Code refactor: move more memory allocation logic into device API.
[blender.git] / intern / cycles / device / device_denoising.cpp
1 /*
2  * Copyright 2011-2017 Blender Foundation
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  * http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16
17 #include "device/device_denoising.h"
18
19 #include "kernel/filter/filter_defines.h"
20
21 CCL_NAMESPACE_BEGIN
22
23 void DenoisingTask::init_from_devicetask(const DeviceTask &task)
24 {
25         radius = task.denoising_radius;
26         nlm_k_2 = powf(2.0f, lerp(-5.0f, 3.0f, task.denoising_strength));
27         if(task.denoising_relative_pca) {
28                 pca_threshold = -powf(10.0f, lerp(-8.0f, 0.0f, task.denoising_feature_strength));
29         }
30         else {
31                 pca_threshold = powf(10.0f, lerp(-5.0f, 3.0f, task.denoising_feature_strength));
32         }
33
34         render_buffer.pass_stride = task.pass_stride;
35         render_buffer.denoising_data_offset  = task.pass_denoising_data;
36         render_buffer.denoising_clean_offset = task.pass_denoising_clean;
37
38         /* Expand filter_area by radius pixels and clamp the result to the extent of the neighboring tiles */
39         rect = make_int4(max(tiles->x[0], filter_area.x - radius),
40                          max(tiles->y[0], filter_area.y - radius),
41                          min(tiles->x[3], filter_area.x + filter_area.z + radius),
42                          min(tiles->y[3], filter_area.y + filter_area.w + radius));
43 }
44
45 void DenoisingTask::tiles_from_rendertiles(RenderTile *rtiles)
46 {
47         tiles = (TilesInfo*) tiles_mem.alloc(sizeof(TilesInfo)/sizeof(int));
48
49         device_ptr buffers[9];
50         for(int i = 0; i < 9; i++) {
51                 buffers[i] = rtiles[i].buffer;
52                 tiles->offsets[i] = rtiles[i].offset;
53                 tiles->strides[i] = rtiles[i].stride;
54         }
55         tiles->x[0] = rtiles[3].x;
56         tiles->x[1] = rtiles[4].x;
57         tiles->x[2] = rtiles[5].x;
58         tiles->x[3] = rtiles[5].x + rtiles[5].w;
59         tiles->y[0] = rtiles[1].y;
60         tiles->y[1] = rtiles[4].y;
61         tiles->y[2] = rtiles[7].y;
62         tiles->y[3] = rtiles[7].y + rtiles[7].h;
63
64         render_buffer.offset = rtiles[4].offset;
65         render_buffer.stride = rtiles[4].stride;
66         render_buffer.ptr    = rtiles[4].buffer;
67
68         functions.set_tiles(buffers);
69 }
70
71 bool DenoisingTask::run_denoising()
72 {
73         /* Allocate denoising buffer. */
74         buffer.passes = 14;
75         buffer.w = align_up(rect.z - rect.x, 4);
76         buffer.h = rect.w - rect.y;
77         buffer.pass_stride = align_up(buffer.w * buffer.h, divide_up(device->mem_address_alignment(), sizeof(float)));
78         buffer.mem.alloc_to_device(buffer.pass_stride * buffer.passes);
79
80         device_ptr null_ptr = (device_ptr) 0;
81
82         /* Prefilter shadow feature. */
83         {
84                 device_sub_ptr unfiltered_a   (buffer.mem, 0,                    buffer.pass_stride);
85                 device_sub_ptr unfiltered_b   (buffer.mem, 1*buffer.pass_stride, buffer.pass_stride);
86                 device_sub_ptr sample_var     (buffer.mem, 2*buffer.pass_stride, buffer.pass_stride);
87                 device_sub_ptr sample_var_var (buffer.mem, 3*buffer.pass_stride, buffer.pass_stride);
88                 device_sub_ptr buffer_var     (buffer.mem, 5*buffer.pass_stride, buffer.pass_stride);
89                 device_sub_ptr filtered_var   (buffer.mem, 6*buffer.pass_stride, buffer.pass_stride);
90                 device_sub_ptr nlm_temporary_1(buffer.mem, 7*buffer.pass_stride, buffer.pass_stride);
91                 device_sub_ptr nlm_temporary_2(buffer.mem, 8*buffer.pass_stride, buffer.pass_stride);
92                 device_sub_ptr nlm_temporary_3(buffer.mem, 9*buffer.pass_stride, buffer.pass_stride);
93
