diff options
| -rw-r--r-- | numpy/core/code_generators/generate_umath.py | 2 | ||||
| -rw-r--r-- | numpy/core/setup.py | 10 | ||||
| -rw-r--r-- | numpy/core/src/umath/loops.h.src | 18 | ||||
| -rw-r--r-- | numpy/core/src/umath/loops_hyperbolic.dispatch.c.src | 395 | ||||
| -rw-r--r-- | numpy/core/src/umath/loops_umath_fp.dispatch.c.src | 8 |
5 files changed, 428 insertions, 5 deletions
diff --git a/numpy/core/code_generators/generate_umath.py b/numpy/core/code_generators/generate_umath.py index 054150b28..673a8e3b5 100644 --- a/numpy/core/code_generators/generate_umath.py +++ b/numpy/core/code_generators/generate_umath.py @@ -745,7 +745,7 @@ defdict = { docstrings.get('numpy.core.umath.tanh'), None, TD('e', f='tanh', astype={'e': 'f'}), - TD('fd', dispatch=[('loops_umath_fp', 'fd')]), + TD('fd', dispatch=[('loops_hyperbolic', 'fd')]), TD(inexact, f='tanh', astype={'e': 'f'}), TD(P, f='tanh'), ), diff --git a/numpy/core/setup.py b/numpy/core/setup.py index 3bf6f6250..100f2fc80 100644 --- a/numpy/core/setup.py +++ b/numpy/core/setup.py @@ -1003,6 +1003,7 @@ def configuration(parent_package='',top_path=None): join('src', 'umath', 'loops_trigonometric.dispatch.c.src'), join('src', 'umath', 'loops_umath_fp.dispatch.c.src'), join('src', 'umath', 'loops_exponent_log.dispatch.c.src'), + join('src', 'umath', 'loops_hyperbolic.dispatch.c.src'), join('src', 'umath', 'matmul.h.src'), join('src', 'umath', 'matmul.c.src'), join('src', 'umath', 'clip.h'), @@ -1033,8 +1034,17 @@ def configuration(parent_package='',top_path=None): svml_path = join('numpy', 'core', 'src', 'umath', 'svml') svml_objs = [] + # we have converted the following into universal intrinsics + # so we can bring the benefits of performance for all platforms + # not just for avx512 on linux without performance/accuracy regression, + # actually the other way around, better performance and + # after all maintainable code. + svml_filter = ( + 'svml_z0_tanh_d_la.s', 'svml_z0_tanh_s_la.s' + ) if can_link_svml() and check_svml_submodule(svml_path): svml_objs = glob.glob(svml_path + '/**/*.s', recursive=True) + svml_objs = [o for o in svml_objs if not o.endswith(svml_filter)] config.add_extension('_multiarray_umath', # Forcing C language even though we have C++ sources. diff --git a/numpy/core/src/umath/loops.h.src b/numpy/core/src/umath/loops.h.src index 3eafbdf66..19ae340f6 100644 --- a/numpy/core/src/umath/loops.h.src +++ b/numpy/core/src/umath/loops.h.src @@ -210,6 +210,24 @@ NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@kind@, /**end repeat**/ #ifndef NPY_DISABLE_OPTIMIZATION + #include "loops_hyperbolic.dispatch.h" +#endif +/**begin repeat + * #TYPE = FLOAT, DOUBLE# + */ +/**begin repeat1 + * #func = tanh# + */ +NPY_CPU_DISPATCH_DECLARE(NPY_NO_EXPORT void @TYPE@_@func@, + (char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))) +/**end