1 | /* |
2 | Simple DirectMedia Layer |
3 | Copyright (C) 1997-2025 Sam Lantinga <slouken@libsdl.org> |
4 | |
5 | This software is provided 'as-is', without any express or implied |
6 | warranty. In no event will the authors be held liable for any damages |
7 | arising from the use of this software. |
8 | |
9 | Permission is granted to anyone to use this software for any purpose, |
10 | including commercial applications, and to alter it and redistribute it |
11 | freely, subject to the following restrictions: |
12 | |
13 | 1. The origin of this software must not be misrepresented; you must not |
14 | claim that you wrote the original software. If you use this software |
15 | in a product, an acknowledgment in the product documentation would be |
16 | appreciated but is not required. |
17 | 2. Altered source versions must be plainly marked as such, and must not be |
18 | misrepresented as being the original software. |
19 | 3. This notice may not be removed or altered from any source distribution. |
20 | */ |
21 | #include "SDL_internal.h" |
22 | |
23 | #include "SDL_sysaudio.h" |
24 | |
25 | #include "SDL_audioresample.h" |
26 | |
27 | // SDL's resampler uses a "bandlimited interpolation" algorithm: |
28 | // https://ccrma.stanford.edu/~jos/resample/ |
29 | |
30 | // TODO: Support changing this at runtime? |
31 | #if defined(SDL_SSE_INTRINSICS) || defined(SDL_NEON_INTRINSICS) |
32 | // In <current year>, SSE is basically mandatory anyway |
33 | // We want RESAMPLER_SAMPLES_PER_FRAME to be a multiple of 4, to make SIMD easier |
34 | #define RESAMPLER_ZERO_CROSSINGS 6 |
35 | #else |
36 | #define RESAMPLER_ZERO_CROSSINGS 5 |
37 | #endif |
38 | |
39 | #define RESAMPLER_SAMPLES_PER_FRAME (RESAMPLER_ZERO_CROSSINGS * 2) |
40 | |
41 | // For a given srcpos, `srcpos + frame` are sampled, where `-RESAMPLER_ZERO_CROSSINGS < frame <= RESAMPLER_ZERO_CROSSINGS`. |
42 | // Note, when upsampling, it is also possible to start sampling from `srcpos = -1`. |
43 | #define RESAMPLER_MAX_PADDING_FRAMES (RESAMPLER_ZERO_CROSSINGS + 1) |
44 | |
45 | // More bits gives more precision, at the cost of a larger table. |
46 | #define RESAMPLER_BITS_PER_ZERO_CROSSING 3 |
47 | #define RESAMPLER_SAMPLES_PER_ZERO_CROSSING (1 << RESAMPLER_BITS_PER_ZERO_CROSSING) |
48 | #define RESAMPLER_FILTER_INTERP_BITS (32 - RESAMPLER_BITS_PER_ZERO_CROSSING) |
49 | #define RESAMPLER_FILTER_INTERP_RANGE (1 << RESAMPLER_FILTER_INTERP_BITS) |
50 | |
51 | // ResampleFrame is just a vector/matrix/matrix multiplication. |
52 | // It performs cubic interpolation of the filter, then multiplies that with the input. |
53 | // dst = [1, frac, frac^2, frac^3] * filter * src |
54 | |
55 | // Cubic Polynomial |
56 | typedef union Cubic |
57 | { |
58 | float v[4]; |
59 | |
60 | #ifdef SDL_SSE_INTRINSICS |
61 | // Aligned loads can be used directly as memory operands for mul/add |
62 | __m128 v128; |
63 | #endif |
64 | |
65 | #ifdef SDL_NEON_INTRINSICS |
66 | float32x4_t v128; |
67 | #endif |
68 | |
69 | } Cubic; |
70 | |
71 | static void ResampleFrame_Generic(const float *src, float *dst, const Cubic *filter, float frac, int chans) |
72 | { |
73 | const float frac2 = frac * frac; |