94                 nlm_state.temporary_1_ptr = *nlm_temporary_1;
95                 nlm_state.temporary_2_ptr = *nlm_temporary_2;
96                 nlm_state.temporary_3_ptr = *nlm_temporary_3;
97
98                 /* Get the A/B unfiltered passes, the combined sample variance, the estimated variance of the sample variance and the buffer variance. */
99                 functions.divide_shadow(*unfiltered_a, *unfiltered_b, *sample_var, *sample_var_var, *buffer_var);
100
101                 /* Smooth the (generally pretty noisy) buffer variance using the spatial information from the sample variance. */
102                 nlm_state.set_parameters(6, 3, 4.0f, 1.0f);
103                 functions.non_local_means(*buffer_var, *sample_var, *sample_var_var, *filtered_var);
104
105                 /* Reuse memory, the previous data isn't needed anymore. */
106                 device_ptr filtered_a = *buffer_var,
107                            filtered_b = *sample_var;
108                 /* Use the smoothed variance to filter the two shadow half images using each other for weight calculation. */
109                 nlm_state.set_parameters(5, 3, 1.0f, 0.25f);
110                 functions.non_local_means(*unfiltered_a, *unfiltered_b, *filtered_var, filtered_a);
111                 functions.non_local_means(*unfiltered_b, *unfiltered_a, *filtered_var, filtered_b);
112
113                 device_ptr residual_var = *sample_var_var;
114                 /* Estimate the residual variance between the two filtered halves. */
115                 functions.combine_halves(filtered_a, filtered_b, null_ptr, residual_var, 2, rect);
116
117                 device_ptr final_a = *unfiltered_a,
118                            final_b = *unfiltered_b;
119                 /* Use the residual variance for a second filter pass. */
120                 nlm_state.set_parameters(4, 2, 1.0f, 0.5f);
121                 functions.non_local_means(filtered_a, filtered_b, residual_var, final_a);
122                 functions.non_local_means(filtered_b, filtered_a, residual_var, final_b);
123
124                 /* Combine the two double-filtered halves to a final shadow feature. */
125                 device_sub_ptr shadow_pass(buffer.mem, 4*buffer.pass_stride, buffer.pass_stride);
126                 functions.combine_halves(final_a, final_b, *shadow_pass, null_ptr, 0, rect);
127         }
128
129         /* Prefilter general features. */
130         {
131                 device_sub_ptr unfiltered     (buffer.mem,  8*buffer.pass_stride, buffer.pass_stride);
132                 device_sub_ptr variance       (buffer.mem,  9*buffer.pass_stride, buffer.pass_stride);
133                 device_sub_ptr nlm_temporary_1(buffer.mem, 10*buffer.pass_stride, buffer.pass_stride);
134                 device_sub_ptr nlm_temporary_2(buffer.mem, 11*buffer.pass_stride, buffer.pass_stride);
135                 device_sub_ptr nlm_temporary_3(buffer.mem, 12*buffer.pass_stride, buffer.pass_stride);
136
137                 nlm_state.temporary_1_ptr = *nlm_temporary_1;
138                 nlm_state.temporary_2_ptr = *nlm_temporary_2;
139                 nlm_state.temporary_3_ptr = *nlm_temporary_3;
140
141                 int mean_from[]     = { 0, 1, 2, 12, 6,  7, 8 };
142                 int variance_from[] = { 3, 4, 5, 13, 9, 10, 11};
143                 int pass_to[]       = { 1, 2, 3, 0,  5,  6,  7};
144                 for(int pass = 0; pass < 7; pass++) {
145                         device_sub_ptr feature_pass(buffer.mem, pass_to[pass]*buffer.pass_stride, buffer.pass_stride);
146                         /* Get the unfiltered pass and its variance from the RenderBuffers. */
147                         functions.get_feature(mean_from[pass], variance_from[pass], *unfiltered, *variance);
148                         /* Smooth the pass and store the result in the denoising buffers. */
149                         nlm_state.set_parameters(2, 2, 1.0f, 0.25f);