repeat1**/ +/**end repeat**/ + +/**end repeat1**/ +/**end repeat**/ + +// SVML +#ifndef NPY_DISABLE_OPTIMIZATION #include "loops_umath_fp.dispatch.h" #endif diff --git a/numpy/core/src/umath/loops_hyperbolic.dispatch.c.src b/numpy/core/src/umath/loops_hyperbolic.dispatch.c.src new file mode 100644 index 000000000..0fda06fff --- /dev/null +++ b/numpy/core/src/umath/loops_hyperbolic.dispatch.c.src @@ -0,0 +1,395 @@ +/*@targets + ** $maxopt baseline + ** (avx2 fma3) AVX512_SKX + ** vsx2 + ** neon_vfpv4 + **/ +#include "numpy/npy_math.h" +#include "simd/simd.h" +#include "loops_utils.h" +#include "loops.h" + +#if NPY_SIMD_FMA3 // native support +/* + * NOTE: The following implementation of tanh(f32, f64) have been converted from + * Intel SVML to universal intrinsics, and the original code can be found in: + * + * - https://github.com/numpy/SVML/blob/main/linux/avx512/svml_z0_tanh_d_la.s + * - https://github.com/numpy/SVML/blob/main/linux/avx512/svml_z0_tanh_s_la.s + * + * ALGORITHM DESCRIPTION: + * + * NOTE: Since the hyperbolic tangent function is odd + * (tanh(x) = -tanh(-x)), below algorithm deals with the absolute + * value of the argument |x|: tanh(x) = sign(x) * tanh(|x|) + * + * We use a table lookup method to compute tanh(|x|). + * The basic idea is to split the input range into a number of subintervals + * and to approximate tanh(.) with a polynomial on each of them. + * + * IEEE SPECIAL CONDITIONS: + * x = [+,-]0, r = [+,-]0 + * x = +Inf, r = +1 + * x = -Inf, r = -1 + * x = QNaN, r = QNaN + * x = SNaN, r = QNaN + * + * + * ALGORITHM DETAILS + * We handle special values in a callout function, aside from main path + * computations. "Special" for this algorithm are: + * INF, NAN, |x| > HUGE_THRESHOLD + * + * + * Main path computations are organized as follows: + * Actually we split the interval [0, SATURATION_THRESHOLD) + * into a number of subintervals. On each subinterval we approximate tanh(.) + * with a minimax polynomial of pre-defined degree. Polynomial coefficients + * are computed beforehand and stored in table. We also use + * + * y := |x| + B, + * + * here B depends on subinterval and is used to make argument + * closer to zero. + * We also add large fake interval [SATURATION_THRESHOLD, HUGE_THRESHOLD], + * where 1.0 + 0.0*y + 0.0*y^2 ... coefficients are stored - just to + * preserve main path computation logic but return 1.0 for all arguments. + * + * Hence reconstruction looks as follows: + * we extract proper polynomial and range reduction coefficients + * (Pj and B), corresponding to subinterval, to which |x| belongs, + * and return + * + * r := sign(x) * (P0 + P1 * y + ... + Pn * y^n) + * + * NOTE: we use multiprecision technique to multiply and sum the first + * K terms of the polynomial. So Pj, j = 