1 // Three-lobed Lanczos, the most common choice.
2 // Note that if you change this, the accuracy for LANCZOS_TABLE_SIZE
3 // needs to be recomputed.
4 #define LANCZOS_RADIUS 3.0f
12 #include <Eigen/Sparse>
13 #include <Eigen/SparseQR>
14 #include <Eigen/OrderingMethods>
16 #include "effect_chain.h"
17 #include "effect_util.h"
20 #include "resample_effect.h"
23 using namespace Eigen;
33 return 1.0f - fabs(x);
39 float lanczos_weight(float x)
41 if (fabs(x) > LANCZOS_RADIUS) {
44 return sinc(M_PI * x) * sinc((M_PI / LANCZOS_RADIUS) * x);
48 // The weight function can be expensive to compute over and over again
49 // (which will happen during e.g. a zoom), but it is also easy to interpolate
50 // linearly. We compute the right half of the function (in the range of
51 // 0..LANCZOS_RADIUS), with two guard elements for easier interpolation, and
52 // linearly interpolate to get our function.
54 // We want to scale the table so that the maximum error is always smaller
55 // than 1e-6. As per http://www-solar.mcs.st-andrews.ac.uk/~clare/Lectures/num-analysis/Numan_chap3.pdf,
56 // the error for interpolating a function linearly between points [a,b] is
58 // e = 1/2 (x-a)(x-b) f''(u_x)
60 // for some point u_x in [a,b] (where f(x) is our Lanczos function; we're
61 // assuming LANCZOS_RADIUS=3 from here on). Obviously this is bounded by
62 // f''(x) over the entire range. Numeric optimization shows the maximum of
63 // |f''(x)| to be in x=1.09369819474562880, with the value 2.40067758733152381.
64 // So if the steps between consecutive values are called d, we get
66 // |e| <= 1/2 (d/2)^2 2.4007
69 // Solve for e = 1e-6 yields a step size of 0.0027, which to cover the range
70 // 0..3 needs 1109 steps. We round up to the next power of two, just to be sure.
72 // You need to call lanczos_table_init_done before the first call to
73 // lanczos_weight_cached.
74 #define LANCZOS_TABLE_SIZE 2048
75 bool lanczos_table_init_done = false;
76 float lanczos_table[LANCZOS_TABLE_SIZE + 2];
78 void init_lanczos_table()
80 for (unsigned i = 0; i < LANCZOS_TABLE_SIZE + 2; ++i) {
81 lanczos_table[i] = lanczos_weight(float(i) * (LANCZOS_RADIUS / LANCZOS_TABLE_SIZE));
83 lanczos_table_init_done = true;
86 float lanczos_weight_cached(float x)
89 if (x > LANCZOS_RADIUS) {
92 float table_pos = x * (LANCZOS_TABLE_SIZE / LANCZOS_RADIUS);
93 unsigned table_pos_int = int(table_pos); // Truncate towards zero.
94 float table_pos_frac = table_pos - table_pos_int;
95 assert(table_pos < LANCZOS_TABLE_SIZE + 2);
96 return lanczos_table[table_pos_int] +
97 table_pos_frac * (lanczos_table[table_pos_int + 1] - lanczos_table[table_pos_int]);
100 // Euclid's algorithm, from Wikipedia.
101 unsigned gcd(unsigned a, unsigned b)
111 template<class DestFloat>
112 unsigned combine_samples(const Tap<float> *src, Tap<DestFloat> *dst, float num_subtexels, float inv_num_subtexels, unsigned num_src_samples, unsigned max_samples_saved, float pos1_pos2_diff, float inv_pos1_pos2_diff)
114 // Cut off near-zero values at both sides.
115 unsigned num_samples_saved = 0;
116 while (num_samples_saved < max_samples_saved &&
117 num_src_samples > 0 &&
118 fabs(src[0].weight) < 1e-6) {
123 while (num_samples_saved < max_samples_saved &&
124 num_src_samples > 0 &&
125 fabs(src[num_src_samples - 1].weight) < 1e-6) {
130 for (unsigned i = 0, j = 0; i < num_src_samples; ++i, ++j) {
131 // Copy the sample directly; it will be overwritten later if we can combine.
