Merge branch 'master' into blender2.8
[blender.git] / intern / cycles / render / buffers.cpp
1 /*
2  * Copyright 2011-2013 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 <stdlib.h>
18
19 #include "render/buffers.h"
20 #include "device/device.h"
21
22 #include "util/util_debug.h"
23 #include "util/util_foreach.h"
24 #include "util/util_hash.h"
25 #include "util/util_image.h"
26 #include "util/util_math.h"
27 #include "util/util_opengl.h"
28 #include "util/util_time.h"
29 #include "util/util_types.h"
30
31 CCL_NAMESPACE_BEGIN
32
33 /* Buffer Params */
34
35 BufferParams::BufferParams()
36 {
37         width = 0;
38         height = 0;
39
40         full_x = 0;
41         full_y = 0;
42         full_width = 0;
43         full_height = 0;
44
45         denoising_data_pass = false;
46         denoising_clean_pass = false;
47
48         Pass::add(PASS_COMBINED, passes);
49 }
50
51 void BufferParams::get_offset_stride(int& offset, int& stride)
52 {
53         offset = -(full_x + full_y*width);
54         stride = width;
55 }
56
57 bool BufferParams::modified(const BufferParams& params)
58 {
59         return !(full_x == params.full_x
60                 && full_y == params.full_y
61                 && width == params.width
62                 && height == params.height
63                 && full_width == params.full_width
64                 && full_height == params.full_height
65                 && Pass::equals(passes, params.passes));
66 }
67
68 int BufferParams::get_passes_size()
69 {
70         int size = 0;
71
72         for(size_t i = 0; i < passes.size(); i++)
73                 size += passes[i].components;
74
75         if(denoising_data_pass) {
76                 size += DENOISING_PASS_SIZE_BASE;
77                 if(denoising_clean_pass) size += DENOISING_PASS_SIZE_CLEAN;
78         }
79
80         return align_up(size, 4);
81 }
82
83 int BufferParams::get_denoising_offset()
84 {
85         int offset = 0;
86
87         for(size_t i = 0; i < passes.size(); i++)
88                 offset += passes[i].components;
89
90         return offset;
91 }
92
93 /* Render Buffer Task */
94
95 RenderTile::RenderTile()
96 {
97         x = 0;
98         y = 0;
99         w = 0;
100         h = 0;
101
102         sample = 0;
103         start_sample = 0;
104         num_samples = 0;
105         resolution = 0;
106
107         offset = 0;
108         stride = 0;
109
110         buffer = 0;
111
112         buffers = NULL;
113 }
114
115 /* Render Buffers */
116
117 RenderBuffers::RenderBuffers(Device *device)
118 : buffer(device, "RenderBuffers", MEM_READ_WRITE),
119   map_neighbor_copied(false)
120 {
121 }
122
123 RenderBuffers::~RenderBuffers()
124 {
125         buffer.free();
126 }
127
128 void RenderBuffers::reset(BufferParams& params_)
129 {
130         params = params_;
131
132         /* re-allocate buffer */
133         buffer.alloc(params.width*params.height*params.get_passes_size());
134         buffer.zero_to_device();
135 }
136
137 void RenderBuffers::zero()
138 {
139         buffer.zero_to_device();
140 }
141
142 bool RenderBuffers::copy_from_device()
143 {
144         if(!buffer.device_pointer)
145                 return false;
146
147         buffer.copy_from_device(0, params.width * params.get_passes_size(), params.height);
148
149         return true;
150 }
151
152 bool RenderBuffers::get_denoising_pass_rect(int offset, float exposure, int sample, int components, float *pixels)
153 {
154         float invsample = 1.0f/sample;
155         float scale = invsample;
156         bool variance = (offset == DENOISING_PASS_NORMAL_VAR) ||
157                         (offset == DENOISING_PASS_ALBEDO_VAR) ||
158                         (offset == DENOISING_PASS_DEPTH_VAR) ||
159                         (offset == DENOISING_PASS_COLOR_VAR);
160
161         if(offset == DENOISING_PASS_COLOR) {
162                 scale *= exposure;
163         }
164         else if(offset == DENOISING_PASS_COLOR_VAR) {
165                 scale *= exposure*exposure;
166         }
167
168         offset += params.get_denoising_offset();
169         int pass_stride = params.get_passes_size();
170         int size = params.width*params.height;
171
172         if(variance) {
173                 /* Approximate variance as E[x^2] - 1/N * (E[x])^2, since online variance
174                  * update does not work efficiently with atomics in the kernel. */
175                 int mean_offset = offset - components;
176                 float *mean = buffer.data() + mean_offset;
177                 float *var = buffer.data() + offset;
178                 assert(mean_offset >= 0);
179
180                 if(components == 1) {
181                         for(int i = 0; i < size; i++, mean += pass_stride, var += pass_stride, pixels++) {
182                                 pixels[0] = max(0.0f, var[0] - mean[0]*mean[0]*invsample)*scale;
183                         }
184                 }
185                 else if(components == 3) {
186                         for(int i = 0; i < size; i++, mean += pass_stride, var += pass_stride, pixels += 3) {
187                                 pixels[0] = max(0.0f, var[0] - mean[0]*mean[0]*invsample)*scale;
188                                 pixels[1] = max(0.0f, var[1] - mean[1]*mean[1]*invsample)*scale;
189                                 pixels[2] = max(0.0f, var[2] - mean[2]*mean[2]*invsample)*scale;
190                         }
191                 }
192                 else {
193                         return false;
194                 }
195         }
196         else {
197                 float *in = buffer.data() + offset;
198
199                 if(components == 1) {
200                         for(int i = 0; i < size; i++, in += pass_stride, pixels++) {
201                                 pixels[0] = in[0]*scale;
202                         }
203                 }
204                 else if(components == 3) {
