Commit: b10c64bd2f1dae93c38e2d1cc656ea08151ab704
Author: Lukas Stockner
Date: Wed Jul 4 14:22:38 2018 +0200
Branches: master
https://developer.blender.org/rBb10c64bd2f1dae93c38e2d1cc656ea08151ab704
Cycles Denoising: Split main function into logical steps
===================================================================
M intern/cycles/device/device_cpu.cpp
M intern/cycles/device/device_cuda.cpp
M intern/cycles/device/device_denoising.cpp
M intern/cycles/device/device_denoising.h
M intern/cycles/device/opencl/opencl_base.cpp
M intern/cycles/device/opencl/opencl_mega.cpp
M intern/cycles/device/opencl/opencl_split.cpp
===================================================================
diff --git a/intern/cycles/device/device_cpu.cpp
b/intern/cycles/device/device_cpu.cpp
index 6be60f8bbb6..81a084c03b2 100644
--- a/intern/cycles/device/device_cpu.cpp
+++ b/intern/cycles/device/device_cpu.cpp
@@ -732,8 +732,6 @@ public:
task.map_neighbor_tiles(rtiles, this);
denoising.tiles_from_rendertiles(rtiles);
- denoising.init_from_devicetask(task);
-
denoising.run_denoising();
task.unmap_neighbor_tiles(rtiles, this);
@@ -766,7 +764,7 @@ public:
}
RenderTile tile;
- DenoisingTask denoising(this);
+ DenoisingTask denoising(this, task);
while(task.acquire_tile(this, tile)) {
if(tile.task == RenderTile::PATH_TRACE) {
diff --git a/intern/cycles/device/device_cuda.cpp
b/intern/cycles/device/device_cuda.cpp
index bf5a95dd233..34c64feb80a 100644
--- a/intern/cycles/device/device_cuda.cpp
+++ b/intern/cycles/device/device_cuda.cpp
@@ -1632,8 +1632,6 @@ public:
task.map_neighbor_tiles(rtiles, this);
denoising.tiles_from_rendertiles(rtiles);
- denoising.init_from_devicetask(task);
-
denoising.run_denoising();
task.unmap_neighbor_tiles(rtiles, this);
@@ -2074,7 +2072,7 @@ public:
/* keep rendering tiles until done */
RenderTile tile;
- DenoisingTask denoising(this);
+ DenoisingTask denoising(this, *task);
while(task->acquire_tile(this, tile)) {
if(tile.task == RenderTile::PATH_TRACE) {
diff --git a/intern/cycles/device/device_denoising.cpp
b/intern/cycles/device/device_denoising.cpp
index 644cf6cd10e..4d2ba508aec 100644
--- a/intern/cycles/device/device_denoising.cpp
+++ b/intern/cycles/device/device_denoising.cpp
@@ -20,12 +20,24 @@
CCL_NAMESPACE_BEGIN
-DenoisingTask::DenoisingTask(Device *device)
+DenoisingTask::DenoisingTask(Device *device, const DeviceTask &task)
: tiles_mem(device, "denoising tiles_mem", MEM_READ_WRITE),
storage(device),
buffer(device),
device(device)
{
+ radius = task.denoising_radius;
+ nlm_k_2 = powf(2.0f, lerp(-5.0f, 3.0f, task.denoising_strength));
+ if(task.denoising_relative_pca) {
+ pca_threshold = -powf(10.0f, lerp(-8.0f, 0.0f,
task.denoising_feature_strength));
+ }
+ else {
+ pca_threshold = powf(10.0f, lerp(-5.0f, 3.0f,
task.denoising_feature_strength));
+ }
+
+ render_buffer.pass_stride = task.pass_stride;
+ render_buffer.denoising_data_offset = task.pass_denoising_data;
+ render_buffer.denoising_clean_offset = task.pass_denoising_clean;
}
DenoisingTask::~DenoisingTask()
@@ -41,26 +53,6 @@ DenoisingTask::~DenoisingTask()
tiles_mem.free();
}
-void DenoisingTask::init_from_devicetask(const DeviceTask &task)
-{
- radius = task.denoising_radius;
- nlm_k_2 = powf(2.0f, lerp(-5.0f, 3.0f, task.denoising_strength));
- if(task.denoising_relative_pca) {
- pca_threshold = -powf(10.0f, lerp(-8.0f, 0.0f,
task.denoising_feature_strength));
- }
- else {
- pca_threshold = powf(10.0f, lerp(-5.0f, 3.0f,
task.denoising_feature_strength));
- }
-
- render_buffer.pass_stride = task.pass_stride;
- render_buffer.denoising_data_offset = task.pass_denoising_data;
