given tile size for tiled IO. */
+static bool configure_image_spec_from_buffer(ImageSpec *image_spec,
+                                             const BufferParams &buffer_params,
+                                             const int2 tile_size = make_int2(0, 0))
+{
+  const std::vector<std::string> channel_names = exr_channel_names_for_passes(buffer_params);
+  const int num_channels = channel_names.size();
+
+  *image_spec = ImageSpec(
+      buffer_params.width, buffer_params.height, num_channels, TypeDesc::FLOAT);
+
+  image_spec->channelnames = move(channel_names);
+
+  if (!buffer_params_to_image_spec_atttributes(image_spec, buffer_params)) {
+    return false;
+  }
+
+  if (tile_size.x != 0 || tile_size.y != 0) {
+    DCHECK_GT(tile_size.x, 0);
+    DCHECK_GT(tile_size.y, 0);
+
+    image_spec->tile_width = min(TileManager::IMAGE_TILE_SIZE, tile_size.x);
+    image_spec->tile_height = min(TileManager::IMAGE_TILE_SIZE, tile_size.y);
+  }
+
+  return true;
+}
+
+/* --------------------------------------------------------------------
+ * Tile Manager.
+ */
+
+TileManager::TileManager()
+{
+  /* Use process ID to separate different processes.
+   * To ensure uniqueness from within a process use combination of object address and instance
+   * index. This solves problem of possible object re-allocation at the same time, and solves
+   * possible conflict when the counter overflows while there are still active instances of the
+   * class. */
+  const int tile_manager_id = g_instance_index.fetch_add(1, std::memory_order_relaxed);
+  tile_file_unique_part_ = to_string(system_self_process_id()) + "-" +
+                           to_string(reinterpret_cast<uintptr_t>(this)) + "-" +
+                           to_string(tile_manager_id);
+}
+
+TileManager::~TileManager()
+{
+}
+
+int TileManager::compute_render_tile_size(const int suggested_tile_size) const
+{
+  /* Must be a multiple of IMAGE_TILE_SIZE so that we can write render tiles into the image file
+   * aligned on image tile boundaries. We can't set IMAGE_TILE_SIZE equal to the render tile size
+   * because too big tile size leads to integer overflow inside OpenEXR. */
+  return (suggested_tile_size <= IMAGE_TILE_SIZE) ? suggested_tile_size :
+                                                    align_up(suggested_tile_size, IMAGE_TILE_SIZE);
+}
+
+void TileManager::reset_scheduling(const BufferParams &params, int2 tile_size)
+{
+  VLOG(3) << "Using tile size of " << tile_size;
+
+  close_tile_output();
+
+  tile_size_ = tile_size;
+
+  tile_state_.num_tiles_x = divide_up(params.width, tile_size_.x);
+  tile_state_.num_tiles_y = divide_up(params.height, tile_size_.y);
+  tile_state_.num_tiles = tile_state_.num_tiles_x * tile_state_.num_tiles_y;
+
+  tile_state_.next_tile_index = 0;
+
+  tile_state_.current_tile = Tile();
+}
+
+void TileManager::update(const BufferParams &params, const Scene *scene)
+{
+  DCHECK_NE(params.pass_stride, -1);
+
+  buffer_params_ = params;
+
+  /* TODO(sergey): Proper Error handling, so that if configuration has failed we don't attempt to
+   * write to a partially configured file. */
+  configure_image_spec_from_buffer(&write_state_.image_spec, buffer_params_, tile_size_);
+
+  const DenoiseParams denoise_params = scene->integrator->get_denoise_params();
+  const AdaptiveSampling adaptive_sampling = scene->integrator->get_adaptive_sampling();
+
+  node_to_image_spec_atttributes(
+      &write_state_.image_spec, &denoise_params, ATTR_DENOISE_SOCKET_PREFIX);
+
+  if (adaptive_sampling.use) {
+    overscan_ = 4;
+  }
+  else {
+    overscan_ = 0;
+  }
+}
+
+bool TileManager::done()
+{
+  return tile_state_.next_tile_index == tile_state_.num_tiles;
+}
+
+bool TileManager::next()
+{
+  if (done()) {
+    return false;
+  }
+
+  tile_state_.current_tile = get_tile_for_index(tile_state_.next_tile_index);
+
+  ++tile_state_.next_tile_index;
+
+  return true;
+}
+
+Tile TileManager::get_tile_for_index(int index) const