74 | const float frac3 = frac * frac2; |
75 | |
76 | int i, chan; |
77 | float scales[RESAMPLER_SAMPLES_PER_FRAME]; |
78 | |
79 | for (i = 0; i < RESAMPLER_SAMPLES_PER_FRAME; ++i, ++filter) { |
80 | scales[i] = filter->v[0] + (filter->v[1] * frac) + (filter->v[2] * frac2) + (filter->v[3] * frac3); |
81 | } |
82 | |
83 | for (chan = 0; chan < chans; ++chan) { |
84 | float out = 0.0f; |
85 | |
86 | for (i = 0; i < RESAMPLER_SAMPLES_PER_FRAME; ++i) { |
87 | out += src[i * chans + chan] * scales[i]; |
88 | } |
89 | |
90 | dst[chan] = out; |
91 | } |
92 | } |
93 | |
94 | static void ResampleFrame_Mono(const float *src, float *dst, const Cubic *filter, float frac, int chans) |
95 | { |
96 | const float frac2 = frac * frac; |
97 | const float frac3 = frac * frac2; |
98 | |
99 | int i; |
100 | float out = 0.0f; |
101 | |
102 | for (i = 0; i < RESAMPLER_SAMPLES_PER_FRAME; ++i, ++filter) { |
103 | // Interpolate between the nearest two filters |
104 | const float scale = filter->v[0] + (filter->v[1] * frac) + (filter->v[2] * frac2) + (filter->v[3] * frac3); |
105 | |
106 | out += src[i] * scale; |
107 | } |
108 | |
109 | dst[0] = out; |
110 | } |
111 | |
112 | static void ResampleFrame_Stereo(const float *src, float *dst, const Cubic *filter, float frac, int chans) |
113 | { |
114 | const float frac2 = frac * frac; |
115 | const float frac3 = frac * frac2; |
116 | |
117 | int i; |
118 | float out0 = 0.0f; |
119 | float out1 = 0.0f; |
120 | |
121 | for (i = 0; i < RESAMPLER_SAMPLES_PER_FRAME; ++i, ++filter) { |
122 | // Interpolate between the nearest two filters |
123 | const float scale = filter->v[0] + (filter->v[1] * frac) + (filter->v[2] * frac2) + (filter->v[3] * frac3); |
124 | |
125 | out0 += src[i * 2 + 0] * scale; |
126 | out1 += src[i * 2 + 1] * scale; |
127 | } |
128 | |
129 | dst[0] = out0; |
130 | dst[1] = out1; |
131 | } |
132 | |
133 | #ifdef SDL_SSE_INTRINSICS |
134 | #define sdl_madd_ps(a, b, c) _mm_add_ps(a, _mm_mul_ps(b, c)) // Not-so-fused multiply-add |
135 | |
136 | static void SDL_TARGETING("sse" ) ResampleFrame_Generic_SSE(const float *src, float *dst, const Cubic *filter, float frac, int chans) |
137 | { |
138 | #if RESAMPLER_SAMPLES_PER_FRAME != 12 |
139 | #error Invalid samples per frame |
140 | #endif |
141 | |
142 | __m128 f0, f1, f2; |
143 | |
144 | { |
145 | const __m128 frac1 = _mm_set1_ps(frac); |
146 | const __m128 frac2 = _mm_mul_ps(frac1, frac1); |
147 | const __m128 frac3 = _mm_mul_ps(frac1, frac2); |
148 | |
149 | // Transposed in SetupAudioResampler |
150 | // Explicitly use _mm_load_ps to workaround ICE in GCC 4.9.4 accessing Cubic.v128 |
151 | #define X(out) \ |
152 | out = _mm_load_ps(filter[0].v); \ |
153 | out = sdl_madd_ps(out, frac1, _mm_load_ps(filter[1].v)); \ |
154 | out = sdl_madd_ps(out, frac2, _mm_load_ps(filter[2].v)); \ |
155 | out = sdl_madd_ps(out, frac3, _mm_load_ps(filter[3].v)); \ |
156 | filter += 4 |
157 | |
158 | X(f0); |
159 | X(f1); |
160 | X(f2); |
161 | |
162 | #undef X |
163 | } |
164 | |
165 | if (chans == 2) { |
166 | // Duplicate each of the filter elements and multiply by the input |
167 | // Use two accumulators to improve throughput |
