/
SDL_audiotypecvt.c
1430 lines (1212 loc) · 53.5 KB
1
2
/*
Simple DirectMedia Layer
3
Copyright (C) 1997-2018 Sam Lantinga <slouken@libsdl.org>
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
This software is provided 'as-is', without any express or implied
warranty. In no event will the authors be held liable for any damages
arising from the use of this software.
Permission is granted to anyone to use this software for any purpose,
including commercial applications, and to alter it and redistribute it
freely, subject to the following restrictions:
1. The origin of this software must not be misrepresented; you must not
claim that you wrote the original software. If you use this software
in a product, an acknowledgment in the product documentation would be
appreciated but is not required.
2. Altered source versions must be plainly marked as such, and must not be
misrepresented as being the original software.
3. This notice may not be removed or altered from any source distribution.
*/
#include "../SDL_internal.h"
#include "SDL_audio.h"
#include "SDL_audio_c.h"
25
#include "SDL_cpuinfo.h"
26
#include "SDL_assert.h"
27
28
29
30
#ifdef __ARM_NEON__
#define HAVE_NEON_INTRINSICS 1
#endif
31
32
33
#ifdef __SSE2__
#define HAVE_SSE2_INTRINSICS 1
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
#endif
#if defined(__x86_64__) && HAVE_SSE2_INTRINSICS
#define NEED_SCALAR_CONVERTER_FALLBACKS 0 /* x86_64 guarantees SSE2. */
#elif __MACOSX__ && HAVE_SSE2_INTRINSICS
#define NEED_SCALAR_CONVERTER_FALLBACKS 0 /* Mac OS X/Intel guarantees SSE2. */
#elif defined(__ARM_ARCH) && (__ARM_ARCH >= 8) && HAVE_NEON_INTRINSICS
#define NEED_SCALAR_CONVERTER_FALLBACKS 0 /* ARMv8+ promise NEON. */
#elif defined(__APPLE__) && defined(__ARM_ARCH) && (__ARM_ARCH >= 7) && HAVE_NEON_INTRINSICS
#define NEED_SCALAR_CONVERTER_FALLBACKS 0 /* All Apple ARMv7 chips promise NEON support. */
#endif
/* Set to zero if platform is guaranteed to use a SIMD codepath here. */
#ifndef NEED_SCALAR_CONVERTER_FALLBACKS
#define NEED_SCALAR_CONVERTER_FALLBACKS 1
#endif
/* Function pointers set to a CPU-specific implementation. */
SDL_AudioFilter SDL_Convert_S8_to_F32 = NULL;
SDL_AudioFilter SDL_Convert_U8_to_F32 = NULL;
SDL_AudioFilter SDL_Convert_S16_to_F32 = NULL;
SDL_AudioFilter SDL_Convert_U16_to_F32 = NULL;
SDL_AudioFilter SDL_Convert_S32_to_F32 = NULL;
SDL_AudioFilter SDL_Convert_F32_to_S8 = NULL;
SDL_AudioFilter SDL_Convert_F32_to_U8 = NULL;
SDL_AudioFilter SDL_Convert_F32_to_S16 = NULL;
SDL_AudioFilter SDL_Convert_F32_to_U16 = NULL;
SDL_AudioFilter SDL_Convert_F32_to_S32 = NULL;
64
65
#define DIVBY128 0.0078125f
#define DIVBY32768 0.000030517578125f
66
#define DIVBY8388607 0.00000011920930376163766f
67
68
69
70
71
#if NEED_SCALAR_CONVERTER_FALLBACKS
static void SDLCALL
SDL_Convert_S8_to_F32_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
72
{
73
const Sint8 *src = ((const Sint8 *) (cvt->buf + cvt->len_cvt)) - 1;
74
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 4)) - 1;
75
76
int i;
77
LOG_DEBUG_CONVERT("AUDIO_S8", "AUDIO_F32");
78
79
for (i = cvt->len_cvt; i; --i, --src, --dst) {
80
*dst = ((float) *src) * DIVBY128;
81
82
83
84
}
cvt->len_cvt *= 4;
if (cvt->filters[++cvt->filter_index]) {
85
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
86
87
88
}
}
89
90
static void SDLCALL
SDL_Convert_U8_to_F32_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
91
{
92
93
const Uint8 *src = ((const Uint8 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 4)) - 1;
94
95
int i;
96
LOG_DEBUG_CONVERT("AUDIO_U8", "AUDIO_F32");
97
98
for (i = cvt->len_cvt; i; --i, --src, --dst) {
99
*dst = (((float) *src) * DIVBY128) - 1.0f;
100
101
102
103
}
cvt->len_cvt *= 4;
if (cvt->filters[++cvt->filter_index]) {
104
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
105
106
107
}
}
108
109
static void SDLCALL
SDL_Convert_S16_to_F32_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
110
{
111
112
const Sint16 *src = ((const Sint16 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 2)) - 1;
113
114
int i;
115
LOG_DEBUG_CONVERT("AUDIO_S16", "AUDIO_F32");
116
117
for (i = cvt->len_cvt / sizeof (Sint16); i; --i, --src, --dst) {
118
*dst = ((float) *src) * DIVBY32768;
119
120
121
122
}
cvt->len_cvt *= 2;
if (cvt->filters[++cvt->filter_index]) {
123
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
