diff options
Diffstat (limited to 'src/core/NEON/SVEMath.inl')
-rw-r--r-- | src/core/NEON/SVEMath.inl | 481 |
1 files changed, 481 insertions, 0 deletions
diff --git a/src/core/NEON/SVEMath.inl b/src/core/NEON/SVEMath.inl new file mode 100644 index 0000000000..fdf94f0859 --- /dev/null +++ b/src/core/NEON/SVEMath.inl @@ -0,0 +1,481 @@ +/* + * Copyright (c) 2020-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#ifndef ACL_SRC_CORE_NEON_SVEMATH_INL +#define ACL_SRC_CORE_NEON_SVEMATH_INL + +#include <cmath> +#include <limits> + +#if defined(__ARM_FEATURE_SVE) && defined(ARM_COMPUTE_ENABLE_SVE) + +#ifndef M_PI +#define M_PI (3.14159265358979323846) +#endif // M_PI + +namespace arm_compute +{ +inline svfloat32_t svtaylor_poly_f32_z(svbool_t pg, + svfloat32_t x, + svfloat32_t coeff_1, + svfloat32_t coeff_2, + svfloat32_t coeff_3, + svfloat32_t coeff_4, + svfloat32_t coeff_5, + svfloat32_t coeff_6, + svfloat32_t coeff_7, + svfloat32_t coeff_8) +{ + const auto A = svmla_f32_z(pg, coeff_1, coeff_5, x); + const auto B = svmla_f32_z(pg, coeff_3, coeff_7, x); + const auto C = svmla_f32_z(pg, coeff_2, coeff_6, x); + const auto D = svmla_f32_z(pg, coeff_4, coeff_8, x); + const auto x2 = svmul_f32_z(pg, x, x); + const auto x4 = svmul_f32_z(pg, x2, x2); + const auto res = svmla_f32_z(pg, svmla_f32_z(pg, A, B, x2), svmla_f32_z(pg, C, D, x2), x4); + return res; +} + +inline svfloat16_t svtaylor_poly_f16_z(svbool_t pg, + svfloat16_t x, + svfloat16_t coeff_1, + svfloat16_t coeff_2, + svfloat16_t coeff_3, + svfloat16_t coeff_4, + svfloat16_t coeff_5, + svfloat16_t coeff_6, + svfloat16_t coeff_7, + svfloat16_t coeff_8) +{ + const auto A = svmla_f16_z(pg, coeff_1, coeff_5, x); + const auto B = svmla_f16_z(pg, coeff_3, coeff_7, x); + const auto C = svmla_f16_z(pg, coeff_2, coeff_6, x); + const auto D = svmla_f16_z(pg, coeff_4, coeff_8, x); + const auto x2 = svmul_f16_z(pg, x, x); + const auto x4 = svmul_f16_z(pg, x2, x2); + const auto res = svmla_f16_z(pg, svmla_f16_z(pg, A, B, x2), svmla_f16_z(pg, C, D, x2), x4); + return res; +} + +inline svfloat16_t svinv_f16_z(svbool_t pg, svfloat16_t x) +{ + auto recip = svrecpe_f16(x); + recip = svmul_f16_z(pg, svrecps_f16(x, recip), recip); + recip = svmul_f16_z(pg, svrecps_f16(x, recip), recip); + return recip; +} + +inline svfloat32_t svinv_f32_z(svbool_t pg, svfloat32_t x) +{ + auto recip = svrecpe_f32(x); + recip = svmul_f32_z(pg, svrecps_f32(x, recip), recip); + recip = svmul_f32_z(pg, svrecps_f32(x, recip), recip); + return recip; +} + +static const uint32_t svexp_f32_coeff[] = { + 0x3f7ffff6, // x^1: 0x1.ffffecp-1f + 0x3efffedb, // x^2: 0x1.fffdb6p-2f + 0x3e2aaf33, // x^3: 0x1.555e66p-3f + 0x3d2b9f17, // x^4: 0x1.573e2ep-5f + 0x3c072010, // x^5: 0x1.0e4020p-7f +}; + +inline svfloat32_t svexp_f32_z(svbool_t pg, svfloat32_t x) +{ + const auto c1 = svreinterpret_f32_u32(svdup_n_u32(svexp_f32_coeff[0])); + const auto c2 = svreinterpret_f32_u32(svdup_n_u32(svexp_f32_coeff[1])); + const auto c3 = svreinterpret_f32_u32(svdup_n_u32(svexp_f32_coeff[2])); + const auto c4 = svreinterpret_f32_u32(svdup_n_u32(svexp_f32_coeff[3])); + const auto c5 = svreinterpret_f32_u32(svdup_n_u32(svexp_f32_coeff[4])); + + const auto shift = svreinterpret_f32_u32(svdup_n_u32(0x4b00007f)); // 2^23 + 127 = 0x1.0000fep23f + const auto inv_ln2 = svreinterpret_f32_u32(svdup_n_u32(0x3fb8aa3b)); // 1 / ln(2) = 0x1.715476p+0f + const auto neg_ln2_hi = + svreinterpret_f32_u32(svdup_n_u32(0xbf317200)); // -ln(2) from bits -1 to -19: -0x1.62e400p-1f + const auto neg_ln2_lo = + svreinterpret_f32_u32(svdup_n_u32(0xb5bfbe8e)); // -ln(2) from bits -20 to -42: -0x1.7f7d1cp-20f + + const auto inf = svdup_n_f32(std::numeric_limits<float>::infinity()); + const auto max_input = svdup_n_f32(88.37f); // Approximately ln(2^127.5) + const auto zero = svdup_n_f32(0.f); + const auto min_input = svdup_n_f32(-86.64f); // Approximately ln(2^-125) + + // Range reduction: + // e^x = 2^n * e^r + // where: + // n = floor(x / ln(2)) + // r = x - n * ln(2) + // + // By adding x / ln(2) with 2^23 + 127 (shift): + // * As FP32 fraction part only has 23-bits, the addition of 2^23 + 127 forces decimal part + // of x / ln(2) out of the result. The integer part of x / ln(2) (i.e. n) + 127 will occupy + // the whole fraction part of z in FP32 format. + // Subtracting 2^23 + 127 (shift) from z will result in the integer part of x / ln(2) + // (i.e. n) because the decimal part has been pushed out and lost. + // * The addition of 127 makes the FP32 fraction part of z ready to be used as the exponent + // in FP32 format. Left shifting z by 23 bits will result in 2^n. + const auto z = svmla_f32_z(pg, shift, x, inv_ln2); + const auto n = svsub_f32_z(pg, z, shift); + const auto scale = svreinterpret_f32_u32(svlsl_n_u32_z(pg, svreinterpret_u32_f32(z), 23)); // 2^n + + // The calculation of n * ln(2) is done using 2 steps to achieve accuracy beyond FP32. + // This outperforms longer Taylor series (3-4 tabs) both in term of accuracy and performance. + const auto r_hi = svmla_f32_z(pg, x, n, neg_ln2_hi); + const auto r = svmla_f32_z(pg, r_hi, n, neg_ln2_lo); + + // Compute the truncated Taylor series of e^r. + // poly = scale * (1 + c1 * r + c2 * r^2 + c3 * r^3 + c4 * r^4 + c5 * r^5) + const auto r2 = svmul_f32_z(pg, r, r); + + const auto p1 = svmul_f32_z(pg, c1, r); + const auto p23 = svmla_f32_z(pg, c2, c3, r); + const auto p45 = svmla_f32_z(pg, c4, c5, r); + const auto p2345 = svmla_f32_z(pg, p23, p45, r2); + const auto p12345 = svmla_f32_z(pg, p1, p2345, r2); + + auto poly = svmla_f32_z(pg, scale, p12345, scale); + + // Handle underflow and overflow. + poly = svsel_f32(svcmplt_f32(pg, x, min_input), zero, poly); + poly = svsel_f32(svcmpgt_f32(pg, x, max_input), inf, poly); + + return poly; +} + +inline svfloat16_t svexp_f16_z(svbool_t pg, svfloat16_t x) +{ + auto bottom = svcvt_f32_z(pg, x); + auto pg_top = svptrue_b16(); + auto top = svcvt_f32_z(pg_top, svreinterpret_f16(svrevh_z(svptrue_b16(), svreinterpret_u32(x)))); + + bottom = svexp_f32_z(pg, bottom); + top = svexp_f32_z(pg_top, top); + + return svtrn1(svcvt_f16_z(pg, bottom), svcvt_f16_z(pg_top, top)); +} + +#ifdef ARM_COMPUTE_ENABLE_SVE2 + +inline svfloat16_t svexp_f16_z_sve2(svbool_t pg, svfloat16_t x) +{ + auto bottom = svcvt_f32_z(pg, x); + auto top = svcvtlt_f32_x(pg, x); + auto pg_top = pg; + + bottom = svexp_f32_z(pg, bottom); + top = svexp_f32_z(pg_top, top); + + return svcvtnt_f16_m(svcvt_f16_z(pg, bottom), pg_top, top); +} + +#endif // ARM_COMPUTE_ENABLE_SVE2 + +inline svfloat32_t svtanh_f32_z(svbool_t pg, svfloat32_t val) +{ + const svfloat32_t CONST_1 = svdup_n_f32(1.f); + const svfloat32_t CONST_2 = svdup_n_f32(2.f); + const svfloat32_t CONST_MIN_TANH = svdup_n_f32(-10.f); + const svfloat32_t CONST_MAX_TANH = svdup_n_f32(10.f); + + svfloat32_t x = svmin_f32_z(pg, svmax_f32_z(pg, val, CONST_MIN_TANH), CONST_MAX_TANH); + svfloat32_t exp2x = svexp_f32_z(pg, svmul_f32_z(pg, CONST_2, x)); + svfloat32_t num = svsub_f32_z(pg, exp2x, CONST_1); + svfloat32_t den = svadd_f32_z(pg, exp2x, CONST_1); + svfloat32_t tanh = svdiv_f32_z(pg, num, den); + return tanh; +} + +inline svfloat16_t svtanh_f16_z(svbool_t pg, svfloat16_t val) +{ + const svfloat16_t CONST_1 = svdup_n_f16(1.f); + const svfloat16_t CONST_2 = svdup_n_f16(2.f); + const svfloat16_t CONST_MIN_TANH = svdup_n_f16(-10.f); + const svfloat16_t CONST_MAX_TANH = svdup_n_f16(10.f); + + const svfloat16_t x = svmin_f16_z(pg, svmax_f16_z(pg, val, CONST_MIN_TANH), CONST_MAX_TANH); + const svfloat16_t exp2x = svexp_f16_z(pg, svmul_f16_z(pg, CONST_2, x)); + const svfloat16_t num = svsub_f16_z(pg, exp2x, CONST_1); + const svfloat16_t den = svadd_f16_z(pg, exp2x, CONST_1); + const svfloat16_t tanh = svdiv_f16_z(pg, num, den); + return tanh; +} + +inline svfloat32_t svlog_f32_z(svbool_t pg, svfloat32_t x) +{ + /** Logarithm polynomial coefficients */ + const svfloat32_t log_tab_1 = svdup_n_f32(-2.29561495781f); + const svfloat32_t log_tab_2 = svdup_n_f32(-2.47071170807f); + const svfloat32_t log_tab_3 = svdup_n_f32(-5.68692588806f); + const svfloat32_t log_tab_4 = svdup_n_f32(-0.165253549814f); + const