1 // Three-lobed Lanczos, the most common choice.
2 #define LANCZOS_RADIUS 3.0
11 #include "effect_chain.h"
12 #include "effect_util.h"
14 #include "resample_effect.h"
26 return 1.0f - fabs(x);
32 float lanczos_weight(float x, float a)
37 return sinc(M_PI * x) * sinc(M_PI * x / a);
41 // Euclid's algorithm, from Wikipedia.
42 unsigned gcd(unsigned a, unsigned b)
52 unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsigned max_samples_saved)
54 unsigned num_samples_saved = 0;
55 for (unsigned i = 0, j = 0; i < num_src_samples; ++i, ++j) {
56 // Copy the sample directly; it will be overwritten later if we can combine.
58 dst[j * 2 + 0] = src[i * 2 + 0];
59 dst[j * 2 + 1] = src[i * 2 + 1];
62 if (i == num_src_samples - 1) {
63 // Last sample; cannot combine.
66 assert(num_samples_saved <= max_samples_saved);
67 if (num_samples_saved == max_samples_saved) {
68 // We could maybe save more here, but other rows can't, so don't bother.
72 float w1 = src[i * 2 + 0];
73 float w2 = src[(i + 1) * 2 + 0];
75 // Differing signs; cannot combine.
79 float pos1 = src[i * 2 + 1];
80 float pos2 = src[(i + 1) * 2 + 1];
83 float offset, total_weight, sum_sq_error;
84 combine_two_samples(w1, w2, &offset, &total_weight, &sum_sq_error);
86 // If the interpolation error is larger than that of about sqrt(2) of
87 // a level at 8-bit precision, don't combine. (You'd think 1.0 was enough,
88 // but since the artifacts are not really random, they can get quite
89 // visible. On the other hand, going to 0.25f, I can see no change at
90 // all with 8-bit output, so it would not seem to be worth it.)
91 if (sum_sq_error > 0.5f / (256.0f * 256.0f)) {
95 // OK, we can combine this and the next sample.
97 dst[j * 2 + 0] = total_weight;
98 dst[j * 2 + 1] = pos1 + offset * (pos2 - pos1);
101 ++i; // Skip the next sample.
104 return num_samples_saved;
109 ResampleEffect::ResampleEffect()
112 offset_x(0.0f), offset_y(0.0f),
113 zoom_x(1.0f), zoom_y(1.0f),
114 zoom_center_x(0.5f), zoom_center_y(0.5f)
116 register_int("width", &output_width);
117 register_int("height", &output_height);
119 // The first blur pass will forward resolution information to us.
120 hpass = new SingleResamplePassEffect(this);
121 CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL));
122 vpass = new SingleResamplePassEffect(NULL);
123 CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL));
128 void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self)
130 Node *hpass_node = graph->add_node(hpass);
131 Node *vpass_node = graph->add_node(vpass);
132 graph->connect_nodes(hpass_node, vpass_node);
133 graph->replace_receiver(self, hpass_node);
134 graph->replace_sender(self, vpass_node);
135 self->disabled = true;
138 // We get this information forwarded from the first blur pass,
139 // since we are not part of the chain ourselves.
140 void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsigned height)
142 assert(input_num == 0);
146 input_height = height;
150 void ResampleEffect::update_size()
153 ok |= hpass->set_int("input_width", input_width);
154 ok |= hpass->set_int("input_height", input_height);
155 ok |= hpass->set_int("output_width", output_width);
156 ok |= hpass->set_int("output_height", input_height);
158 ok |= vpass->set_int("input_width", output_width);
159 ok |= vpass->set_int("input_height", input_height);
160 ok |= vpass->set_int("output_width", output_width);
161 ok |= vpass->set_int("output_height", output_height);
165 // The offset added due to zoom may have changed with the size.
166 update_offset_and_zoom();
169 void ResampleEffect::update_offset_and_zoom()
173 // Zoom from the right origin. (zoom_center is given in normalized coordinates,
175 float extra_offset_x = zoom_center_x * (1.0f - 1.0f / zoom_x) * input_width;
176 float extra_offset_y = (1.0f - zoom_center_y) * (1.0f - 1.0f / zoom_y) * input_height;
178 ok |= hpass->set_float("offset", extra_offset_x + offset_x);
179 ok |= vpass->set_float("offset", extra_offset_y - offset_y); // Compensate for the bottom-left origin.
180 ok |= hpass->set_float("zoom", zoom_x);
181 ok |= vpass->set_float("zoom", zoom_y);
186 bool ResampleEffect::set_float(const string &key, float value) {
187 if (key == "width") {
188 output_width = value;
192 if (key == "height") {
193 output_height = value;
199 update_offset_and_zoom();
204 update_offset_and_zoom();
207 if (key == "zoom_x") {
212 update_offset_and_zoom();
215 if (key == "zoom_y") {
220 update_offset_and_zoom();
223 if (key == "zoom_center_x") {
224 zoom_center_x = value;
225 update_offset_and_zoom();
228 if (key == "zoom_center_y") {
229 zoom_center_y = value;
230 update_offset_and_zoom();
236 SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
238 direction(HORIZONTAL),
243 last_input_width(-1),
244 last_input_height(-1),
245 last_output_width(-1),
246 last_output_height(-1),
247 last_offset(0.0 / 0.0), // NaN.