124
125
126
}
}
127
128
static void SDLCALL
SDL_Convert_U16_to_F32_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
129
{
130
131
const Uint16 *src = ((const Uint16 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 2)) - 1;
132
133
int i;
134
LOG_DEBUG_CONVERT("AUDIO_U16", "AUDIO_F32");
135
136
for (i = cvt->len_cvt / sizeof (Uint16); i; --i, --src, --dst) {
137
*dst = (((float) *src) * DIVBY32768) - 1.0f;
138
139
140
141
}
cvt->len_cvt *= 2;
if (cvt->filters[++cvt->filter_index]) {
142
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
143
144
145
}
}
146
147
static void SDLCALL
SDL_Convert_S32_to_F32_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
148
{
149
const Sint32 *src = (const Sint32 *) cvt->buf;
150
float *dst = (float *) cvt->buf;
151
152
int i;
153
LOG_DEBUG_CONVERT("AUDIO_S32", "AUDIO_F32");
154
155
for (i = cvt->len_cvt / sizeof (Sint32); i; --i, ++src, ++dst) {
156
*dst = ((float) (*src>>8)) * DIVBY8388607;
157
158
159
}
if (cvt->filters[++cvt->filter_index]) {
160
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
161
162
163
}
}
164
165
static void SDLCALL
SDL_Convert_F32_to_S8_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
166
{
167
168
const float *src = (const float *) cvt->buf;
Sint8 *dst = (Sint8 *) cvt->buf;
169
170
int i;
171
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_S8");
172
173
for (i = cvt->len_cvt / sizeof (float); i; --i, ++src, ++dst) {
174
const float sample = *src;
175
if (sample >= 1.0f) {
176
*dst = 127;
177
} else if (sample <= -1.0f) {
178
*dst = -128;
179
180
181
} else {
*dst = (Sint8)(sample * 127.0f);
}
182
183
}
184
cvt->len_cvt /= 4;
185
if (cvt->filters[++cvt->filter_index]) {
186
cvt->filters[cvt->filter_index](cvt, AUDIO_S8);
187
188
189
}
}
190
191
static void SDLCALL
SDL_Convert_F32_to_U8_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
192
{
193
194
const float *src = (const float *) cvt->buf;
Uint8 *dst = (Uint8 *) cvt->buf;
195
196
int i;
197
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_U8");
198
199
for (i = cvt->len_cvt / sizeof (float); i; --i, ++src, ++dst) {
200
const float sample = *src;
201
if (sample >= 1.0f) {
202
*dst = 255;
203
} else if (sample <= -1.0f) {
204
205
206
207
*dst = 0;
} else {
*dst = (Uint8)((sample + 1.0f) * 127.0f);
}
208
209
}
210
cvt->len_cvt /= 4;
211
if (cvt->filters[++cvt->filter_index]) {
212
cvt->filters[cvt->filter_index](cvt, AUDIO_U8);
213
214
215
}
}
216
217
static void SDLCALL
SDL_Convert_F32_to_S16_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
218
{
219
220
const float *src = (const float *) cvt->buf;
Sint16 *dst = (Sint16 *) cvt->buf;
221
222
int i;
223
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_S16");
224
225
for (i = cvt->len_cvt / sizeof (float); i; --i, ++src, ++dst) {
226
const float sample = *src;
227
if (sample >= 1.0f) {
228
*dst = 32767;
229
} else if (sample <= -1.0f) {
230
*dst = -32768;
231
232
233
} else {
*dst = (Sint16)(sample * 32767.0f);
}
234
235
}
236
cvt->len_cvt /= 2;
237
if (cvt->filters[++cvt->filter_index]) {
238
cvt->filters[cvt->filter_index](cvt, AUDIO_S16SYS);
239
240
241
}
}
242
243
static void SDLCALL
SDL_Convert_F32_to_U16_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
244
{
245
246
const float *src = (const float *) cvt->buf;
Uint16 *dst = (Uint16 *) cvt->buf;
247
248
int i;
249
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_U16");
250
251
for (i = cvt->len_cvt / sizeof (float); i; --i, ++src, ++dst) {
252
const float sample = *src;
253
if (sample >= 1.0f) {
254
*dst = 65535;
255
} else if (sample <= -1.0f) {
256
257
258
259
*dst = 0;
} else {
*dst = (Uint16)((sample + 1.0f) * 32767.0f);
}
260
261
}
262
cvt->len_cvt /= 2;
263
if (cvt->filters[++cvt->filter_index]) {
264
cvt->filters[cvt->filter_index](cvt, AUDIO_U16SYS);
265
266
267
}
}
268
269
static void SDLCALL
SDL_Convert_F32_to_S32_Scalar(SDL_AudioCVT *cvt, SDL_AudioFormat format)
270
{
271
272
const float *src = (const float *) cvt->buf;
Sint32 *dst = (Sint32 *) cvt->buf;
273
274
int i;
275
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_S32");
276
277
for (i = cvt->len_cvt / sizeof (float); i; --i, ++src, ++dst) {
278
const float sample = *src;
279
if (sample >= 1.0f) {
280
*dst = 2147483647;
281
} else if (sample <= -1.0f) {
282
*dst = (Sint32) -2147483648LL;
283
} else {
