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"
21 return 1.0f - fabs(x);
27 float lanczos_weight(float x, float a)
32 return sinc(M_PI * x) * sinc(M_PI * x / a);
36 // Euclid's algorithm, from Wikipedia.
37 unsigned gcd(unsigned a, unsigned b)
47 unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsigned max_samples_saved)
49 unsigned num_samples_saved = 0;
50 for (unsigned i = 0, j = 0; i < num_src_samples; ++i, ++j) {
51 // Copy the sample directly; it will be overwritten later if we can combine.
53 dst[j * 2 + 0] = src[i * 2 + 0];
54 dst[j * 2 + 1] = src[i * 2 + 1];
57 if (i == num_src_samples - 1) {
58 // Last sample; cannot combine.
61 assert(num_samples_saved <= max_samples_saved);
62 if (num_samples_saved == max_samples_saved) {
63 // We could maybe save more here, but other rows can't, so don't bother.
67 float w1 = src[i * 2 + 0];
68 float w2 = src[(i + 1) * 2 + 0];
70 // Differing signs; cannot combine.
74 float pos1 = src[i * 2 + 1];
75 float pos2 = src[(i + 1) * 2 + 1];
78 float offset, total_weight, sum_sq_error;
79 combine_two_samples(w1, w2, &offset, &total_weight, &sum_sq_error);
81 // If the interpolation error is larger than that of about sqrt(2) of
82 // a level at 8-bit precision, don't combine. (You'd think 1.0 was enough,
83 // but since the artifacts are not really random, they can get quite
84 // visible. On the other hand, going to 0.25f, I can see no change at
85 // all with 8-bit output, so it would not seem to be worth it.)
86 if (sum_sq_error > 0.5f / (256.0f * 256.0f)) {
90 // OK, we can combine this and the next sample.
92 dst[j * 2 + 0] = total_weight;
93 dst[j * 2 + 1] = pos1 + offset * (pos2 - pos1);
96 ++i; // Skip the next sample.
99 return num_samples_saved;
104 ResampleEffect::ResampleEffect()
108 register_int("width", &output_width);
109 register_int("height", &output_height);
111 // The first blur pass will forward resolution information to us.
112 hpass = new SingleResamplePassEffect(this);
113 CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL));
114 vpass = new SingleResamplePassEffect(NULL);
115 CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL));
120 void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self)
122 Node *hpass_node = graph->add_node(hpass);
123 Node *vpass_node = graph->add_node(vpass);
124 graph->connect_nodes(hpass_node, vpass_node);
125 graph->replace_receiver(self, hpass_node);
126 graph->replace_sender(self, vpass_node);
127 self->disabled = true;
130 // We get this information forwarded from the first blur pass,
131 // since we are not part of the chain ourselves.
132 void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsigned height)
134 assert(input_num == 0);
138 input_height = height;
142 void ResampleEffect::update_size()
145 ok |= hpass->set_int("input_width", input_width);
146 ok |= hpass->set_int("input_height", input_height);
147 ok |= hpass->set_int("output_width", output_width);
148 ok |= hpass->set_int("output_height", input_height);
150 ok |= vpass->set_int("input_width", output_width);
151 ok |= vpass->set_int("input_height", input_height);
152 ok |= vpass->set_int("output_width", output_width);
153 ok |= vpass->set_int("output_height", output_height);
158 bool ResampleEffect::set_float(const std::string &key, float value) {
159 if (key == "width") {
160 output_width = value;
164 if (key == "height") {
165 output_height = value;
172 SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
174 direction(HORIZONTAL),
177 last_input_width(-1),
178 last_input_height(-1),
179 last_output_width(-1),
180 last_output_height(-1)
182 register_int("direction", (int *)&direction);
183 register_int("input_width", &input_width);
184 register_int("input_height", &input_height);
185 register_int("output_width", &output_width);
186 register_int("output_height", &output_height);
188 glGenTextures(1, &texnum);
191 SingleResamplePassEffect::~SingleResamplePassEffect()
193 glDeleteTextures(1, &texnum);
196 std::string SingleResamplePassEffect::output_fragment_shader()
199 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
200 return buf + read_file("resample_effect.frag");
203 // Using vertical scaling as an example:
205 // Generally out[y] = w0 * in[yi] + w1 * in[yi + 1] + w2 * in[yi + 2] + ...
207 // Obviously, yi will depend on y (in a not-quite-linear way), but so will
208 // the weights w0, w1, w2, etc.. The easiest way of doing this is to encode,
209 // for each sample, the weight and the yi value, e.g. <yi, w0>, <yi + 1, w1>,
210 // and so on. For each y, we encode these along the x-axis (since that is spare),
211 // so out[0] will read from parameters <x,y> = <0,0>, <1,0>, <2,0> and so on.
213 // For horizontal scaling, we fill in the exact same texture;
214 // the shader just interprets it differently.
215 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
217 unsigned src_size, dst_size;
218 if (direction == SingleResamplePassEffect::HORIZONTAL) {
219 assert(input_height == output_height);
220 src_size = input_width;
221 dst_size = output_width;
222 } else if (direction == SingleResamplePassEffect::VERTICAL) {
223 assert(input_width == output_width);
224 src_size = input_height;
225 dst_size = output_height;
231 // For many resamplings (e.g. 640 -> 1280), we will end up with the same
232 // set of samples over and over again in a loop. Thus, we can compute only
233 // the first such loop, and then ask the card to repeat the texture for us.