248 last_zoom(0.0 / 0.0) // NaN.
250 register_int("direction", (int *)&direction);
251 register_int("input_width", &input_width);
252 register_int("input_height", &input_height);
253 register_int("output_width", &output_width);
254 register_int("output_height", &output_height);
255 register_float("offset", &offset);
256 register_float("zoom", &zoom);
258 glGenTextures(1, &texnum);
261 SingleResamplePassEffect::~SingleResamplePassEffect()
263 glDeleteTextures(1, &texnum);
266 string SingleResamplePassEffect::output_fragment_shader()
269 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
270 return buf + read_file("resample_effect.frag");
273 // Using vertical scaling as an example:
275 // Generally out[y] = w0 * in[yi] + w1 * in[yi + 1] + w2 * in[yi + 2] + ...
277 // Obviously, yi will depend on y (in a not-quite-linear way), but so will
278 // the weights w0, w1, w2, etc.. The easiest way of doing this is to encode,
279 // for each sample, the weight and the yi value, e.g. <yi, w0>, <yi + 1, w1>,
280 // and so on. For each y, we encode these along the x-axis (since that is spare),
281 // so out[0] will read from parameters <x,y> = <0,0>, <1,0>, <2,0> and so on.
283 // For horizontal scaling, we fill in the exact same texture;
284 // the shader just interprets it differently.
285 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
287 unsigned src_size, dst_size;
288 if (direction == SingleResamplePassEffect::HORIZONTAL) {
289 assert(input_height == output_height);
290 src_size = input_width;
291 dst_size = output_width;
292 } else if (direction == SingleResamplePassEffect::VERTICAL) {
293 assert(input_width == output_width);
294 src_size = input_height;
295 dst_size = output_height;
300 // For many resamplings (e.g. 640 -> 1280), we will end up with the same
301 // set of samples over and over again in a loop. Thus, we can compute only
302 // the first such loop, and then ask the card to repeat the texture for us.
303 // This is both easier on the texture cache and lowers our CPU cost for
304 // generating the kernel somewhat.
305 float scaling_factor;
306 if (fabs(zoom - 1.0f) < 1e-6) {
307 num_loops = gcd(src_size, dst_size);
308 scaling_factor = float(dst_size) / float(src_size);
310 // If zooming is enabled (ie., zoom != 1), we turn off the looping.
311 // We _could_ perhaps do it for rational zoom levels (especially
312 // things like 2:1), but it doesn't seem to be worth it, given that
313 // the most common use case would seem to be varying the zoom
314 // from frame to frame.
316 scaling_factor = zoom * float(dst_size) / float(src_size);
318 slice_height = 1.0f / num_loops;
319 unsigned dst_samples = dst_size / num_loops;
321 // Sample the kernel in the right place. A diagram with a triangular kernel
322 // (corresponding to linear filtering, and obviously with radius 1)
323 // for easier ASCII art drawing:
329 // x---x---x x x---x---x---x
331 // Scaling up (in this case, 2x) means sampling more densely:
337 // x-x-x-x-x-x x x x-x-x-x-x-x-x-x
339 // When scaling up, any destination pixel will only be influenced by a few
340 // (in this case, two) neighboring pixels, and more importantly, the number
341 // will not be influenced by the scaling factor. (Note, however, that the
342 // pixel centers have moved, due to OpenGL's center-pixel convention.)
343 // The only thing that changes is the weights themselves, as the sampling
344 // points are at different distances from the original pixels.
346 // Scaling down is a different story:
352 // --x------ x --x-------x--
354 // Again, the pixel centers have moved in a maybe unintuitive fashion,
355 // although when you consider that there are multiple source pixels around,
356 // it's not so bad as at first look:
362 // --x-------x-------x-------x--
364 // As you can see, the new pixels become averages of the two neighboring old
365 // ones (the situation for Lanczos is of course more complex).
367 // Anyhow, in this case we clearly need to look at more source pixels
368 // to compute the destination pixel, and how many depend on the scaling factor.
369 // Thus, the kernel width will vary with how much we scale.
370 float radius_scaling_factor = min(scaling_factor, 1.0f);
371 int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
372 int src_samples = int_radius * 2 + 1;
373 float *weights = new float[dst_samples * src_samples * 2];
374 float subpixel_offset = offset - lrintf(offset); // The part not covered by whole_pixel_offset.
375 assert(subpixel_offset >= -0.5f && subpixel_offset <= 0.5f);
376 for (unsigned y = 0; y < dst_samples; ++y) {
377 // Find the point around which we want to sample the source image,
378 // compensating for differing pixel centers as the scale changes.