284
*dst = ((Sint32)(sample * 8388607.0f)) << 8;
285
}
286
287
288
}
if (cvt->filters[++cvt->filter_index]) {
289
cvt->filters[cvt->filter_index](cvt, AUDIO_S32SYS);
290
291
}
}
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
#endif
#if HAVE_SSE2_INTRINSICS
static void SDLCALL
SDL_Convert_S8_to_F32_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Sint8 *src = ((const Sint8 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 4)) - 1;
int i;
LOG_DEBUG_CONVERT("AUDIO_S8", "AUDIO_F32 (using SSE2)");
/* Get dst aligned to 16 bytes (since buffer is growing, we don't have to worry about overreading from src) */
for (i = cvt->len_cvt; i && (((size_t) (dst-15)) & 15); --i, --src, --dst) {
307
*dst = ((float) *src) * DIVBY128;
308
309
310
311
312
313
314
315
316
317
}
src -= 15; dst -= 15; /* adjust to read SSE blocks from the start. */
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
const __m128i *mmsrc = (const __m128i *) src;
const __m128i zero = _mm_setzero_si128();
318
const __m128 divby128 = _mm_set1_ps(DIVBY128);
319
320
321
322
323
324
325
while (i >= 16) { /* 16 * 8-bit */
const __m128i bytes = _mm_load_si128(mmsrc); /* get 16 sint8 into an XMM register. */
/* treat as int16, shift left to clear every other sint16, then back right with sign-extend. Now sint16. */
const __m128i shorts1 = _mm_srai_epi16(_mm_slli_epi16(bytes, 8), 8);
/* right-shift-sign-extend gets us sint16 with the other set of values. */
const __m128i shorts2 = _mm_srai_epi16(bytes, 8);
/* unpack against zero to make these int32, shift to make them sign-extend, convert to float, multiply. Whew! */
326
327
328
329
const __m128 floats1 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_slli_epi32(_mm_unpacklo_epi16(shorts1, zero), 16), 16)), divby128);
const __m128 floats2 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_slli_epi32(_mm_unpacklo_epi16(shorts2, zero), 16), 16)), divby128);
const __m128 floats3 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_slli_epi32(_mm_unpackhi_epi16(shorts1, zero), 16), 16)), divby128);
const __m128 floats4 = _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_slli_epi32(_mm_unpackhi_epi16(shorts2, zero), 16), 16)), divby128);
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
/* Interleave back into correct order, store. */
_mm_store_ps(dst, _mm_unpacklo_ps(floats1, floats2));
_mm_store_ps(dst+4, _mm_unpackhi_ps(floats1, floats2));
_mm_store_ps(dst+8, _mm_unpacklo_ps(floats3, floats4));
_mm_store_ps(dst+12, _mm_unpackhi_ps(floats3, floats4));
i -= 16; mmsrc--; dst -= 16;
}
src = (const Sint8 *) mmsrc;
}
src += 15; dst += 15; /* adjust for any scalar finishing. */
/* Finish off any leftovers with scalar operations. */
while (i) {
345
*dst = ((float) *src) * DIVBY128;
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
i--; src--; dst--;
}
cvt->len_cvt *= 4;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}
static void SDLCALL
SDL_Convert_U8_to_F32_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Uint8 *src = ((const Uint8 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 4)) - 1;
int i;
LOG_DEBUG_CONVERT("AUDIO_U8", "AUDIO_F32 (using SSE2)");
/* Get dst aligned to 16 bytes (since buffer is growing, we don't have to worry about overreading from src) */
for (i = cvt->len_cvt; i && (((size_t) (dst-15)) & 15); --i, --src, --dst) {
366
*dst = (((float) *src) * DIVBY128) - 1.0f;
367
368
369
370
371
372
373
374
375
376
}
src -= 15; dst -= 15; /* adjust to read SSE blocks from the start. */
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
const __m128i *mmsrc = (const __m128i *) src;
const __m128i zero = _mm_setzero_si128();
377
const __m128 divby128 = _mm_set1_ps(DIVBY128);
378
379
380
381
382
383
384
385
386
const __m128 minus1 = _mm_set1_ps(-1.0f);
while (i >= 16) { /* 16 * 8-bit */
const __m128i bytes = _mm_load_si128(mmsrc); /* get 16 uint8 into an XMM register. */
/* treat as int16, shift left to clear every other sint16, then back right with zero-extend. Now uint16. */
const __m128i shorts1 = _mm_srli_epi16(_mm_slli_epi16(bytes, 8), 8);
/* right-shift-zero-extend gets us uint16 with the other set of values. */
const __m128i shorts2 = _mm_srli_epi16(bytes, 8);
/* unpack against zero to make these int32, convert to float, multiply, add. Whew! */
/* Note that AVX2 can do floating point multiply+add in one instruction, fwiw. SSE2 cannot. */