150                         functions.non_local_means(*unfiltered, *unfiltered, *variance, *feature_pass);
151                 }
152         }
153
154         /* Copy color passes. */
155         {
156                 int mean_from[]     = {20, 21, 22};
157                 int variance_from[] = {23, 24, 25};
158                 int mean_to[]       = { 8,  9, 10};
159                 int variance_to[]   = {11, 12, 13};
160                 int num_color_passes = 3;
161
162                 device_only_memory<float> temp_color(device, "Denoising temporary color");
163                 temp_color.alloc_to_device(3*buffer.pass_stride);
164
165                 for(int pass = 0; pass < num_color_passes; pass++) {
166                         device_sub_ptr color_pass(temp_color, pass*buffer.pass_stride, buffer.pass_stride);
167                         device_sub_ptr color_var_pass(buffer.mem, variance_to[pass]*buffer.pass_stride, buffer.pass_stride);
168                         functions.get_feature(mean_from[pass], variance_from[pass], *color_pass, *color_var_pass);
169                 }
170
171                 {
172                         device_sub_ptr depth_pass    (buffer.mem,                                 0,   buffer.pass_stride);
173                         device_sub_ptr color_var_pass(buffer.mem, variance_to[0]*buffer.pass_stride, 3*buffer.pass_stride);
174                         device_sub_ptr output_pass   (buffer.mem,     mean_to[0]*buffer.pass_stride, 3*buffer.pass_stride);
175                         functions.detect_outliers(temp_color.device_pointer, *color_var_pass, *depth_pass, *output_pass);
176                 }
177
178                 temp_color.free();
179         }
180
181         storage.w = filter_area.z;
182         storage.h = filter_area.w;
183         storage.transform.alloc_to_device(storage.w*storage.h*TRANSFORM_SIZE);
184         storage.rank.alloc_to_device(storage.w*storage.h);
185
186         functions.construct_transform();
187
188         device_only_memory<float> temporary_1(device, "Denoising NLM temporary 1");
189         device_only_memory<float> temporary_2(device, "Denoising NLM temporary 2");
190         temporary_1.alloc_to_device(buffer.w*buffer.h);
191         temporary_2.alloc_to_device(buffer.w*buffer.h);
192         reconstruction_state.temporary_1_ptr = temporary_1.device_pointer;
193         reconstruction_state.temporary_2_ptr = temporary_2.device_pointer;
194
195         storage.XtWX.alloc_to_device(storage.w*storage.h*XTWX_SIZE);
196         storage.XtWY.alloc_to_device(storage.w*storage.h*XTWY_SIZE);
197
198         reconstruction_state.filter_rect = make_int4(filter_area.x-rect.x, filter_area.y-rect.y, storage.w, storage.h);
199         int tile_coordinate_offset = filter_area.y*render_buffer.stride + filter_area.x;
200         reconstruction_state.buffer_params = make_int4(render_buffer.offset + tile_coordinate_offset,
201                                                        render_buffer.stride,
202                                                        render_buffer.pass_stride,
203                                                        render_buffer.denoising_clean_offset);
204         reconstruction_state.source_w = rect.z-rect.x;
205         reconstruction_state.source_h = rect.w-rect.y;
206
207         {
208                 device_sub_ptr color_ptr    (buffer.mem,  8*buffer.pass_stride, 3*buffer.pass_stride);
209                 device_sub_ptr color_var_ptr(buffer.mem, 11*buffer.pass_stride, 3*buffer.pass_stride);
210                 functions.reconstruct(*color_ptr, *color_var_ptr, render_buffer.ptr);
211         }
212
213         storage.XtWX.free();
214         storage.XtWY.free();
215         storage.transform.free();
216         storage.rank.free();
217         temporary_1.free();
218         temporary_2.free();
219         buffer.mem.free();
220         tiles_mem.free();
221         return true;
222 }
223
224 CCL_NAMESPACE_END