168 | __m128 out0 = _mm_mul_ps(_mm_loadu_ps(src + 0), _mm_unpacklo_ps(f0, f0)); |
169 | __m128 out1 = _mm_mul_ps(_mm_loadu_ps(src + 4), _mm_unpackhi_ps(f0, f0)); |
170 | out0 = sdl_madd_ps(out0, _mm_loadu_ps(src + 8), _mm_unpacklo_ps(f1, f1)); |
171 | out1 = sdl_madd_ps(out1, _mm_loadu_ps(src + 12), _mm_unpackhi_ps(f1, f1)); |
172 | out0 = sdl_madd_ps(out0, _mm_loadu_ps(src + 16), _mm_unpacklo_ps(f2, f2)); |
173 | out1 = sdl_madd_ps(out1, _mm_loadu_ps(src + 20), _mm_unpackhi_ps(f2, f2)); |
174 | |
175 | // Add the accumulators together |
176 | __m128 out = _mm_add_ps(out0, out1); |
177 | |
178 | // Add the lower and upper pairs together |
179 | out = _mm_add_ps(out, _mm_movehl_ps(out, out)); |
180 | |
181 | // Store the result |
182 | _mm_storel_pi((__m64 *)dst, out); |
183 | return; |
184 | } |
185 | |
186 | if (chans == 1) { |
187 | // Multiply the filter by the input |
188 | __m128 out = _mm_mul_ps(f0, _mm_loadu_ps(src + 0)); |
189 | out = sdl_madd_ps(out, f1, _mm_loadu_ps(src + 4)); |
190 | out = sdl_madd_ps(out, f2, _mm_loadu_ps(src + 8)); |
191 | |
192 | // Horizontal sum |
193 | __m128 shuf = _mm_shuffle_ps(out, out, _MM_SHUFFLE(2, 3, 0, 1)); |
194 | out = _mm_add_ps(out, shuf); |
195 | out = _mm_add_ss(out, _mm_movehl_ps(shuf, out)); |
196 | |
197 | _mm_store_ss(dst, out); |
198 | return; |
199 | } |
200 | |
201 | int chan = 0; |
202 | |
203 | // Process 4 channels at once |
204 | for (; chan + 4 <= chans; chan += 4) { |
205 | const float *in = &src[chan]; |
206 | __m128 out0 = _mm_setzero_ps(); |
207 | __m128 out1 = _mm_setzero_ps(); |
208 | |
209 | #define X(a, b, out) \ |
210 | out = sdl_madd_ps(out, _mm_loadu_ps(in), _mm_shuffle_ps(a, a, _MM_SHUFFLE(b, b, b, b))); \ |
211 | in += chans |
212 | |
213 | #define Y(a) \ |
214 | X(a, 0, out0); \ |
215 | X(a, 1, out1); \ |
216 | X(a, 2, out0); \ |
217 | X(a, 3, out1) |
218 | |
219 | Y(f0); |
220 | Y(f1); |
221 | Y(f2); |
222 | |
223 | #undef X |
224 | #undef Y |
225 | |
226 | // Add the accumulators together |
227 | __m128 out = _mm_add_ps(out0, out1); |
228 | |
229 | _mm_storeu_ps(&dst[chan], out); |
230 | } |
231 | |
232 | // Process the remaining channels one at a time. |
233 | // Channel counts 1,2,4,8 are already handled above, leaving 3,5,6,7 to deal with (looping 3,1,2,3 times). |
234 | // Without vgatherdps (AVX2), this gets quite messy. |
235 | for (; chan < chans; ++chan) { |
236 | const float *in = &src[chan]; |
237 | __m128 v0, v1, v2; |
238 | |
239 | #define X(x) \ |
240 | x = _mm_unpacklo_ps(_mm_load_ss(in), _mm_load_ss(in + chans)); \ |
241 | in += chans + chans; \ |
242 | x = _mm_movelh_ps(x, _mm_unpacklo_ps(_mm_load_ss(in), _mm_load_ss(in + chans))); \ |
243 | in += chans + chans |
244 | |
245 | X(v0); |
246 | X(v1); |
247 | X(v2); |
248 | |
249 | #undef X |
250 | |
251 | __m128 out = _mm_mul_ps(f0, v0); |
252 | out = sdl_madd_ps(out, f1, v1); |
253 | out = sdl_madd_ps(out, f2, v2); |
254 | |
255 | // Horizontal sum |
256 | __m128 shuf = _mm_shuffle_ps(out, out, _MM_SHUFFLE(2, 3, 0, 1)); |
257 | out = _mm_add_ps(out, shuf); |
258 | out = _mm_add_ss(out, _mm_movehl_ps(shuf, out)); |
259 | |
260 | _mm_store_ss(&dst[chan], out); |