0..K are stored in + * table each as a pair of target precision numbers (Pj and PLj) to + * achieve wider than target precision. + * + */ +#if NPY_SIMD_F64 +static void +simd_tanh_f64(const double *src, npy_intp ssrc, double *dst, npy_intp sdst, npy_intp len) +{ + static const npy_uint64 NPY_DECL_ALIGNED(NPY_SIMD_WIDTH) lut16x18[] = { + // 0 + 0x0ull, 0x3fcc000000000000ull, 0x3fd4000000000000ull, 0x3fdc000000000000ull, + 0x3fe4000000000000ull, 0x3fec000000000000ull, 0x3ff4000000000000ull, 0x3ffc000000000000ull, + 0x4004000000000000ull, 0x400c000000000000ull, 0x4014000000000000ull, 0x401c000000000000ull, + 0x4024000000000000ull, 0x402c000000000000ull, 0x4034000000000000ull, 0x0ull, + // 1 + 0x0ull, 0x3fcb8fd0416a7c92ull, 0x3fd35f98a0ea650eull, 0x3fda5729ee488037ull, + 0x3fe1bf47eabb8f95ull, 0x3fe686650b8c2015ull, 0x3feb2523bb6b2deeull, 0x3fee1fbf97e33527ull, + 0x3fef9258260a71c2ull, 0x3feff112c63a9077ull, 0x3fefff419668df11ull, 0x3feffffc832750f2ull, + 0x3feffffffdc96f35ull, 0x3fefffffffffcf58ull, 0x3ff0000000000000ull, 0x3ff0000000000000ull, + // 2 + 0x3ff0000000000000ull, 0x3fee842ca3f08532ull, 0x3fed11574af58f1bull, 0x3fea945b9c24e4f9ull, + 0x3fe6284c3374f815ull, 0x3fe02500a09f8d6eull, 0x3fd1f25131e3a8c0ull, 0x3fbd22ca1c24a139ull, + 0x3f9b3afe1fba5c76ull, 0x3f6dd37d19b22b21ull, 0x3f27ccec13a9ef96ull, 0x3ecbe6c3f33250aeull, + 0x3e41b4865394f75full, 0x3d8853f01bda5f28ull, 0x3c73953c0197ef58ull, 0x0ull, + // 3 + 0xbbf0b3ea3fdfaa19ull, 0xbfca48aaeb53bc21ull, 0xbfd19921f4329916ull, 0xbfd5e0f09bef8011ull, + 0xbfd893b59c35c882ull, 0xbfd6ba7cb7576538ull, 0xbfce7291743d7555ull, 0xbfbb6d85a01efb80ull, + 0xbf9addae58c7141aull, 0xbf6dc59376c7aa19ull, 0xbf27cc5e74677410ull, 0xbecbe6c0e8b4cc87ull, + 0xbe41b486526b0565ull, 0xbd8853f01bef63a4ull, 0xbc73955be519be31ull, 0x0ull, + // 4 + 0xbfd5555555555555ull, 0xbfd183afc292ba11ull, 0xbfcc1a4b039c9bfaull, 0xbfc16e1e6d8d0be6ull, + 0xbf92426c751e48a2ull, 0x3fb4f152b2bad124ull, 0x3fbbba40cbef72beull, 0x3fb01ba038be6a3dull, + 0x3f916df44871efc8ull, 0x3f63c6869dfc8870ull, 0x3f1fb9aef915d828ull, 0x3ec299d1e27c6e11ull, + 0x3e379b5ddcca334cull, 0x3d8037f57bc62c9aull, 0x3c6a2d4b50a2cff7ull, 0x0ull, + // 5 + 0xbce6863ee44ed636ull, 0x3fc04dcd0476c75eull, 0x3fc43d3449a80f08ull, 0x3fc5c26f3699b7e7ull, + 0x3fc1a686f6ab2533ull, 0x3faf203c316ce730ull, 0xbf89c7a02788557cull, 0xbf98157e26e0d541ull, + 0xbf807b55c1c7d278ull, 0xbf53a18d5843190full, 0xbf0fb6bbc89b1a5bull, 0xbeb299c9c684a963ull, + 0xbe279b5dd4fb3d01ull, 0xbd7037f57ae72aa6ull, 0xbc5a2ca2bba78e86ull, 0x0ull, + // 6 + 0x3fc1111111112ab5ull, 0x3fb5c19efdfc08adull, 0x3fa74c98dc34fbacull, 0xbf790d6a8eff0a77ull, + 0xbfac3c021789a786ull, 0xbfae2196b7326859ull, 0xbf93a7a011ff8c2aull, 0x3f6e4709c7e8430eull, + 0x3f67682afa611151ull, 0x3f3ef2ee77717cbfull, 0x3ef95a4482f180b7ull, 0x3e9dc2c27da3b603ull, + 