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 3.0
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;
39 return 1.0f - fabs(x);
45 float lanczos_weight(float x)
47 if (fabs(x) > LANCZOS_RADIUS) {
50 return sinc(M_PI * x) * sinc((M_PI / LANCZOS_RADIUS) * x);
54 // The weight function can be expensive to compute over and over again
55 // (which will happen during e.g. a zoom), but it is also easy to interpolate
56 // linearly. We compute the right half of the function (in the range of
57 // 0..LANCZOS_RADIUS), with two guard elements for easier interpolation, and
58 // linearly interpolate to get our function.
60 // We want to scale the table so that the maximum error is always smaller
61 // than 1e-6. As per http://www-solar.mcs.st-andrews.ac.uk/~clare/Lectures/num-analysis/Numan_chap3.pdf,
62 // the error for interpolating a function linearly between points [a,b] is
64 // e = 1/2 (x-a)(x-b) f''(u_x)
66 // for some point u_x in [a,b] (where f(x) is our Lanczos function; we're
67 // assuming LANCZOS_RADIUS=3 from here on). Obviously this is bounded by
68 // f''(x) over the entire range. Numeric optimization shows the maximum of
69 // |f''(x)| to be in x=1.09369819474562880, with the value 2.40067758733152381.
70 // So if the steps between consecutive values are called d, we get
72 // |e| <= 1/2 (d/2)^2 2.4007
75 // Solve for e = 1e-6 yields a step size of 0.0027, which to cover the range
76 // 0..3 needs 1109 steps. We round up to the next power of two, just to be sure.
78 // You need to call lanczos_table_init_done before the first call to
79 // lanczos_weight_cached.
80 #define LANCZOS_TABLE_SIZE 2048
81 bool lanczos_table_init_done = false;
82 float lanczos_table[LANCZOS_TABLE_SIZE + 2];
84 void init_lanczos_table()
86 for (unsigned i = 0; i < LANCZOS_TABLE_SIZE + 2; ++i) {
87 lanczos_table[i] = lanczos_weight(float(i) * (LANCZOS_RADIUS / LANCZOS_TABLE_SIZE));
89 lanczos_table_init_done = true;
92 float lanczos_weight_cached(float x)
95 if (x > LANCZOS_RADIUS) {
98 float table_pos = x * (LANCZOS_TABLE_SIZE / LANCZOS_RADIUS);
99 int table_pos_int = int(table_pos); // Truncate towards zero.
100 float table_pos_frac = table_pos - table_pos_int;
101 assert(table_pos < LANCZOS_TABLE_SIZE + 2);
102 return lanczos_table[table_pos_int] +
103 table_pos_frac * (lanczos_table[table_pos_int + 1] - lanczos_table[table_pos_int]);
106 // Euclid's algorithm, from Wikipedia.
107 unsigned gcd(unsigned a, unsigned b)
117 template<class DestFloat>
118 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)
120 // Cut off near-zero values at both sides.
121 unsigned num_samples_saved = 0;
122 while (num_samples_saved < max_samples_saved &&
123 num_src_samples > 0 &&
124 fabs(src[0].weight) < 1e-6) {
129 while (num_samples_saved < max_samples_saved &&
130 num_src_samples > 0 &&
131 fabs(src[num_src_samples - 1].weight) < 1e-6) {
136 for (unsigned i = 0, j = 0; i < num_src_samples; ++i, ++j) {
137 // Copy the sample directly; it will be overwritten later if we can combine.
139 dst[j].weight = convert_float<float, DestFloat>(src[i].weight);
140 dst[j].pos = convert_float<float, DestFloat>(src[i].pos);
143 if (i == num_src_samples - 1) {
144 // Last sample; cannot combine.
147 assert(num_samples_saved <= max_samples_saved);
148 if (num_samples_saved == max_samples_saved) {
149 // We could maybe save more here, but other rows can't, so don't bother.