+{
+  /* TODO(sergey): Consider using hilbert spiral, or. maybe, even configurable. Not sure this
+   * brings a lot of value since this is only applicable to BIG tiles. */
+
+  const int tile_index_y = index / tile_state_.num_tiles_x;
+  const int tile_index_x = index - tile_index_y * tile_state_.num_tiles_x;
+
+  const int tile_window_x = tile_index_x * tile_size_.x;
+  const int tile_window_y = tile_index_y * tile_size_.y;
+
+  Tile tile;
+
+  tile.x = max(0, tile_window_x - overscan_);
+  tile.y = max(0, tile_window_y - overscan_);
+
+  tile.window_x = tile_window_x - tile.x;
+  tile.window_y = tile_window_y - tile.y;
+  tile.window_width = min(tile_size_.x, buffer_params_.width - tile_window_x);
+  tile.window_height = min(tile_size_.y, buffer_params_.height - tile_window_y);
+
+  tile.width = min(buffer_params_.width - tile.x, tile.window_x + tile.window_width + overscan_);
+  tile.height = min(buffer_params_.height - tile.y,
+                    tile.window_y + tile.window_height + overscan_);
+
+  return tile;
+}
+
+const Tile &TileManager::get_current_tile() const
+{
+  return tile_state_.current_tile;
+}
+
+const int2 TileManager::get_size() const
+{
+  return make_int2(buffer_params_.width, buffer_params_.height);
+}
+
+bool TileManager::open_tile_output()
+{
+  write_state_.filename = path_temp_get("cycles-tile-buffer-" + tile_file_unique_part_ + "-" +
+                                        to_string(write_state_.tile_file_index) + ".exr");
+
+  write_state_.tile_out = ImageOutput::create(write_state_.filename);
+  if (!write_state_.tile_out) {
+    LOG(ERROR) << "Error creating image output for " << write_state_.filename;
+    return false;
+  }
+
+  if (!write_state_.tile_out->supports("tiles")) {
+    LOG(ERROR) << "Progress tile file format does not support tiling.";
+    return false;
+  }
+
+  if (!write_state_.tile_out->open(write_state_.filename, write_state_.image_spec)) {
+    LOG(ERROR) << "Error opening tile file: " << write_state_.tile_out->geterror();
+    write_state_.tile_out = nullptr;
+    return false;
+  }
+
+  write_state_.num_tiles_written = 0;
+
+  VLOG(3) << "Opened tile file " << write_state_.filename;
+
+  return true;
+}
+
+bool TileManager::close_tile_output()
+{
+  if (!write_state_.tile_out) {
+    return true;
+  }
+
+  const bool success = write_state_.tile_out->close();
+  write_state_.tile_out = nullptr;
+
+  if (!success) {
+    LOG(ERROR) << "Error closing tile file.";
+    return false;
+  }
+
+  VLOG(3) << "Tile output is closed.";
+
+  return true;
+}
+
+bool TileManager::write_tile(const RenderBuffers &tile_buffers)
+{
+  if (!write_state_.tile_out) {
+    if (!open_tile_output()) {
+      return false;
+    }
+  }
+
+  DCHECK_EQ(tile_buffers.params.pass_stride, buffer_params_.pass_stride);
+
+  vector<float> pixel_storage;
+
+  const BufferParams &tile_params = tile_buffers.params;
+
+  const int tile_x = tile_params.full_x - buffer_params_.full_x + tile_params.window_x;
+  const int tile_y = tile_params.full_y - buffer_params_.full_y + tile_params.window_y;
+
+  const int64_t pass_stride = tile_params.pass_stride;
+  const int64_t tile_row_stride = tile_params.width * pass_stride;
+
+  const int64_t xstride = pass_stride * sizeof(float);
+  const int64_t ystride = xstride * tile_params.width;
+  const int64_t zstride = ystride * tile_params.height;
+
+  const float *pixels = tile_buffers.buffer.data() + tile_params.window_x * pass_stride +
+                        tile_params.window_y * tile_row_stride;
+
+  VLOG(3) << "Write tile at " << tile_x << ", " << tile_y;
+
+  /* The image tile sizes in the OpenEXR file are different from the size of our big tiles. The
+   * write_tiles() method expects a contiguous image region that will be split into tiles
+   * internally. OpenEXR expects the size of this region to be a multiple of the tile size,
+   * however OpenImageIO automatically adds the required padding.