261 | } |
262 | } |
263 | |
264 | #undef sdl_madd_ps |
265 | #endif |
266 | |
267 | #ifdef SDL_NEON_INTRINSICS |
268 | static void ResampleFrame_Generic_NEON(const float *src, float *dst, const Cubic *filter, float frac, int chans) |
269 | { |
270 | #if RESAMPLER_SAMPLES_PER_FRAME != 12 |
271 | #error Invalid samples per frame |
272 | #endif |
273 | |
274 | float32x4_t f0, f1, f2; |
275 | |
276 | { |
277 | const float32x4_t frac1 = vdupq_n_f32(frac); |
278 | const float32x4_t frac2 = vmulq_f32(frac1, frac1); |
279 | const float32x4_t frac3 = vmulq_f32(frac1, frac2); |
280 | |
281 | // Transposed in SetupAudioResampler |
282 | #define X(out) \ |
283 | out = vmlaq_f32(vmlaq_f32(vmlaq_f32(filter[0].v128, filter[1].v128, frac1), filter[2].v128, frac2), filter[3].v128, frac3); \ |
284 | filter += 4 |
285 | |
286 | X(f0); |
287 | X(f1); |
288 | X(f2); |
289 | |
290 | #undef X |
291 | } |
292 | |
293 | if (chans == 2) { |
294 | float32x4x2_t g0 = vzipq_f32(f0, f0); |
295 | float32x4x2_t g1 = vzipq_f32(f1, f1); |
296 | float32x4x2_t g2 = vzipq_f32(f2, f2); |
297 | |
298 | // Duplicate each of the filter elements and multiply by the input |
299 | // Use two accumulators to improve throughput |
300 | float32x4_t out0 = vmulq_f32(vld1q_f32(src + 0), g0.val[0]); |
301 | float32x4_t out1 = vmulq_f32(vld1q_f32(src + 4), g0.val[1]); |
302 | out0 = vmlaq_f32(out0, vld1q_f32(src + 8), g1.val[0]); |
303 | out1 = vmlaq_f32(out1, vld1q_f32(src + 12), g1.val[1]); |
304 | out0 = vmlaq_f32(out0, vld1q_f32(src + 16), g2.val[0]); |
305 | out1 = vmlaq_f32(out1, vld1q_f32(src + 20), g2.val[1]); |
306 | |
307 | // Add the accumulators together |
308 | out0 = vaddq_f32(out0, out1); |
309 | |
310 | // Add the lower and upper pairs together |
311 | float32x2_t out = vadd_f32(vget_low_f32(out0), vget_high_f32(out0)); |
312 | |
313 | // Store the result |
314 | vst1_f32(dst, out); |
315 | return; |
316 | } |
317 | |
318 | if (chans == 1) { |
319 | // Multiply the filter by the input |
320 | float32x4_t out = vmulq_f32(f0, vld1q_f32(src + 0)); |
321 | out = vmlaq_f32(out, f1, vld1q_f32(src + 4)); |
322 | out = vmlaq_f32(out, f2, vld1q_f32(src + 8)); |
323 | |
324 | // Horizontal sum |
325 | float32x2_t sum = vadd_f32(vget_low_f32(out), vget_high_f32(out)); |
326 | sum = vpadd_f32(sum, sum); |
327 | |
328 | vst1_lane_f32(dst, sum, 0); |
329 | return; |
330 | } |
331 | |
332 | int chan = 0; |
333 | |
334 | // Process 4 channels at once |
335 | for (; chan + 4 <= chans; chan += 4) { |
336 | const float *in = &src[chan]; |
337 | float32x4_t out0 = vdupq_n_f32(0); |
338 | float32x4_t out1 = vdupq_n_f32(0); |
339 | |
340 | #define X(a, b, out) \ |
341 | out = vmlaq_f32(out, vld1q_f32(in), vdupq_lane_f32(a, b)); \ |
342 | in += chans |
343 | |
344 | #define Y(a) \ |
345 | X(vget_low_f32(a), 0, out0); \ |
346 | X(vget_low_f32(a), 1, out1); \ |
347 | X(vget_high_f32(a), 0, out0); \ |
348 | X(vget_high_f32(a), 1, out1) |
349 | |
350 | Y(f0); |
351 | Y(f1); |
352 | Y(f2); |
353 | |
354 | #undef X |
355 | #undef Y |
356 | |
357 | // Add the accumulators together |
358 | float32x4_t out = vaddq_f32(out0, out1); |
359 | |