153 float w1 = src[i].weight;
154 float w2 = src[i + 1].weight;
155 if (w1 * w2 < 0.0f) {
156 // Differing signs; cannot combine.
160 float pos1 = src[i].pos;
161 float pos2 = src[i + 1].pos;
164 DestFloat pos, total_weight;
166 combine_two_samples(w1, w2, pos1, pos2, num_subtexels, inv_num_subtexels, &pos, &total_weight, &sum_sq_error);
168 // If the interpolation error is larger than that of about sqrt(2) of
169 // a level at 8-bit precision, don't combine. (You'd think 1.0 was enough,
170 // but since the artifacts are not really random, they can get quite
171 // visible. On the other hand, going to 0.25f, I can see no change at
172 // all with 8-bit output, so it would not seem to be worth it.)
173 if (sum_sq_error > 0.5f / (255.0f * 255.0f)) {
177 // OK, we can combine this and the next sample.
179 dst[j].weight = total_weight;
183 ++i; // Skip the next sample.
186 return num_samples_saved;
189 // Normalize so that the sum becomes one. Note that we do it twice;
190 // this sometimes helps a tiny little bit when we have many samples.
192 void normalize_sum(Tap<T>* vals, unsigned num)
194 for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
196 for (unsigned i = 0; i < num; ++i) {
197 sum += to_fp64(vals[i].weight);
199 double inv_sum = 1.0 / sum;
200 for (unsigned i = 0; i < num; ++i) {
201 vals[i].weight = from_fp64<T>(to_fp64(vals[i].weight) * inv_sum);
206 // Make use of the bilinear filtering in the GPU to reduce the number of samples
207 // we need to make. This is a bit more complex than BlurEffect since we cannot combine
208 // two neighboring samples if their weights have differing signs, so we first need to
209 // figure out the maximum number of samples. Then, we downconvert all the weights to
210 // that number -- we could have gone for a variable-length system, but this is simpler,
211 // and the gains would probably be offset by the extra cost of checking when to stop.
213 // The greedy strategy for combining samples is optimal.
214 template<class DestFloat>
215 unsigned combine_many_samples(const Tap<float> *weights, unsigned src_size, unsigned src_samples, unsigned dst_samples, Tap<DestFloat> **bilinear_weights)
217 float num_subtexels = src_size / movit_texel_subpixel_precision;
218 float inv_num_subtexels = movit_texel_subpixel_precision / src_size;
220 unsigned max_samples_saved = UINT_MAX;
221 for (unsigned y = 0; y < dst_samples && max_samples_saved > 0; ++y) {
222 unsigned num_samples_saved = combine_samples<DestFloat>(weights + y * src_samples, NULL, num_subtexels, inv_num_subtexels, src_samples, max_samples_saved);
223 max_samples_saved = min(max_samples_saved, num_samples_saved);
226 // Now that we know the right width, actually combine the samples.
227 unsigned src_bilinear_samples = src_samples - max_samples_saved;
228 *bilinear_weights = new Tap<DestFloat>[dst_samples * src_bilinear_samples];
229 for (unsigned y = 0; y < dst_samples; ++y) {
230 Tap<DestFloat> *bilinear_weights_ptr = *bilinear_weights + y * src_bilinear_samples;
231 unsigned num_samples_saved = combine_samples(
232 weights + y * src_samples,
233 bilinear_weights_ptr,
238 assert(num_samples_saved == max_samples_saved);
239 normalize_sum(bilinear_weights_ptr, src_bilinear_samples);
241 return src_bilinear_samples;
244 // Compute the sum of squared errors between the ideal weights (which are
245 // assumed to fall exactly on pixel centers) and the weights that result
246 // from sampling at <bilinear_weights>. The primary reason for the difference
247 // is inaccuracy in the sampling positions, both due to limited precision
248 // in storing them (already inherent in sending them in as fp16_int_t)
249 // and in subtexel sampling precision (which we calculate in this function).
