1 // Three-lobed Lanczos, the most common choice.
2 #define LANCZOS_RADIUS 3.0
11 #include "effect_chain.h"
12 #include "effect_util.h"
13 #include "resample_effect.h"
25 return 1.0f - fabs(x);
31 float lanczos_weight(float x, float a)
36 return sinc(M_PI * x) * sinc(M_PI * x / a);
40 // Euclid's algorithm, from Wikipedia.
41 unsigned gcd(unsigned a, unsigned b)
51 unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsigned max_samples_saved)
53 unsigned num_samples_saved = 0;
54 for (unsigned i = 0, j = 0; i < num_src_samples; ++i, ++j) {
55 // Copy the sample directly; it will be overwritten later if we can combine.
57 dst[j * 2 + 0] = src[i * 2 + 0];
58 dst[j * 2 + 1] = src[i * 2 + 1];
61 if (i == num_src_samples - 1) {
62 // Last sample; cannot combine.
65 assert(num_samples_saved <= max_samples_saved);
66 if (num_samples_saved == max_samples_saved) {
67 // We could maybe save more here, but other rows can't, so don't bother.
71 float w1 = src[i * 2 + 0];
72 float w2 = src[(i + 1) * 2 + 0];
74 // Differing signs; cannot combine.
78 float pos1 = src[i * 2 + 1];
79 float pos2 = src[(i + 1) * 2 + 1];
82 float offset, total_weight, sum_sq_error;
83 combine_two_samples(w1, w2, &offset, &total_weight, &sum_sq_error);
85 // If the interpolation error is larger than that of about sqrt(2) of
86 // a level at 8-bit precision, don't combine. (You'd think 1.0 was enough,
87 // but since the artifacts are not really random, they can get quite
88 // visible. On the other hand, going to 0.25f, I can see no change at
89 // all with 8-bit output, so it would not seem to be worth it.)
90 if (sum_sq_error > 0.5f / (256.0f * 256.0f)) {
94 // OK, we can combine this and the next sample.
96 dst[j * 2 + 0] = total_weight;
97 dst[j * 2 + 1] = pos1 + offset * (pos2 - pos1);
100 ++i; // Skip the next sample.
103 return num_samples_saved;
108 ResampleEffect::ResampleEffect()
112 register_int("width", &output_width);
113 register_int("height", &output_height);
115 // The first blur pass will forward resolution information to us.
116 hpass = new SingleResamplePassEffect(this);
117 CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL));
118 vpass = new SingleResamplePassEffect(NULL);
119 CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL));
124 void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self)
126 Node *hpass_node = graph->add_node(hpass);
127 Node *vpass_node = graph->add_node(vpass);
128 graph->connect_nodes(hpass_node, vpass_node);
129 graph->replace_receiver(self, hpass_node);
130 graph->replace_sender(self, vpass_node);
131 self->disabled = true;
134 // We get this information forwarded from the first blur pass,
135 // since we are not part of the chain ourselves.
136 void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsigned height)
138 assert(input_num == 0);
142 input_height = height;
146 void ResampleEffect::update_size()
149 ok |= hpass->set_int("input_width", input_width);
150 ok |= hpass->set_int("input_height", input_height);
151 ok |= hpass->set_int("output_width", output_width);
152 ok |= hpass->set_int("output_height", input_height);
154 ok |= vpass->set_int("input_width", output_width);
155 ok |= vpass->set_int("input_height", input_height);
156 ok |= vpass->set_int("output_width", output_width);
157 ok |= vpass->set_int("output_height", output_height);
162 bool ResampleEffect::set_float(const string &key, float value) {
163 if (key == "width") {
164 output_width = value;
168 if (key == "height") {
169 output_height = value;
176 SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
178 direction(HORIZONTAL),
181 last_input_width(-1),
182 last_input_height(-1),
183 last_output_width(-1),
184 last_output_height(-1)
186 register_int("direction", (int *)&direction);
187 register_int("input_width", &input_width);
188 register_int("input_height", &input_height);
189 register_int("output_width", &output_width);
190 register_int("output_height", &output_height);
192 glGenTextures(1, &texnum);
195 SingleResamplePassEffect::~SingleResamplePassEffect()
197 glDeleteTextures(1, &texnum);
200 string SingleResamplePassEffect::output_fragment_shader()
203 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
204 return buf + read_file("resample_effect.frag");
207 // Using vertical scaling as an example:
209 // Generally out[y] = w0 * in[yi] + w1 * in[yi + 1] + w2 * in[yi + 2] + ...
