Fix adaptive sampling artifacts on tile boundaries

Implement an overscan support for tiles, so that adaptive sampling can
rely on the pixels neighbourhood.

Differential Revision: https://developer.blender.org/D12599

6 files changed, 62 insertions, 38 deletions

diff --git a/intern/cycles/integrator/denoiser_oidn.cpp b/intern/cycles/integrator/denoiser_oidn.cpp
index 7fc2b2b1892..ee3b62668a7 100644
--- a/intern/cycles/integrator/denoiser_oidn.cpp
+++ b/intern/cycles/integrator/denoiser_oidn.cpp
@@ -289,7 +289,13 @@ class OIDNDenoiseContext {
      * pixels. */
     const PassAccessorCPU pass_accessor(pass_access_info, 1.0f, num_samples_);
 
-    pass_accessor.get_render_tile_pixels(render_buffers_, buffer_params_, destination);
+    BufferParams buffer_params = buffer_params_;
+    buffer_params.window_x = 0;
+    buffer_params.window_y = 0;
+    buffer_params.window_width = buffer_params.width;
+    buffer_params.window_height = buffer_params.height;
+
+    pass_accessor.get_render_tile_pixels(render_buffers_, buffer_params, destination);
   }
 
   /* Read pass pixels using PassAccessor into a temporary buffer which is owned by the pass.. */
diff --git a/intern/cycles/integrator/pass_accessor_cpu.cpp b/intern/cycles/integrator/pass_accessor_cpu.cpp
index 3c6691f6d43..80908271ff6 100644
--- a/intern/cycles/integrator/pass_accessor_cpu.cpp
+++ b/intern/cycles/integrator/pass_accessor_cpu.cpp
@@ -99,17 +99,22 @@ inline void PassAccessorCPU::run_get_pass_kernel_processor_float(
 {
   DCHECK_EQ(destination.stride, 0) << "Custom stride for float destination is not implemented.";
 
-  const float *buffer_data = render_buffers->buffer.data();
+  const int64_t pass_stride = buffer_params.pass_stride;
+  const int64_t buffer_row_stride = buffer_params.stride * buffer_params.pass_stride;
+
+  const float *window_data = render_buffers->buffer.data() + buffer_params.window_x * pass_stride +
+                             buffer_params.window_y * buffer_row_stride;
+
   const int pixel_stride = destination.pixel_stride ? destination.pixel_stride :
                                                       destination.num_components;
 
-  tbb::parallel_for(0, buffer_params.height, [&](int64_t y) {
-    int64_t pixel_index = y * buffer_params.width;
-    for (int64_t x = 0; x < buffer_params.width; ++x, ++pixel_index) {
-      const int64_t input_pixel_offset = pixel_index * buffer_params.pass_stride;
-      const float *buffer = buffer_data + input_pixel_offset;
-      float *pixel = destination.pixels + (pixel_index + destination.offset) * pixel_stride;
+  tbb::parallel_for(0, buffer_params.window_height, [&](int64_t y) {
+    const float *buffer = window_data + y * buffer_row_stride;
+    float *pixel = destination.pixels +
+                   (y * buffer_params.width + destination.offset) * pixel_stride;
 
+    for (int64_t x = 0; x < buffer_params.window_width;
+         ++x, buffer += pass_stride, pixel += pixel_stride) {
       processor(kfilm_convert, buffer, pixel);
     }
   });
@@ -123,26 +128,28 @@ inline void PassAccessorCPU::run_get_pass_kernel_processor_half_rgba(
     const Destination &destination,
     const Processor &processor) const
 {
-  const float *buffer_data = render_buffers->buffer.data();
+  const int64_t pass_stride = buffer_params.pass_stride;
+  const int64_t buffer_row_stride = buffer_params.stride * buffer_params.pass_stride;
+
+  const float *window_data = render_buffers->buffer.data() + buffer_params.window_x * pass_stride +
+                             buffer_params.window_y * buffer_row_stride;
 
   half4 *dst_start = destination.pixels_half_rgba + destination.offset;
   const int destination_stride = destination.stride != 0 ? destination.stride :
                                                            buffer_params.width;
 