387
388
389
390
const __m128 floats1 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(shorts1, zero)), divby128), minus1);
const __m128 floats2 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(shorts2, zero)), divby128), minus1);
const __m128 floats3 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(shorts1, zero)), divby128), minus1);
const __m128 floats4 = _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(shorts2, zero)), divby128), minus1);
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
/* Interleave back into correct order, store. */
_mm_store_ps(dst, _mm_unpacklo_ps(floats1, floats2));
_mm_store_ps(dst+4, _mm_unpackhi_ps(floats1, floats2));
_mm_store_ps(dst+8, _mm_unpacklo_ps(floats3, floats4));
_mm_store_ps(dst+12, _mm_unpackhi_ps(floats3, floats4));
i -= 16; mmsrc--; dst -= 16;
}
src = (const Uint8 *) mmsrc;
}
src += 15; dst += 15; /* adjust for any scalar finishing. */
/* Finish off any leftovers with scalar operations. */
while (i) {
406
*dst = (((float) *src) * DIVBY128) - 1.0f;
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
i--; src--; dst--;
}
cvt->len_cvt *= 4;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}
static void SDLCALL
SDL_Convert_S16_to_F32_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Sint16 *src = ((const Sint16 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 2)) - 1;
int i;
LOG_DEBUG_CONVERT("AUDIO_S16", "AUDIO_F32 (using SSE2)");
/* Get dst aligned to 16 bytes (since buffer is growing, we don't have to worry about overreading from src) */
for (i = cvt->len_cvt / sizeof (Sint16); i && (((size_t) (dst-7)) & 15); --i, --src, --dst) {
427
*dst = ((float) *src) * DIVBY32768;
428
429
430
431
432
433
434
435
}
src -= 7; dst -= 7; /* adjust to read SSE blocks from the start. */
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
436
const __m128 divby32768 = _mm_set1_ps(DIVBY32768);
437
438
439
440
441
442
443
while (i >= 8) { /* 8 * 16-bit */
const __m128i ints = _mm_load_si128((__m128i const *) src); /* get 8 sint16 into an XMM register. */
/* treat as int32, shift left to clear every other sint16, then back right with sign-extend. Now sint32. */
const __m128i a = _mm_srai_epi32(_mm_slli_epi32(ints, 16), 16);
/* right-shift-sign-extend gets us sint32 with the other set of values. */
const __m128i b = _mm_srai_epi32(ints, 16);
/* Interleave these back into the right order, convert to float, multiply, store. */
444
445
_mm_store_ps(dst, _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi32(a, b)), divby32768));
_mm_store_ps(dst+4, _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi32(a, b)), divby32768));
446
447
448
449
450
451
452
453
i -= 8; src -= 8; dst -= 8;
}
}
src += 7; dst += 7; /* adjust for any scalar finishing. */
/* Finish off any leftovers with scalar operations. */
while (i) {
454
*dst = ((float) *src) * DIVBY32768;
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
i--; src--; dst--;
}
cvt->len_cvt *= 2;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}
static void SDLCALL
SDL_Convert_U16_to_F32_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Uint16 *src = ((const Uint16 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 2)) - 1;
int i;
LOG_DEBUG_CONVERT("AUDIO_U16", "AUDIO_F32 (using SSE2)");
/* Get dst aligned to 16 bytes (since buffer is growing, we don't have to worry about overreading from src) */
for (i = cvt->len_cvt / sizeof (Sint16); i && (((size_t) (dst-7)) & 15); --i, --src, --dst) {
475
*dst = (((float) *src) * DIVBY32768) - 1.0f;
476
477
478
479
480
481
482
483
}
src -= 7; dst -= 7; /* adjust to read SSE blocks from the start. */
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
484
const __m128 divby32768 = _mm_set1_ps(DIVBY32768);
485
486
487
488
489
490
491
492
const __m128 minus1 = _mm_set1_ps(1.0f);
while (i >= 8) { /* 8 * 16-bit */
const __m128i ints = _mm_load_si128((__m128i const *) src); /* get 8 sint16 into an XMM register. */
/* treat as int32, shift left to clear every other sint16, then back right with zero-extend. Now sint32. */
const __m128i a = _mm_srli_epi32(_mm_slli_epi32(ints, 16), 16);
/* right-shift-sign-extend gets us sint32 with the other set of values. */
const __m128i b = _mm_srli_epi32(ints, 16);
/* Interleave these back into the right order, convert to float, multiply, store. */
