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
13 #include <Eigen/Sparse>
14 #include <Eigen/SparseQR>
15 #include <Eigen/OrderingMethods>
17 #include "effect_chain.h"
18 #include "effect_util.h"
21 #include "resample_effect.h"
24 using namespace Eigen;
34 return 1.0f - fabs(x);
40 float lanczos_weight(float x)
42 if (fabs(x) > LANCZOS_RADIUS) {
45 return sinc(M_PI * x) * sinc((M_PI / LANCZOS_RADIUS) * x);
49 // The weight function can be expensive to compute over and over again
50 // (which will happen during e.g. a zoom), but it is also easy to interpolate
51 // linearly. We compute the right half of the function (in the range of
52 // 0..LANCZOS_RADIUS), with two guard elements for easier interpolation, and
53 // linearly interpolate to get our function.
55 // We want to scale the table so that the maximum error is always smaller
56 // than 1e-6. As per http://www-solar.mcs.st-andrews.ac.uk/~clare/Lectures/num-analysis/Numan_chap3.pdf,
57 // the error for interpolating a function linearly between points [a,b] is
59 // e = 1/2 (x-a)(x-b) f''(u_x)
61 // for some point u_x in [a,b] (where f(x) is our Lanczos function; we're
62 // assuming LANCZOS_RADIUS=3 from here on). Obviously this is bounded by
63 // f''(x) over the entire range. Numeric optimization shows the maximum of
64 // |f''(x)| to be in x=1.09369819474562880, with the value 2.40067758733152381.
65 // So if the steps between consecutive values are called d, we get
67 // |e| <= 1/2 (d/2)^2 2.4007
70 // Solve for e = 1e-6 yields a step size of 0.0027, which to cover the range
71 // 0..3 needs 1109 steps. We round up to the next power of two, just to be sure.
73 // You need to call lanczos_table_init_done before the first call to
74 // lanczos_weight_cached.
75 #define LANCZOS_TABLE_SIZE 2048
76 static once_flag lanczos_table_init_done;
77 float lanczos_table[LANCZOS_TABLE_SIZE + 2];
79 void init_lanczos_table()
81 for (unsigned i = 0; i < LANCZOS_TABLE_SIZE + 2; ++i) {
82 lanczos_table[i] = lanczos_weight(float(i) * (LANCZOS_RADIUS / LANCZOS_TABLE_SIZE));
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);
201 void normalize_sum(T* vals, unsigned num)
203 for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) {
205 for (unsigned i = 0; i < num; ++i) {
206 sum += to_fp32(vals[i]);
208 float inv_sum = 1.0 / sum;
209 for (unsigned i = 0; i < num; ++i) {
210 vals[i] = from_fp32<T>(to_fp32(vals[i]) * inv_sum);
215 // Make use of the bilinear filtering in the GPU to reduce the number of samples
216 // we need to make. This is a bit more complex than BlurEffect since we cannot combine
217 // two neighboring samples if their weights have differing signs, so we first need to
218 // figure out the maximum number of samples. Then, we downconvert all the weights to
219 // that number -- we could have gone for a variable-length system, but this is simpler,
220 // and the gains would probably be offset by the extra cost of checking when to stop.
222 // The greedy strategy for combining samples is optimal.
223 template<class DestFloat>
224 unsigned combine_many_samples(const Tap<float> *weights, unsigned src_size, unsigned src_samples, unsigned dst_samples, unique_ptr<Tap<DestFloat>[]> *bilinear_weights)
226 float num_subtexels = src_size / movit_texel_subpixel_precision;
227 float inv_num_subtexels = movit_texel_subpixel_precision / src_size;
228 float pos1_pos2_diff = 1.0f / src_size;
229 float inv_pos1_pos2_diff = src_size;
231 unsigned max_samples_saved = UINT_MAX;
232 for (unsigned y = 0; y < dst_samples && max_samples_saved > 0; ++y) {
233 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);
234 max_samples_saved = min(max_samples_saved, num_samples_saved);
237 // Now that we know the right width, actually combine the samples.