-  tbb::parallel_for(0, buffer_params.height, [&](int64_t y) {
-    int64_t pixel_index = y * buffer_params.width;
-    half4 *dst_row_start = dst_start + y * destination_stride;
-    for (int64_t x = 0; x < buffer_params.width; ++x, ++pixel_index) {
-      const int64_t input_pixel_offset = pixel_index * buffer_params.pass_stride;
-      const float *buffer = buffer_data + input_pixel_offset;
+  tbb::parallel_for(0, buffer_params.window_height, [&](int64_t y) {
+    const float *buffer = window_data + y * buffer_row_stride;
+    half4 *pixel = dst_start + y * destination_stride;
+    for (int64_t x = 0; x < buffer_params.window_width; ++x, buffer += pass_stride, ++pixel) {
 
-      float pixel[4];
-      processor(kfilm_convert, buffer, pixel);
+      float pixel_rgba[4];
+      processor(kfilm_convert, buffer, pixel_rgba);
 
-      film_apply_pass_pixel_overlays_rgba(kfilm_convert, buffer, pixel);
+      film_apply_pass_pixel_overlays_rgba(kfilm_convert, buffer, pixel_rgba);
 
-      half4 *pixel_half_rgba = dst_row_start + x;
-      float4_store_half(&pixel_half_rgba->x, make_float4(pixel[0], pixel[1], pixel[2], pixel[3]));
+      float4_store_half(&pixel->x,
+                        make_float4(pixel_rgba[0], pixel_rgba[1], pixel_rgba[2], pixel_rgba[3]));
     }
   });
 }
diff --git a/intern/cycles/integrator/pass_accessor_gpu.cpp b/intern/cycles/integrator/pass_accessor_gpu.cpp
index eb80ba99655..7b01d061708 100644
--- a/intern/cycles/integrator/pass_accessor_gpu.cpp
+++ b/intern/cycles/integrator/pass_accessor_gpu.cpp
@@ -43,10 +43,13 @@ void PassAccessorGPU::run_film_convert_kernels(DeviceKernel kernel,
   KernelFilmConvert kfilm_convert;
   init_kernel_film_convert(&kfilm_convert, buffer_params, destination);
 
-  const int work_size = buffer_params.width * buffer_params.height;
+  const int work_size = buffer_params.window_width * buffer_params.window_height;
 
   const int destination_stride = destination.stride != 0 ? destination.stride :
-                                                           buffer_params.width;
+                                                           buffer_params.window_width;
+
+  const int offset = buffer_params.window_x * buffer_params.pass_stride +
+                     buffer_params.window_y * buffer_params.stride * buffer_params.pass_stride;
 
   if (destination.d_pixels) {
     DCHECK_EQ(destination.stride, 0) << "Custom stride for float destination is not implemented.";
@@ -55,8 +58,8 @@ void PassAccessorGPU::run_film_convert_kernels(DeviceKernel kernel,
                     const_cast<device_ptr *>(&destination.d_pixels),
                     const_cast<device_ptr *>(&render_buffers->buffer.device_pointer),
                     const_cast<int *>(&work_size),
-                    const_cast<int *>(&buffer_params.width),
-                    const_cast<int *>(&buffer_params.offset),
+                    const_cast<int *>(&buffer_params.window_width),
+                    const_cast<int *>(&offset),
                     const_cast<int *>(&buffer_params.stride),
                     const_cast<int *>(&destination.offset),
                     const_cast<int *>(&destination_stride)};
@@ -70,8 +73,8 @@ void PassAccessorGPU::run_film_convert_kernels(DeviceKernel kernel,
                     const_cast<device_ptr *>(&destination.d_pixels_half_rgba),
                     const_cast<device_ptr *>(&render_buffers->buffer.device_pointer),
                     const_cast<int *>(&work_size),
-                    const_cast<int *>(&buffer_params.width),
-                    const_cast<int *>(&buffer_params.offset),
+                    const_cast<int *>(&buffer_params.window_width),
+                    const_cast<int *>(&offset),
                     const_cast<int *>(&buffer_params.stride),
                     const_cast<int *>(&destination.offset),
                     const_cast<int *>(&destination_stride)};
diff --git a/intern/cycles/integrator/path_trace.cpp b/intern/cycles/integrator/path_trace.cpp
index 7624b244175..3ea5c3c64b8 100644
--- a/intern/cycles/integrator/path_trace.cpp
+++ b/intern/cycles/integrator/path_trace.cpp
@@ -282,6 +282,12 @@ static BufferParams scale_buffer_params(const BufferParams &params, int resoluti
 