360 | vst1q_f32(&dst[chan], out); |
361 | } |
362 | |
363 | // Process the remaining channels one at a time. |
364 | // Channel counts 1,2,4,8 are already handled above, leaving 3,5,6,7 to deal with (looping 3,1,2,3 times). |
365 | for (; chan < chans; ++chan) { |
366 | const float *in = &src[chan]; |
367 | float32x4_t v0, v1, v2; |
368 | |
369 | #define X(x) \ |
370 | x = vld1q_dup_f32(in); \ |
371 | in += chans; \ |
372 | x = vld1q_lane_f32(in, x, 1); \ |
373 | in += chans; \ |
374 | x = vld1q_lane_f32(in, x, 2); \ |
375 | in += chans; \ |
376 | x = vld1q_lane_f32(in, x, 3); \ |
377 | in += chans |
378 | |
379 | X(v0); |
380 | X(v1); |
381 | X(v2); |
382 | |
383 | #undef X |
384 | |
385 | float32x4_t out = vmulq_f32(f0, v0); |
386 | out = vmlaq_f32(out, f1, v1); |
387 | out = vmlaq_f32(out, f2, v2); |
388 | |
389 | // Horizontal sum |
390 | float32x2_t sum = vadd_f32(vget_low_f32(out), vget_high_f32(out)); |
391 | sum = vpadd_f32(sum, sum); |
392 | |
393 | vst1_lane_f32(&dst[chan], sum, 0); |
394 | } |
395 | } |
396 | #endif |
397 | |
398 | // Calculate the cubic equation which passes through all four points. |
399 | // https://en.wikipedia.org/wiki/Ordinary_least_squares |
400 | // https://en.wikipedia.org/wiki/Polynomial_regression |
401 | static void CubicLeastSquares(Cubic *coeffs, float y0, float y1, float y2, float y3) |
402 | { |
403 | // Least squares matrix for xs = [0, 1/3, 2/3, 1] |
404 | // [ 1.0 0.0 0.0 0.0 ] |
405 | // [ -5.5 9.0 -4.5 1.0 ] |
406 | // [ 9.0 -22.5 18.0 -4.5 ] |
407 | // [ -4.5 13.5 -13.5 4.5 ] |
408 | |
409 | coeffs->v[0] = y0; |
410 | coeffs->v[1] = -5.5f * y0 + 9.0f * y1 - 4.5f * y2 + y3; |
411 | coeffs->v[2] = 9.0f * y0 - 22.5f * y1 + 18.0f * y2 - 4.5f * y3; |
412 | coeffs->v[3] = -4.5f * y0 + 13.5f * y1 - 13.5f * y2 + 4.5f * y3; |
413 | } |
414 | |
415 | // Zeroth-order modified Bessel function of the first kind |
416 | // https://mathworld.wolfram.com/ModifiedBesselFunctionoftheFirstKind.html |
417 | static float BesselI0(float x) |
418 | { |
419 | float sum = 0.0f; |
420 | float i = 1.0f; |
421 | float t = 1.0f; |
422 | x *= x * 0.25f; |
423 | |
424 | while (t >= sum * SDL_FLT_EPSILON) { |
425 | sum += t; |
426 | t *= x / (i * i); |
427 | ++i; |
428 | } |
429 | |
430 | return sum; |
431 | } |
432 | |
433 | // Pre-calculate 180 degrees of sin(pi * x) / pi |
434 | // The speedup from this isn't huge, but it also avoids precision issues. |
435 | // If sinf isn't available, SDL_sinf just calls SDL_sin. |
436 | // Know what SDL_sin(SDL_PI_F) equals? Not quite zero. |
437 | static void SincTable(float *table, int len) |
438 | { |
439 | int i; |
440 | |
441 | for (i = 0; i < len; ++i) { |
442 | table[i] = SDL_sinf(i * (SDL_PI_F / len)) / SDL_PI_F; |
443 | } |
444 | } |
445 | |
446 | // Calculate Sinc(x/y), using a lookup table |
447 | static float Sinc(const float *table, int x, int y) |
448 | { |
449 | float s = table[x % y]; |
450 | s = ((x / y) & 1) ? -s : s; |
451 | return (s * y) / x; |
452 | } |
453 | |
454 | static Cubic ResamplerFilter[RESAMPLER_SAMPLES_PER_ZERO_CROSSING][RESAMPLER_SAMPLES_PER_FRAME]; |
455 | |
456 | static void GenerateResamplerFilter(void) |
457 | { |
458 | enum |