238 unsigned src_bilinear_samples = src_samples - max_samples_saved;
239 bilinear_weights->reset(new Tap<DestFloat>[dst_samples * src_bilinear_samples]);
240 for (unsigned y = 0; y < dst_samples; ++y) {
241 Tap<DestFloat> *bilinear_weights_ptr = bilinear_weights->get() + y * src_bilinear_samples;
242 unsigned num_samples_saved = combine_samples(
243 weights + y * src_samples,
244 bilinear_weights_ptr,
251 assert(num_samples_saved == max_samples_saved);
252 normalize_sum(bilinear_weights_ptr, src_bilinear_samples);
254 return src_bilinear_samples;
257 // Compute the sum of squared errors between the ideal weights (which are
258 // assumed to fall exactly on pixel centers) and the weights that result
259 // from sampling at <bilinear_weights>. The primary reason for the difference
260 // is inaccuracy in the sampling positions, both due to limited precision
261 // in storing them (already inherent in sending them in as fp16_int_t)
262 // and in subtexel sampling precision (which we calculate in this function).
264 double compute_sum_sq_error(const Tap<float>* weights, unsigned num_weights,
265 const Tap<T>* bilinear_weights, unsigned num_bilinear_weights,
268 // Find the effective range of the bilinear-optimized kernel.
269 // Due to rounding of the positions, this is not necessarily the same
270 // as the intended range (ie., the range of the original weights).
271 int lower_pos = int(floor(to_fp32(bilinear_weights[0].pos) * size - 0.5f));
272 int upper_pos = int(ceil(to_fp32(bilinear_weights[num_bilinear_weights - 1].pos) * size - 0.5f)) + 2;
273 lower_pos = min<int>(lower_pos, lrintf(weights[0].pos * size - 0.5f));
274 upper_pos = max<int>(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5f) + 1);
276 float* effective_weights = new float[upper_pos - lower_pos];
277 for (int i = 0; i < upper_pos - lower_pos; ++i) {
278 effective_weights[i] = 0.0f;
281 // Now find the effective weights that result from this sampling.
282 for (unsigned i = 0; i < num_bilinear_weights; ++i) {
283 const float pixel_pos = to_fp32(bilinear_weights[i].pos) * size - 0.5f;
284 const int x0 = int(floor(pixel_pos)) - lower_pos;
285 const int x1 = x0 + 1;
286 const float f = lrintf((pixel_pos - (x0 + lower_pos)) / movit_texel_subpixel_precision) * movit_texel_subpixel_precision;
290 assert(x0 < upper_pos - lower_pos);
291 assert(x1 < upper_pos - lower_pos);
293 effective_weights[x0] += to_fp32(bilinear_weights[i].weight) * (1.0f - f);
294 effective_weights[x1] += to_fp32(bilinear_weights[i].weight) * f;
297 // Subtract the desired weights to get the error.
298 for (unsigned i = 0; i < num_weights; ++i) {
299 const int x = lrintf(weights[i].pos * size - 0.5f) - lower_pos;
301 assert(x < upper_pos - lower_pos);
303 effective_weights[x] -= weights[i].weight;
306 double sum_sq_error = 0.0;
307 for (unsigned i = 0; i < num_weights; ++i) {
308 sum_sq_error += effective_weights[i] * effective_weights[i];
311 delete[] effective_weights;
317 ResampleEffect::ResampleEffect()
320 offset_x(0.0f), offset_y(0.0f),
321 zoom_x(1.0f), zoom_y(1.0f),
322 zoom_center_x(0.5f), zoom_center_y(0.5f)
324 register_int("width", &output_width);
325 register_int("height", &output_height);
327 if (movit_compute_shaders_supported) {
328 // The effect will forward resolution information to us.
329 compute_effect_owner.reset(new ResampleComputeEffect(this));
330 compute_effect = compute_effect_owner.get();
332 // The first blur pass will forward resolution information to us.