   scaled_params.width = max(1, params.width / resolution_divider);
   scaled_params.height = max(1, params.height / resolution_divider);
+
+  scaled_params.window_x = params.window_x / resolution_divider;
+  scaled_params.window_y = params.window_y / resolution_divider;
+  scaled_params.window_width = params.window_width / resolution_divider;
+  scaled_params.window_height = params.window_height / resolution_divider;
+
   scaled_params.full_x = params.full_x / resolution_divider;
   scaled_params.full_y = params.full_y / resolution_divider;
   scaled_params.full_width = params.full_width / resolution_divider;
@@ -1005,12 +1011,12 @@ bool PathTrace::set_render_tile_pixels(PassAccessor &pass_accessor,
 int2 PathTrace::get_render_tile_size() const
 {
   if (full_frame_state_.render_buffers) {
-    return make_int2(full_frame_state_.render_buffers->params.width,
-                     full_frame_state_.render_buffers->params.height);
+    return make_int2(full_frame_state_.render_buffers->params.window_width,
+                     full_frame_state_.render_buffers->params.window_height);
   }
 
   const Tile &tile = tile_manager_.get_current_tile();
-  return make_int2(tile.width, tile.height);
+  return make_int2(tile.window_width, tile.window_height);
 }
 
 int2 PathTrace::get_render_tile_offset() const
@@ -1020,7 +1026,7 @@ int2 PathTrace::get_render_tile_offset() const
   }
 
   const Tile &tile = tile_manager_.get_current_tile();
-  return make_int2(tile.x, tile.y);
+  return make_int2(tile.x + tile.window_x, tile.y + tile.window_y);
 }
 
 int2 PathTrace::get_render_size() const
diff --git a/intern/cycles/integrator/path_trace_work.cpp b/intern/cycles/integrator/path_trace_work.cpp
index c29177907c9..f626beb0aaa 100644
--- a/intern/cycles/integrator/path_trace_work.cpp
+++ b/intern/cycles/integrator/path_trace_work.cpp
@@ -191,8 +191,10 @@ PassAccessor::Destination PathTraceWork::get_display_destination_template(
   PassAccessor::Destination destination(film_->get_display_pass());
 
   const int2 display_texture_size = display->get_texture_size();
-  const int texture_x = effective_buffer_params_.full_x - effective_full_params_.full_x;
-  const int texture_y = effective_buffer_params_.full_y - effective_full_params_.full_y;
+  const int texture_x = effective_buffer_params_.full_x - effective_full_params_.full_x +
+                        effective_buffer_params_.window_x;
+  const int texture_y = effective_buffer_params_.full_y - effective_full_params_.full_y +
+                        effective_buffer_params_.window_y;
 
   destination.offset = texture_y * display_texture_size.x + texture_x;
   destination.stride = display_texture_size.x;
diff --git a/intern/cycles/integrator/path_trace_work_gpu.cpp b/intern/cycles/integrator/path_trace_work_gpu.cpp
index 7babc9d09fa..c29b0fb039e 100644
--- a/intern/cycles/integrator/path_trace_work_gpu.cpp
+++ b/intern/cycles/integrator/path_trace_work_gpu.cpp
@@ -737,13 +737,13 @@ void PathTraceWorkGPU::copy_to_display_naive(PathTraceDisplay *display,
 {
   const int full_x = effective_buffer_params_.full_x;
   const int full_y = effective_buffer_params_.full_y;
-  const int width = effective_buffer_params_.width;
-  const int height = effective_buffer_params_.height;
-  const int final_width = buffers_->params.width;
-  const int final_height = buffers_->params.height;
+  const int width = effective_buffer_params_.window_width;
+  const int height = effective_buffer_params_.window_height;
+  const int final_width = buffers_->params.window_width;
+  const int final_height = buffers_->params.window_height;
 
-  const int texture_x = full_x - effective_full_params_.full_x;
-  const int texture_y = full_y - effective_full_params_.full_y;
+  const int texture_x = full_x - effective_full_params_.full_x + effective_buffer_params_.window_x;
+  const int texture_y = full_y - effective_full_params_.full_y + effective_buffer_params_.window_y;
 
   /* Re-allocate display memory if needed, and make sure the device pointer is allocated.
    *