459 | { |
460 | // Generate samples at 3x the target resolution, so that we have samples at [0, 1/3, 2/3, 1] of each position |
461 | TABLE_SAMPLES_PER_ZERO_CROSSING = RESAMPLER_SAMPLES_PER_ZERO_CROSSING * 3, |
462 | TABLE_SIZE = RESAMPLER_ZERO_CROSSINGS * TABLE_SAMPLES_PER_ZERO_CROSSING, |
463 | }; |
464 | |
465 | // if dB > 50, beta=(0.1102 * (dB - 8.7)), according to Matlab. |
466 | const float dB = 80.0f; |
467 | const float beta = 0.1102f * (dB - 8.7f); |
468 | const float bessel_beta = BesselI0(beta); |
469 | const float lensqr = TABLE_SIZE * TABLE_SIZE; |
470 | |
471 | int i, j; |
472 | |
473 | float sinc[TABLE_SAMPLES_PER_ZERO_CROSSING]; |
474 | SincTable(sinc, TABLE_SAMPLES_PER_ZERO_CROSSING); |
475 | |
476 | // Generate one wing of the filter |
477 | // https://en.wikipedia.org/wiki/Kaiser_window |
478 | // https://en.wikipedia.org/wiki/Whittaker%E2%80%93Shannon_interpolation_formula |
479 | float filter[TABLE_SIZE + 1]; |
480 | filter[0] = 1.0f; |
481 | |
482 | for (i = 1; i <= TABLE_SIZE; ++i) { |
483 | float b = BesselI0(beta * SDL_sqrtf((lensqr - (i * i)) / lensqr)) / bessel_beta; |
484 | float s = Sinc(sinc, i, TABLE_SAMPLES_PER_ZERO_CROSSING); |
485 | filter[i] = b * s; |
486 | } |
487 | |
488 | // Generate the coefficients for each point |
489 | // When interpolating, the fraction represents how far we are between input samples, |
490 | // so we need to align the filter by "moving" it to the right. |
491 | // |
492 | // For the left wing, this means interpolating "forwards" (away from the center) |
493 | // For the right wing, this means interpolating "backwards" (towards the center) |
494 | // |
495 | // The center of the filter is at the end of the left wing (RESAMPLER_ZERO_CROSSINGS - 1) |
496 | // The left wing is the filter, but reversed |
497 | // The right wing is the filter, but offset by 1 |
498 | // |
499 | // Since the right wing is offset by 1, this just means we interpolate backwards |
500 | // between the same points, instead of forwards |
501 | // interp(p[n], p[n+1], t) = interp(p[n+1], p[n+1-1], 1 - t) = interp(p[n+1], p[n], 1 - t) |
502 | for (i = 0; i < RESAMPLER_SAMPLES_PER_ZERO_CROSSING; ++i) { |
503 | for (j = 0; j < RESAMPLER_ZERO_CROSSINGS; ++j) { |
504 | const float *ys = &filter[((j * RESAMPLER_SAMPLES_PER_ZERO_CROSSING) + i) * 3]; |
505 | |
506 | Cubic *fwd = &ResamplerFilter[i][RESAMPLER_ZERO_CROSSINGS - j - 1]; |
507 | Cubic *rev = &ResamplerFilter[RESAMPLER_SAMPLES_PER_ZERO_CROSSING - i - 1][RESAMPLER_ZERO_CROSSINGS + j]; |
508 | |
509 | // Calculate the cubic equation of the 4 points |
510 | CubicLeastSquares(fwd, ys[0], ys[1], ys[2], ys[3]); |
511 | CubicLeastSquares(rev, ys[3], ys[2], ys[1], ys[0]); |
512 | } |
513 | } |
514 | } |
515 | |
516 | typedef void (*ResampleFrameFunc)(const float *src, float *dst, const Cubic *filter, float frac, int chans); |
517 | static ResampleFrameFunc ResampleFrame[8]; |
518 | |
519 | // Transpose 4x4 floats |
520 | static void Transpose4x4(Cubic *data) |
521 | { |
522 | int i, j; |
523 | |
524 | Cubic temp[4] = { data[0], data[1], data[2], data[3] }; |
525 | |
526 | for (i = 0; i < 4; ++i) { |
527 | for (j = 0; j < 4; ++j) { |
