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.
132 if (dst != nullptr) {
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.
172 if (dst != nullptr) {
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, nullptr, 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()
303 : owns_effects(true),
306 offset_x(0.0f), offset_y(0.0f),
307 zoom_x(1.0f), zoom_y(1.0f),
308 zoom_center_x(0.5f), zoom_center_y(0.5f)
310 register_int("width", &output_width);
311 register_int("height", &output_height);
313 // The first blur pass will forward resolution information to us.
314 hpass = new SingleResamplePassEffect(this);
315 CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL));
316 vpass = new SingleResamplePassEffect(nullptr);
317 CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL));
322 ResampleEffect::~ResampleEffect()
330 void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self)
332 Node *hpass_node = graph->add_node(hpass);
333 Node *vpass_node = graph->add_node(vpass);
334 graph->connect_nodes(hpass_node, vpass_node);
335 graph->replace_receiver(self, hpass_node);
336 graph->replace_sender(self, vpass_node);
337 self->disabled = true;
338 owns_effects = false;
341 // We get this information forwarded from the first blur pass,
342 // since we are not part of the chain ourselves.
343 void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsigned height)
345 assert(input_num == 0);
349 input_height = height;
353 void ResampleEffect::update_size()
356 ok |= hpass->set_int("input_width", input_width);
357 ok |= hpass->set_int("input_height", input_height);
358 ok |= hpass->set_int("output_width", output_width);
359 ok |= hpass->set_int("output_height", input_height);
361 ok |= vpass->set_int("input_width", output_width);
362 ok |= vpass->set_int("input_height", input_height);
363 ok |= vpass->set_int("output_width", output_width);
364 ok |= vpass->set_int("output_height", output_height);
368 // The offset added due to zoom may have changed with the size.
369 update_offset_and_zoom();
372 void ResampleEffect::update_offset_and_zoom()
376 // Zoom from the right origin. (zoom_center is given in normalized coordinates,
378 float extra_offset_x = zoom_center_x * (1.0f - 1.0f / zoom_x) * input_width;
379 float extra_offset_y = (1.0f - zoom_center_y) * (1.0f - 1.0f / zoom_y) * input_height;
381 ok |= hpass->set_float("offset", extra_offset_x + offset_x);
382 ok |= vpass->set_float("offset", extra_offset_y - offset_y); // Compensate for the bottom-left origin.
383 ok |= hpass->set_float("zoom", zoom_x);
384 ok |= vpass->set_float("zoom", zoom_y);
389 bool ResampleEffect::set_float(const string &key, float value) {
390 if (key == "width") {
391 output_width = value;
395 if (key == "height") {
396 output_height = value;
402 update_offset_and_zoom();
407 update_offset_and_zoom();
410 if (key == "zoom_x") {
415 update_offset_and_zoom();
418 if (key == "zoom_y") {
423 update_offset_and_zoom();
426 if (key == "zoom_center_x") {
427 zoom_center_x = value;
428 update_offset_and_zoom();
431 if (key == "zoom_center_y") {
432 zoom_center_y = value;
433 update_offset_and_zoom();
439 SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
441 direction(HORIZONTAL),
446 last_input_width(-1),
447 last_input_height(-1),
448 last_output_width(-1),
449 last_output_height(-1),
450 last_offset(0.0 / 0.0), // NaN.
451 last_zoom(0.0 / 0.0), // NaN.
452 last_texture_width(-1), last_texture_height(-1)
454 register_int("direction", (int *)&direction);
455 register_int("input_width", &input_width);
456 register_int("input_height", &input_height);
457 register_int("output_width", &output_width);
458 register_int("output_height", &output_height);
459 register_float("offset", &offset);
460 register_float("zoom", &zoom);
461 register_uniform_sampler2d("sample_tex", &uniform_sample_tex);
462 register_uniform_int("num_samples", &uniform_num_samples);
463 register_uniform_float("num_loops", &uniform_num_loops);
464 register_uniform_float("slice_height", &uniform_slice_height);
465 register_uniform_float("sample_x_scale", &uniform_sample_x_scale);
466 register_uniform_float("sample_x_offset", &uniform_sample_x_offset);
467 register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset);
469 glGenTextures(1, &texnum);
471 if (!lanczos_table_init_done) {
472 // Could in theory race between two threads if we are unlucky,
473 // but that is harmless, since they'll write the same data.
474 init_lanczos_table();
478 SingleResamplePassEffect::~SingleResamplePassEffect()
480 glDeleteTextures(1, &texnum);
483 string SingleResamplePassEffect::output_fragment_shader()
486 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
487 return buf + read_file("resample_effect.frag");
490 // Using vertical scaling as an example:
492 // Generally out[y] = w0 * in[yi] + w1 * in[yi + 1] + w2 * in[yi + 2] + ...