svfloat32_t log_tab_5 = svdup_n_f32(5.17591238022f); + const svfloat32_t log_tab_6 = svdup_n_f32(0.844007015228f); + const svfloat32_t log_tab_7 = svdup_n_f32(4.58445882797f); + const svfloat32_t log_tab_8 = svdup_n_f32(0.0141278216615f); + + const auto CONST_127 = svdup_n_s32(127); // 127 + const auto CONST_LN2 = svdup_n_f32(0.6931471805f); // ln(2) + + // Extract exponent + auto m = svsub_s32_z(pg, svasr_n_s32_z(pg, svreinterpret_s32_f32(x), 23), CONST_127); + auto val = svreinterpret_f32_s32(svsub_s32_z(pg, svreinterpret_s32_f32(x), svlsl_n_s32_z(pg, m, 23))); + + // Polynomial Approximation + auto poly = svtaylor_poly_f32_z(pg, val, log_tab_1, log_tab_2, log_tab_3, log_tab_4, log_tab_5, log_tab_6, + log_tab_7, log_tab_8); + + // Reconstruct + poly = svmla_f32_z(pg, poly, svcvt_f32_s32_z(pg, m), CONST_LN2); + + return poly; +} + +inline svfloat16_t svlog_f16_z(svbool_t pg, svfloat16_t x) +{ + auto bottom = svcvt_f32_z(pg, x); + auto pg_top = svptrue_b16(); + auto top = svcvt_f32_z(pg_top, svreinterpret_f16(svrevh_z(svptrue_b16(), svreinterpret_u32(x)))); + + bottom = svlog_f32_z(pg, bottom); + top = svlog_f32_z(pg_top, top); + + return svtrn1(svcvt_f16_z(pg, bottom), svcvt_f16_z(pg_top, top)); +} + +#ifdef ARM_COMPUTE_ENABLE_SVE2 + +inline svfloat16_t svlog_f16_z_sve2(svbool_t pg, svfloat16_t x) +{ + auto bottom = svcvt_f32_z(pg, x); + auto top = svcvtlt_f32_x(pg, x); + auto pg_top = pg; + + bottom = svlog_f32_z(pg, bottom); + top = svlog_f32_z(pg_top, top); + + return svcvtnt_f16_m(svcvt_f16_z(pg, bottom), pg_top, top); +} + +#endif // ARM_COMPUTE_ENABLE_SVE2 + +inline svfloat32_t svsin_f32_z(svbool_t pg, svfloat32_t val) +{ + using ScalarType = float; + using IntType = uint32_t; + + constexpr float te_sin_coeff2 = 0.166666666666f; // 1/(2*3) + constexpr float te_sin_coeff3 = 0.05f; // 1/(4*5) + constexpr float te_sin_coeff4 = 0.023809523810f; // 1/(6*7) + constexpr float te_sin_coeff5 = 0.013888888889f; // 1/(8*9) + + const auto pi_v = wrapper::svdup_n(ScalarType(M_PI)); + const auto pio2_v = wrapper::svdup_n(ScalarType(M_PI / 2)); + const auto ipi_v = wrapper::svdup_n(ScalarType(1 / M_PI)); + + //Find positive or negative + const auto c_v = svabs_z(pg, wrapper::svcvt_z<int32_t>(pg, svmul_z(pg, val, ipi_v))); + const auto sign_v = svcmple(pg, val, wrapper::svdup_n(ScalarType(0))); + const auto odd_v = svcmpne(pg, svand_z(pg, wrapper::svreinterpret<IntType>(c_v), wrapper::svdup_n(IntType(1))), + wrapper::svdup_n(IntType(0))); + + auto neg_v = sveor_z(pg, odd_v, sign_v); + + //Modulus a - (n * int(a*(1/n))) + auto ma = svsub_z(pg, svabs_z(pg, val), svmul_z(pg, pi_v, wrapper::svcvt_z<ScalarType>(pg, c_v))); + const auto reb_v = svcmpge(pg, ma, pio2_v); + + //Rebase a between 0 and pi/2 + ma = svsel(reb_v, svsub_z(pg, pi_v, ma), ma); + + //Taylor series + const auto ma2 = svmul_z(pg, ma, ma); + + //2nd elem: x^3 / 3! + auto elem = svmul_z(pg, svmul_z(pg, ma, ma2), wrapper::svdup_n(ScalarType(te_sin_coeff2))); + auto res = svsub_z(pg, ma, elem); + + //3rd elem: x^5 / 5! + elem = svmul_z(pg, svmul_z(pg, elem, ma2), wrapper::svdup_n(ScalarType(te_sin_coeff3))); + res = svadd_z(pg, res, elem); + + //4th elem: x^7 / 7!float32x2_t vsin_f32(float32x2_t val) + elem = svmul_z(pg, svmul_z(pg, elem, ma2), wrapper::svdup_n(ScalarType(te_sin_coeff4))); + res = svsub_z(pg, res, elem); + + //5th elem: x^9 / 9! + elem = svmul_z(pg, svmul_z(pg, elem, ma2), wrapper::svdup_n(ScalarType(te_sin_coeff5))); + res = svadd_z(pg, res, elem); + + //Change of sign + res = svneg_m(res, neg_v, res); + return res; +} + +inline svfloat16_t svsin_f16_z(svbool_t pg, svfloat16_t val) +{ + auto bottom = svcvt_f32_z(pg, val); + auto pg_top = svptrue_b16(); + auto top = svcvt_f32_z(pg_top, svreinterpret_f16(svrevh_z(svptrue_b16(), svreinterpret_u32(val)))); + + bottom = svsin_f32_z(pg, bottom); + top = svsin_f32_z(pg_top, top); + + return svtrn1(svcvt_f16_z(pg, bottom), svcvt_f16_z(pg_top, top)); +} + +#ifdef ARM_COMPUTE_ENABLE_SVE2 + +inline svfloat16_t svsin_f16_z_sve2(svbool_t pg, svfloat16_t val) +{ + auto bottom = svcvt_f32_z(pg, val); + auto top = svcvtlt_f32_x(pg, val); + auto pg_top = pg; + + bottom = svsin_f32_z(pg, bottom); + top = svsin_f32_z(pg_top, top); + + return svcvtnt_f16_m(svcvt_f16_z(pg, bottom), pg_top, top); +} + +#endif // ARM_COMPUTE_ENABLE_SVE2 + +inline svfloat32_t svpow_f32_z(svbool_t pg, svfloat32_t a, svfloat32_t b) +{ + return svexp_f32_z(pg, svmul_z(pg, b, svlog_f32_z(pg, a))); +} + +inline svfloat16_t svpow_f16_z(svbool_t pg, svfloat16_t a, svfloat16_t b) +{ + auto a_bottom = svcvt_f32_z(pg, a); + auto b_bottom = svcvt_f32_z(pg, b); + + auto pg_top = svptrue_b16(); + auto a_top = svcvt_f32_z(pg_top, svreinterpret_f16(svrevh_z(svptrue_b16(), svreinterpret_u32(a)))); + auto b_top = svcvt_f32_z(pg_top, svreinterpret_f16(svrevh_z(svptrue_b16(), svreinterpret_u32(b)))); + + auto res_bottom = svpow_f32_z(pg, a_bottom, b_bottom); + auto res_top = svpow_f32_z(pg_top, a_top, b_top); + + return svtrn1(svcvt_f16_z(pg, res_bottom), svcvt_f16_z(pg_top, res_top)); +} + +#ifdef ARM_COMPUTE_ENABLE_SVE2 + +inline svfloat16_t svpow_f16_z_sve2(svbool_t pg, svfloat16_t a, svfloat16_t b) +{ + auto a_bottom = svcvt_f32_z(pg, a); + auto b_bottom = svcvt_f32_z(pg, b); + + auto pg_top = pg; + auto a_top = svcvtlt_f32_x(pg, a); + auto b_top = svcvtlt_f32_x(pg, b); + + auto res_bottom = svpow_f32_z(pg, a_bottom, b_bottom); + auto