133 dst[j].weight = convert_float<float, DestFloat>(src[i].weight);
134 dst[j].pos = convert_float<float, DestFloat>(src[i].pos);
137 if (i == num_src_samples - 1) {
138 // Last sample; cannot combine.
141 assert(num_samples_saved <= max_samples_saved);
142 if (num_samples_saved == max_samples_saved) {
143 // We could maybe save more here, but other rows can't, so don't bother.
147 float w1 = src[i].weight;
148 float w2 = src[i + 1].weight;
149 if (w1 * w2 < 0.0f) {
150 // Differing signs; cannot combine.
154 float pos1 = src[i].pos;
155 float pos2 = src[i + 1].pos;
158 DestFloat pos, total_weight;
160 combine_two_samples(w1, w2, pos1, pos1_pos2_diff, inv_pos1_pos2_diff, num_subtexels, inv_num_subtexels, &pos, &total_weight, &sum_sq_error);
162 // If the interpolation error is larger than that of about sqrt(2) of
163 // a level at 8-bit precision, don't combine. (You'd think 1.0 was enough,
164 // but since the artifacts are not really random, they can get quite
165 // visible. On the other hand, going to 0.25f, I can see no change at
166 // all with 8-bit output, so it would not seem to be worth it.)
167 if (sum_sq_error > 0.5f / (255.0f * 255.0f)) {
171 // OK, we can combine this and the next sample.
173 dst[j].weight = total_weight;
177 ++i; // Skip the next sample.
180 return num_samples_saved;
183 // Normalize so that the sum becomes one. Note that we do it twice;
184 // this sometimes helps a tiny little bit when we have many samples.
186 void normalize_sum(Tap<T>* vals, unsigned num)
188 for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
190 for (unsigned i = 0; i < num; ++i) {
191 sum += to_fp32(vals[i].weight);
193 float inv_sum = 1.0 / sum;
194 for (unsigned i = 0; i < num; ++i) {
195 vals[i].weight = from_fp32<T>(to_fp32(vals[i].weight) * inv_sum);
200 // Make use of the bilinear filtering in the GPU to reduce the number of samples
201 // we need to make. This is a bit more complex than BlurEffect since we cannot combine
202 // two neighboring samples if their weights have differing signs, so we first need to
203 // figure out the maximum number of samples. Then, we downconvert all the weights to
204 // that number -- we could have gone for a variable-length system, but this is simpler,
205 // and the gains would probably be offset by the extra cost of checking when to stop.
207 // The greedy strategy for combining samples is optimal.
208 template<class DestFloat>
209 unsigned combine_many_samples(const Tap<float> *weights, unsigned src_size, unsigned src_samples, unsigned dst_samples, unique_ptr<Tap<DestFloat>[]> *bilinear_weights)
211 float num_subtexels = src_size / movit_texel_subpixel_precision;
212 float inv_num_subtexels = movit_texel_subpixel_precision / src_size;
213 float pos1_pos2_diff = 1.0f / src_size;
214 float inv_pos1_pos2_diff = src_size;
216 unsigned max_samples_saved = UINT_MAX;
217 for (unsigned y = 0; y < dst_samples && max_samples_saved > 0; ++y) {
218 unsigned num_samples_saved = combine_samples<DestFloat>(weights + y * src_samples, NULL, num_subtexels, inv_num_subtexels, src_samples, max_samples_saved, pos1_pos2_diff, inv_pos1_pos2_diff);
219 max_samples_saved = min(max_samples_saved, num_samples_saved);
222 // Now that we know the right width, actually combine the samples.
223 unsigned src_bilinear_samples = src_samples - max_samples_saved;
224 bilinear_weights->reset(new Tap<DestFloat>[dst_samples * src_bilinear_samples]);
225 for (unsigned y = 0; y < dst_samples; ++y) {
226 Tap<DestFloat> *bilinear_weights_ptr = bilinear_weights->get() + y * src_bilinear_samples;
227 unsigned num_samples_saved = combine_samples(
228 weights + y * src_samples,
229 bilinear_weights_ptr,
236 assert(num_samples_saved == max_samples_saved);
237 normalize_sum(bilinear_weights_ptr, src_bilinear_samples);
239 return src_bilinear_samples;
242 // Compute the sum of squared errors between the ideal weights (which are
243 // assumed to fall exactly on pixel centers) and the weights that result
244 // from sampling at <bilinear_weights>. The primary reason for the difference
245 // is inaccuracy in the sampling positions, both due to limited precision
246 // in storing them (already inherent in sending them in as fp16_int_t)
247 // and in subtexel sampling precision (which we calculate in this function).