333 hpass_owner.reset(new SingleResamplePassEffect(this));
334 hpass = hpass_owner.get();
335 CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL));
336 vpass_owner.reset(new SingleResamplePassEffect(this));
337 vpass = vpass_owner.get();
338 CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL));
344 ResampleEffect::~ResampleEffect()
348 void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self)
350 if (compute_effect != nullptr) {
351 Node *compute_node = graph->add_node(compute_effect_owner.release());
352 graph->replace_receiver(self, compute_node);
353 graph->replace_sender(self, compute_node);
355 Node *hpass_node = graph->add_node(hpass_owner.release());
356 Node *vpass_node = graph->add_node(vpass_owner.release());
357 graph->connect_nodes(hpass_node, vpass_node);
358 graph->replace_receiver(self, hpass_node);
359 graph->replace_sender(self, vpass_node);
361 self->disabled = true;
364 // We get this information forwarded from the first blur pass,
365 // since we are not part of the chain ourselves.
366 void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsigned height)
368 assert(input_num == 0);
372 input_height = height;
376 void ResampleEffect::update_size()
379 if (compute_effect != nullptr) {
380 ok |= compute_effect->set_int("input_width", input_width);
381 ok |= compute_effect->set_int("input_height", input_height);
382 ok |= compute_effect->set_int("output_width", output_width);
383 ok |= compute_effect->set_int("output_height", output_height);
385 ok |= hpass->set_int("input_width", input_width);
386 ok |= hpass->set_int("input_height", input_height);
387 ok |= hpass->set_int("output_width", output_width);
388 ok |= hpass->set_int("output_height", input_height);
390 ok |= vpass->set_int("input_width", output_width);
391 ok |= vpass->set_int("input_height", input_height);
392 ok |= vpass->set_int("output_width", output_width);
393 ok |= vpass->set_int("output_height", output_height);
397 // The offset added due to zoom may have changed with the size.
398 update_offset_and_zoom();
401 void ResampleEffect::update_offset_and_zoom()
405 // Zoom from the right origin. (zoom_center is given in normalized coordinates,
407 float extra_offset_x = zoom_center_x * (1.0f - 1.0f / zoom_x) * input_width;
408 float extra_offset_y = (1.0f - zoom_center_y) * (1.0f - 1.0f / zoom_y) * input_height;
410 if (compute_effect != nullptr) {
411 ok |= compute_effect->set_float("offset_x", extra_offset_x + offset_x);
412 ok |= compute_effect->set_float("offset_y", extra_offset_y - offset_y); // Compensate for the bottom-left origin.
413 ok |= compute_effect->set_float("zoom_x", zoom_x);
414 ok |= compute_effect->set_float("zoom_y", zoom_y);
416 ok |= hpass->set_float("offset", extra_offset_x + offset_x);
417 ok |= vpass->set_float("offset", extra_offset_y - offset_y); // Compensate for the bottom-left origin.
418 ok |= hpass->set_float("zoom", zoom_x);
419 ok |= vpass->set_float("zoom", zoom_y);
425 bool ResampleEffect::set_float(const string &key, float value) {
426 if (key == "width") {
427 output_width = value;
431 if (key == "height") {
432 output_height = value;
438 update_offset_and_zoom();
443 update_offset_and_zoom();
446 if (key == "zoom_x") {
451 update_offset_and_zoom();
454 if (key == "zoom_y") {
459 update_offset_and_zoom();
462 if (key == "zoom_center_x") {
463 zoom_center_x = value;
464 update_offset_and_zoom();
467 if (key == "zoom_center_y") {
468 zoom_center_y = value;
469 update_offset_and_zoom();
475 SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent)
477 direction(HORIZONTAL),
482 last_input_width(-1),
483 last_input_height(-1),
484 last_output_width(-1),
485 last_output_height(-1),
486 last_offset(0.0 / 0.0), // NaN.
487 last_zoom(0.0 / 0.0) // NaN.