+   *
+   * The only thing we have to ensure is that the tile_x and tile_y are a multiple of the
+   * image tile size, which happens in compute_render_tile_size. */
+  if (!write_state_.tile_out->write_tiles(tile_x,
+                                          tile_x + tile_params.window_width,
+                                          tile_y,
+                                          tile_y + tile_params.window_height,
+                                          0,
+                                          1,
+                                          TypeDesc::FLOAT,
+                                          pixels,
+                                          xstride,
+                                          ystride,
+                                          zstride)) {
+    LOG(ERROR) << "Error writing tile " << write_state_.tile_out->geterror();
+    return false;
+  }
+
+  ++write_state_.num_tiles_written;
+
+  return true;
+}
+
+void TileManager::finish_write_tiles()
+{
+  if (!write_state_.tile_out) {
+    /* None of the tiles were written hence the file was not created.
+     * Avoid creation of fully empty file since it is redundant. */
+    return;
+  }
+
+  /* EXR expects all tiles to present in file. So explicitly write missing tiles as all-zero. */
+  if (write_state_.num_tiles_written < tile_state_.num_tiles) {
+    vector<float> pixel_storage(tile_size_.x * tile_size_.y * buffer_params_.pass_stride);
+
+    for (int tile_index = write_state_.num_tiles_written; tile_index < tile_state_.num_tiles;
+         ++tile_index) {
+      const Tile tile = get_tile_for_index(tile_index);
+
+      const int tile_x = tile.x + tile.window_x;
+      const int tile_y = tile.y + tile.window_y;
+
+      VLOG(3) << "Write dummy tile at " << tile_x << ", " << tile_y;
+
+      write_state_.tile_out->write_tiles(tile_x,
+                                         tile_x + tile.window_width,
+                                         tile_y,
+                                         tile_y + tile.window_height,
+                                         0,
+                                         1,
+                                         TypeDesc::FLOAT,
+                                         pixel_storage.data());
+    }
+  }
+
+  close_tile_output();
+
+  if (full_buffer_written_cb) {
+    full_buffer_written_cb(write_state_.filename);
+  }
+
+  /* Advance the counter upon explicit finish of the file.
+   * Makes it possible to re-use tile manager for another scene, and avoids unnecessary increments
+   * of the tile-file-within-session index. */
+  ++write_state_.tile_file_index;
+
+  write_state_.filename = "";
+}
+
+bool TileManager::read_full_buffer_from_disk(const string_view filename,
+                                             RenderBuffers *buffers,
+                                             DenoiseParams *denoise_params)
+{
+  unique_ptr<ImageInput> in(ImageInput::open(filename));
+  if (!in) {
+    LOG(ERROR) << "Error opening tile file " << filename;
+    return false;
+  }
+
+  const ImageSpec &image_spec = in->spec();
+
+  BufferParams buffer_params;
+  if (!buffer_params_from_image_spec_atttributes(&buffer_params, image_spec)) {
+    return false;
+  }
+  buffers->reset(buffer_params);
+
+  if (!node_from_image_spec_atttributes(denoise_params, image_spec, ATTR_DENOISE_SOCKET_PREFIX)) {
+    return false;
+  }
+
+  if (!in->read_image(TypeDesc::FLOAT, buffers->buffer.data())) {
+    LOG(ERROR) << "Error reading pixels from the tile file " << in->geterror();
+    return false;
+  }
+
+  if (!in->close()) {
+    LOG(ERROR) << "Error closing tile file " << in->geterror();
+    return false;
+  }
+
+  return true;
+}
+
+CCL_NAMESPACE_END
diff --git a/intern/cycles/session/tile.h b/intern/cycles/session/tile.h
new file mode 100644
index 00000000000..37a02081a53
--- /dev/null
+++ b/intern/cycles/session/tile.h
@@ -0,0 +1,182 @@
+/*
+ * 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.
+ */
+
+#pragma once
+
+#include "session/buffers.h"
+#include "util/util_image.h"
+#include "util/util_string.h"
+#include "util/util_unique_ptr.h"
+
+CCL_NAMESPACE_BEGIN
+
+class DenoiseParams;
+class Scene;
+
+/* --------------------------------------------------------------------
+ * Tile.
+ */
+
+class Tile {
+ public:
+  int x = 0, y = 0;
+  int width = 0, height = 0;
+
+  int window_x = 0, window_y = 0;
+  int window_width = 0, window_height = 0;
+
+  Tile()
+  {
+  }
+};
+
+/* --------------------------------------------------------------------
+ * Tile Manager.