205                         for(int i = 0; i < size; i++, in += pass_stride, pixels += 3) {
206                                 pixels[0] = in[0]*scale;
207                                 pixels[1] = in[1]*scale;
208                                 pixels[2] = in[2]*scale;
209                         }
210                 }
211                 else {
212                         return false;
213                 }
214         }
215
216         return true;
217 }
218
219 bool RenderBuffers::get_pass_rect(PassType type, float exposure, int sample, int components, float *pixels)
220 {
221         int pass_offset = 0;
222
223         for(size_t j = 0; j < params.passes.size(); j++) {
224                 Pass& pass = params.passes[j];
225
226                 if(pass.type != type) {
227                         pass_offset += pass.components;
228                         continue;
229                 }
230
231                 float *in = buffer.data() + pass_offset;
232                 int pass_stride = params.get_passes_size();
233
234                 float scale = (pass.filter)? 1.0f/(float)sample: 1.0f;
235                 float scale_exposure = (pass.exposure)? scale*exposure: scale;
236
237                 int size = params.width*params.height;
238
239                 if(components == 1) {
240                         assert(pass.components == components);
241
242                         /* scalar */
243                         if(type == PASS_DEPTH) {
244                                 for(int i = 0; i < size; i++, in += pass_stride, pixels++) {
245                                         float f = *in;
246                                         pixels[0] = (f == 0.0f)? 1e10f: f*scale_exposure;
247                                 }
248                         }
249                         else if(type == PASS_MIST) {
250                                 for(int i = 0; i < size; i++, in += pass_stride, pixels++) {
251                                         float f = *in;
252                                         pixels[0] = saturate(f*scale_exposure);
253                                 }
254                         }
255 #ifdef WITH_CYCLES_DEBUG
256                         else if(type == PASS_BVH_TRAVERSED_NODES ||
257                                 type == PASS_BVH_TRAVERSED_INSTANCES ||
258                                 type == PASS_BVH_INTERSECTIONS ||
259                                 type == PASS_RAY_BOUNCES)
260                         {
261                                 for(int i = 0; i < size; i++, in += pass_stride, pixels++) {
262                                         float f = *in;
263                                         pixels[0] = f*scale;
264                                 }
265                         }
266 #endif
267                         else {
268                                 for(int i = 0; i < size; i++, in += pass_stride, pixels++) {
269                                         float f = *in;
270                                         pixels[0] = f*scale_exposure;
271                                 }
272                         }
273                 }
274                 else if(components == 3) {
275                         assert(pass.components == 4);
276
277                         /* RGBA */
278                         if(type == PASS_SHADOW) {
279                                 for(int i = 0; i < size; i++, in += pass_stride, pixels += 3) {
280                                         float4 f = make_float4(in[0], in[1], in[2], in[3]);
281                                         float invw = (f.w > 0.0f)? 1.0f/f.w: 1.0f;
282
283                                         pixels[0] = f.x*invw;
284                                         pixels[1] = f.y*invw;
285                                         pixels[2] = f.z*invw;
286                                 }
287                         }
288                         else if(pass.divide_type != PASS_NONE) {
289                                 /* RGB lighting passes that need to divide out color */
290                                 pass_offset = 0;
291                                 for(size_t k = 0; k < params.passes.size(); k++) {
292                                         Pass& color_pass = params.passes[k];
293                                         if(color_pass.type == pass.divide_type)
294                                                 break;
295                                         pass_offset += color_pass.components;
296                                 }
297
298                                 float *in_divide = buffer.data() + pass_offset;
299
300                                 for(int i = 0; i < size; i++, in += pass_stride, in_divide += pass_stride, pixels += 3) {
301                                         float3 f = make_float3(in[0], in[1], in[2]);
302                                         float3 f_divide = make_float3(in_divide[0], in_divide[1], in_divide[2]);
303
304                                         f = safe_divide_even_color(f*exposure, f_divide);
305
306                                         pixels[0] = f.x;
307                                         pixels[1] = f.y;
308                                         pixels[2] = f.z;
309                                 }
310                         }
311                         else {
312                                 /* RGB/vector */
313                                 for(int i = 0; i < size; i++, in += pass_stride, pixels += 3) {
314                                         float3 f = make_float3(in[0], in[1], in[2]);
315
316                                         pixels[0] = f.x*scale_exposure;
317                                         pixels[1] = f.y*scale_exposure;
318                                         pixels[2] = f.z*scale_exposure;
319                                 }
320                         }