251 double compute_sum_sq_error(const Tap<float>* weights, unsigned num_weights,
252 const Tap<T>* bilinear_weights, unsigned num_bilinear_weights,
255 // Find the effective range of the bilinear-optimized kernel.
256 // Due to rounding of the positions, this is not necessarily the same
257 // as the intended range (ie., the range of the original weights).
258 int lower_pos = int(floor(to_fp64(bilinear_weights[0].pos) * size - 0.5));
259 int upper_pos = int(ceil(to_fp64(bilinear_weights[num_bilinear_weights - 1].pos) * size - 0.5)) + 2;
260 lower_pos = min<int>(lower_pos, lrintf(weights[0].pos * size - 0.5));
261 upper_pos = max<int>(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5) + 1);
263 float* effective_weights = new float[upper_pos - lower_pos];
264 for (int i = 0; i < upper_pos - lower_pos; ++i) {
265 effective_weights[i] = 0.0f;
268 // Now find the effective weights that result from this sampling.
269 for (unsigned i = 0; i < num_bilinear_weights; ++i) {
270 const float pixel_pos = to_fp64(bilinear_weights[i].pos) * size - 0.5f;
271 const int x0 = int(floor(pixel_pos)) - lower_pos;
272 const int x1 = x0 + 1;
273 const float f = lrintf((pixel_pos - (x0 + lower_pos)) / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;
277 assert(x0 < upper_pos - lower_pos);
278 assert(x1 < upper_pos - lower_pos);
280 effective_weights[x0] += to_fp64(bilinear_weights[i].weight) * (1.0 - f);
281 effective_weights[x1] += to_fp64(bilinear_weights[i].weight) * f;
284 // Subtract the desired weights to get the error.
285 for (unsigned i = 0; i < num_weights; ++i) {
286 const int x = lrintf(weights[i].pos * size - 0.5f) - lower_pos;
288 assert(x < upper_pos - lower_pos);
290 effective_weights[x] -= weights[i].weight;
293 double sum_sq_error = 0.0;
294 for (unsigned i = 0; i < num_weights; ++i) {
295 sum_sq_error += effective_weights[i] * effective_weights[i];
298 delete[] effective_weights;
304 ResampleEffect::ResampleEffect()
307 offset_x(0.0f), offset_y(0.0f),
308 zoom_x(1.0f), zoom_y(1.0f),
309 zoom_center_x(0.5f), zoom_center_y(0.5f)
311 register_int("width", &output_width);
312 register_int("height", &output_height);
314 // The first blur pass will forward resolution information to us.
315 hpass = new SingleResamplePassEffect(this);
316 CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL));
317 vpass = new SingleResamplePassEffect(NULL);
318 CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL));
323 void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self)
325 Node *hpass_node = graph->add_node(hpass);
326 Node *vpass_node = graph->add_node(vpass);
327 graph->connect_nodes(hpass_node, vpass_node);
328 graph->replace_receiver(self, hpass_node);
329 graph->replace_sender(self, vpass_node);
330 self->disabled = true;
333 // We get this information forwarded from the first blur pass,
334 // since we are not part of the chain ourselves.
335 void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsigned height)
337 assert(input_num == 0);
341 input_height = height;
345 void ResampleEffect::update_size()
348 ok |= hpass->set_int("input_width", input_width);
349 ok |= hpass->set_int("input_height", input_height);
350 ok |= hpass->set_int("output_width", output_width);
351 ok |= hpass->set_int("output_height", input_height);
353 ok |= vpass->set_int("input_width", output_width);
354 ok |= vpass->set_int("input_height", input_height);
355 ok |= vpass->set_int("output_width", output_width);
356 ok |= vpass->set_int("output_height", output_height);
360 // The offset added due to zoom may have changed with the size.