249 double compute_sum_sq_error(const Tap<float>* weights, unsigned num_weights,
250 const Tap<T>* bilinear_weights, unsigned num_bilinear_weights,
253 // Find the effective range of the bilinear-optimized kernel.
254 // Due to rounding of the positions, this is not necessarily the same
255 // as the intended range (ie., the range of the original weights).
256 int lower_pos = int(floor(to_fp32(bilinear_weights[0].pos) * size - 0.5f));
257 int upper_pos = int(ceil(to_fp32(bilinear_weights[num_bilinear_weights - 1].pos) * size - 0.5f)) + 2;
258 lower_pos = min<int>(lower_pos, lrintf(weights[0].pos * size - 0.5f));
259 upper_pos = max<int>(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5f) + 1);
261 float* effective_weights = new float[upper_pos - lower_pos];
262 for (int i = 0; i < upper_pos - lower_pos; ++i) {
263 effective_weights[i] = 0.0f;
266 // Now find the effective weights that result from this sampling.
267 for (unsigned i = 0; i < num_bilinear_weights; ++i) {
268 const float pixel_pos = to_fp32(bilinear_weights[i].pos) * size - 0.5f;
269 const int x0 = int(floor(pixel_pos)) - lower_pos;
270 const int x1 = x0 + 1;
271 const float f = lrintf((pixel_pos - (x0 + lower_pos)) / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;
275 assert(x0 < upper_pos - lower_pos);
276 assert(x1 < upper_pos - lower_pos);
278 effective_weights[x0] += to_fp32(bilinear_weights[i].weight) * (1.0f - f);
279 effective_weights[x1] += to_fp32(bilinear_weights[i].weight) * f;
282 // Subtract the desired weights to get the error.
283 for (unsigned i = 0; i < num_weights; ++i) {
284 const int x = lrintf(weights[i].pos * size - 0.5f) - lower_pos;
286 assert(x < upper_pos - lower_pos);
288 effective_weights[x] -= weights[i].weight;
291 double sum_sq_error = 0.0;
292 for (unsigned i = 0; i < num_weights; ++i) {
293 sum_sq_error += effective_weights[i] * effective_weights[i];
296 delete[] effective_weights;
302 ResampleEffect::ResampleEffect()
305 offset_x(0.0f), offset_y(0.0f),
306 zoom_x(1.0f), zoom_y(1.0f),
307 zoom_center_x(0.5f), zoom_center_y(0.5f)
309 register_int("width", &output_width);
310 register_int("height", &output_height);
312 // The first blur pass will forward resolution information to us.
313 hpass = new SingleResamplePassEffect(this);
314 CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL));
315 vpass = new SingleResamplePassEffect(NULL);
316 CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL));
321 void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self)
323 Node *hpass_node = graph->add_node(hpass);
324 Node *vpass_node = graph->add_node(vpass);
325 graph->connect_nodes(hpass_node, vpass_node);
326 graph->replace_receiver(self, hpass_node);
327 graph->replace_sender(self, vpass_node);
328 self->disabled = true;
331 // We get this information forwarded from the first blur pass,
332 // since we are not part of the chain ourselves.
333 void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsigned height)
335 assert(input_num == 0);
339 input_height = height;
343 void ResampleEffect::update_size()
346 ok |= hpass->set_int("input_width", input_width);
347 ok |= hpass->set_int("input_height", input_height);
348 ok |= hpass->set_int("output_width", output_width);
349 ok |= hpass->set_int("output_height", input_height);
351 ok |= vpass->set_int("input_width", output_width);
352 ok |= vpass->set_int("input_height", input_height);
353 ok |= vpass->set_int("output_width", output_width);
354 ok |= vpass->set_int("output_height", output_height);
358 // The offset added due to zoom may have changed with the size.