493
494
_mm_store_ps(dst, _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi32(a, b)), divby32768), minus1));
_mm_store_ps(dst+4, _mm_add_ps(_mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi32(a, b)), divby32768), minus1));
495
496
497
498
499
500
501
502
i -= 8; src -= 8; dst -= 8;
}
}
src += 7; dst += 7; /* adjust for any scalar finishing. */
/* Finish off any leftovers with scalar operations. */
while (i) {
503
*dst = (((float) *src) * DIVBY32768) - 1.0f;
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
i--; src--; dst--;
}
cvt->len_cvt *= 2;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}
static void SDLCALL
SDL_Convert_S32_to_F32_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Sint32 *src = (const Sint32 *) cvt->buf;
float *dst = (float *) cvt->buf;
int i;
LOG_DEBUG_CONVERT("AUDIO_S32", "AUDIO_F32 (using SSE2)");
/* Get dst aligned to 16 bytes */
for (i = cvt->len_cvt / sizeof (Sint32); i && (((size_t) dst) & 15); --i, ++src, ++dst) {
524
*dst = ((float) (*src>>8)) * DIVBY8388607;
525
526
527
528
529
530
531
}
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
SDL_assert(!i || ((((size_t) src) & 15) == 0));
{
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
532
const __m128 divby8388607 = _mm_set1_ps(DIVBY8388607);
533
534
const __m128i *mmsrc = (const __m128i *) src;
while (i >= 4) { /* 4 * sint32 */
535
/* shift out lowest bits so int fits in a float32. Small precision loss, but much faster. */
536
_mm_store_ps(dst, _mm_mul_ps(_mm_cvtepi32_ps(_mm_srai_epi32(_mm_load_si128(mmsrc), 8)), divby8388607));
537
538
539
540
541
542
543
i -= 4; mmsrc++; dst += 4;
}
src = (const Sint32 *) mmsrc;
}
/* Finish off any leftovers with scalar operations. */
while (i) {
544
*dst = ((float) (*src>>8)) * DIVBY8388607;
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
i--; src++; dst++;
}
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}
static void SDLCALL
SDL_Convert_F32_to_S8_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const float *src = (const float *) cvt->buf;
Sint8 *dst = (Sint8 *) cvt->buf;
int i;
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_S8 (using SSE2)");
/* Get dst aligned to 16 bytes */
for (i = cvt->len_cvt / sizeof (float); i && (((size_t) dst) & 15); --i, ++src, ++dst) {
564
565
566
567
568
569
570
571
const float sample = *src;
if (sample >= 1.0f) {
*dst = 127;
} else if (sample <= -1.0f) {
*dst = -128;
} else {
*dst = (Sint8)(sample * 127.0f);
}
572
573
574
575
576
577
578
}
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
579
580
const __m128 one = _mm_set1_ps(1.0f);
const __m128 negone = _mm_set1_ps(-1.0f);
581
582
583
const __m128 mulby127 = _mm_set1_ps(127.0f);
__m128i *mmdst = (__m128i *) dst;
while (i >= 16) { /* 16 * float32 */
584
585
586
587
const __m128i ints1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src)), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints2 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+4)), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints3 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+8)), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints4 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+12)), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
588
589
590
591
592
593
594
595
_mm_store_si128(mmdst, _mm_packs_epi16(_mm_packs_epi32(ints1, ints2), _mm_packs_epi32(ints3, ints4))); /* pack down, store out. */
i -= 16; src += 16; mmdst++;
}
dst = (Sint8 *) mmdst;
}
/* Finish off any leftovers with scalar operations. */
while (i) {
596
597
598
599
600
601
602
603
const float sample = *src;
if (sample >= 1.0f) {
*dst = 127;
} else if (sample <= -1.0f) {
*dst = -128;
} else {
*dst = (Sint8)(sample * 127.0f);
}
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
i--; src++; dst++;
}
cvt->len_cvt /= 4;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_S8);
}
}
static void SDLCALL
SDL_Convert_F32_to_U8_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const float *src = (const float *) cvt->buf;
Uint8 *dst = (Uint8 *) cvt->buf;
int i;
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_U8 (using SSE2)");
/* Get dst aligned to 16 bytes */
for (i = cvt->len_cvt / sizeof (float); i && (((size_t) dst) & 15); --i, ++src, ++dst) {
624
625
626
627
628
629
630
631
const float sample = *src;