361 update_offset_and_zoom();
364 void ResampleEffect::update_offset_and_zoom()
368 // Zoom from the right origin. (zoom_center is given in normalized coordinates,
370 float extra_offset_x = zoom_center_x * (1.0f - 1.0f / zoom_x) * input_width;
371 float extra_offset_y = (1.0f - zoom_center_y) * (1.0f - 1.0f / zoom_y) * input_height;
373 ok |= hpass->set_float("offset", extra_offset_x + offset_x);
374 ok |= vpass->set_float("offset", extra_offset_y - offset_y); // Compensate for the bottom-left origin.
375 ok |= hpass->set_float("zoom", zoom_x);
376 ok |= vpass->set_float("zoom", zoom_y);
381 bool ResampleEffect::set_float(const string &key, float value) {
382 if (key == "width") {
383 output_width = value;
387 if (key == "height") {
388 output_height = value;
394 update_offset_and_zoom();
399 update_offset_and_zoom();
402 if (key == "zoom_x") {
407 update_offset_and_zoom();
410 if (key == "zoom_y") {
415 update_offset_and_zoom();
418 if (key == "zoom_center_x") {
419 zoom_center_x = value;
420 update_offset_and_zoom();
423 if (key == "zoom_center_y") {
424 zoom_center_y = value;
425 update_offset_and_zoom();
431 SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
433 direction(HORIZONTAL),
438 last_input_width(-1),
439 last_input_height(-1),
440 last_output_width(-1),
441 last_output_height(-1),
442 last_offset(0.0 / 0.0), // NaN.
443 last_zoom(0.0 / 0.0), // NaN.
444 last_texture_width(-1), last_texture_height(-1)
446 register_int("direction", (int *)&direction);
447 register_int("input_width", &input_width);
448 register_int("input_height", &input_height);
449 register_int("output_width", &output_width);
450 register_int("output_height", &output_height);
451 register_float("offset", &offset);
452 register_float("zoom", &zoom);
453 register_uniform_sampler2d("sample_tex", &uniform_sample_tex);
454 register_uniform_int("num_samples", &uniform_num_samples);
455 register_uniform_float("num_loops", &uniform_num_loops);
456 register_uniform_float("slice_height", &uniform_slice_height);
457 register_uniform_float("sample_x_scale", &uniform_sample_x_scale);
458 register_uniform_float("sample_x_offset", &uniform_sample_x_offset);
459 register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset);
461 glGenTextures(1, &texnum);
463 if (!lanczos_table_init_done) {
464 // Could in theory race between two threads if we are unlucky,
465 // but that is harmless, since they'll write the same data.
466 init_lanczos_table();
470 SingleResamplePassEffect::~SingleResamplePassEffect()
472 glDeleteTextures(1, &texnum);
475 string SingleResamplePassEffect::output_fragment_shader()
478 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
479 return buf + read_file("resample_effect.frag");
482 // Using vertical scaling as an example:
484 // Generally out[y] = w0 * in[yi] + w1 * in[yi + 1] + w2 * in[yi + 2] + ...
486 // Obviously, yi will depend on y (in a not-quite-linear way), but so will
487 // the weights w0, w1, w2, etc.. The easiest way of doing this is to encode,
488 // for each sample, the weight and the yi value, e.g. <yi, w0>, <yi + 1, w1>,
489 // and so on. For each y, we encode these along the x-axis (since that is spare),
490 // so out[0] will read from parameters <x,y> = <0,0>, <1,0>, <2,0> and so on.
492 // For horizontal scaling, we fill in the exact same texture;
493 // the shader just interprets it differently.
494 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
496 unsigned src_size, dst_size;
497 if (direction == SingleResamplePassEffect::HORIZONTAL) {
498 assert(input_height == output_height);
499 src_size = input_width;
500 dst_size = output_width;
501 } else if (direction == SingleResamplePassEffect::VERTICAL) {
502 assert(input_width == output_width);
503 src_size = input_height;
504 dst_size = output_height;
509 // For many resamplings (e.g. 640 -> 1280), we will end up with the same
510 // set of samples over and over again in a loop. Thus, we can compute only
511 // the first such loop, and then ask the card to repeat the texture for us.