528 | data[i].v[j] = temp[j].v[i]; |
529 | } |
530 | } |
531 | } |
532 | |
533 | static void SetupAudioResampler(void) |
534 | { |
535 | int i, j; |
536 | bool transpose = false; |
537 | |
538 | GenerateResamplerFilter(); |
539 | |
540 | #ifdef SDL_SSE_INTRINSICS |
541 | if (SDL_HasSSE()) { |
542 | for (i = 0; i < 8; ++i) { |
543 | ResampleFrame[i] = ResampleFrame_Generic_SSE; |
544 | } |
545 | transpose = true; |
546 | } else |
547 | #endif |
548 | #ifdef SDL_NEON_INTRINSICS |
549 | if (SDL_HasNEON()) { |
550 | for (i = 0; i < 8; ++i) { |
551 | ResampleFrame[i] = ResampleFrame_Generic_NEON; |
552 | } |
553 | transpose = true; |
554 | } else |
555 | #endif |
556 | { |
557 | for (i = 0; i < 8; ++i) { |
558 | ResampleFrame[i] = ResampleFrame_Generic; |
559 | } |
560 | |
561 | ResampleFrame[0] = ResampleFrame_Mono; |
562 | ResampleFrame[1] = ResampleFrame_Stereo; |
563 | } |
564 | |
565 | if (transpose) { |
566 | // Transpose each set of 4 coefficients, to reduce work when resampling |
567 | for (i = 0; i < RESAMPLER_SAMPLES_PER_ZERO_CROSSING; ++i) { |
568 | for (j = 0; j + 4 <= RESAMPLER_SAMPLES_PER_FRAME; j += 4) { |
569 | Transpose4x4(&ResamplerFilter[i][j]); |
570 | } |
571 | } |
572 | } |
573 | } |
574 | |
575 | void SDL_SetupAudioResampler(void) |
576 | { |
577 | static SDL_InitState init; |
578 | |
579 | if (SDL_ShouldInit(&init)) { |
580 | SetupAudioResampler(); |
581 | SDL_SetInitialized(&init, true); |
582 | } |
583 | } |
584 | |
585 | Sint64 SDL_GetResampleRate(int src_rate, int dst_rate) |
586 | { |
587 | SDL_assert(src_rate > 0); |
588 | SDL_assert(dst_rate > 0); |
589 | |
590 | Sint64 numerator = (Sint64)src_rate << 32; |
591 | Sint64 denominator = (Sint64)dst_rate; |
592 | |
593 | // Generally it's expected that `dst_frames = (src_frames * dst_rate) / src_rate` |
594 | // To match this as closely as possible without infinite precision, always round up the resample rate. |
595 | // For example, without rounding up, a sample ratio of 2:3 would have `sample_rate = 0xAAAAAAAA` |
596 | // After 3 frames, the position would be 0x1.FFFFFFFE, meaning we haven't fully consumed the second input frame. |
597 | // By rounding up to 0xAAAAAAAB, we would instead reach 0x2.00000001, fulling consuming the second frame. |
598 | // Technically you could say this is kicking the can 0x100000000 steps down the road, but I'm fine with that :) |
599 | // sample_rate = div_ceil(numerator, denominator) |
600 | Sint64 sample_rate = ((numerator - 1) / denominator) + 1; |
601 | |
602 | SDL_assert(sample_rate > 0); |
603 | |
604 | return sample_rate; |
605 | } |
606 | |
607 | int SDL_GetResamplerHistoryFrames(void) |
608 | { |
609 | // Even if we aren't currently resampling, make sure to keep enough history in case we need to later. |
610 | |
611 | return RESAMPLER_MAX_PADDING_FRAMES; |
612 | } |
613 | |
614 | int SDL_GetResamplerPaddingFrames(Sint64 resample_rate) |
615 | { |
616 | // This must always be <= SDL_GetResamplerHistoryFrames() |
617 | |
618 | return resample_rate ? RESAMPLER_MAX_PADDING_FRAMES : 0; |
619 | } |
620 | |
621 | // These are not general purpose. They do not check for all possible underflow/overflow |