379 float center_src_y = (y + 0.5f) / scaling_factor - 0.5f;
380 int base_src_y = lrintf(center_src_y);
382 // Now sample <int_radius> pixels on each side around that point.
383 for (int i = 0; i < src_samples; ++i) {
384 int src_y = base_src_y + i - int_radius;
385 float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y - subpixel_offset), LANCZOS_RADIUS);
386 weights[(y * src_samples + i) * 2 + 0] = weight * radius_scaling_factor;
387 weights[(y * src_samples + i) * 2 + 1] = (src_y + 0.5) / float(src_size);
391 // Now make use of the bilinear filtering in the GPU to reduce the number of samples
392 // we need to make. This is a bit more complex than BlurEffect since we cannot combine
393 // two neighboring samples if their weights have differing signs, so we first need to
394 // figure out the maximum number of samples. Then, we downconvert all the weights to
395 // that number -- we could have gone for a variable-length system, but this is simpler,
396 // and the gains would probably be offset by the extra cost of checking when to stop.
398 // The greedy strategy for combining samples is optimal.
399 src_bilinear_samples = 0;
400 for (unsigned y = 0; y < dst_samples; ++y) {
401 unsigned num_samples_saved = combine_samples(weights + (y * src_samples) * 2, NULL, src_samples, UINT_MAX);
402 src_bilinear_samples = max<int>(src_bilinear_samples, src_samples - num_samples_saved);
405 // Now that we know the right width, actually combine the samples.
406 float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2];
407 fp16_int_t *bilinear_weights_fp16 = new fp16_int_t[dst_samples * src_bilinear_samples * 2];
408 for (unsigned y = 0; y < dst_samples; ++y) {
409 float *bilinear_weights_ptr = bilinear_weights + (y * src_bilinear_samples) * 2;
410 fp16_int_t *bilinear_weights_fp16_ptr = bilinear_weights_fp16 + (y * src_bilinear_samples) * 2;
411 unsigned num_samples_saved = combine_samples(
412 weights + (y * src_samples) * 2,
413 bilinear_weights_ptr,
415 src_samples - src_bilinear_samples);
416 assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
419 for (int i = 0; i < src_bilinear_samples; ++i) {
420 bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 0]);
421 bilinear_weights_fp16_ptr[i * 2 + 1] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 1]);
424 // Normalize so that the sum becomes one. Note that we do it twice;
425 // this sometimes helps a tiny little bit when we have many samples.
426 for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
428 for (int i = 0; i < src_bilinear_samples; ++i) {
429 sum += fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]);
431 for (int i = 0; i < src_bilinear_samples; ++i) {
432 bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16(
433 fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]) / sum);
438 // Encode as a two-component texture. Note the GL_REPEAT.
439 glActiveTexture(GL_TEXTURE0 + *sampler_num);
441 glBindTexture(GL_TEXTURE_2D, texnum);
443 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
445 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
447 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
449 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_HALF_FLOAT, bilinear_weights_fp16);
453 delete[] bilinear_weights;
454 delete[] bilinear_weights_fp16;
457 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
459 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
461 assert(input_width > 0);
462 assert(input_height > 0);
463 assert(output_width > 0);
464 assert(output_height > 0);
466 if (input_width != last_input_width ||
467 input_height != last_input_height ||
468 output_width != last_output_width ||
469 output_height != last_output_height ||
470 offset != last_offset ||
472 update_texture(glsl_program_num, prefix, sampler_num);
473 last_input_width = input_width;
474 last_input_height = input_height;
475 last_output_width = output_width;
476 last_output_height = output_height;
477 last_offset = offset;
481 glActiveTexture(GL_TEXTURE0 + *sampler_num);
483 glBindTexture(GL_TEXTURE_2D, texnum);
486 set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num);
488 set_uniform_int(glsl_program_num, prefix, "num_samples", src_bilinear_samples);
489 set_uniform_float(glsl_program_num, prefix, "num_loops", num_loops);
490 set_uniform_float(glsl_program_num, prefix, "slice_height", slice_height);
492 // Instructions for how to convert integer sample numbers to positions in the weight texture.
493 set_uniform_float(glsl_program_num, prefix, "sample_x_scale", 1.0f / src_bilinear_samples);
494 set_uniform_float(glsl_program_num, prefix, "sample_x_offset", 0.5f / src_bilinear_samples);
496 float whole_pixel_offset;
497 if (direction == SingleResamplePassEffect::VERTICAL) {
498 whole_pixel_offset = lrintf(offset) / float(input_height);
500 whole_pixel_offset = lrintf(offset) / float(input_width);
502 set_uniform_float(glsl_program_num, prefix, "whole_pixel_offset", whole_pixel_offset);
504 // We specifically do not want mipmaps on the input texture;
505 // they break minification.
506 Node *self = chain->find_node_for_effect(this);
507 glActiveTexture(chain->get_input_sampler(self, 0));
509 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);