0x3e12e2afd9f7433eull, 0x3d59f320348679baull, 0x3c44b61d9bbcc940ull, 0x0ull, + // 7 + 0xbda1ea19ddddb3b4ull, 0xbfb0b8df995ce4dfull, 0xbfb2955cf41e8164ull, 0xbfaf9d05c309f7c6ull, + 0xbf987d27ccff4291ull, 0x3f8b2ca62572b098ull, 0x3f8f1cf6c7f5b00aull, 0x3f60379811e43dd5ull, + 0xbf4793826f78537eull, 0xbf2405695e36240full, 0xbee0e08de39ce756ull, 0xbe83d709ba5f714eull, + 0xbdf92e3fc5ee63e0ull, 0xbd414cc030f2110eull, 0xbc2ba022e8d82a87ull, 0x0ull, + // 8 + 0xbfaba1ba1990520bull, 0xbf96e37bba52f6fcull, 0x3ecff7df18455399ull, 0x3f97362834d33a4eull, + 0x3f9e7f8380184b45ull, 0x3f869543e7c420d4ull, 0xbf7326bd4914222aull, 0xbf5fc15b0a9d98faull, + 0x3f14cffcfa69fbb6ull, 0x3f057e48e5b79d10ull, 0x3ec33b66d7d77264ull, 0x3e66ac4e578b9b10ull, + 0x3ddcc74b8d3d5c42ull, 0x3d23c589137f92b4ull, 0x3c107f8e2c8707a1ull, 0x0ull, + // 9 + 0xbe351ca7f096011full, 0x3f9eaaf3320c3851ull, 0x3f9cf823fe761fc1ull, 0x3f9022271754ff1full, + 0xbf731fe77c9c60afull, 0xbf84a6046865ec7dull, 0xbf4ca3f1f2b9192bull, 0x3f4c77dee0afd227ull, + 0x3f04055bce68597aull, 0xbee2bf0cb4a71647ull, 0xbea31eaafe73efd5ull, 0xbe46abb02c4368edull, + 0xbdbcc749ca8079ddull, 0xbd03c5883836b9d2ull, 0xbbf07a5416264aecull, 0x0ull, + // 10 + 0x3f9664f94e6ac14eull, 0xbf94d3343bae39ddull, 0xbf7bc748e60df843ull, 0xbf8c89372b43ba85ull, + 0xbf8129a092de747aull, 0x3f60c85b4d538746ull, 0x3f5be9392199ec18ull, 0xbf2a0c68a4489f10ull, + 0xbf00462601dc2faaull, 0x3eb7b6a219dea9f4ull, 0x3e80cbcc8d4c5c8aull, 0x3e2425bb231a5e29ull, + 0x3d9992a4beac8662ull, 0x3ce191ba5ed3fb67ull, 0x3bc892450bad44c4ull, 0x0ull, + // 11 + 0xbea8c4c1fd7852feull, 0xbfccce16b1046f13ull, 0xbf81a16f224bb7b6ull, 0xbf62cbf00406bc09ull, + 0x3f75b29bb02cf69bull, 0x3f607df0f9f90c17ull, 0xbf4b852a6e0758d5ull, 0xbf0078c63d1b8445ull, + 0x3eec12eadd55be7aull, 0xbe6fa600f593181bull, 0xbe5a3c935dce3f7dull, 0xbe001c6d95e3ae96ull, + 0xbd74755a00ea1fd3ull, 0xbcbc1c6c063bb7acull, 0xbba3be9a4460fe00ull, 0x0ull, + // 12 + 0xbf822404577aa9ddull, 0x403d8b07f7a82aa3ull, 0xbf9f44ab92fbab0aull, 0x3fb2eac604473d6aull, + 0x3f45f87d903aaac8ull, 0xbf5e104671036300ull, 0x3f19bc98ddf0f340ull, 0x3f0d4304bc9246e8ull, + 0xbed13c415f7b9d41ull, 0xbe722b8d9720cdb0ull, 0x3e322666d739bec0ull, 0x3dd76a553d7e7918ull, + 0x3d4de0fa59416a39ull, 0x3c948716cf3681b4ull, 0x3b873f9f2d2fda99ull, 0x0ull, + // 13 + 0xbefdd99a221ed573ull, 0x4070593a3735bab4ull, 0xbfccab654e44835eull, 0x3fd13ed80037dbacull, + 0xbf6045b9076cc487ull, 0x3f2085ee7e8ac170ull, 0x3f23524622610430ull, 0xbeff12a6626911b4ull, + 0x3eab9008bca408afull, 0x3e634df71865f620ull, 0xbe05bb1bcf83ca73ull, 0xbdaf2ac143fb6762ull, + 0xbd23eae52a3dbf57ull, 0xbc6b5e3e9ca0955eull, 0xbb5eca68e2c1ba2eull, 0x0ull, + // 14 + 0x3f6e3be689423841ull, 0xc0d263511f5baac1ull, 0x40169f73b15ebe5cull, 0xc025c1dd41cd6cb5ull, + 0xbf58fd89fe05e0d1ull, 0x3f73f7af01d5af7aull, 