622 | SDL_FORCE_INLINE bool ResamplerAdd(Sint64 a, Sint64 b, Sint64 *ret) |
623 | { |
624 | if ((b > 0) && (a > SDL_MAX_SINT64 - b)) { |
625 | return false; |
626 | } |
627 | |
628 | *ret = a + b; |
629 | return true; |
630 | } |
631 | |
632 | SDL_FORCE_INLINE bool ResamplerMul(Sint64 a, Sint64 b, Sint64 *ret) |
633 | { |
634 | if ((b > 0) && (a > SDL_MAX_SINT64 / b)) { |
635 | return false; |
636 | } |
637 | |
638 | *ret = a * b; |
639 | return true; |
640 | } |
641 | |
642 | Sint64 SDL_GetResamplerInputFrames(Sint64 output_frames, Sint64 resample_rate, Sint64 resample_offset) |
643 | { |
644 | // Calculate the index of the last input frame, then add 1. |
645 | // ((((output_frames - 1) * resample_rate) + resample_offset) >> 32) + 1 |
646 | |
647 | Sint64 output_offset; |
648 | if (!ResamplerMul(output_frames, resample_rate, &output_offset) || |
649 | !ResamplerAdd(output_offset, -resample_rate + resample_offset + 0x100000000, &output_offset)) { |
650 | output_offset = SDL_MAX_SINT64; |
651 | } |
652 | |
653 | Sint64 input_frames = (Sint64)(Sint32)(output_offset >> 32); |
654 | input_frames = SDL_max(input_frames, 0); |
655 | |
656 | return input_frames; |
657 | } |
658 | |
659 | Sint64 SDL_GetResamplerOutputFrames(Sint64 input_frames, Sint64 resample_rate, Sint64 *inout_resample_offset) |
660 | { |
661 | Sint64 resample_offset = *inout_resample_offset; |
662 | |
663 | // input_offset = (input_frames << 32) - resample_offset; |
664 | Sint64 input_offset; |
665 | if (!ResamplerMul(input_frames, 0x100000000, &input_offset) || |
666 | !ResamplerAdd(input_offset, -resample_offset, &input_offset)) { |
667 | input_offset = SDL_MAX_SINT64; |
668 | } |
669 | |
670 | // output_frames = div_ceil(input_offset, resample_rate) |
671 | Sint64 output_frames = (input_offset > 0) ? ((input_offset - 1) / resample_rate) + 1 : 0; |
672 | |
673 | *inout_resample_offset = (output_frames * resample_rate) - input_offset; |
674 | |
675 | return output_frames; |
676 | } |
677 | |
678 | void SDL_ResampleAudio(int chans, const float *src, int inframes, float *dst, int outframes, |
679 | Sint64 resample_rate, Sint64 *inout_resample_offset) |
680 | { |
681 | int i; |
682 | Sint64 srcpos = *inout_resample_offset; |
683 | ResampleFrameFunc resample_frame = ResampleFrame[chans - 1]; |
684 | |
685 | SDL_assert(resample_rate > 0); |
686 | |
687 | src -= (RESAMPLER_ZERO_CROSSINGS - 1) * chans; |
688 | |
689 | for (i = 0; i < outframes; ++i) { |
690 | int srcindex = (int)(Sint32)(srcpos >> 32); |
691 | Uint32 srcfraction = (Uint32)(srcpos & 0xFFFFFFFF); |
692 | srcpos += resample_rate; |
693 | |
694 | SDL_assert(srcindex >= -1 && srcindex < inframes); |
695 | |
696 | const Cubic *filter = ResamplerFilter[srcfraction >> RESAMPLER_FILTER_INTERP_BITS]; |
697 | const float frac = (float)(srcfraction & (RESAMPLER_FILTER_INTERP_RANGE - 1)) * (1.0f / RESAMPLER_FILTER_INTERP_RANGE); |
698 | |
699 | const float *frame = &src[srcindex * chans]; |
700 | resample_frame(frame, dst, filter, frac, chans); |
701 | |
702 | dst += chans; |
703 | } |
704 | |
705 | *inout_resample_offset = srcpos - ((Sint64)inframes << 32); |
706 | } |
707 | |