Fix T78881: Cycles OpenImageDenoise not using albedo and normal correctly
Properly normalize buffers now. Also expose option to not use albedo and normal just like OptiX.
This commit is contained in:
parent
2b5e21fe00
commit
6e74a8b69f
|
@ -212,7 +212,7 @@ def enum_denoiser(self, context):
|
|||
items += enum_openimagedenoise_denoiser(self, context)
|
||||
return items
|
||||
|
||||
enum_denoising_optix_input_passes = (
|
||||
enum_denoising_input_passes = (
|
||||
('RGB', "Color", "Use only color as input", 1),
|
||||
('RGB_ALBEDO', "Color + Albedo", "Use color and albedo data as input", 2),
|
||||
('RGB_ALBEDO_NORMAL', "Color + Albedo + Normal", "Use color, albedo and normal data as input", 3),
|
||||
|
@ -1451,11 +1451,18 @@ class CyclesRenderLayerSettings(bpy.types.PropertyGroup):
|
|||
|
||||
denoising_optix_input_passes: EnumProperty(
|
||||
name="Input Passes",
|
||||
description="Passes handed over to the OptiX denoiser (this can have different effects on the denoised image)",
|
||||
items=enum_denoising_optix_input_passes,
|
||||
description="Passes used by the denoiser to distinguish noise from shader and geometry detail",
|
||||
items=enum_denoising_input_passes,
|
||||
default='RGB_ALBEDO',
|
||||
)
|
||||
|
||||
denoising_openimagedenoise_input_passes: EnumProperty(
|
||||
name="Input Passes",
|
||||
description="Passes used by the denoiser to distinguish noise from shader and geometry detail",
|
||||
items=enum_denoising_input_passes,
|
||||
default='RGB_ALBEDO_NORMAL',
|
||||
)
|
||||
|
||||
use_pass_crypto_object: BoolProperty(
|
||||
name="Cryptomatte Object",
|
||||
description="Render cryptomatte object pass, for isolating objects in compositing",
|
||||
|
|
|
@ -1008,6 +1008,7 @@ class CYCLES_RENDER_PT_denoising(CyclesButtonsPanel, Panel):
|
|||
col.prop(cycles_view_layer, "denoising_optix_input_passes")
|
||||
return
|
||||
elif denoiser == 'OPENIMAGEDENOISE':
|
||||
col.prop(cycles_view_layer, "denoising_openimagedenoise_input_passes")
|
||||
return
|
||||
|
||||
col.prop(cycles_view_layer, "denoising_radius", text="Radius")
|
||||
|
|
|
@ -954,7 +954,13 @@ DenoiseParams BlenderSync::get_denoise_params(BL::Scene &b_scene,
|
|||
denoising.strength = get_float(clayer, "denoising_strength");
|
||||
denoising.feature_strength = get_float(clayer, "denoising_feature_strength");
|
||||
denoising.relative_pca = get_boolean(clayer, "denoising_relative_pca");
|
||||
denoising.optix_input_passes = get_enum(clayer, "denoising_optix_input_passes");
|
||||
|
||||
denoising.input_passes = (DenoiserInput)get_enum(
|
||||
clayer,
|
||||
(denoising.type == DENOISER_OPTIX) ? "denoising_optix_input_passes" :
|
||||
"denoising_openimagedenoise_input_passes",
|
||||
DENOISER_INPUT_NUM,
|
||||
DENOISER_INPUT_RGB_ALBEDO_NORMAL);
|
||||
|
||||
denoising.store_passes = get_boolean(clayer, "denoising_store_passes");
|
||||
}
|
||||
|
|
|
@ -951,12 +951,13 @@ class CPUDevice : public Device {
|
|||
|
||||
void denoise_openimagedenoise_buffer(DeviceTask &task,
|
||||
float *buffer,
|
||||
size_t offset,
|
||||
size_t stride,
|
||||
size_t x,
|
||||
size_t y,
|
||||
size_t w,
|
||||
size_t h)
|
||||
const size_t offset,
|
||||
const size_t stride,
|
||||
const size_t x,
|
||||
const size_t y,
|
||||
const size_t w,
|
||||
const size_t h,
|
||||
const float scale)
|
||||
{
|
||||
#ifdef WITH_OPENIMAGEDENOISE
|
||||
assert(openimagedenoise_supported());
|
||||
|
@ -982,31 +983,65 @@ class CPUDevice : public Device {
|
|||
}
|