0xbf1e40bdead17e6bull, 0x3ee224cd6c4513e5ull, + 0xbe24b645e68eeaa3ull, 0xbe4abfebfb72bc83ull, 0x3dd51c38f8695ed3ull, 0x3d8313ac38c6832bull, + 0x3cf7787935626685ull, 0x3c401ffc49c6bc29ull, 0xbabf0b21acfa52abull, 0x0ull, + // 15 + 0xbf2a1306713a4f3aull, 0xc1045e509116b066ull, 0x4041fab9250984ceull, 0xc0458d090ec3de95ull, + 0xbf74949d60113d63ull, 0x3f7c9fd6200d0adeull, 0x3f02cd40e0ad0a9full, 0xbe858ab8e019f311ull, + 0xbe792fa6323b7cf8ull, 0x3e2df04d67876402ull, 0xbd95c72be95e4d2cull, 0xbd55a89c30203106ull, + 0xbccad6b3bb9eff65ull, 0xbc12705ccd3dd884ull, 0xba8e0a4c47ae75f5ull, 0x0ull, + // 16 + 0xbf55d7e76dc56871ull, 0x41528c38809c90c7ull, 0xc076d57fb5190b02ull, 0x4085f09f888f8adaull, + 0x3fa246332a2fcba5ull, 0xbfb29d851a896fcdull, 0x3ed9065ae369b212ull, 0xbeb8e1ba4c98a030ull, + 0x3e6ffd0766ad4016ull, 0xbe0c63c29f505f5bull, 0xbd7fab216b9e0e49ull, 0x3d2826b62056aa27ull, + 0x3ca313e31762f523ull, 0x3bea37aa21895319ull, 0x3ae5c7f1fd871496ull, 0x0ull, + // 17 + 0x3f35e67ab76a26e7ull, 0x41848ee0627d8206ull, 0xc0a216d618b489ecull, 0x40a5b89107c8af4full, + 0x3fb69d8374520edaull, 0xbfbded519f981716ull, 0xbef02d288b5b3371ull, 0x3eb290981209c1a6ull, + 0xbe567e924bf5ff6eull, 0x3de3f7f7de6b0eb6ull, 0x3d69ed18bae3ebbcull, 0xbcf7534c4f3dfa71ull, + 0xbc730b73f1eaff20ull, 0xbbba2cff8135d462ull, 0xbab5a71b5f7d9035ull, 0x0ull + }; + const int nlanes = npyv_nlanes_f64; + for (; len > 0; len -= nlanes, src += ssrc*nlanes, dst += sdst*nlanes) { + npyv_f64 x; + if (ssrc == 1) { + x = npyv_load_tillz_f64(src, len); + } else { + x = npyv_loadn_tillz_f64(src, ssrc, len); + } + npyv_s64 denorm = npyv_and_s64(npyv_reinterpret_s64_f64(x), npyv_setall_s64(0x7ff8000000000000ll)); + // |x| > HUGE_THRESHOLD, INF and NaNs are filtered out to callout. + npyv_b64 callout_m = npyv_cmple_s64(denorm, npyv_setall_s64(0x7fe0000000000000ll)); + npyv_s64 idxs = npyv_sub_s64(denorm, npyv_setall_s64(0x3fc0000000000000ll)); + // no native 64-bit for max/min and its fine to use 32-bit max/min + // since we're not crossing 32-bit edge + npyv_s32 idxl = npyv_max_s32(npyv_reinterpret_s32_s64(idxs), npyv_zero_s32()); + idxl = npyv_min_s32(idxl, npyv_setall_s32(0x780000)); + npyv_u64 idx = npyv_shri_u64(npyv_reinterpret_u64_s32(idxl), 51); + + npyv_f64 b = npyv_lut16_f64((const double*)lut16x18 + 16*0, idx); + npyv_f64 c0 = npyv_lut16_f64((const double*)lut16x18 + 1*16, idx); + npyv_f64 c1 = npyv_lut16_f64((const double*)lut16x18 + 2*16, idx); + npyv_f64 c2 = npyv_lut16_f64((const double*)lut16x18 + 3*16, idx); + npyv_f64 c3 = npyv_lut16_f64((const double*)lut16x18 + 4*16, idx); + npyv_f64 c4 = npyv_lut16_f64((const double*)lut16x18 + 5*16, idx); + npyv_f64 c5 = npyv_lut16_f64((const double*)lut16x18 + 6*16, idx); + npyv_f64 c6 = npyv_lut16_f64((const double*)lut16x18 + 7*16, idx); + npyv_f64 c7 = npyv_lut16_f64((const