25#ifndef __always_inline
26#define __always_inline inline __attribute__((always_inline))
48#define SSE_BUTTERFLY(M0, M1, M2, M3, M4) \
50 M3 = _mm_adds_epi16(M0, M2); \
51 M4 = _mm_subs_epi16(M1, M2); \
52 M0 = _mm_subs_epi16(M0, M2); \
53 M1 = _mm_adds_epi16(M1, M2); \
54 M2 = _mm_max_epi16(M3, M4); \
55 M3 = _mm_or_si128(_mm_cmpgt_epi16(M3, M4), _mm_cmpeq_epi16(M3, M4)); \
56 M4 = _mm_max_epi16(M0, M1); \
57 M1 = _mm_or_si128(_mm_cmpgt_epi16(M0, M1), _mm_cmpeq_epi16(M0, M1)); \
74#define _I8_SHUFFLE_MASK 15, 14, 11, 10, 7, 6, 3, 2, 13, 12, 9, 8, 5, 4, 1, 0
76#define SSE_DEINTERLEAVE_K5(M0, M1, M2, M3) \
78 M2 = _mm_set_epi8(_I8_SHUFFLE_MASK); \
79 M0 = _mm_shuffle_epi8(M0, M2); \
80 M1 = _mm_shuffle_epi8(M1, M2); \
81 M2 = _mm_unpacklo_epi64(M0, M1); \
82 M3 = _mm_unpackhi_epi64(M0, M1); \
99#define SSE_DEINTERLEAVE_K7(M0, M1, M2, M3, M4, M5, M6, M7, \
100 M8, M9, M10, M11, M12, M13, M14, M15) \
102 M8 = _mm_set_epi8(_I8_SHUFFLE_MASK); \
103 M0 = _mm_shuffle_epi8(M0, M8); \
104 M1 = _mm_shuffle_epi8(M1, M8); \
105 M2 = _mm_shuffle_epi8(M2, M8); \
106 M3 = _mm_shuffle_epi8(M3, M8); \
107 M4 = _mm_shuffle_epi8(M4, M8); \
108 M5 = _mm_shuffle_epi8(M5, M8); \
109 M6 = _mm_shuffle_epi8(M6, M8); \
110 M7 = _mm_shuffle_epi8(M7, M8); \
111 M8 = _mm_unpacklo_epi64(M0, M1); \
112 M9 = _mm_unpackhi_epi64(M0, M1); \
113 M10 = _mm_unpacklo_epi64(M2, M3); \
114 M11 = _mm_unpackhi_epi64(M2, M3); \
115 M12 = _mm_unpacklo_epi64(M4, M5); \
116 M13 = _mm_unpackhi_epi64(M4, M5); \
117 M14 = _mm_unpacklo_epi64(M6, M7); \
118 M15 = _mm_unpackhi_epi64(M6, M7); \
131#define SSE_BRANCH_METRIC_N2(M0, M1, M2, M3, M4, M6, M7) \
133 M0 = _mm_sign_epi16(M4, M0); \
134 M1 = _mm_sign_epi16(M4, M1); \
135 M2 = _mm_sign_epi16(M4, M2); \
136 M3 = _mm_sign_epi16(M4, M3); \
137 M6 = _mm_hadds_epi16(M0, M1); \
138 M7 = _mm_hadds_epi16(M2, M3); \
153#define SSE_BRANCH_METRIC_N4(M0, M1, M2, M3, M4, M5) \
155 M0 = _mm_sign_epi16(M4, M0); \
156 M1 = _mm_sign_epi16(M4, M1); \
157 M2 = _mm_sign_epi16(M4, M2); \
158 M3 = _mm_sign_epi16(M4, M3); \
159 M0 = _mm_hadds_epi16(M0, M1); \
160 M1 = _mm_hadds_epi16(M2, M3); \
161 M5 = _mm_hadds_epi16(M0, M1); \
177#if defined(HAVE_SSE4_1) || defined(HAVE_SSE41)
178#define SSE_MINPOS(M0, M1) \
180 if (sse41_supported) { \
181 M0 = _mm_minpos_epu16(M0); \
183 M1 = _mm_shuffle_epi32(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
184 M0 = _mm_min_epi16(M0, M1); \
185 M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
186 M0 = _mm_min_epi16(M0, M1); \
187 M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 0, 1)); \
188 M0 = _mm_min_epi16(M0, M1); \