211 // Obviously, yi will depend on y (in a not-quite-linear way), but so will
212 // the weights w0, w1, w2, etc.. The easiest way of doing this is to encode,
213 // for each sample, the weight and the yi value, e.g. <yi, w0>, <yi + 1, w1>,
214 // and so on. For each y, we encode these along the x-axis (since that is spare),
215 // so out[0] will read from parameters <x,y> = <0,0>, <1,0>, <2,0> and so on.
217 // For horizontal scaling, we fill in the exact same texture;
218 // the shader just interprets it differently.
219 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
221 unsigned src_size, dst_size;
222 if (direction == SingleResamplePassEffect::HORIZONTAL) {
223 assert(input_height == output_height);
224 src_size = input_width;
225 dst_size = output_width;
226 } else if (direction == SingleResamplePassEffect::VERTICAL) {
227 assert(input_width == output_width);
228 src_size = input_height;
229 dst_size = output_height;
234 // For many resamplings (e.g. 640 -> 1280), we will end up with the same
235 // set of samples over and over again in a loop. Thus, we can compute only
236 // the first such loop, and then ask the card to repeat the texture for us.
237 // This is both easier on the texture cache and lowers our CPU cost for
238 // generating the kernel somewhat.
239 num_loops = gcd(src_size, dst_size);
240 slice_height = 1.0f / num_loops;
241 unsigned dst_samples = dst_size / num_loops;
243 // Sample the kernel in the right place. A diagram with a triangular kernel
244 // (corresponding to linear filtering, and obviously with radius 1)
245 // for easier ASCII art drawing:
251 // x---x---x x x---x---x---x
253 // Scaling up (in this case, 2x) means sampling more densely:
259 // x-x-x-x-x-x x x x-x-x-x-x-x-x-x
261 // When scaling up, any destination pixel will only be influenced by a few
262 // (in this case, two) neighboring pixels, and more importantly, the number
263 // will not be influenced by the scaling factor. (Note, however, that the
264 // pixel centers have moved, due to OpenGL's center-pixel convention.)
265 // The only thing that changes is the weights themselves, as the sampling
266 // points are at different distances from the original pixels.
268 // Scaling down is a different story:
274 // --x------ x --x-------x--
276 // Again, the pixel centers have moved in a maybe unintuitive fashion,
277 // although when you consider that there are multiple source pixels around,
278 // it's not so bad as at first look:
284 // --x-------x-------x-------x--
286 // As you can see, the new pixels become averages of the two neighboring old
287 // ones (the situation for Lanczos is of course more complex).
289 // Anyhow, in this case we clearly need to look at more source pixels
290 // to compute the destination pixel, and how many depend on the scaling factor.
291 // Thus, the kernel width will vary with how much we scale.
292 float radius_scaling_factor = min(float(dst_size) / float(src_size), 1.0f);
293 int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
294 int src_samples = int_radius * 2 + 1;
295 float *weights = new float[dst_samples * src_samples * 2];
296 for (unsigned y = 0; y < dst_samples; ++y) {
297 // Find the point around which we want to sample the source image,
298 // compensating for differing pixel centers as the scale changes.