489 register_int("direction", (int *)&direction);
490 register_int("input_width", &input_width);
491 register_int("input_height", &input_height);
492 register_int("output_width", &output_width);
493 register_int("output_height", &output_height);
494 register_float("offset", &offset);
495 register_float("zoom", &zoom);
496 register_uniform_sampler2d("sample_tex", &uniform_sample_tex);
497 register_uniform_int("num_samples", &uniform_num_samples);
498 register_uniform_float("num_loops", &uniform_num_loops);
499 register_uniform_float("slice_height", &uniform_slice_height);
500 register_uniform_float("sample_x_scale", &uniform_sample_x_scale);
501 register_uniform_float("sample_x_offset", &uniform_sample_x_offset);
502 register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset);
504 call_once(lanczos_table_init_done, init_lanczos_table);
507 SingleResamplePassEffect::~SingleResamplePassEffect()
511 string SingleResamplePassEffect::output_fragment_shader()
514 sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL));
515 return buf + read_file("resample_effect.frag");
518 // Using vertical scaling as an example:
520 // Generally out[y] = w0 * in[yi] + w1 * in[yi + 1] + w2 * in[yi + 2] + ...
522 // Obviously, yi will depend on y (in a not-quite-linear way), but so will
523 // the weights w0, w1, w2, etc.. The easiest way of doing this is to encode,
524 // for each sample, the weight and the yi value, e.g. <yi, w0>, <yi + 1, w1>,
525 // and so on. For each y, we encode these along the x-axis (since that is spare),
526 // so out[0] will read from parameters <x,y> = <0,0>, <1,0>, <2,0> and so on.
528 // For horizontal scaling, we fill in the exact same texture;
529 // the shader just interprets it differently.
530 void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
532 unsigned src_size, dst_size;
533 if (direction == SingleResamplePassEffect::HORIZONTAL) {
534 assert(input_height == output_height);
535 src_size = input_width;
536 dst_size = output_width;
537 } else if (direction == SingleResamplePassEffect::VERTICAL) {
538 assert(input_width == output_width);
539 src_size = input_height;
540 dst_size = output_height;
545 ScalingWeights weights = calculate_bilinear_scaling_weights(src_size, dst_size, zoom, offset, BilinearFormatConstraints::ALLOW_FP16_AND_FP32);
546 src_bilinear_samples = weights.src_bilinear_samples;
547 num_loops = weights.num_loops;
548 slice_height = 1.0f / weights.num_loops;
550 // Encode as a two-component texture. Note the GL_REPEAT.
551 glActiveTexture(GL_TEXTURE0 + *sampler_num);
553 glBindTexture(GL_TEXTURE_2D, tex.get_texnum());
556 GLenum type, internal_format;
558 assert((weights.bilinear_weights_fp16 == nullptr) != (weights.bilinear_weights_fp32 == nullptr));
559 if (weights.bilinear_weights_fp32 != nullptr) {
561 internal_format = GL_RG32F;
562 pixels = weights.bilinear_weights_fp32.get();
564 type = GL_HALF_FLOAT;
565 internal_format = GL_RG16F;
566 pixels = weights.bilinear_weights_fp16.get();
569 tex.update(weights.src_bilinear_samples, weights.dst_samples, internal_format, GL_RG, type, pixels);
572 ResampleComputeEffect::ResampleComputeEffect(ResampleEffect *parent)
580 last_input_width(-1),
581 last_input_height(-1),
582 last_output_width(-1),
583 last_output_height(-1),
584 last_offset_x(0.0 / 0.0), // NaN.
585 last_offset_y(0.0 / 0.0), // NaN.
586 last_zoom_x(0.0 / 0.0), // NaN.
587 last_zoom_y(0.0 / 0.0) // NaN.