- render_buffer.denoising_clean_offset = task.pass_denoising_clean;
-
- /* Expand filter_area by radius pixels and clamp the result to the
extent of the neighboring tiles */
- rect = rect_from_shape(filter_area.x, filter_area.y, filter_area.z,
filter_area.w);
- rect = rect_expand(rect, radius);
- rect = rect_clip(rect, make_int4(tiles->x[0], tiles->y[0], tiles->x[3],
tiles->y[3]));
-}
void DenoisingTask::tiles_from_rendertiles(RenderTile *rtiles)
{
@@ -88,120 +80,142 @@ void DenoisingTask::tiles_from_rendertiles(RenderTile
*rtiles)
functions.set_tiles(buffers);
}
-bool DenoisingTask::run_denoising()
+void DenoisingTask::setup_denoising_buffer()
{
- /* Allocate denoising buffer. */
+ /* Expand filter_area by radius pixels and clamp the result to the
extent of the neighboring tiles */
+ rect = rect_from_shape(filter_area.x, filter_area.y, filter_area.z,
filter_area.w);
+ rect = rect_expand(rect, radius);
+ rect = rect_clip(rect, make_int4(tiles->x[0], tiles->y[0], tiles->x[3],
tiles->y[3]));
+
buffer.passes = 14;
buffer.width = rect.z - rect.x;
buffer.stride = align_up(buffer.width, 4);
buffer.h = rect.w - rect.y;
- buffer.pass_stride = align_up(buffer.stride * buffer.h,
divide_up(device->mem_sub_ptr_alignment(), sizeof(float)));
- buffer.mem.alloc_to_device(buffer.pass_stride * buffer.passes, false);
+ int alignment_floats = divide_up(device->mem_sub_ptr_alignment(),
sizeof(float));
+ buffer.pass_stride = align_up(buffer.stride * buffer.h,
alignment_floats);
+ /* Pad the total size by four floats since the SIMD kernels might go a
bit over the end. */
+ int mem_size = align_up(buffer.pass_stride * buffer.passes + 4,
alignment_floats);
+ buffer.mem.alloc_to_device(mem_size, false);
+}
+void DenoisingTask::prefilter_shadowing()
+{
device_ptr null_ptr = (device_ptr) 0;
- /* Prefilter shadow feature. */
- {
- device_sub_ptr unfiltered_a (buffer.mem, 0,
buffer.pass_stride);
- device_sub_ptr unfiltered_b (buffer.mem,
1*buffer.pass_stride, buffer.pass_stride);
- device_sub_ptr sample_var (buffer.mem,
2*buffer.pass_stride, buffer.pass_stride);
- device_sub_ptr sample_var_var (buffer.mem,
3*buffer.pass_stride, buffer.pass_stride);
- device_sub_ptr buffer_var (buffer.mem,
5*buffer.pass_stride, buffer.pass_stride);
- device_sub_ptr filtered_var (buffer.mem,
6*buffer.pass_stride, buffer.pass_stride);
- device_sub_ptr nlm_temporary_1(buffer.mem,
7*buffer.pass_stride, buffer.pass_stride);
- device_sub_ptr nlm_temporary_2(buffer.mem,
8*buffer.pass_stride, buffer.pass_stride);
- device_sub_ptr nlm_temporary_3(buffer.mem,
9*buffer.pass_stride, buffer.pass_stride);
-
- nlm_state.temporary_1_ptr = *nlm_temporary_1;
- nlm_state.temporary_2_ptr = *nlm_temporary_2;
- nlm_state.temporary_3_ptr = *nlm_temporary_3;
-
- /* Get the A/B unfiltered passes, the combined sample variance,
the estimated variance of the sample variance and the buffer variance. */
- functions.divide_shadow(*unfiltered_a, *unfiltered_b,
*sample_var, *sample_var_var, *buffer_var);
-
- /* Smooth the (generally pretty noisy) buffer variance using
the spatial information from the sample variance. */
- nlm_state.set_parameters(6, 3, 4.0f, 1.0f);
- functions.non_local_means(*buffer_var, *sample_var,
*sample_var_var, *filtered_var);
-
- /* Reuse memory, the previous data isn't needed anymore. */
- device_ptr filtered_a = *buffer_var,
- filtered_b = *sample_var;
- /* Use the smoothed variance to filter the two shadow half
images using each other for weight calculation. */
- nlm_state.set_parameters(5, 3, 1.0f, 0.25f);
- functions.non_local_means(*unfiltered_a, *unfiltered_b,
*filtered_var, filtered_a);