299 float center_src_y = (y + 0.5f) * float(src_size) / float(dst_size) - 0.5f;
300 int base_src_y = lrintf(center_src_y);
302 // Now sample <int_radius> pixels on each side around that point.
303 for (int i = 0; i < src_samples; ++i) {
304 int src_y = base_src_y + i - int_radius;
305 float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y), LANCZOS_RADIUS);
306 weights[(y * src_samples + i) * 2 + 0] = weight * radius_scaling_factor;
307 weights[(y * src_samples + i) * 2 + 1] = (src_y + 0.5) / float(src_size);
312 // Now make use of the bilinear filtering in the GPU to reduce the number of samples
313 // we need to make. This is a bit more complex than BlurEffect since we cannot combine
314 // two neighboring samples if their weights have differing signs, so we first need to
315 // figure out the maximum number of samples. Then, we downconvert all the weights to
316 // that number -- we could have gone for a variable-length system, but this is simpler,
317 // and the gains would probably be offset by the extra cost of checking when to stop.
319 // The greedy strategy for combining samples is optimal.
320 src_bilinear_samples = 0;
321 for (unsigned y = 0; y < dst_samples; ++y) {
322 unsigned num_samples_saved = combine_samples(weights + (y * src_samples) * 2, NULL, src_samples, UINT_MAX);
323 src_bilinear_samples = max<int>(src_bilinear_samples, src_samples - num_samples_saved);
326 // Now that we know the right width, actually combine the samples.
327 float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2];
328 for (unsigned y = 0; y < dst_samples; ++y) {
329 unsigned num_samples_saved = combine_samples(
330 weights + (y * src_samples) * 2,
331 bilinear_weights + (y * src_bilinear_samples) * 2,
333 src_samples - src_bilinear_samples);
334 assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
336 // Normalize so that the sum becomes one. Note that we do it twice;
337 // this sometimes helps a tiny little bit when we have many samples.
338 for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
340 for (int i = 0; i < src_bilinear_samples; ++i) {
341 sum += bilinear_weights[(y * src_bilinear_samples + i) * 2 + 0];
343 for (int i = 0; i < src_bilinear_samples; ++i) {
344 bilinear_weights[(y * src_bilinear_samples + i) * 2 + 0] /= sum;
349 // Encode as a two-component texture. Note the GL_REPEAT.
350 glActiveTexture(GL_TEXTURE0 + *sampler_num);
352 glBindTexture(GL_TEXTURE_2D, texnum);
354 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
356 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
358 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
360 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_FLOAT, bilinear_weights);
364 delete[] bilinear_weights;
367 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
369 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
371 assert(input_width > 0);
372 assert(input_height > 0);
373 assert(output_width > 0);
374 assert(output_height > 0);
376 if (input_width != last_input_width ||
377 input_height != last_input_height ||
378 output_width != last_output_width ||
379 output_height != last_output_height) {
380 update_texture(glsl_program_num, prefix, sampler_num);
381 last_input_width = input_width;
382 last_input_height = input_height;
383 last_output_width = output_width;
384 last_output_height = output_height;
387 glActiveTexture(GL_TEXTURE0 + *sampler_num);
389 glBindTexture(GL_TEXTURE_2D, texnum);
392 set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num);
394 set_uniform_int(glsl_program_num, prefix, "num_samples", src_bilinear_samples);
395 set_uniform_float(glsl_program_num, prefix, "num_loops", num_loops);
396 set_uniform_float(glsl_program_num, prefix, "slice_height", slice_height);
398 // Instructions for how to convert integer sample numbers to positions in the weight texture.
399 set_uniform_float(glsl_program_num, prefix, "sample_x_scale", 1.0f / src_bilinear_samples);
400 set_uniform_float(glsl_program_num, prefix, "sample_x_offset", 0.5f / src_bilinear_samples);
402 // We specifically do not want mipmaps on the input texture;
403 // they break minification.
404 glActiveTexture(GL_TEXTURE0);
406 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);