494 // Obviously, yi will depend on y (in a not-quite-linear way), but so will
495 // the weights w0, w1, w2, etc.. The easiest way of doing this is to encode,
496 // for each sample, the weight and the yi value, e.g. <yi, w0>, <yi + 1, w1>,
497 // and so on. For each y, we encode these along the x-axis (since that is spare),
498 // so out[0] will read from parameters <x,y> = <0,0>, <1,0>, <2,0> and so on.
500 // For horizontal scaling, we fill in the exact same texture;
501 // the shader just interprets it differently.
502 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
504 unsigned src_size, dst_size;
505 if (direction == SingleResamplePassEffect::HORIZONTAL) {
506 assert(input_height == output_height);
507 src_size = input_width;
508 dst_size = output_width;
509 } else if (direction == SingleResamplePassEffect::VERTICAL) {
510 assert(input_width == output_width);
511 src_size = input_height;
512 dst_size = output_height;
517 ScalingWeights weights = calculate_scaling_weights(src_size, dst_size, zoom, offset);
518 src_bilinear_samples = weights.src_bilinear_samples;
519 num_loops = weights.num_loops;
520 slice_height = 1.0f / weights.num_loops;
522 // Encode as a two-component texture. Note the GL_REPEAT.
523 glActiveTexture(GL_TEXTURE0 + *sampler_num);
525 glBindTexture(GL_TEXTURE_2D, texnum);
527 if (last_texture_width == -1) {
528 // Need to set this state the first time.
529 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
531 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
533 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
537 GLenum type, internal_format;
539 assert((weights.bilinear_weights_fp16 == nullptr) != (weights.bilinear_weights_fp32 == nullptr));
540 if (weights.bilinear_weights_fp32 != nullptr) {
542 internal_format = GL_RG32F;
543 pixels = weights.bilinear_weights_fp32.get();
545 type = GL_HALF_FLOAT;
546 internal_format = GL_RG16F;
547 pixels = weights.bilinear_weights_fp16.get();
550 if (int(weights.src_bilinear_samples) == last_texture_width &&
551 int(weights.dst_samples) == last_texture_height &&
552 internal_format == last_texture_internal_format) {
553 // Texture dimensions and type are unchanged; it is more efficient
554 // to just update it rather than making an entirely new texture.
555 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, weights.src_bilinear_samples, weights.dst_samples, GL_RG, type, pixels);
557 glTexImage2D(GL_TEXTURE_2D, 0, internal_format, weights.src_bilinear_samples, weights.dst_samples, 0, GL_RG, type, pixels);
558 last_texture_width = weights.src_bilinear_samples;
559 last_texture_height = weights.dst_samples;
560 last_texture_internal_format = internal_format;
565 ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset)
567 if (!lanczos_table_init_done) {
568 // Only needed if run from outside ResampleEffect.
569 init_lanczos_table();
572 // For many resamplings (e.g. 640 -> 1280), we will end up with the same
573 // set of samples over and over again in a loop. Thus, we can compute only
574 // the first such loop, and then ask the card to repeat the texture for us.
575 // This is both easier on the texture cache and lowers our CPU cost for
576 // generating the kernel somewhat.
577 float scaling_factor;
579 if (fabs(zoom - 1.0f) < 1e-6) {
580 num_loops = gcd(src_size, dst_size);
581 scaling_factor = float(dst_size) / float(src_size);
583 // If zooming is enabled (ie., zoom != 1), we turn off the looping.
584 // We _could_ perhaps do it for rational zoom levels (especially
585 // things like 2:1), but it doesn't seem to be worth it, given that
586 // the most common use case would seem to be varying the zoom
587 // from frame to frame.
589 scaling_factor = zoom * float(dst_size) / float(src_size);
591 unsigned dst_samples = dst_size / num_loops;
593 // Sample the kernel in the right place. A diagram with a triangular kernel
594 // (corresponding to linear filtering, and obviously with radius 1)
595 // for easier ASCII art drawing:
601 // x---x---x x x---x---x---x
603 // Scaling up (in this case, 2x) means sampling more densely:
609 // x-x-x-x-x-x x x x-x-x-x-x-x-x-x
611 // When scaling up, any destination pixel will only be influenced by a few
612 // (in this case, two) neighboring pixels, and more importantly, the number
613 // will not be influenced by the scaling factor. (Note, however, that the
614 // pixel centers have moved, due to OpenGL's center-pixel convention.)
615 // The only thing that changes is the weights themselves, as the sampling
616 // points are at different distances from the original pixels.
618 // Scaling down is a different story:
624 // --x------ x --x-------x--
626 // Again, the pixel centers have moved in a maybe unintuitive fashion,
627 // although when you consider that there are multiple source pixels around,
628 // it's not so bad as at first look:
634 // --x-------x-------x-------x--
636 // As you can see, the new pixels become averages of the two neighboring old
637 // ones (the situation for Lanczos is of course more complex).