321                 }
322                 else if(components == 4) {
323                         assert(pass.components == components);
324
325                         /* RGBA */
326                         if(type == PASS_SHADOW) {
327                                 for(int i = 0; i < size; i++, in += pass_stride, pixels += 4) {
328                                         float4 f = make_float4(in[0], in[1], in[2], in[3]);
329                                         float invw = (f.w > 0.0f)? 1.0f/f.w: 1.0f;
330
331                                         pixels[0] = f.x*invw;
332                                         pixels[1] = f.y*invw;
333                                         pixels[2] = f.z*invw;
334                                         pixels[3] = 1.0f;
335                                 }
336                         }
337                         else if(type == PASS_MOTION) {
338                                 /* need to normalize by number of samples accumulated for motion */
339                                 pass_offset = 0;
340                                 for(size_t k = 0; k < params.passes.size(); k++) {
341                                         Pass& color_pass = params.passes[k];
342                                         if(color_pass.type == PASS_MOTION_WEIGHT)
343                                                 break;
344                                         pass_offset += color_pass.components;
345                                 }
346
347                                 float *in_weight = buffer.data() + pass_offset;
348
349                                 for(int i = 0; i < size; i++, in += pass_stride, in_weight += pass_stride, pixels += 4) {
350                                         float4 f = make_float4(in[0], in[1], in[2], in[3]);
351                                         float w = in_weight[0];
352                                         float invw = (w > 0.0f)? 1.0f/w: 0.0f;
353
354                                         pixels[0] = f.x*invw;
355                                         pixels[1] = f.y*invw;
356                                         pixels[2] = f.z*invw;
357                                         pixels[3] = f.w*invw;
358                                 }
359                         }
360                         else {
361                                 for(int i = 0; i < size; i++, in += pass_stride, pixels += 4) {
362                                         float4 f = make_float4(in[0], in[1], in[2], in[3]);
363
364                                         pixels[0] = f.x*scale_exposure;
365                                         pixels[1] = f.y*scale_exposure;
366                                         pixels[2] = f.z*scale_exposure;
367
368                                         /* clamp since alpha might be > 1.0 due to russian roulette */
369                                         pixels[3] = saturate(f.w*scale);
370                                 }
371                         }
372                 }
373
374                 return true;
375         }
376
377         return false;
378 }
379
380 /* Display Buffer */
381
382 DisplayBuffer::DisplayBuffer(Device *device, bool linear)
383 : draw_width(0),
384   draw_height(0),
385   transparent(true), /* todo: determine from background */
386   half_float(linear),
387   rgba_byte(device, "display buffer byte"),
388   rgba_half(device, "display buffer half")
389 {
390 }
391
392 DisplayBuffer::~DisplayBuffer()
393 {
394         rgba_byte.free();
395         rgba_half.free();
396 }
397
398 void DisplayBuffer::reset(BufferParams& params_)
399 {
400         draw_width = 0;
401         draw_height = 0;
402
403         params = params_;
404
405         /* allocate display pixels */
406         if(half_float) {
407                 rgba_half.alloc_to_device(params.width, params.height);
408         }
409         else {
410                 rgba_byte.alloc_to_device(params.width, params.height);
411         }
412 }
413
414 void DisplayBuffer::draw_set(int width, int height)
415 {
416         assert(width <= params.width && height <= params.height);
417
418         draw_width = width;
419         draw_height = height;
420 }
421
422 void DisplayBuffer::draw(Device *device, const DeviceDrawParams& draw_params)
423 {
424         if(draw_width != 0 && draw_height != 0) {
425                 device_memory& rgba = (half_float)? (device_memory&)rgba_half:
426                                                     (device_memory&)rgba_byte;
427
428                 device->draw_pixels(
429                             rgba, 0,
430                             draw_width, draw_height, params.width, params.height,
431                             params.full_x, params.full_y, params.full_width, params.full_height,
432                             transparent, draw_params);
433         }
434 }
435
436 bool DisplayBuffer::draw_ready()
437 {
438         return (draw_width != 0 && draw_height != 0);
439 }
440
441 void DisplayBuffer::write(const string& filename)
442 {
443         int w = draw_width;
444         int h = draw_height;
445
446         if(w == 0 || h == 0)
447                 return;
448         
449         if(half_float)
450                 return;
451
452         /* read buffer from device */
453         uchar4 *pixels = rgba_byte.copy_from_device(0, w, h);
454
455         /* write image */
456         ImageOutput *out = ImageOutput::create(filename);
457         ImageSpec spec(w, h, 4, TypeDesc::UINT8);
458
459         out->open(filename, spec);
460
461         /* conversion for different top/bottom convention */
462         out->write_image(TypeDesc::UINT8,
463                 (uchar*)(pixels + (h-1)*w),
464                 AutoStride,
465                 -w*sizeof(uchar4),
466                 AutoStride);
467
468         out->close();
469
470         delete out;
471 }
472
473 CCL_NAMESPACE_END
474