192#define SSE_MINPOS(M0, M1) \
194 M1 = _mm_shuffle_epi32(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
195 M0 = _mm_min_epi16(M0, M1); \
196 M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
197 M0 = _mm_min_epi16(M0, M1); \
198 M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 0, 1)); \
199 M0 = _mm_min_epi16(M0, M1); \
215#define SSE_NORMALIZE_K5(M0, M1, M2, M3) \
217 M2 = _mm_min_epi16(M0, M1); \
220 M0 = _mm_subs_epi16(M0, M2); \
221 M1 = _mm_subs_epi16(M1, M2); \
235#define SSE_NORMALIZE_K7(M0, M1, M2, M3, M4, M5, M6, M7, M8, M9, M10, M11) \
237 M8 = _mm_min_epi16(M0, M1); \
238 M9 = _mm_min_epi16(M2, M3); \
239 M10 = _mm_min_epi16(M4, M5); \
240 M11 = _mm_min_epi16(M6, M7); \
241 M8 = _mm_min_epi16(M8, M9); \
242 M10 = _mm_min_epi16(M10, M11); \
243 M8 = _mm_min_epi16(M8, M10); \
246 M0 = _mm_subs_epi16(M0, M8); \
247 M1 = _mm_subs_epi16(M1, M8); \
248 M2 = _mm_subs_epi16(M2, M8); \
249 M3 = _mm_subs_epi16(M3, M8); \
250 M4 = _mm_subs_epi16(M4, M8); \
251 M5 = _mm_subs_epi16(M5, M8); \
252 M6 = _mm_subs_epi16(M6, M8); \
253 M7 = _mm_subs_epi16(M7, M8); \
263 const int16_t *out, int16_t *sums, int16_t *paths,
int norm)
265 __m128i m0, m1, m2, m3, m4, m5, m6;
268 m2 = _mm_castpd_si128(_mm_loaddup_pd((
double const *) val));
271 m0 = _mm_load_si128((__m128i *) &out[0]);
272 m1 = _mm_load_si128((__m128i *) &out[8]);
275 m0 = _mm_sign_epi16(m2, m0);
276 m1 = _mm_sign_epi16(m2, m1);
277 m2 = _mm_hadds_epi16(m0, m1);
280 m0 = _mm_load_si128((__m128i *) &sums[0]);
281 m1 = _mm_load_si128((__m128i *) &sums[8]);
291 _mm_store_si128((__m128i *) &sums[0], m2);
292 _mm_store_si128((__m128i *) &sums[8], m6);
293 _mm_store_si128((__m128i *) &paths[0], m5);
294 _mm_store_si128((__m128i *) &paths[8], m4);
305 const int16_t *out, int16_t *sums, int16_t *paths,
int norm)
307 __m128i m0, m1, m2, m3, m4, m5, m6;
310 m4 = _mm_castpd_si128(_mm_loaddup_pd((
double const *) val));
313 m0 = _mm_load_si128((__m128i *) &out[0]);
314 m1 = _mm_load_si128((__m128i *) &out[8]);
315 m2 = _mm_load_si128((__m128i *) &out[16]);
316 m3 = _mm_load_si128((__m128i *) &out[24]);
321 m0 = _mm_load_si128((__m128i *) &sums[0]);
322 m1 = _mm_load_si128((__m128i *) &sums[8]);
332 _mm_store_si128((__m128i *) &sums[0], m2);
333 _mm_store_si128((__m128i *) &sums[8], m6);
334 _mm_store_si128((__m128i *) &paths[0], m5);
335 _mm_store_si128((__m128i *) &paths[8], m4);
345 const int16_t *out, int16_t *sums, int16_t *paths,
int norm)
347 __m128i m0, m1, m2, m3, m4, m5, m6, m7, m8,
348 m9, m10, m11, m12, m13, m14, m15;
351 m0 = _mm_load_si128((__m128i *) &sums[0]);