589 register_int("input_width", &input_width);
590 register_int("input_height", &input_height);
591 register_int("output_width", &output_width);
592 register_int("output_height", &output_height);
593 register_float("offset_x", &offset_x);
594 register_float("offset_y", &offset_y);
595 register_float("zoom_x", &zoom_x);
596 register_float("zoom_y", &zoom_y);
597 register_uniform_sampler2d("sample_tex_horizontal", &uniform_sample_tex_horizontal);
598 register_uniform_sampler2d("sample_tex_vertical", &uniform_sample_tex_vertical);
599 register_uniform_int("num_horizontal_samples", &uniform_num_horizontal_samples);
600 register_uniform_int("num_vertical_samples", &uniform_num_vertical_samples);
601 register_uniform_int("vertical_int_radius", &uniform_vertical_int_radius);
602 register_uniform_float("inv_vertical_scaling_factor", &uniform_inv_vertical_scaling_factor);
603 register_uniform_int("output_samples_per_block", &uniform_output_samples_per_block);
604 register_uniform_int("num_horizontal_filters", &uniform_num_horizontal_filters);
605 register_uniform_int("num_vertical_filters", &uniform_num_vertical_filters);
606 register_uniform_float("slice_height", &uniform_slice_height);
607 register_uniform_float("horizontal_whole_pixel_offset", &uniform_horizontal_whole_pixel_offset);
608 register_uniform_int("vertical_whole_pixel_offset", &uniform_vertical_whole_pixel_offset);
609 register_uniform_float("inv_input_height", &uniform_inv_input_height);
610 register_uniform_float("input_texcoord_y_adjust", &uniform_input_texcoord_y_adjust);
612 call_once(lanczos_table_init_done, init_lanczos_table);
615 ResampleComputeEffect::~ResampleComputeEffect()
619 string ResampleComputeEffect::output_fragment_shader()
622 return buf + read_file("resample_effect.comp");
625 // The compute shader does horizontal scaling first, using exactly the same
626 // two-component texture format as in the two-pass version (see the comments
627 // on ResampleComputeEffect). The vertical scaling calculates the offset values
628 // in the shader, so we only store a one-component texture with the weights
630 void ResampleComputeEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
632 ScalingWeights horiz_weights = calculate_bilinear_scaling_weights(input_width, output_width, zoom_x, offset_x, BilinearFormatConstraints::ALLOW_FP32_ONLY);
633 ScalingWeights vert_weights = calculate_raw_scaling_weights(input_height, output_height, zoom_y, offset_y);
634 uniform_vertical_int_radius = vert_weights.int_radius;
635 vertical_scaling_factor = vert_weights.scaling_factor;
636 uniform_inv_vertical_scaling_factor = 1.0f / vert_weights.scaling_factor;
637 src_horizontal_bilinear_samples = horiz_weights.src_bilinear_samples;
638 src_vertical_samples = vert_weights.src_bilinear_samples;
639 uniform_num_horizontal_filters = horiz_weights.dst_samples;
640 uniform_num_vertical_filters = vert_weights.dst_samples;
641 slice_height = 1.0f / horiz_weights.num_loops;
643 // Encode as a two-component texture. Note the GL_REPEAT.
644 glActiveTexture(GL_TEXTURE0 + *sampler_num);
646 glBindTexture(GL_TEXTURE_2D, tex_horiz.get_texnum());
649 tex_horiz.update(horiz_weights.src_bilinear_samples, horiz_weights.dst_samples, GL_RG32F, GL_RG, GL_FLOAT, horiz_weights.bilinear_weights_fp32.get());
651 glActiveTexture(GL_TEXTURE0 + *sampler_num + 1);
653 glBindTexture(GL_TEXTURE_2D, tex_vert.get_texnum());
656 // Storing the vertical weights as fp16 instead of fp32 saves a few
657 // percent on NVIDIA, and it doesn't seem to hurt quality any.
658 // (The horizontal weights is a different story, since the offsets
659 // can get large and are fairly accuracy-sensitive. Also, they are
660 // loaded only once per workgroup, at the very beginning.)
661 tex_vert.update(vert_weights.src_bilinear_samples, vert_weights.dst_samples, GL_R16F, GL_RED, GL_HALF_FLOAT, vert_weights.raw_weights.get());
663 // Figure out how many output samples each compute shader block is going to output.