double*)lut16x18 + 8*16, idx); + npyv_f64 c8 = npyv_lut16_f64((const double*)lut16x18 + 9*16, idx); + npyv_f64 c9 = npyv_lut16_f64((const double*)lut16x18 + 10*16, idx); + npyv_f64 c10 = npyv_lut16_f64((const double*)lut16x18 + 11*16, idx); + npyv_f64 c11 = npyv_lut16_f64((const double*)lut16x18 + 12*16, idx); + npyv_f64 c12 = npyv_lut16_f64((const double*)lut16x18 + 13*16, idx); + npyv_f64 c13 = npyv_lut16_f64((const double*)lut16x18 + 14*16, idx); + npyv_f64 c14 = npyv_lut16_f64((const double*)lut16x18 + 15*16, idx); + npyv_f64 c15 = npyv_lut16_f64((const double*)lut16x18 + 16*16, idx); + npyv_f64 c16 = npyv_lut16_f64((const double*)lut16x18 + 17*16, idx); + + // no need to zerofy nans or avoid FP exceptions by NO_EXC like SVML does + // since we're clearing the FP status anyway. + npyv_f64 sign = npyv_and_f64(x, npyv_reinterpret_f64_s64(npyv_setall_s64(0x8000000000000000ull))); + npyv_f64 y = npyv_sub_f64(npyv_abs_f64(x), b); + npyv_f64 r = npyv_muladd_f64(c16, y, c15); + r = npyv_muladd_f64(r, y, c14); + r = npyv_muladd_f64(r, y, c13); + r = npyv_muladd_f64(r, y, c12); + r = npyv_muladd_f64(r, y, c11); + r = npyv_muladd_f64(r, y, c10); + r = npyv_muladd_f64(r, y, c9); + r = npyv_muladd_f64(r, y, c8); + r = npyv_muladd_f64(r, y, c7); + r = npyv_muladd_f64(r, y, c6); + r = npyv_muladd_f64(r, y, c5); + r = npyv_muladd_f64(r, y, c4); + r = npyv_muladd_f64(r, y, c3); + r = npyv_muladd_f64(r, y, c2); + r = npyv_muladd_f64(r, y, c1); + r = npyv_muladd_f64(r, y, c0); + r = npyv_or_f64(r, sign); + + npy_uint64 callout = npyv_tobits_b64(callout_m); + if (NPY_UNLIKELY(callout != ((1 << nlanes) - 1))) { + double NPY_DECL_ALIGNED(NPY_SIMD_WIDTH) back2c[npyv_nlanes_f64]; + npyv_storea_f64(back2c, x); + for (unsigned i = 0; i < npyv_nlanes_f64; ++i) { + if ((callout >> i) & 1) { + continue; + } + back2c[i] = npy_tanh(back2c[i]); + } + r = npyv_select_f64(callout_m, r, npyv_loada_f64(back2c)); + } + if (sdst == 1) { + npyv_store_till_f64(dst, len, r); + } else { + npyv_storen_till_f64(dst, sdst, len, r); + } + } +} +#endif // NPY_SIMD_F64 +static void +simd_tanh_f32(const float *src, npy_intp ssrc, float *dst, npy_intp sdst, npy_intp len) +{ + static const npy_uint32 NPY_DECL_ALIGNED(NPY_SIMD_WIDTH) lut32x8[] = { + // 0 + 0x0, 0x3d700000, 0x3d900000, 0x3db00000, 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000, + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, 0x40d00000, 0x40f00000, 0x41100000, 0x0, + // 1 + 0x0, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0, + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000, + // 2 + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360, + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, 0x3717b0da, 0x35a43bce, 0x338306c6, 0x0, + // 3 + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92, + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, 