if (sample >= 1.0f) {
*dst = 255;
} else if (sample <= -1.0f) {
*dst = 0;
} else {
*dst = (Uint8)((sample + 1.0f) * 127.0f);
}
632
633
634
635
636
637
638
}
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
639
640
const __m128 one = _mm_set1_ps(1.0f);
const __m128 negone = _mm_set1_ps(-1.0f);
641
642
643
const __m128 mulby127 = _mm_set1_ps(127.0f);
__m128i *mmdst = (__m128i *) dst;
while (i >= 16) { /* 16 * float32 */
644
645
646
647
const __m128i ints1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_add_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src)), one), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints2 = _mm_cvtps_epi32(_mm_mul_ps(_mm_add_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+4)), one), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints3 = _mm_cvtps_epi32(_mm_mul_ps(_mm_add_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+8)), one), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints4 = _mm_cvtps_epi32(_mm_mul_ps(_mm_add_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+12)), one), one), mulby127)); /* load 4 floats, clamp, convert to sint32 */
648
649
650
651
652
653
654
655
_mm_store_si128(mmdst, _mm_packus_epi16(_mm_packs_epi32(ints1, ints2), _mm_packs_epi32(ints3, ints4))); /* pack down, store out. */
i -= 16; src += 16; mmdst++;
}
dst = (Uint8 *) mmdst;
}
/* Finish off any leftovers with scalar operations. */
while (i) {
656
657
658
659
660
661
662
663
const float sample = *src;
if (sample >= 1.0f) {
*dst = 255;
} else if (sample <= -1.0f) {
*dst = 0;
} else {
*dst = (Uint8)((sample + 1.0f) * 127.0f);
}
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
i--; src++; dst++;
}
cvt->len_cvt /= 4;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_U8);
}
}
static void SDLCALL
SDL_Convert_F32_to_S16_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const float *src = (const float *) cvt->buf;
Sint16 *dst = (Sint16 *) cvt->buf;
int i;
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_S16 (using SSE2)");
/* Get dst aligned to 16 bytes */
for (i = cvt->len_cvt / sizeof (float); i && (((size_t) dst) & 15); --i, ++src, ++dst) {
684
685
686
687
688
689
690
691
const float sample = *src;
if (sample >= 1.0f) {
*dst = 32767;
} else if (sample <= -1.0f) {
*dst = -32768;
} else {
*dst = (Sint16)(sample * 32767.0f);
}
692
693
694
695
696
697
698
}
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
699
700
const __m128 one = _mm_set1_ps(1.0f);
const __m128 negone = _mm_set1_ps(-1.0f);
701
702
703
const __m128 mulby32767 = _mm_set1_ps(32767.0f);
__m128i *mmdst = (__m128i *) dst;
while (i >= 8) { /* 8 * float32 */
704
705
const __m128i ints1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src)), one), mulby32767)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints2 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+4)), one), mulby32767)); /* load 4 floats, clamp, convert to sint32 */
706
707
708
709
710
711
712
713
_mm_store_si128(mmdst, _mm_packs_epi32(ints1, ints2)); /* pack to sint16, store out. */
i -= 8; src += 8; mmdst++;
}
dst = (Sint16 *) mmdst;
}
/* Finish off any leftovers with scalar operations. */
while (i) {
714
715
716
717
718
719
720
721
const float sample = *src;
if (sample >= 1.0f) {
*dst = 32767;
} else if (sample <= -1.0f) {
*dst = -32768;
} else {
*dst = (Sint16)(sample * 32767.0f);
}
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
i--; src++; dst++;
}
cvt->len_cvt /= 2;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_S16SYS);
}
}
static void SDLCALL
SDL_Convert_F32_to_U16_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const float *src = (const float *) cvt->buf;
Uint16 *dst = (Uint16 *) cvt->buf;
int i;
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_U16 (using SSE2)");
/* Get dst aligned to 16 bytes */
for (i = cvt->len_cvt / sizeof (float); i && (((size_t) dst) & 15); --i, ++src, ++dst) {
742
743
744
745
746
747
748
749
const float sample = *src;
if (sample >= 1.0f) {
*dst = 65535;
} else if (sample <= -1.0f) {
*dst = 0;
} else {
*dst = (Uint16)((sample + 1.0f) * 32767.0f);
}
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
}
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
/* This calculates differently than the scalar path because SSE2 can't
pack int32 data down to unsigned int16. _mm_packs_epi32 does signed