664 int usable_input_samples_per_block = 128 - 2 * uniform_vertical_int_radius;
665 int output_samples_per_block = int(floor(usable_input_samples_per_block * vertical_scaling_factor));
666 if (output_samples_per_block < 1) {
667 output_samples_per_block = 1;
669 uniform_output_samples_per_block = output_samples_per_block;
674 ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset)
676 // Only needed if run from outside ResampleEffect.
677 call_once(lanczos_table_init_done, init_lanczos_table);
679 // For many resamplings (e.g. 640 -> 1280), we will end up with the same
680 // set of samples over and over again in a loop. Thus, we can compute only
681 // the first such loop, and then ask the card to repeat the texture for us.
682 // This is both easier on the texture cache and lowers our CPU cost for
683 // generating the kernel somewhat.
684 float scaling_factor;
686 if (fabs(zoom - 1.0f) < 1e-6) {
687 num_loops = gcd(src_size, dst_size);
688 scaling_factor = float(dst_size) / float(src_size);
690 // If zooming is enabled (ie., zoom != 1), we turn off the looping.
691 // We _could_ perhaps do it for rational zoom levels (especially
692 // things like 2:1), but it doesn't seem to be worth it, given that
693 // the most common use case would seem to be varying the zoom
694 // from frame to frame.
696 scaling_factor = zoom * float(dst_size) / float(src_size);
698 unsigned dst_samples = dst_size / num_loops;
700 // Sample the kernel in the right place. A diagram with a triangular kernel
701 // (corresponding to linear filtering, and obviously with radius 1)
702 // for easier ASCII art drawing:
708 // x---x---x x x---x---x---x
710 // Scaling up (in this case, 2x) means sampling more densely:
716 // x-x-x-x-x-x x x x-x-x-x-x-x-x-x
718 // When scaling up, any destination pixel will only be influenced by a few
719 // (in this case, two) neighboring pixels, and more importantly, the number
720 // will not be influenced by the scaling factor. (Note, however, that the
721 // pixel centers have moved, due to OpenGL's center-pixel convention.)
722 // The only thing that changes is the weights themselves, as the sampling
723 // points are at different distances from the original pixels.
725 // Scaling down is a different story:
731 // --x------ x --x-------x--
733 // Again, the pixel centers have moved in a maybe unintuitive fashion,
734 // although when you consider that there are multiple source pixels around,
735 // it's not so bad as at first look:
741 // --x-------x-------x-------x--
743 // As you can see, the new pixels become averages of the two neighboring old
744 // ones (the situation for Lanczos is of course more complex).
746 // Anyhow, in this case we clearly need to look at more source pixels
747 // to compute the destination pixel, and how many depend on the scaling factor.
748 // Thus, the kernel width will vary with how much we scale.
749 float radius_scaling_factor = min(scaling_factor, 1.0f);
750 const int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor);
751 const int src_samples = int_radius * 2 + 1;
752 unique_ptr<Tap<float>[]> weights(new Tap<float>[dst_samples * src_samples]);
753 float subpixel_offset = offset - lrintf(offset); // The part not covered by whole_pixel_offset.
754 assert(subpixel_offset >= -0.5f && subpixel_offset <= 0.5f);
755 float inv_scaling_factor = 1.0f / scaling_factor;
756 for (unsigned y = 0; y < dst_samples; ++y) {
757 // Find the point around which we want to sample the source image,
758 // compensating for differing pixel centers as the scale changes.
759 float center_src_y = (y + 0.5f) * inv_scaling_factor - 0.5f;
760 int base_src_y = lrintf(center_src_y);
762 // Now sample <int_radius> pixels on each side around that point.