234 // This is both easier on the texture cache and lowers our CPU cost for
235 // generating the kernel somewhat.
236 num_loops = gcd(src_size, dst_size);
237 slice_height = 1.0f / num_loops;
238 unsigned dst_samples = dst_size / num_loops;
240 // Sample the kernel in the right place. A diagram with a triangular kernel
241 // (corresponding to linear filtering, and obviously with radius 1)
242 // for easier ASCII art drawing:
248 // x---x---x x x---x---x---x
250 // Scaling up (in this case, 2x) means sampling more densely:
256 // x-x-x-x-x-x x x x-x-x-x-x-x-x-x
258 // When scaling up, any destination pixel will only be influenced by a few
259 // (in this case, two) neighboring pixels, and more importantly, the number
260 // will not be influenced by the scaling factor. (Note, however, that the
261 // pixel centers have moved, due to OpenGL's center-pixel convention.)
262 // The only thing that changes is the weights themselves, as the sampling
263 // points are at different distances from the original pixels.
265 // Scaling down is a different story:
271 // --x------ x --x-------x--
273 // Again, the pixel centers have moved in a maybe unintuitive fashion,
274 // although when you consider that there are multiple source pixels around,
275 // it's not so bad as at first look:
281 // --x-------x-------x-------x--
283 // As you can see, the new pixels become averages of the two neighboring old
284 // ones (the situation for Lanczos is of course more complex).
286 // Anyhow, in this case we clearly need to look at more source pixels
287 // to compute the destination pixel, and how many depend on the scaling factor.
288 // Thus, the kernel width will vary with how much we scale.
289 float radius_scaling_factor = std::min(float(dst_size) / float(src_size), 1.0f);
290 int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
291 int src_samples = int_radius * 2 + 1;
292 float *weights = new float[dst_samples * src_samples * 2];
293 for (unsigned y = 0; y < dst_samples; ++y) {
294 // Find the point around which we want to sample the source image,
295 // compensating for differing pixel centers as the scale changes.
296 float center_src_y = (y + 0.5f) * float(src_size) / float(dst_size) - 0.5f;
297 int base_src_y = lrintf(center_src_y);
299 // Now sample <int_radius> pixels on each side around that point.
300 for (int i = 0; i < src_samples; ++i) {
301 int src_y = base_src_y + i - int_radius;
302 float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y), LANCZOS_RADIUS);
303 weights[(y * src_samples + i) * 2 + 0] = weight * radius_scaling_factor;
304 weights[(y * src_samples + i) * 2 + 1] = (src_y + 0.5) / float(src_size);
308 // Now make use of the bilinear filtering in the GPU to reduce the number of samples
309 // we need to make. This is a bit more complex than BlurEffect since we cannot combine
310 // two neighboring samples if their weights have differing signs, so we first need to
311 // figure out the maximum number of samples. Then, we downconvert all the weights to
312 // that number -- we could have gone for a variable-length system, but this is simpler,
313 // and the gains would probably be offset by the extra cost of checking when to stop.
315 // The greedy strategy for combining samples is optimal.
316 src_bilinear_samples = 0;
317 for (unsigned y = 0; y < dst_samples; ++y) {
318 unsigned num_samples_saved = combine_samples(weights + (y * src_samples) * 2, NULL, src_samples, UINT_MAX);
319 src_bilinear_samples = std::max<int>(src_bilinear_samples, src_samples - num_samples_saved);
322 // Now that we know the right width, actually combine the samples.
323 float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2];
324 for (unsigned y = 0; y < dst_samples; ++y) {
325 unsigned num_samples_saved = combine_samples(
326 weights + (y * src_samples) * 2,
327 bilinear_weights + (y * src_bilinear_samples) * 2,
329 src_samples - src_bilinear_samples);
330 assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
333 // Encode as a two-component texture. Note the GL_REPEAT.
334 glActiveTexture(GL_TEXTURE0 + *sampler_num);
336 glBindTexture(GL_TEXTURE_2D, texnum);
338 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
340 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
342 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
344 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_FLOAT, bilinear_weights);
348 delete[] bilinear_weights;
351 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
353 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
355 assert(input_width > 0);
356 assert(input_height > 0);
357 assert(output_width > 0);
358 assert(output_height > 0);
360 if (input_width != last_input_width ||
361 input_height != last_input_height ||
362 output_width != last_output_width ||
363 output_height != last_output_height) {
364 update_texture(glsl_program_num, prefix, sampler_num);
365 last_input_width = input_width;
366 last_input_height = input_height;
367 last_output_width = output_width;
368 last_output_height = output_height;
371 glActiveTexture(GL_TEXTURE0 + *sampler_num);
373 glBindTexture(GL_TEXTURE_2D, texnum);
376 set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num);
378 set_uniform_int(glsl_program_num, prefix, "num_samples", src_bilinear_samples);
379 set_uniform_float(glsl_program_num, prefix, "num_loops", num_loops);
380 set_uniform_float(glsl_program_num, prefix, "slice_height", slice_height);
382 // Instructions for how to convert integer sample numbers to positions in the weight texture.
383 set_uniform_float(glsl_program_num, prefix, "sample_x_scale", 1.0f / src_bilinear_samples);
384 set_uniform_float(glsl_program_num, prefix, "sample_x_offset", 0.5f / src_bilinear_samples);
386 // We specifically do not want mipmaps on the input texture;
387 // they break minification.
388 glActiveTexture(GL_TEXTURE0);
390 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);