res_top = svpow_f32_z(pg_top, a_top, b_top); + + return svcvtnt_f16_m(svcvt_f16_z(pg, res_bottom), pg_top, res_top); +} + +#endif // ARM_COMPUTE_ENABLE_SVE2 + +#if defined(ARM_COMPUTE_ENABLE_SVE2) +template <> +inline svuint8_t convert_float_to_int<svuint8_t>(const svfloat32_t &in_0, + const svfloat32_t &in_1, + const svfloat32_t &in_2, + const svfloat32_t &in_3) +{ + svuint8_t out; + const auto all_true_pg = svptrue_b32(); + auto tmp_0 = svcvt_u32_f32_z(all_true_pg, in_0); + auto tmp_1 = svcvt_u32_f32_z(all_true_pg, in_1); + auto tmp_2 = svcvt_u32_f32_z(all_true_pg, in_2); + auto tmp_3 = svcvt_u32_f32_z(all_true_pg, in_3); + + auto tmp_16_0 = svqxtnt_u32(svqxtnb_u32(tmp_0), tmp_1); + auto tmp_16_1 = svqxtnt_u32(svqxtnb_u32(tmp_2), tmp_3); + + auto tmp_16_uzp_0 = svuzp1(tmp_16_0, tmp_16_0); + auto tmp_16_uzp_1 = svuzp2(tmp_16_0, tmp_16_0); + auto tmp_16_uzp_2 = svuzp1(tmp_16_1, tmp_16_1); + auto tmp_16_uzp_3 = svuzp2(tmp_16_1, tmp_16_1); + + auto pg = svwhilelt_b16_s32(0, svcnth() / 2); + + tmp_16_0 = svsplice(pg, tmp_16_uzp_0, tmp_16_uzp_1); + tmp_16_1 = svsplice(pg, tmp_16_uzp_2, tmp_16_uzp_3); + + out = svqxtnt_u16(svqxtnb_u16(tmp_16_0), tmp_16_1); + + auto out_uzp_0 = svuzp1(out, out); + auto out_uzp_1 = svuzp2(out, out); + + pg = svwhilelt_b8_s32(0, svcntb() / 2); + out = svsplice(pg, out_uzp_0, out_uzp_1); + + return out; +} + +template <> +inline svint8_t convert_float_to_int<svint8_t>(const svfloat32_t &in_0, + const svfloat32_t &in_1, + const svfloat32_t &in_2, + const svfloat32_t &in_3) +{ + svint8_t out; + const auto all_true_pg = svptrue_b32(); + auto tmp_0 = svcvt_s32_f32_z(all_true_pg, in_0); + auto tmp_1 = svcvt_s32_f32_z(all_true_pg, in_1); + auto tmp_2 = svcvt_s32_f32_z(all_true_pg, in_2); + auto tmp_3 = svcvt_s32_f32_z(all_true_pg, in_3); + + auto tmp_16_0 = svqxtnt_s32(svqxtnb_s32(tmp_0), tmp_1); + auto tmp_16_1 = svqxtnt_s32(svqxtnb_s32(tmp_2), tmp_3); + + auto tmp_16_uzp_0 = svuzp1(tmp_16_0, tmp_16_0); + auto tmp_16_uzp_1 = svuzp2(tmp_16_0, tmp_16_0); + auto tmp_16_uzp_2 = svuzp1(tmp_16_1, tmp_16_1); + auto tmp_16_uzp_3 = svuzp2(tmp_16_1, tmp_16_1); + + auto pg = svwhilelt_b16_s32(0, svcnth() / 2); + + tmp_16_0 = svsplice(pg, tmp_16_uzp_0, tmp_16_uzp_1); + tmp_16_1 = svsplice(pg, tmp_16_uzp_2, tmp_16_uzp_3); + + out = svqxtnt_s16(svqxtnb_s16(tmp_16_0), tmp_16_1); + + auto out_uzp_0 = svuzp1(out, out); + auto out_uzp_1 = svuzp2(out, out); + + pg = svwhilelt_b8_s32(0, svcntb() / 2); + out = svsplice(pg, out_uzp_0, out_uzp_1); + + return out; +} +#endif /* defined(ARM_COMPUTE_ENABLE_SVE2) */ + +} // namespace arm_compute +#endif /* defined(ARM_COMPUTE_ENABLE_SVE) */ + +#endif // ACL_SRC_CORE_NEON_SVEMATH_INL |