512 // This is both easier on the texture cache and lowers our CPU cost for
513 // generating the kernel somewhat.
514 float scaling_factor;
515 if (fabs(zoom - 1.0f) < 1e-6) {
516 num_loops = gcd(src_size, dst_size);
517 scaling_factor = float(dst_size) / float(src_size);
519 // If zooming is enabled (ie., zoom != 1), we turn off the looping.
520 // We _could_ perhaps do it for rational zoom levels (especially
521 // things like 2:1), but it doesn't seem to be worth it, given that
522 // the most common use case would seem to be varying the zoom
523 // from frame to frame.
525 scaling_factor = zoom * float(dst_size) / float(src_size);
527 slice_height = 1.0f / num_loops;
528 unsigned dst_samples = dst_size / num_loops;
530 // Sample the kernel in the right place. A diagram with a triangular kernel
531 // (corresponding to linear filtering, and obviously with radius 1)
532 // for easier ASCII art drawing:
538 // x---x---x x x---x---x---x
540 // Scaling up (in this case, 2x) means sampling more densely:
546 // x-x-x-x-x-x x x x-x-x-x-x-x-x-x
548 // When scaling up, any destination pixel will only be influenced by a few
549 // (in this case, two) neighboring pixels, and more importantly, the number
550 // will not be influenced by the scaling factor. (Note, however, that the
551 // pixel centers have moved, due to OpenGL's center-pixel convention.)
552 // The only thing that changes is the weights themselves, as the sampling
553 // points are at different distances from the original pixels.
555 // Scaling down is a different story:
561 // --x------ x --x-------x--
563 // Again, the pixel centers have moved in a maybe unintuitive fashion,
564 // although when you consider that there are multiple source pixels around,
565 // it's not so bad as at first look:
571 // --x-------x-------x-------x--
573 // As you can see, the new pixels become averages of the two neighboring old
574 // ones (the situation for Lanczos is of course more complex).
576 // Anyhow, in this case we clearly need to look at more source pixels
577 // to compute the destination pixel, and how many depend on the scaling factor.
578 // Thus, the kernel width will vary with how much we scale.
579 float radius_scaling_factor = min(scaling_factor, 1.0f);
580 int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
581 int src_samples = int_radius * 2 + 1;
582 Tap<float> *weights = new Tap<float>[dst_samples * src_samples];
583 float subpixel_offset = offset - lrintf(offset); // The part not covered by whole_pixel_offset.
584 assert(subpixel_offset >= -0.5f && subpixel_offset <= 0.5f);
585 for (unsigned y = 0; y < dst_samples; ++y) {
586 // Find the point around which we want to sample the source image,
587 // compensating for differing pixel centers as the scale changes.
588 float center_src_y = (y + 0.5f) / scaling_factor - 0.5f;
589 int base_src_y = lrintf(center_src_y);
591 // Now sample <int_radius> pixels on each side around that point.
592 for (int i = 0; i < src_samples; ++i) {
593 int src_y = base_src_y + i - int_radius;
594 float weight = lanczos_weight_cached(radius_scaling_factor * (src_y - center_src_y - subpixel_offset));
595 weights[y * src_samples + i].weight = weight * radius_scaling_factor;
596 weights[y * src_samples + i].pos = (src_y + 0.5) / float(src_size);
600 // Now make use of the bilinear filtering in the GPU to reduce the number of samples
601 // we need to make. Try fp16 first; if it's not accurate enough, we go to fp32.
602 // Our tolerance level for total error is a bit higher than the one for invididual
603 // samples, since one would assume overall errors in the shape don't matter as much.