saturation, so that would corrupt our data. _mm_packus_epi32 exists,
but not before SSE 4.1. So we convert from float to sint16, packing
that down with legit signed saturation, and then xor the top bit
against 1. This results in the correct unsigned 16-bit value, even
though it looks like dark magic. */
const __m128 mulby32767 = _mm_set1_ps(32767.0f);
const __m128i topbit = _mm_set1_epi16(-32768);
766
767
const __m128 one = _mm_set1_ps(1.0f);
const __m128 negone = _mm_set1_ps(-1.0f);
768
769
__m128i *mmdst = (__m128i *) dst;
while (i >= 8) { /* 8 * float32 */
770
771
const __m128i ints1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src)), one), mulby32767)); /* load 4 floats, clamp, convert to sint32 */
const __m128i ints2 = _mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src+4)), one), mulby32767)); /* load 4 floats, clamp, convert to sint32 */
772
773
774
775
776
777
778
779
_mm_store_si128(mmdst, _mm_xor_si128(_mm_packs_epi32(ints1, ints2), topbit)); /* pack to sint16, xor top bit, store out. */
i -= 8; src += 8; mmdst++;
}
dst = (Uint16 *) mmdst;
}
/* Finish off any leftovers with scalar operations. */
while (i) {
780
781
782
783
784
785
786
787
const float sample = *src;
if (sample >= 1.0f) {
*dst = 65535;
} else if (sample <= -1.0f) {
*dst = 0;
} else {
*dst = (Uint16)((sample + 1.0f) * 32767.0f);
}
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
i--; src++; dst++;
}
cvt->len_cvt /= 2;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_U16SYS);
}
}
static void SDLCALL
SDL_Convert_F32_to_S32_SSE2(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const float *src = (const float *) cvt->buf;
Sint32 *dst = (Sint32 *) cvt->buf;
int i;
LOG_DEBUG_CONVERT("AUDIO_F32", "AUDIO_S32 (using SSE2)");
/* Get dst aligned to 16 bytes */
for (i = cvt->len_cvt / sizeof (float); i && (((size_t) dst) & 15); --i, ++src, ++dst) {
808
809
810
811
const float sample = *src;
if (sample >= 1.0f) {
*dst = 2147483647;
} else if (sample <= -1.0f) {
812
*dst = (Sint32) -2147483648LL;
813
814
815
} else {
*dst = ((Sint32)(sample * 8388607.0f)) << 8;
}
816
817
818
819
820
821
822
}
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
SDL_assert(!i || ((((size_t) src) & 15) == 0));
{
/* Aligned! Do SSE blocks as long as we have 16 bytes available. */
823
824
const __m128 one = _mm_set1_ps(1.0f);
const __m128 negone = _mm_set1_ps(-1.0f);
825
const __m128 mulby8388607 = _mm_set1_ps(8388607.0f);
826
827
__m128i *mmdst = (__m128i *) dst;
while (i >= 4) { /* 4 * float32 */
828
_mm_store_si128(mmdst, _mm_slli_epi32(_mm_cvtps_epi32(_mm_mul_ps(_mm_min_ps(_mm_max_ps(negone, _mm_load_ps(src)), one), mulby8388607)), 8)); /* load 4 floats, clamp, convert to sint32 */
829
830
831
832
833
834
835
i -= 4; src += 4; mmdst++;
}
dst = (Sint32 *) mmdst;
}
/* Finish off any leftovers with scalar operations. */
while (i) {
836
837
838
839
const float sample = *src;
if (sample >= 1.0f) {
*dst = 2147483647;
} else if (sample <= -1.0f) {
840
*dst = (Sint32) -2147483648LL;
841
842
843
} else {
*dst = ((Sint32)(sample * 8388607.0f)) << 8;
}
844
845
846
847
848
849
850
851
852
853
i--; src++; dst++;
}
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_S32SYS);
}
}
#endif
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#if HAVE_NEON_INTRINSICS
static void SDLCALL
SDL_Convert_S8_to_F32_NEON(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Sint8 *src = ((const Sint8 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 4)) - 1;
int i;
LOG_DEBUG_CONVERT("AUDIO_S8", "AUDIO_F32 (using NEON)");
/* Get dst aligned to 16 bytes (since buffer is growing, we don't have to worry about overreading from src) */
for (i = cvt->len_cvt; i && (((size_t) (dst-15)) & 15); --i, --src, --dst) {
*dst = ((float) *src) * DIVBY128;
}
src -= 15; dst -= 15; /* adjust to read NEON blocks from the start. */
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do NEON blocks as long as we have 16 bytes available. */
const int8_t *mmsrc = (const int8_t *) src;
const float32x4_t divby128 = vdupq_n_f32(DIVBY128);
while (i >= 16) { /* 16 * 8-bit */
const int8x16_t bytes = vld1q_s8(mmsrc); /* get 16 sint8 into a NEON register. */
const int16x8_t int16hi = vmovl_s8(vget_high_s8(bytes)); /* convert top 8 bytes to 8 int16 */
const int16x8_t int16lo = vmovl_s8(vget_low_s8(bytes)); /* convert bottom 8 bytes to 8 int16 */