359 update_offset_and_zoom();
362 void ResampleEffect::update_offset_and_zoom()
366 // Zoom from the right origin. (zoom_center is given in normalized coordinates,
368 float extra_offset_x = zoom_center_x * (1.0f - 1.0f / zoom_x) * input_width;
369 float extra_offset_y = (1.0f - zoom_center_y) * (1.0f - 1.0f / zoom_y) * input_height;
371 ok |= hpass->set_float("offset", extra_offset_x + offset_x);
372 ok |= vpass->set_float("offset", extra_offset_y - offset_y); // Compensate for the bottom-left origin.
373 ok |= hpass->set_float("zoom", zoom_x);
374 ok |= vpass->set_float("zoom", zoom_y);
379 bool ResampleEffect::set_float(const string &key, float value) {
380 if (key == "width") {
381 output_width = value;
385 if (key == "height") {
386 output_height = value;
392 update_offset_and_zoom();
397 update_offset_and_zoom();
400 if (key == "zoom_x") {
405 update_offset_and_zoom();
408 if (key == "zoom_y") {
413 update_offset_and_zoom();
416 if (key == "zoom_center_x") {
417 zoom_center_x = value;
418 update_offset_and_zoom();
421 if (key == "zoom_center_y") {
422 zoom_center_y = value;
423 update_offset_and_zoom();
429 SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
431 direction(HORIZONTAL),
436 last_input_width(-1),
437 last_input_height(-1),
438 last_output_width(-1),
439 last_output_height(-1),
440 last_offset(0.0 / 0.0), // NaN.
441 last_zoom(0.0 / 0.0), // NaN.
442 last_texture_width(-1), last_texture_height(-1)
444 register_int("direction", (int *)&direction);
445 register_int("input_width", &input_width);
446 register_int("input_height", &input_height);
447 register_int("output_width", &output_width);
448 register_int("output_height", &output_height);
449 register_float("offset", &offset);
450 register_float("zoom", &zoom);
451 register_uniform_sampler2d("sample_tex", &uniform_sample_tex);
452 register_uniform_int("num_samples", &uniform_num_samples);
453 register_uniform_float("num_loops", &uniform_num_loops);
454 register_uniform_float("slice_height", &uniform_slice_height);
455 register_uniform_float("sample_x_scale", &uniform_sample_x_scale);
456 register_uniform_float("sample_x_offset", &uniform_sample_x_offset);
457 register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset);
459 glGenTextures(1, &texnum);
461 if (!lanczos_table_init_done) {
462 // Could in theory race between two threads if we are unlucky,
463 // but that is harmless, since they'll write the same data.
464 init_lanczos_table();
468 SingleResamplePassEffect::~SingleResamplePassEffect()
470 glDeleteTextures(1, &texnum);
473 string SingleResamplePassEffect::output_fragment_shader()
476 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
477 return buf + read_file("resample_effect.frag");
480 // Using vertical scaling as an example:
482 // Generally out[y] = w0 * in[yi] + w1 * in[yi + 1] + w2 * in[yi + 2] + ...
484 // Obviously, yi will depend on y (in a not-quite-linear way), but so will
485 // the weights w0, w1, w2, etc.. The easiest way of doing this is to encode,
486 // for each sample, the weight and the yi value, e.g. <yi, w0>, <yi + 1, w1>,
487 // and so on. For each y, we encode these along the x-axis (since that is spare),
488 // so out[0] will read from parameters <x,y> = <0,0>, <1,0>, <2,0> and so on.
490 // For horizontal scaling, we fill in the exact same texture;
491 // the shader just interprets it differently.
492 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
494 unsigned src_size, dst_size;
495 if (direction == SingleResamplePassEffect::HORIZONTAL) {
496 assert(input_height == output_height);
497 src_size = input_width;
498 dst_size = output_width;
499 } else if (direction == SingleResamplePassEffect::VERTICAL) {
500 assert(input_width == output_width);
501 src_size = input_height;
502 dst_size = output_height;
507 ScalingWeights weights = calculate_scaling_weights(src_size, dst_size, zoom, offset);
508 src_bilinear_samples = weights.src_bilinear_samples;
509 num_loops = weights.num_loops;
510 slice_height = 1.0f / weights.num_loops;
512 // Encode as a two-component texture. Note the GL_REPEAT.