- functions.non_local_means(*unfiltered_b, *unfiltered_a,
*filtered_var, filtered_b);
-
- device_ptr residual_var = *sample_var_var;
- /* Estimate the residual variance between the two filtered
halves. */
- functions.combine_halves(filtered_a, filtered_b, null_ptr,
residual_var, 2, rect);
-
- device_ptr final_a = *unfiltered_a,
- final_b = *unfiltered_b;
- /* Use the residual variance for a second filter pass. */
- nlm_state.set_parameters(4, 2, 1.0f, 0.5f);
- functions.non_local_means(filtered_a, filtered_b, residual_var,
final_a);
- functions.non_local_means(filtered_b, filtered_a, residual_var,
final_b);
-
- /* Combine the two double-filtered halves to a final shadow
feature. */
- device_sub_ptr shadow_pass(buffer.mem, 4*buffer.pass_stride,
buffer.pass_stride);
- functions.combine_halves(final_a, final_b, *shadow_pass,
null_ptr, 0, rect);
- }
+ device_sub_ptr unfiltered_a (buffer.mem, 0,
buffer.pass_stride);
+ device_sub_ptr unfiltered_b (buffer.mem, 1*buffer.pass_stride,
buffer.pass_stride);
+ device_sub_ptr sample_var (buffer.mem, 2*buffer.pass_stride,
buffer.pass_stride);
+ device_sub_ptr sample_var_var (buffer.mem, 3*buffer.pass_stride,
buffer.pass_stride);
+ device_sub_ptr buffer_var (buffer.mem, 5*buffer.pass_stride,
buffer.pass_stride);
+ device_sub_ptr filtered_var (buffer.mem, 6*buffer.pass_stride,
buffer.pass_stride);
+ device_sub_ptr nlm_temporary_1(buffer.mem, 7*buffer.pass_stride,
buffer.pass_stride);
+ device_sub_ptr nlm_temporary_2(buffer.mem, 8*buffer.pass_stride,
buffer.pass_stride);
+ device_sub_ptr nlm_temporary_3(buffer.mem, 9*buffer.pass_stride,
buffer.pass_stride);
+
+ nlm_state.temporary_1_ptr = *nlm_temporary_1;
+ nlm_state.temporary_2_ptr = *nlm_temporary_2;
+ nlm_state.temporary_3_ptr = *nlm_temporary_3;
+
+ /* Get the A/B unfiltered passes, the combined sample variance, the
estimated variance of the sample variance and the buffer variance. */
+ functions.divide_shadow(*unfiltered_a, *unfiltered_b, *sample_var,
*sample_var_var, *buffer_var);
+
+ /* Smooth the (generally pretty noisy) buffer variance using the
spatial information from the sample variance. */
+ nlm_state.set_parameters(6, 3, 4.0f, 1.0f);
+ functions.non_local_means(*buffer_var, *sample_var, *sample_var_var,
*filtered_var);
+
+ /* Reuse memory, the previous data isn't needed anymore. */
+ device_ptr filtered_a = *buffer_var,
+ filtered_b = *sample_var;
+ /* Use the smoothed variance to filter the two shadow half images using
each other for weight calculation. */
+ nlm_state.set_parameters(5, 3, 1.0f, 0.25f);
+ functions.non_local_means(*unfiltered_a, *unfiltered_b, *filtered_var,
filtered_a);
+ functions.non_local_means(*unfiltered_b, *unfiltered_a, *filtered_var,
filtered_b);
+
+ device_ptr residual_var = *sample_var_var;
+ /* Estimate the residual variance between the two filtered halves. */
+ functions.combine_halves(filtered_a, filtered_b, null_ptr,
residual_var, 2, rect);
+
+ device_ptr final_a = *unfiltered_a,
+ final_b = *unfiltered_b;
+ /* Use the residual variance for a second filter pass. */
+ nlm_state.set_parameters(4, 2, 1.0f, 0.5f);
+ functions.non_local_means(filtered_a, filtered_b, residual_var,
final_a);
+ functions.non_local_means(filtered_b, filtered_a, residual_var,
final_b);
+
+ /* Combine the two double-filtered halves to a final shadow feature. */
+ device_sub_ptr shadow_pass(buffer.mem, 4*buffer.pass_stride,
buffer.pass_stride);
+ functions.combine_halves(fi
@@ Diff output truncated at 10240 characters. @@
_______________________________________________
Bf-blender-cvs mailing list
[email protected]
https://lists.blender.org/mailman/listinfo/bf-blender-cvs