/* split int16 to two int32, then convert to float, then multiply to normalize, store. */
vst1q_f32(dst, vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(int16hi))), divby128));
vst1q_f32(dst+4, vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(int16hi))), divby128));
vst1q_f32(dst+8, vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(int16lo))), divby128));
vst1q_f32(dst+12, vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(int16lo))), divby128));
i -= 16; mmsrc -= 16; dst -= 16;
}
src = (const Sint8 *) mmsrc;
}
src += 15; dst += 15; /* adjust for any scalar finishing. */
/* Finish off any leftovers with scalar operations. */
while (i) {
*dst = ((float) *src) * DIVBY128;
i--; src--; dst--;
}
cvt->len_cvt *= 4;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}
static void SDLCALL
SDL_Convert_U8_to_F32_NEON(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Uint8 *src = ((const Uint8 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 4)) - 1;
int i;
LOG_DEBUG_CONVERT("AUDIO_U8", "AUDIO_F32 (using NEON)");
/* Get dst aligned to 16 bytes (since buffer is growing, we don't have to worry about overreading from src) */
for (i = cvt->len_cvt; i && (((size_t) (dst-15)) & 15); --i, --src, --dst) {
*dst = (((float) *src) * DIVBY128) - 1.0f;
}
src -= 15; dst -= 15; /* adjust to read NEON blocks from the start. */
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do NEON blocks as long as we have 16 bytes available. */
const uint8_t *mmsrc = (const uint8_t *) src;
const float32x4_t divby128 = vdupq_n_f32(DIVBY128);
const float32x4_t one = vdupq_n_f32(1.0f);
while (i >= 16) { /* 16 * 8-bit */
const uint8x16_t bytes = vld1q_u8(mmsrc); /* get 16 uint8 into a NEON register. */
const uint16x8_t uint16hi = vmovl_u8(vget_high_u8(bytes)); /* convert top 8 bytes to 8 uint16 */
const uint16x8_t uint16lo = vmovl_u8(vget_low_u8(bytes)); /* convert bottom 8 bytes to 8 uint16 */
/* split uint16 to two uint32, then convert to float, then multiply to normalize, subtract to adjust for sign, store. */
vst1q_f32(dst, vmlsq_f32(vcvtq_f32_u32(vmovl_u16(vget_high_u16(uint16hi))), divby128, one));
vst1q_f32(dst+4, vmlsq_f32(vcvtq_f32_u32(vmovl_u16(vget_low_u16(uint16hi))), divby128, one));
vst1q_f32(dst+8, vmlsq_f32(vcvtq_f32_u32(vmovl_u16(vget_high_u16(uint16lo))), divby128, one));
vst1q_f32(dst+12, vmlsq_f32(vcvtq_f32_u32(vmovl_u16(vget_low_u16(uint16lo))), divby128, one));
i -= 16; mmsrc -= 16; dst -= 16;
}
src = (const Uint8 *) mmsrc;
}
src += 15; dst += 15; /* adjust for any scalar finishing. */
/* Finish off any leftovers with scalar operations. */
while (i) {
*dst = (((float) *src) * DIVBY128) - 1.0f;
i--; src--; dst--;
}
cvt->len_cvt *= 4;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}
static void SDLCALL
SDL_Convert_S16_to_F32_NEON(SDL_AudioCVT *cvt, SDL_AudioFormat format)
{
const Sint16 *src = ((const Sint16 *) (cvt->buf + cvt->len_cvt)) - 1;
float *dst = ((float *) (cvt->buf + cvt->len_cvt * 2)) - 1;
int i;
LOG_DEBUG_CONVERT("AUDIO_S16", "AUDIO_F32 (using NEON)");
/* Get dst aligned to 16 bytes (since buffer is growing, we don't have to worry about overreading from src) */
for (i = cvt->len_cvt / sizeof (Sint16); i && (((size_t) (dst-7)) & 15); --i, --src, --dst) {
*dst = ((float) *src) * DIVBY32768;
}
src -= 7; dst -= 7; /* adjust to read NEON blocks from the start. */
SDL_assert(!i || ((((size_t) dst) & 15) == 0));
/* Make sure src is aligned too. */
if ((((size_t) src) & 15) == 0) {
/* Aligned! Do NEON blocks as long as we have 16 bytes available. */
const float32x4_t divby32768 = vdupq_n_f32(DIVBY32768);
while (i >= 8) { /* 8 * 16-bit */
const int16x8_t ints = vld1q_s16((int16_t const *) src); /* get 8 sint16 into a NEON register. */
/* split int16 to two int32, then convert to float, then multiply to normalize, store. */
vst1q_f32(dst, vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(ints))), divby32768));
vst1q_f32(dst+4, vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(ints))), divby32768));
i -= 8; src -= 8; dst -= 8;
}
}
src += 7; dst += 7; /* adjust for any scalar finishing. */
/* Finish off any leftovers with scalar operations. */
while (i) {
*dst = ((float) *src) * DIVBY32768;
i--; src--; dst--;
}
cvt->len_cvt *= 2;
if (cvt->filters[++cvt->filter_index]) {
cvt->filters[cvt->filter_index](cvt, AUDIO_F32SYS);
}
}