352 m1 = _mm_load_si128((__m128i *) &sums[8]);
353 m2 = _mm_load_si128((__m128i *) &sums[16]);
354 m3 = _mm_load_si128((__m128i *) &sums[24]);
355 m4 = _mm_load_si128((__m128i *) &sums[32]);
356 m5 = _mm_load_si128((__m128i *) &sums[40]);
357 m6 = _mm_load_si128((__m128i *) &sums[48]);
358 m7 = _mm_load_si128((__m128i *) &sums[56]);
362 m8, m9, m10, m11, m12, m13, m14, m15)
365 m7 = _mm_castpd_si128(_mm_loaddup_pd((
double const *) val));
368 m0 = _mm_load_si128((__m128i *) &out[0]);
369 m1 = _mm_load_si128((__m128i *) &out[8]);
370 m2 = _mm_load_si128((__m128i *) &out[16]);
371 m3 = _mm_load_si128((__m128i *) &out[24]);
375 m0 = _mm_load_si128((__m128i *) &out[32]);
376 m1 = _mm_load_si128((__m128i *) &out[40]);
377 m2 = _mm_load_si128((__m128i *) &out[48]);
378 m3 = _mm_load_si128((__m128i *) &out[56]);
386 _mm_store_si128((__m128i *) &paths[0], m0);
387 _mm_store_si128((__m128i *) &paths[8], m2);
388 _mm_store_si128((__m128i *) &paths[32], m9);
389 _mm_store_si128((__m128i *) &paths[40], m11);
395 _mm_store_si128((__m128i *) &paths[16], m0);
396 _mm_store_si128((__m128i *) &paths[24], m9);
397 _mm_store_si128((__m128i *) &paths[48], m13);
398 _mm_store_si128((__m128i *) &paths[56], m15);
402 m7, m11, m0, m8, m9, m10)
404 _mm_store_si128((__m128i *) &sums[0], m4);
405 _mm_store_si128((__m128i *) &sums[8], m5);
406 _mm_store_si128((__m128i *) &sums[16], m6);
407 _mm_store_si128((__m128i *) &sums[24], m7);
408 _mm_store_si128((__m128i *) &sums[32], m1);
409 _mm_store_si128((__m128i *) &sums[40], m3);
410 _mm_store_si128((__m128i *) &sums[48], m2);
411 _mm_store_si128((__m128i *) &sums[56], m11);
420 const int16_t *out, int16_t *sums, int16_t *paths,
int norm)
422 __m128i m0, m1, m2, m3, m4, m5, m6, m7;
423 __m128i m8, m9, m10, m11, m12, m13, m14, m15;
426 m0 = _mm_load_si128((__m128i *) &sums[0]);
427 m1 = _mm_load_si128((__m128i *) &sums[8]);
428 m2 = _mm_load_si128((__m128i *) &sums[16]);
429 m3 = _mm_load_si128((__m128i *) &sums[24]);
430 m4 = _mm_load_si128((__m128i *) &sums[32]);
431 m5 = _mm_load_si128((__m128i *) &sums[40]);
432 m6 = _mm_load_si128((__m128i *) &sums[48]);
433 m7 = _mm_load_si128((__m128i *) &sums[56]);
437 m8, m9, m10, m11, m12, m13, m14, m15)
440 m7 = _mm_castpd_si128(_mm_loaddup_pd((
double const *) val));
443 m0 = _mm_load_si128((__m128i *) &out[0]);
444 m1 = _mm_load_si128((__m128i *) &out[8]);
445 m2 = _mm_load_si128((__m128i *) &out[16]);
446 m3 = _mm_load_si128((__m128i *) &out[24]);
450 m0 = _mm_load_si128((__m128i *) &out[32]);
451 m1 = _mm_load_si128((__m128i *) &out[40]);
452 m2 = _mm_load_si128((__m128i *) &out[48]);