604 const float max_error = 2.0f / (255.0f * 255.0f);
605 Tap<fp16_int_t> *bilinear_weights_fp16;
606 src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp16);
607 Tap<float> *bilinear_weights_fp32 = NULL;
608 bool fallback_to_fp32 = false;
609 double max_sum_sq_error_fp16 = 0.0;
610 for (unsigned y = 0; y < dst_samples; ++y) {
611 double sum_sq_error_fp16 = compute_sum_sq_error(
612 weights + y * src_samples, src_samples,
613 bilinear_weights_fp16 + y * src_bilinear_samples, src_bilinear_samples,
615 max_sum_sq_error_fp16 = std::max(max_sum_sq_error_fp16, sum_sq_error_fp16);
616 if (max_sum_sq_error_fp16 > max_error) {
621 if (max_sum_sq_error_fp16 > max_error) {
622 fallback_to_fp32 = true;
623 src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp32);
626 // Encode as a two-component texture. Note the GL_REPEAT.
627 glActiveTexture(GL_TEXTURE0 + *sampler_num);
629 glBindTexture(GL_TEXTURE_2D, texnum);
631 if (last_texture_width == -1) {
632 // Need to set this state the first time.
633 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
635 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
637 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
641 GLenum type, internal_format;
643 if (fallback_to_fp32) {
645 internal_format = GL_RG32F;
646 pixels = bilinear_weights_fp32;
648 type = GL_HALF_FLOAT;
649 internal_format = GL_RG16F;
650 pixels = bilinear_weights_fp16;
653 if (int(src_bilinear_samples) == last_texture_width &&
654 int(dst_samples) == last_texture_height &&
655 internal_format == last_texture_internal_format) {
656 // Texture dimensions and type are unchanged; it is more efficient
657 // to just update it rather than making an entirely new texture.
658 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, src_bilinear_samples, dst_samples, GL_RG, type, pixels);
660 glTexImage2D(GL_TEXTURE_2D, 0, internal_format, src_bilinear_samples, dst_samples, 0, GL_RG, type, pixels);
661 last_texture_width = src_bilinear_samples;
662 last_texture_height = dst_samples;
663 last_texture_internal_format = internal_format;
668 delete[] bilinear_weights_fp16;
669 delete[] bilinear_weights_fp32;
672 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
674 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
676 assert(input_width > 0);
677 assert(input_height > 0);
678 assert(output_width > 0);
679 assert(output_height > 0);
681 if (input_width != last_input_width ||
682 input_height != last_input_height ||
683 output_width != last_output_width ||
684 output_height != last_output_height ||
685 offset != last_offset ||
687 update_texture(glsl_program_num, prefix, sampler_num);
688 last_input_width = input_width;
689 last_input_height = input_height;
690 last_output_width = output_width;
691 last_output_height = output_height;
692 last_offset = offset;
696 glActiveTexture(GL_TEXTURE0 + *sampler_num);
698 glBindTexture(GL_TEXTURE_2D, texnum);
701 uniform_sample_tex = *sampler_num;
703 uniform_num_samples = src_bilinear_samples;
704 uniform_num_loops = num_loops;
705 uniform_slice_height = slice_height;
707 // Instructions for how to convert integer sample numbers to positions in the weight texture.
708 uniform_sample_x_scale = 1.0f / src_bilinear_samples;
709 uniform_sample_x_offset = 0.5f / src_bilinear_samples;
711 if (direction == SingleResamplePassEffect::VERTICAL) {
712 uniform_whole_pixel_offset = lrintf(offset) / float(input_height);
714 uniform_whole_pixel_offset = lrintf(offset) / float(input_width);
717 // We specifically do not want mipmaps on the input texture;
718 // they break minification.
719 Node *self = chain->find_node_for_effect(this);
720 if (chain->has_input_sampler(self, 0)) {
721 glActiveTexture(chain->get_input_sampler(self, 0));
723 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);