763 float inv_src_size = 1.0 / float(src_size);
764 for (int i = 0; i < src_samples; ++i) {
765 int src_y = base_src_y + i - int_radius;
766 float weight = lanczos_weight_cached(radius_scaling_factor * (src_y - center_src_y - subpixel_offset));
767 weights[y * src_samples + i].weight = weight * radius_scaling_factor;
768 weights[y * src_samples + i].pos = (src_y + 0.5f) * inv_src_size;
773 ret.src_bilinear_samples = src_samples;
774 ret.dst_samples = dst_samples;
775 ret.int_radius = int_radius;
776 ret.scaling_factor = scaling_factor;
777 ret.num_loops = num_loops;
778 ret.bilinear_weights_fp16 = nullptr;
779 ret.bilinear_weights_fp32 = move(weights);
780 ret.raw_weights = nullptr;
786 ScalingWeights calculate_bilinear_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset, BilinearFormatConstraints constraints)
788 ScalingWeights ret = calculate_scaling_weights(src_size, dst_size, zoom, offset);
789 unique_ptr<Tap<float>[]> weights = move(ret.bilinear_weights_fp32);
790 const int src_samples = ret.src_bilinear_samples;
792 // Now make use of the bilinear filtering in the GPU to reduce the number of samples
793 // we need to make. Try fp16 first; if it's not accurate enough, we go to fp32.
794 // Our tolerance level for total error is a bit higher than the one for invididual
795 // samples, since one would assume overall errors in the shape don't matter as much.
796 const float max_error = 2.0f / (255.0f * 255.0f);
797 unique_ptr<Tap<fp16_int_t>[]> bilinear_weights_fp16;
798 unique_ptr<Tap<float>[]> bilinear_weights_fp32;
799 double max_sum_sq_error_fp16 = 0.0;
800 int src_bilinear_samples;
801 if (constraints == BilinearFormatConstraints::ALLOW_FP32_ONLY) {
802 max_sum_sq_error_fp16 = numeric_limits<double>::max();
804 assert(constraints == BilinearFormatConstraints::ALLOW_FP16_AND_FP32);
805 src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, ret.dst_samples, &bilinear_weights_fp16);
806 for (unsigned y = 0; y < ret.dst_samples; ++y) {
807 double sum_sq_error_fp16 = compute_sum_sq_error(
808 weights.get() + y * src_samples, src_samples,
809 bilinear_weights_fp16.get() + y * src_bilinear_samples, src_bilinear_samples,
811 max_sum_sq_error_fp16 = std::max(max_sum_sq_error_fp16, sum_sq_error_fp16);
812 if (max_sum_sq_error_fp16 > max_error) {
818 if (max_sum_sq_error_fp16 > max_error) {
819 bilinear_weights_fp16.reset();
820 src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, ret.dst_samples, &bilinear_weights_fp32);
823 ret.src_bilinear_samples = src_bilinear_samples;
824 ret.bilinear_weights_fp16 = move(bilinear_weights_fp16);
825 ret.bilinear_weights_fp32 = move(bilinear_weights_fp32);
829 // Unlike calculate_bilinear_scaling_weights(), this just converts the weights,
830 // without any combining trickery. Thus, it is also much faster.
831 ScalingWeights calculate_raw_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset)
833 ScalingWeights ret = calculate_scaling_weights(src_size, dst_size, zoom, offset);
834 unique_ptr<Tap<float>[]> weights = move(ret.bilinear_weights_fp32);
835 const int src_samples = ret.src_bilinear_samples;
837 // Convert to fp16 (without any positions, as they are calculated implicitly
838 // by the compute shader) and normalize.
839 unique_ptr<fp16_int_t[]> raw_weights(new fp16_int_t[ret.dst_samples * src_samples]);
840 for (unsigned y = 0; y < ret.dst_samples; ++y) {
841 for (int i = 0; i < src_samples; ++i) {
842 raw_weights[y * src_samples + i] = fp32_to_fp16(weights[y * src_samples + i].weight);
844 normalize_sum(raw_weights.get() + y * src_samples, src_samples);
847 ret.raw_weights = move(raw_weights);
851 void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
853 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
855 assert(input_width > 0);
856 assert(input_height > 0);
857 assert(output_width > 0);
858 assert(output_height > 0);
860 if (input_width != last_input_width ||
861 input_height != last_input_height ||
862 output_width != last_output_width ||
863 output_height != last_output_height ||
864 offset != last_offset ||
866 update_texture(glsl_program_num, prefix, sampler_num);
867 last_input_width = input_width;
868 last_input_height = input_height;
869 last_output_width = output_width;
870 last_output_height = output_height;
871 last_offset = offset;
875 glActiveTexture(GL_TEXTURE0 + *sampler_num);
877 glBindTexture(GL_TEXTURE_2D, tex.get_texnum());
880 uniform_sample_tex = *sampler_num;
882 uniform_num_samples = src_bilinear_samples;
883 uniform_num_loops = num_loops;
884 uniform_slice_height = slice_height;
886 // Instructions for how to convert integer sample numbers to positions in the weight texture.