||||
|
||||
/* Set images with appropriate stride for our interleaved pass storage. */
|
||||
const struct {
|
||||
struct {
|
||||
const char *name;
|
||||
int offset;
|
||||
} passes[] = {{"color", task.pass_denoising_data + DENOISING_PASS_COLOR},
|
||||
{"normal", task.pass_denoising_data + DENOISING_PASS_NORMAL},
|
||||
{"albedo", task.pass_denoising_data + DENOISING_PASS_ALBEDO},
|
||||
{"output", 0},
|
||||
const int offset;
|
||||
const bool scale;
|
||||
const bool use;
|
||||
array<float> scaled_buffer;
|
||||
} passes[] = {{"color", task.pass_denoising_data + DENOISING_PASS_COLOR, false, true},
|
||||
{"albedo",
|
||||
task.pass_denoising_data + DENOISING_PASS_ALBEDO,
|
||||
true,
|
||||
task.denoising.input_passes >= DENOISER_INPUT_RGB_ALBEDO},
|
||||
{"normal",
|
||||
task.pass_denoising_data + DENOISING_PASS_NORMAL,
|
||||
true,
|
||||
task.denoising.input_passes >= DENOISER_INPUT_RGB_ALBEDO_NORMAL},
|
||||
{"output", 0, false, true},
|
||||
{ NULL,
|
||||
0 }};
|
||||
|
||||
for (int i = 0; passes[i].name; i++) {
|
||||
if (!passes[i].use) {
|
||||
continue;
|
||||
}
|
||||
|
||||
const int64_t pixel_offset = offset + x + y * stride;
|
||||
const int64_t buffer_offset = (pixel_offset * task.pass_stride + passes[i].offset) *
|
||||
sizeof(float);
|
||||
const int64_t pixel_stride = task.pass_stride * sizeof(float);
|
||||
const int64_t buffer_offset = (pixel_offset * task.pass_stride + passes[i].offset);
|
||||
const int64_t pixel_stride = task.pass_stride;
|
||||
const int64_t row_stride = stride * pixel_stride;
|
||||
|
||||
if (passes[i].scale && scale != 1.0f) {
|
||||
/* Normalize albedo and normal passes as they are scaled by the number of samples.
|
||||
* For the color passes OIDN will perform autoexposure making it unnecessary. */
|
||||
array<float> &scaled_buffer = passes[i].scaled_buffer;
|
||||
scaled_buffer.resize(w * h * 3);
|
||||
|
||||
for (int y = 0; y < h; y++) {
|
||||
const float *pass_row = buffer + buffer_offset + y * row_stride;
|
||||
float *scaled_row = scaled_buffer.data() + y * w * 3;
|
||||
|
||||
for (int x = 0; x < w; x++) {
|
||||
scaled_row[x * 3 + 0] = pass_row[x * pixel_stride + 0] * scale;
|
||||
scaled_row[x * 3 + 1] = pass_row[x * pixel_stride + 1] * scale;
|
||||
scaled_row[x * 3 + 2] = pass_row[x * pixel_stride + 2] * scale;
|
||||
}
|
||||
}
|
||||
|
||||
oidn_filter.setImage(
|
||||
passes[i].name, scaled_buffer.data(), oidn::Format::Float3, w, h, 0, 0, 0);
|
||||
}
|
||||
else {
|
||||
oidn_filter.setImage(passes[i].name,
|
||||
(char *)buffer + buffer_offset,
|
||||
buffer + buffer_offset,
|
||||
oidn::Format::Float3,
|
||||
w,
|
||||
h,
|
||||
0,
|
||||
pixel_stride,
|
||||
row_stride);
|
||||
pixel_stride * sizeof(float),
|
||||
row_stride * sizeof(float));
|
||||
}
|
||||
}
|
||||
|
||||
/* Execute filter. */
|
||||
|
@ -1021,6 +1056,7 @@ class CPUDevice : public Device {
|
|||
(void)y;
|
||||
(void)w;
|
||||
(void)h;
|
||||
(void)scale;
|
||||
#endif
|
||||
}
|
||||
|
||||
|
@ -1037,7 +1073,8 @@ class CPUDevice : public Device {
|
|||
rtile.x,
|
||||
rtile.y,
|
||||
rtile.w,
|
||||
rtile.h);
|
||||
rtile.h,
|
||||
1.0f / rtile.sample);
|
||||
|
||||
/* todo: it may be possible to avoid this copy, but we have to ensure that
|
||||
* when other code copies data from the device it doesn't overwrite the
|
||||
|
@ -1047,6 +1084,9 @@ class CPUDevice : public Device {
|
|||
else {
|
||||
/* Per-tile denoising. */
|
||||
rtile.sample = rtile.start_sample + rtile.num_samples;