+ */
+
+class TileManager {
+ public:
+  /* This callback is invoked by whenever on-dist tiles storage file is closed after writing. */
+  function<void(string_view)> full_buffer_written_cb;
+
+  TileManager();
+  ~TileManager();
+
+  TileManager(const TileManager &other) = delete;
+  TileManager(TileManager &&other) noexcept = delete;
+  TileManager &operator=(const TileManager &other) = delete;
+  TileManager &operator=(TileManager &&other) = delete;
+
+  /* Reset current progress and start new rendering of the full-frame parameters in tiles of the
+   * given size.
+   * Only touches scheduling-related state of the tile manager. */
+  /* TODO(sergey): Consider using tile area instead of exact size to help dealing with extreme
+   * cases of stretched renders. */
+  void reset_scheduling(const BufferParams &params, int2 tile_size);
+
+  /* Update for the known buffer passes and scene parameters.
+   * Will store all parameters needed for buffers access outside of the scene graph. */
+  void update(const BufferParams &params, const Scene *scene);
+
+  inline int get_num_tiles() const
+  {
+    return tile_state_.num_tiles;
+  }
+
+  inline bool has_multiple_tiles() const
+  {
+    return tile_state_.num_tiles > 1;
+  }
+
+  inline int get_tile_overscan() const
+  {
+    return overscan_;
+  }
+
+  bool next();
+  bool done();
+
+  const Tile &get_current_tile() const;
+  const int2 get_size() const;
+
+  /* Write render buffer of a tile to a file on disk.
+   *
+   * Opens file for write when first tile is written.
+   *
+   * Returns true on success. */
+  bool write_tile(const RenderBuffers &tile_buffers);
+
+  /* Inform the tile manager that no more tiles will be written to disk.
+   * The file will be considered final, all handles to it will be closed. */
+  void finish_write_tiles();
+
+  /* Check whether any tile has been written to disk. */
+  inline bool has_written_tiles() const
+  {
+    return write_state_.num_tiles_written != 0;
+  }
+
+  /* Read full frame render buffer from tiles file on disk.
+   *
+   * Returns true on success. */
+  bool read_full_buffer_from_disk(string_view filename,
+                                  RenderBuffers *buffers,
+                                  DenoiseParams *denoise_params);
+
+  /* Compute valid tile size compatible with image saving. */
+  int compute_render_tile_size(const int suggested_tile_size) const;
+
+  /* Tile size in the image file. */
+  static const int IMAGE_TILE_SIZE = 128;
+
+ protected:
+  /* Get tile configuration for its index.
+   * The tile index must be within [0, state_.tile_state_). */
+  Tile get_tile_for_index(int index) const;
+
+  bool open_tile_output();
+  bool close_tile_output();
+
+  /* Part of an on-disk tile file name which avoids conflicts between several Cycles instances or
+   * several sessions. */
+  string tile_file_unique_part_;
+
+  int2 tile_size_ = make_int2(0, 0);
+
+  /* Number of extra pixels around the actual tile to render. */
+  int overscan_ = 0;
+
+  BufferParams buffer_params_;
+
+  /* Tile scheduling state. */
+  struct {
+    int num_tiles_x = 0;
+    int num_tiles_y = 0;
+    int num_tiles = 0;
+
+    int next_tile_index;
+
+    Tile current_tile;
+  } tile_state_;
+
+  /* State of tiles writing to a file on disk. */
+  struct {
+    /* Index of a tile file used during the current session.
+     * This number is used for the file name construction, making it possible to render several
+     * scenes throughout duration of the session and keep all results available for later read
+     * access. */
+    int tile_file_index = 0;
+
+    string filename;
+
+    /* Specification of the tile image which corresponds to the buffer parameters.
+     * Contains channels configured according to the passes configuration in the path traces.
+     *
+     * Output images are saved using this specification, input images are expected to have matched
+     * specification. */
+    ImageSpec image_spec;
+
+    /* Output handle for the tile file.
+     *
+     * This file can not be closed until all tiles has been provided, so the handle is stored in
+     * the state and is created whenever writing is requested. */
+    unique_ptr<ImageOutput> tile_out;
+
+    int num_tiles_written = 0;
+  } write_state_;
+};
+
+CCL_NAMESPACE_END