513 glActiveTexture(GL_TEXTURE0 + *sampler_num);
515 glBindTexture(GL_TEXTURE_2D, texnum);
517 if (last_texture_width == -1) {
518 // Need to set this state the first time.
519 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
521 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
523 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
527 GLenum type, internal_format;
529 assert((weights.bilinear_weights_fp16 == nullptr) != (weights.bilinear_weights_fp32 == nullptr));
530 if (weights.bilinear_weights_fp32 != nullptr) {
532 internal_format = GL_RG32F;
533 pixels = weights.bilinear_weights_fp32.get();
535 type = GL_HALF_FLOAT;
536 internal_format = GL_RG16F;
537 pixels = weights.bilinear_weights_fp16.get();
540 if (int(weights.src_bilinear_samples) == last_texture_width &&
541 int(weights.dst_samples) == last_texture_height &&
542 internal_format == last_texture_internal_format) {
543 // Texture dimensions and type are unchanged; it is more efficient
544 // to just update it rather than making an entirely new texture.
545 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, weights.src_bilinear_samples, weights.dst_samples, GL_RG, type, pixels);
547 glTexImage2D(GL_TEXTURE_2D, 0, internal_format, weights.src_bilinear_samples, weights.dst_samples, 0, GL_RG, type, pixels);
548 last_texture_width = weights.src_bilinear_samples;
549 last_texture_height = weights.dst_samples;
550 last_texture_internal_format = internal_format;
555 ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset)
557 if (!lanczos_table_init_done) {
558 // Only needed if run from outside ResampleEffect.
559 init_lanczos_table();
562 // For many resamplings (e.g. 640 -> 1280), we will end up with the same
563 // set of samples over and over again in a loop. Thus, we can compute only
564 // the first such loop, and then ask the card to repeat the texture for us.
565 // This is both easier on the texture cache and lowers our CPU cost for
566 // generating the kernel somewhat.
567 float scaling_factor;
569 if (fabs(zoom - 1.0f) < 1e-6) {
570 num_loops = gcd(src_size, dst_size);
571 scaling_factor = float(dst_size) / float(src_size);
573 // If zooming is enabled (ie., zoom != 1), we turn off the looping.
574 // We _could_ perhaps do it for rational zoom levels (especially
575 // things like 2:1), but it doesn't seem to be worth it, given that
576 // the most common use case would seem to be varying the zoom
577 // from frame to frame.
579 scaling_factor = zoom * float(dst_size) / float(src_size);
581 unsigned dst_samples = dst_size / num_loops;
583 // Sample the kernel in the right place. A diagram with a triangular kernel
584 // (corresponding to linear filtering, and obviously with radius 1)
585 // for easier ASCII art drawing:
591 // x---x---x x x---x---x---x
593 // Scaling up (in this case, 2x) means sampling more densely:
599 // x-x-x-x-x-x x x x-x-x-x-x-x-x-x
601 // When scaling up, any destination pixel will only be influenced by a few
602 // (in this case, two) neighboring pixels, and more importantly, the number
603 // will not be influenced by the scaling factor. (Note, however, that the
604 // pixel centers have moved, due to OpenGL's center-pixel convention.)
605 // The only thing that changes is the weights themselves, as the sampling
606 // points are at different distances from the original pixels.
608 // Scaling down is a different story:
614 // --x------ x --x-------x--
616 // Again, the pixel centers have moved in a maybe unintuitive fashion,
617 // although when you consider that there are multiple source pixels around,
618 // it's not so bad as at first look:
624 // --x-------x-------x-------x--
626 // As you can see, the new pixels become averages of the two neighboring old
627 // ones (the situation for Lanczos is of course more complex).
629 // Anyhow, in this case we clearly need to look at more source pixels
630 // to compute the destination pixel, and how many depend on the scaling factor.
631 // Thus, the kernel width will vary with how much we scale.
632 float radius_scaling_factor = min(scaling_factor, 1.0f);
633 int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
634 int src_samples = int_radius * 2 + 1;
635 unique_ptr<Tap<float>[]> weights(new Tap<float>[dst_samples * src_samples]);
636 float subpixel_offset = offset - lrintf(offset); // The part not covered by whole_pixel_offset.