|
||||
const float scale = 1.0f / rtile.sample;
|
||||
const float invscale = rtile.sample;
|
||||
const size_t pass_stride = task.pass_stride;
|
||||
|
||||
/* Map neighboring tiles into one buffer for denoising. */
|
||||
RenderTileNeighbors neighbors(rtile);
|
||||
|
@ -1075,22 +1115,24 @@ class CPUDevice : public Device {
|
|||
const int ymax = min(ntile.y + ntile.h, rect.w);
|
||||
|
||||
const size_t tile_offset = ntile.offset + xmin + ymin * ntile.stride;
|
||||
const float *tile_buffer = (float *)ntile.buffer + tile_offset * task.pass_stride;
|
||||
const float *tile_buffer = (float *)ntile.buffer + tile_offset * pass_stride;
|
||||
|
||||
const size_t merged_stride = rect_size.x;
|
||||
const size_t merged_offset = (xmin - rect.x) + (ymin - rect.y) * merged_stride;
|
||||
float *merged_buffer = merged.data() + merged_offset * task.pass_stride;
|
||||
float *merged_buffer = merged.data() + merged_offset * pass_stride;
|
||||
|
||||
for (int y = ymin; y < ymax; y++) {
|
||||
memcpy(merged_buffer, tile_buffer, sizeof(float) * task.pass_stride * (xmax - xmin));
|
||||
tile_buffer += ntile.stride * task.pass_stride;
|
||||
merged_buffer += merged_stride * task.pass_stride;
|
||||
for (int x = 0; x < pass_stride * (xmax - xmin); x++) {
|
||||
merged_buffer[x] = tile_buffer[x] * scale;
|
||||
}
|
||||
tile_buffer += ntile.stride * pass_stride;
|
||||
merged_buffer += merged_stride * pass_stride;
|
||||
}
|
||||
}
|
||||
|
||||
/* Denoise */
|
||||
denoise_openimagedenoise_buffer(
|
||||
task, merged.data(), 0, rect_size.x, 0, 0, rect_size.x, rect_size.y);
|
||||
task, merged.data(), 0, rect_size.x, 0, 0, rect_size.x, rect_size.y, 1.0f);
|
||||
|
||||
/* Copy back result from merged buffer. */
|
||||
RenderTile &ntile = neighbors.target;
|
||||
|
@ -1101,16 +1143,20 @@ class CPUDevice : public Device {
|
|||
const int ymax = min(ntile.y + ntile.h, rect.w);
|
||||
|
||||
const size_t tile_offset = ntile.offset + xmin + ymin * ntile.stride;
|
||||
float *tile_buffer = (float *)ntile.buffer + tile_offset * task.pass_stride;
|
||||
float *tile_buffer = (float *)ntile.buffer + tile_offset * pass_stride;
|
||||
|
||||
const size_t merged_stride = rect_size.x;
|
||||
const size_t merged_offset = (xmin - rect.x) + (ymin - rect.y) * merged_stride;
|
||||
const float *merged_buffer = merged.data() + merged_offset * task.pass_stride;
|
||||
const float *merged_buffer = merged.data() + merged_offset * pass_stride;
|
||||
|
||||
for (int y = ymin; y < ymax; y++) {
|
||||
memcpy(tile_buffer, merged_buffer, sizeof(float) * task.pass_stride * (xmax - xmin));
|
||||
tile_buffer += ntile.stride * task.pass_stride;
|
||||
merged_buffer += merged_stride * task.pass_stride;
|
||||
for (int x = 0; x < pass_stride * (xmax - xmin); x += pass_stride) {
|
||||
tile_buffer[x + 0] = merged_buffer[x + 0] * invscale;
|
||||
tile_buffer[x + 1] = merged_buffer[x + 1] * invscale;
|
||||
tile_buffer[x + 2] = merged_buffer[x + 2] * invscale;
|
||||
}
|
||||
tile_buffer += ntile.stride * pass_stride;
|
||||
merged_buffer += merged_stride * pass_stride;
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
@ -877,7 +877,7 @@ class OptiXDevice : public CUDADevice {
|
|||
|
||||
# if OPTIX_DENOISER_NO_PIXEL_STRIDE
|
||||
device_only_memory<float> input_rgb(this, "denoiser input rgb");
|
||||
input_rgb.alloc_to_device(rect_size.x * rect_size.y * 3 * task.denoising.optix_input_passes);
|
||||
input_rgb.alloc_to_device(rect_size.x * rect_size.y * 3 * task.denoising.input_passes);