639 // Anyhow, in this case we clearly need to look at more source pixels
640 // to compute the destination pixel, and how many depend on the scaling factor.
641 // Thus, the kernel width will vary with how much we scale.
642 float radius_scaling_factor = min(scaling_factor, 1.0f);
643 int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
644 int src_samples = int_radius * 2 + 1;
645 unique_ptr<Tap<float>[]> weights(new Tap<float>[dst_samples * src_samples]);
646 float subpixel_offset = offset - lrintf(offset); // The part not covered by whole_pixel_offset.
647 assert(subpixel_offset >= -0.5f && subpixel_offset <= 0.5f);
648 for (unsigned y = 0; y < dst_samples; ++y) {
649 // Find the point around which we want to sample the source image,
650 // compensating for differing pixel centers as the scale changes.
651 float center_src_y = (y + 0.5f) / scaling_factor - 0.5f;
652 int base_src_y = lrintf(center_src_y);
654 // Now sample <int_radius> pixels on each side around that point.
655 float inv_src_size = 1.0 / float(src_size);
656 for (int i = 0; i < src_samples; ++i) {
657 int src_y = base_src_y + i - int_radius;
658 float weight = lanczos_weight_cached(radius_scaling_factor * (src_y - center_src_y - subpixel_offset));
659 weights[y * src_samples + i].weight = weight * radius_scaling_factor;
660 weights[y * src_samples + i].pos = (src_y + 0.5f) * inv_src_size;
664 // Now make use of the bilinear filtering in the GPU to reduce the number of samples
665 // we need to make. Try fp16 first; if it's not accurate enough, we go to fp32.
666 // Our tolerance level for total error is a bit higher than the one for invididual
667 // samples, since one would assume overall errors in the shape don't matter as much.
668 const float max_error = 2.0f / (255.0f * 255.0f);
669 unique_ptr<Tap<fp16_int_t>[]> bilinear_weights_fp16;
670 int src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, dst_samples, &bilinear_weights_fp16);
671 unique_ptr<Tap<float>[]> bilinear_weights_fp32 = nullptr;
672 double max_sum_sq_error_fp16 = 0.0;
673 for (unsigned y = 0; y < dst_samples; ++y) {
674 double sum_sq_error_fp16 = compute_sum_sq_error(
675 weights.get() + y * src_samples, src_samples,
676 bilinear_weights_fp16.get() + y * src_bilinear_samples, src_bilinear_samples,
678 max_sum_sq_error_fp16 = std::max(max_sum_sq_error_fp16, sum_sq_error_fp16);
679 if (max_sum_sq_error_fp16 > max_error) {
684 if (max_sum_sq_error_fp16 > max_error) {
685 bilinear_weights_fp16.reset();
686 src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, dst_samples, &bilinear_weights_fp32);
690 ret.src_bilinear_samples = src_bilinear_samples;
691 ret.dst_samples = dst_samples;
692 ret.num_loops = num_loops;
693 ret.bilinear_weights_fp16 = move(bilinear_weights_fp16);
694 ret.bilinear_weights_fp32 = move(bilinear_weights_fp32);
698 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
700 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
702 assert(input_width > 0);
703 assert(input_height > 0);
704 assert(output_width > 0);
705 assert(output_height > 0);
707 if (input_width != last_input_width ||
708 input_height != last_input_height ||
709 output_width != last_output_width ||
710 output_height != last_output_height ||
711 offset != last_offset ||
713 update_texture(glsl_program_num, prefix, sampler_num);
714 last_input_width = input_width;
715 last_input_height = input_height;
716 last_output_width = output_width;
717 last_output_height = output_height;
718 last_offset = offset;
722 glActiveTexture(GL_TEXTURE0 + *sampler_num);
724 glBindTexture(GL_TEXTURE_2D, texnum);
727 uniform_sample_tex = *sampler_num;
729 uniform_num_samples = src_bilinear_samples;
730 uniform_num_loops = num_loops;
731 uniform_slice_height = slice_height;
733 // Instructions for how to convert integer sample numbers to positions in the weight texture.
734 uniform_sample_x_scale = 1.0f / src_bilinear_samples;
735 uniform_sample_x_offset = 0.5f / src_bilinear_samples;
737 if (direction == SingleResamplePassEffect::VERTICAL) {
738 uniform_whole_pixel_offset = lrintf(offset) / float(input_height);
740 uniform_whole_pixel_offset = lrintf(offset) / float(input_width);
743 // We specifically do not want mipmaps on the input texture;
744 // they break minification.
745 Node *self = chain->find_node_for_effect(this);
746 if (chain->has_input_sampler(self, 0)) {
747 glActiveTexture(chain->get_input_sampler(self, 0));
749 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);