453 m3 = _mm_load_si128((__m128i *) &out[56]);
457 m0 = _mm_load_si128((__m128i *) &out[64]);
458 m1 = _mm_load_si128((__m128i *) &out[72]);
459 m2 = _mm_load_si128((__m128i *) &out[80]);
460 m3 = _mm_load_si128((__m128i *) &out[88]);
464 m0 = _mm_load_si128((__m128i *) &out[96]);
465 m1 = _mm_load_si128((__m128i *) &out[104]);
466 m2 = _mm_load_si128((__m128i *) &out[112]);
467 m3 = _mm_load_si128((__m128i *) &out[120]);
475 _mm_store_si128((__m128i *) &paths[0], m0);
476 _mm_store_si128((__m128i *) &paths[8], m2);
477 _mm_store_si128((__m128i *) &paths[32], m9);
478 _mm_store_si128((__m128i *) &paths[40], m11);
484 _mm_store_si128((__m128i *) &paths[16], m0);
485 _mm_store_si128((__m128i *) &paths[24], m9);
486 _mm_store_si128((__m128i *) &paths[48], m13);
487 _mm_store_si128((__m128i *) &paths[56], m15);
491 m7, m11, m0, m8, m9, m10)
493 _mm_store_si128((__m128i *) &sums[0], m4);
494 _mm_store_si128((__m128i *) &sums[8], m5);
495 _mm_store_si128((__m128i *) &sums[16], m6);
496 _mm_store_si128((__m128i *) &sums[24], m7);
497 _mm_store_si128((__m128i *) &sums[32], m1);
498 _mm_store_si128((__m128i *) &sums[40], m3);
499 _mm_store_si128((__m128i *) &sums[48], m2);
500 _mm_store_si128((__m128i *) &sums[56], m11);
#define SSE_BRANCH_METRIC_N4(M0, M1, M2, M3, M4, M5)
Definition: conv_acc_sse_impl.h:153
#define SSE_DEINTERLEAVE_K5(M0, M1, M2, M3)
Definition: conv_acc_sse_impl.h:76
#define __always_inline
Definition: conv_acc_sse_impl.h:26
#define SSE_NORMALIZE_K5(M0, M1, M2, M3)
Definition: conv_acc_sse_impl.h:215
#define SSE_NORMALIZE_K7(M0, M1, M2, M3, M4, M5, M6, M7, M8, M9, M10, M11)
Definition: conv_acc_sse_impl.h:235
static __always_inline void _sse_metrics_k7_n4(const int16_t *val, const int16_t *out, int16_t *sums, int16_t *paths, int norm)
Definition: conv_acc_sse_impl.h:419
#define SSE_BUTTERFLY(M0, M1, M2, M3, M4)
Definition: conv_acc_sse_impl.h:48
#define SSE_BRANCH_METRIC_N2(M0, M1, M2, M3, M4, M6, M7)
Definition: conv_acc_sse_impl.h:131
static __always_inline void _sse_metrics_k5_n2(const int16_t *val, const int16_t *out, int16_t *sums, int16_t *paths, int norm)
Definition: conv_acc_sse_impl.h:262
static __always_inline void _sse_metrics_k7_n2(const int16_t *val, const int16_t *out, int16_t *sums, int16_t *paths, int norm)
Definition: conv_acc_sse_impl.h:344
static __always_inline void _sse_metrics_k5_n4(const int16_t *val, const int16_t *out, int16_t *sums, int16_t *paths, int norm)
Definition: conv_acc_sse_impl.h:304
#define SSE_DEINTERLEAVE_K7(M0, M1, M2, M3, M4, M5, M6, M7, M8, M9, M10, M11, M12, M13, M14, M15)
Definition: conv_acc_sse_impl.h:99