637 assert(subpixel_offset >= -0.5f && subpixel_offset <= 0.5f);
638 for (unsigned y = 0; y < dst_samples; ++y) {
639 // Find the point around which we want to sample the source image,
640 // compensating for differing pixel centers as the scale changes.
641 float center_src_y = (y + 0.5f) / scaling_factor - 0.5f;
642 int base_src_y = lrintf(center_src_y);
644 // Now sample <int_radius> pixels on each side around that point.
645 float inv_src_size = 1.0 / float(src_size);
646 for (int i = 0; i < src_samples; ++i) {
647 int src_y = base_src_y + i - int_radius;
648 float weight = lanczos_weight_cached(radius_scaling_factor * (src_y - center_src_y - subpixel_offset));
649 weights[y * src_samples + i].weight = weight * radius_scaling_factor;
650 weights[y * src_samples + i].pos = (src_y + 0.5f) * inv_src_size;
654 // Now make use of the bilinear filtering in the GPU to reduce the number of samples
655 // we need to make. Try fp16 first; if it's not accurate enough, we go to fp32.
656 // Our tolerance level for total error is a bit higher than the one for invididual
657 // samples, since one would assume overall errors in the shape don't matter as much.
658 const float max_error = 2.0f / (255.0f * 255.0f);
659 unique_ptr<Tap<fp16_int_t>[]> bilinear_weights_fp16;
660 int src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, dst_samples, &bilinear_weights_fp16);
661 unique_ptr<Tap<float>[]> bilinear_weights_fp32 = NULL;
662 double max_sum_sq_error_fp16 = 0.0;
663 for (unsigned y = 0; y < dst_samples; ++y) {
664 double sum_sq_error_fp16 = compute_sum_sq_error(
665 weights.get() + y * src_samples, src_samples,
666 bilinear_weights_fp16.get() + y * src_bilinear_samples, src_bilinear_samples,
668 max_sum_sq_error_fp16 = std::max(max_sum_sq_error_fp16, sum_sq_error_fp16);
669 if (max_sum_sq_error_fp16 > max_error) {
674 if (max_sum_sq_error_fp16 > max_error) {
675 bilinear_weights_fp16.reset();
676 src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, dst_samples, &bilinear_weights_fp32);
680 ret.src_bilinear_samples = src_bilinear_samples;
681 ret.dst_samples = dst_samples;
682 ret.num_loops = num_loops;
683 ret.bilinear_weights_fp16 = move(bilinear_weights_fp16);
684 ret.bilinear_weights_fp32 = move(bilinear_weights_fp32);
688 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
690 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
692 assert(input_width > 0);
693 assert(input_height > 0);
694 assert(output_width > 0);
695 assert(output_height > 0);
697 if (input_width != last_input_width ||
698 input_height != last_input_height ||
699 output_width != last_output_width ||
700 output_height != last_output_height ||
701 offset != last_offset ||
703 update_texture(glsl_program_num, prefix, sampler_num);
704 last_input_width = input_width;
705 last_input_height = input_height;
706 last_output_width = output_width;
707 last_output_height = output_height;
708 last_offset = offset;
712 glActiveTexture(GL_TEXTURE0 + *sampler_num);
714 glBindTexture(GL_TEXTURE_2D, texnum);
717 uniform_sample_tex = *sampler_num;
719 uniform_num_samples = src_bilinear_samples;
720 uniform_num_loops = num_loops;
721 uniform_slice_height = slice_height;
723 // Instructions for how to convert integer sample numbers to positions in the weight texture.
724 uniform_sample_x_scale = 1.0f / src_bilinear_samples;
725 uniform_sample_x_offset = 0.5f / src_bilinear_samples;
727 if (direction == SingleResamplePassEffect::VERTICAL) {
728 uniform_whole_pixel_offset = lrintf(offset) / float(input_height);
730 uniform_whole_pixel_offset = lrintf(offset) / float(input_width);
733 // We specifically do not want mipmaps on the input texture;
734 // they break minification.
735 Node *self = chain->find_node_for_effect(this);
736 if (chain->has_input_sampler(self, 0)) {
737 glActiveTexture(chain->get_input_sampler(self, 0));
739 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);