|
||||
|
||||
void *input_args[] = {&input_rgb.device_pointer,
|
||||
&input_ptr,
|
||||
|
@ -886,7 +886,7 @@ class OptiXDevice : public CUDADevice {
|
|||
&input_stride,
|
||||
&task.pass_stride,
|
||||
const_cast<int *>(pass_offset),
|
||||
&task.denoising.optix_input_passes,
|
||||
&task.denoising.input_passes,
|
||||
&rtile.sample};
|
||||
launch_filter_kernel(
|
||||
"kernel_cuda_filter_convert_to_rgb", rect_size.x, rect_size.y, input_args);
|
||||
|
@ -897,7 +897,7 @@ class OptiXDevice : public CUDADevice {
|
|||
# endif
|
||||
|
||||
const bool recreate_denoiser = (denoiser == NULL) ||
|
||||
(task.denoising.optix_input_passes != denoiser_input_passes);
|
||||
(task.denoising.input_passes != denoiser_input_passes);
|
||||
if (recreate_denoiser) {
|
||||
// Destroy existing handle before creating new one
|
||||
if (denoiser != NULL) {
|
||||
|
@ -906,9 +906,9 @@ class OptiXDevice : public CUDADevice {
|
|||
|
||||
// Create OptiX denoiser handle on demand when it is first used
|
||||
OptixDenoiserOptions denoiser_options;
|
||||
assert(task.denoising.optix_input_passes >= 1 && task.denoising.optix_input_passes <= 3);
|
||||
assert(task.denoising.input_passes >= 1 && task.denoising.input_passes <= 3);
|
||||
denoiser_options.inputKind = static_cast<OptixDenoiserInputKind>(
|
||||
OPTIX_DENOISER_INPUT_RGB + (task.denoising.optix_input_passes - 1));
|
||||
OPTIX_DENOISER_INPUT_RGB + (task.denoising.input_passes - 1));
|
||||
# if OPTIX_ABI_VERSION < 28
|
||||
denoiser_options.pixelFormat = OPTIX_PIXEL_FORMAT_FLOAT3;
|
||||
# endif
|
||||
|
@ -917,7 +917,7 @@ class OptiXDevice : public CUDADevice {
|
|||
optixDenoiserSetModel(denoiser, OPTIX_DENOISER_MODEL_KIND_HDR, NULL, 0));
|
||||
|
||||
// OptiX denoiser handle was created with the requested number of input passes
|
||||
denoiser_input_passes = task.denoising.optix_input_passes;
|
||||
denoiser_input_passes = task.denoising.input_passes;
|
||||
}
|
||||
|
||||
OptixDenoiserSizes sizes = {};
|
||||
|
@ -992,7 +992,7 @@ class OptiXDevice : public CUDADevice {
|
|||
denoiser_state.device_pointer,
|
||||
scratch_offset,
|
||||
input_layers,
|
||||
task.denoising.optix_input_passes,
|
||||
task.denoising.input_passes,
|
||||
overlap_offset.x,
|
||||
overlap_offset.y,
|
||||
output_layers,
|
||||
|
|
|
@ -42,6 +42,14 @@ enum DenoiserType {
|
|||
DENOISER_ALL = ~0,
|
||||
};
|
||||
|
||||
enum DenoiserInput {
|
||||
DENOISER_INPUT_RGB = 1,
|
||||
DENOISER_INPUT_RGB_ALBEDO = 2,
|
||||
DENOISER_INPUT_RGB_ALBEDO_NORMAL = 3,
|
||||
|
||||
DENOISER_INPUT_NUM,
|
||||
};
|
||||
|
||||
typedef int DenoiserTypeMask;
|
||||
|
||||
class DenoiseParams {
|
||||
|
@ -73,10 +81,10 @@ class DenoiseParams {
|
|||
/* Clamp the input to the range of +-1e8. Should be enough for any legitimate data. */
|
||||
bool clamp_input;
|
||||
|
||||
/** Optix Denoiser **/
|
||||
/** OIDN/Optix Denoiser **/
|
||||
|
||||
/* Passes handed over to the OptiX denoiser (default to color + albedo). */
|
||||
int optix_input_passes;
|
||||
/* Passes handed over to the OIDN/OptiX denoiser (default to color + albedo). */
|
||||
DenoiserInput input_passes;
|
||||
|
||||
DenoiseParams()
|
||||
{
|
||||
|
@ -92,7 +100,7 @@ class DenoiseParams {
|
|||
neighbor_frames = 2;
|
||||
clamp_input = true;
|
||||
|
||||
optix_input_passes = 2;
|
||||
input_passes = DENOISER_INPUT_RGB_ALBEDO_NORMAL;
|
||||
|
||||
start_sample = 0;
|
||||
}
|
||||
|
|
Loading…
Reference in New Issue