0xb717b096, 0xb5a43bae, 0xb383012c, 0x0, + // 4 + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e, + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x0, + // 5 + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4, + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x0, + // 6 + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad, + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x0, + // 7 + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0, + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x0 + }; + + const int nlanes = npyv_nlanes_f32; + for (; len > 0; len -= nlanes, src += ssrc*nlanes, dst += sdst*nlanes) { + npyv_f32 x; + if (ssrc == 1) { + x = npyv_load_tillz_f32(src, len); + } else { + x = npyv_loadn_tillz_f32(src, ssrc, len); + } + npyv_s32 denorm = npyv_and_s32(npyv_reinterpret_s32_f32(x), npyv_setall_s32(0x7fe00000)); + // |x| > HUGE_THRESHOLD, INF and NaNs are filtered out to callout. + npyv_b32 callout_m = npyv_cmple_s32(denorm, npyv_setall_s32(0x7f000000)); + + npyv_s32 idxs = npyv_sub_s32(denorm, npyv_setall_s32(0x3d400000)); + idxs = npyv_max_s32(idxs, npyv_zero_s32()); + idxs = npyv_min_s32(idxs, npyv_setall_s32(0x3e00000)); + npyv_u32 idx = npyv_shri_u32(npyv_reinterpret_u32_s32(idxs), 21); + + npyv_f32 b = npyv_lut32_f32((const float*)lut32x8 + 32*0, idx); + npyv_f32 c0 = npyv_lut32_f32((const float*)lut32x8 + 32*1, idx); + npyv_f32 c1 = npyv_lut32_f32((const float*)lut32x8 + 32*2, idx); + npyv_f32 c2 = npyv_lut32_f32((const float*)lut32x8 + 32*3, idx); + npyv_f32 c3 = npyv_lut32_f32((const float*)lut32x8 + 32*4, idx); + npyv_f32 c4 = npyv_lut32_f32((const float*)lut32x8 + 32*5, idx); + npyv_f32 c5 = npyv_lut32_f32((const float*)lut32x8 + 32*6, idx); + npyv_f32 c6 = npyv_lut32_f32((const float*)lut32x8 + 32*7, idx); + + // no need to zerofy nans or avoid FP exceptions by NO_EXC like SVML does + // since we're clearing the FP status anyway. + npyv_f32 sign = npyv_and_f32(x, npyv_reinterpret_f32_u32(npyv_setall_u32(0x80000000))); + npyv_f32 y = npyv_sub_f32(npyv_abs_f32(x), b); + npyv_f32 r = npyv_muladd_f32(c6, y, c5); + r = npyv_muladd_f32(r, y, c4); + r = npyv_muladd_f32(r, y, c3); + r = npyv_muladd_f32(r, y, c2); + r = npyv_muladd_f32(r, y, c1); + r = npyv_muladd_f32(r, y, c0); + r = npyv_or_f32(r, sign); + + npy_uint64 callout = npyv_tobits_b32(callout_m); + if (NPY_UNLIKELY(callout != ((1 << nlanes) - 1))) { + float NPY_DECL_ALIGNED(NPY_SIMD_WIDTH) back2c[npyv_nlanes_f32]; + npyv_storea_f32(back2c, x); + for (unsigned i = 0; i < npyv_nlanes_f32; ++i) { + if ((callout >> i) & 1) { + continue; + } + back2c[i] = npy_tanhf(back2c[i]); + } + r = npyv_select_f32(callout_m, r, npyv_loada_f32(back2c)); + } + if (sdst == 1) { + npyv_store_till_f32(dst, len, r); + } else { + npyv_storen_till_f32(dst, sdst, len, r); + } + } +} +#endif // NPY_SIMD_FMA3 + +/**begin repeat + * #TYPE = FLOAT, DOUBLE# + * #type = float, double# + * #sfx = f32, f64# + * #ssfx = f, # + * #simd = NPY_SIMD_FMA3, NPY_SIMD_FMA3 && NPY_SIMD_F64# + */ +/**begin repeat1 + * #func = tanh# + * #simd_req_clear = 1# + */ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@func@) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data)) +{ + const @type@ *src = (@type@*)args[0]; + @type@ *dst = (@type@*)args[1]; + + const int lsize = sizeof(src[0]); + const npy_intp ssrc = steps[0] / lsize; + const npy_intp sdst = steps[1] / lsize; + npy_intp len = dimensions[0]; + assert(len <= 1 || (steps[0] % lsize == 0 && steps[1] % lsize == 0)); +#if @simd@ + if (is_mem_overlap(src, steps[0], dst, steps[1], len) || + !npyv_loadable_stride_@sfx@(ssrc) || !npyv_storable_stride_@sfx@(sdst) + ) { + for (; len > 0; --len, src += ssrc, dst += sdst) { + simd_@func@_@sfx@(src, 1, dst, 1, 1); + } + } else { + simd_@func@_@sfx@(src, ssrc, dst, sdst, len); + } + npyv_cleanup(); + #if @simd_req_clear@ + npy_clear_floatstatus_barrier((char*)dimensions); + #endif +#else + for (; len > 0; --len, src += ssrc, dst += sdst) { + const @type@ src0 = *src; + *dst = npy_@func@@ssfx@(src0); + } +#endif +} +/**end repeat1**/ +/**end repeat**/ diff --git a/numpy/core/src/umath/loops_umath_fp.dispatch.c.src b/numpy/core/src/umath/loops_umath_fp.dispatch.c.src index a8289fc51..281b4231d 100644 --- a/numpy/core/src/umath/loops_umath_fp.dispatch.c.src +++ b/numpy/core/src/umath/loops_umath_fp.dispatch.c.src @@ -14,8 +14,8 @@ * #func_suffix = f16, 8# */ /**begin repeat1 - * #func = tanh, exp2, log2, log10, expm1, log1p, cbrt, tan, asin, acos, atan, sinh, cosh, asinh, acosh, atanh# - * #default_val = 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0# + * #func = exp2, log2, log10, expm1, log1p, cbrt, tan, asin, acos, atan, sinh, cosh, asinh, acosh, atanh# + * #default_val = 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0# */ static void simd_@func@_@sfx@(const npyv_lanetype_@sfx@ *src, npy_intp ssrc, @@ -83,8 +83,8 @@ simd_@func@_f64(const double *src, npy_intp ssrc, * #sfx = f64, f32# */ /**begin repeat1 - * #func = tanh, exp2, log2, log10, expm1, log1p, cbrt, tan, arcsin, arccos, arctan, sinh, cosh, arcsinh, arccosh, arctanh# - * #intrin = tanh, exp2, log2, log10, expm1, log1p, cbrt, tan, asin, acos, atan, sinh, cosh, asinh, acosh, atanh# + * #func = exp2, log2, log10, expm1, log1p, cbrt, tan, arcsin, arccos, arctan, sinh, cosh, arcsinh, arccosh, arctanh# + * #intrin = exp2, log2, log10, expm1, log1p, cbrt, tan, asin, acos, atan, sinh, cosh, asinh, acosh, atanh# */ NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@func@) (char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data)) |