887 uniform_sample_x_scale = 1.0f / src_bilinear_samples;
888 uniform_sample_x_offset = 0.5f / src_bilinear_samples;
890 if (direction == SingleResamplePassEffect::VERTICAL) {
891 uniform_whole_pixel_offset = lrintf(offset) / float(input_height);
893 uniform_whole_pixel_offset = lrintf(offset) / float(input_width);
897 Support2DTexture::Support2DTexture()
899 glGenTextures(1, &texnum);
901 glBindTexture(GL_TEXTURE_2D, texnum);
903 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
905 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
907 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
911 Support2DTexture::~Support2DTexture()
913 glDeleteTextures(1, &texnum);
917 void Support2DTexture::update(GLint width, GLint height, GLenum internal_format, GLenum format, GLenum type, const GLvoid * data)
919 glBindTexture(GL_TEXTURE_2D, texnum);
921 if (width == last_texture_width &&
922 height == last_texture_height &&
923 internal_format == last_texture_internal_format) {
924 // Texture dimensions and type are unchanged; it is more efficient
925 // to just update it rather than making an entirely new texture.
926 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, data);
929 glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, format, type, data);
931 last_texture_width = width;
932 last_texture_height = height;
933 last_texture_internal_format = internal_format;
937 void ResampleComputeEffect::get_compute_dimensions(unsigned output_width, unsigned output_height,
938 unsigned *x, unsigned *y, unsigned *z) const
941 *y = (output_height + uniform_output_samples_per_block - 1) / uniform_output_samples_per_block;
945 void ResampleComputeEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
947 Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
949 assert(input_width > 0);
950 assert(input_height > 0);
951 assert(output_width > 0);
952 assert(output_height > 0);
954 if (input_width != last_input_width ||
955 input_height != last_input_height ||
956 output_width != last_output_width ||
957 output_height != last_output_height ||
958 offset_x != last_offset_x ||
959 offset_y != last_offset_y ||
960 zoom_x != last_zoom_x ||
961 zoom_x != last_zoom_y) {
962 update_texture(glsl_program_num, prefix, sampler_num);
963 last_input_width = input_width;
964 last_input_height = input_height;
965 last_output_width = output_width;
966 last_output_height = output_height;
967 last_offset_x = offset_x;
968 last_offset_y = offset_y;
969 last_zoom_x = zoom_x;
970 last_zoom_y = zoom_y;
973 glActiveTexture(GL_TEXTURE0 + *sampler_num);
975 glBindTexture(GL_TEXTURE_2D, tex_horiz.get_texnum());
977 uniform_sample_tex_horizontal = *sampler_num;
980 glActiveTexture(GL_TEXTURE0 + *sampler_num);
982 glBindTexture(GL_TEXTURE_2D, tex_vert.get_texnum());
984 uniform_sample_tex_vertical = *sampler_num;
987 uniform_num_horizontal_samples = src_horizontal_bilinear_samples;
988 uniform_num_vertical_samples = src_vertical_samples;
989 uniform_slice_height = slice_height;
991 uniform_horizontal_whole_pixel_offset = lrintf(offset_x) / float(input_width);
992 uniform_vertical_whole_pixel_offset = lrintf(offset_y);
994 uniform_inv_input_height = 1.0f / float(input_height);
995 uniform_input_texcoord_y_adjust = 0.5f / float(input_height);