| 389 | return npyv_select_f32(_mm_or_si128(nfinite_mask, overflow_mask), a, trunc); |
| 390 | } |
| 391 | NPY_FINLINE npyv_f64 npyv_trunc_f64(npyv_f64 a) |
| 392 | { |
| 393 | const __m128d one = _mm_set1_pd(1.0); |
| 394 | const __m128d szero = _mm_set1_pd(-0.0); |
| 395 | const __m128d two_power_52 = _mm_set1_pd(0x10000000000000); |
| 396 | __m128d nan_mask = _mm_cmpunord_pd(a, a); |
| 397 | // eliminate nans to avoid invalid fp errors within cmpge |
| 398 | __m128d abs_x = npyv_abs_f64(_mm_xor_pd(nan_mask, a)); |
| 399 | // round by add magic number 2^52 |
| 400 | // assuming that MXCSR register is set to rounding |
| 401 | __m128d abs_round = _mm_sub_pd(_mm_add_pd(two_power_52, abs_x), two_power_52); |
| 402 | __m128d subtrahend = _mm_and_pd(_mm_cmpgt_pd(abs_round, abs_x), one); |
| 403 | __m128d trunc = _mm_sub_pd(abs_round, subtrahend); |
| 404 | // copysign |
| 405 | trunc = _mm_or_pd(trunc, _mm_and_pd(a, szero)); |
| 406 | // a if |a| >= 2^52 or a == NaN |
| 407 | __m128d mask = _mm_cmpge_pd(abs_x, two_power_52); |
| 408 | mask = _mm_or_pd(mask, nan_mask); |
| 409 | return npyv_select_f64(_mm_castpd_si128(mask), a, trunc); |
| 410 | } |
| 411 | #endif |
| 412 | |
| 413 | // floor |
nothing calls this directly
no test coverage detected