1 /* 2 * Single-precision vector powf function. 3 * 4 * Copyright (c) 2019-2024, Arm Limited. 5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception 6 */ 7 8 #include "v_math.h" 9 #include "test_defs.h" 10 #include "test_sig.h" 11 12 #define Min v_u32 (0x00800000) 13 #define Max v_u32 (0x7f800000) 14 #define Thresh v_u32 (0x7f000000) /* Max - Min. */ 15 #define MantissaMask v_u32 (0x007fffff) 16 17 #define A d->log2_poly 18 #define C d->exp2f_poly 19 20 /* 2.6 ulp ~ 0.5 + 2^24 (128*Ln2*relerr_log2 + relerr_exp2). */ 21 #define Off v_u32 (0x3f35d000) 22 23 #define V_POWF_LOG2_TABLE_BITS 5 24 #define V_EXP2F_TABLE_BITS 5 25 #define Log2IdxMask ((1 << V_POWF_LOG2_TABLE_BITS) - 1) 26 #define Scale ((double) (1 << V_EXP2F_TABLE_BITS)) 27 28 static const struct data 29 { 30 struct 31 { 32 double invc, logc; 33 } log2_tab[1 << V_POWF_LOG2_TABLE_BITS]; 34 float64x2_t log2_poly[4]; 35 uint64_t exp2f_tab[1 << V_EXP2F_TABLE_BITS]; 36 float64x2_t exp2f_poly[3]; 37 } data = { 38 .log2_tab = {{0x1.6489890582816p+0, -0x1.e960f97b22702p-2 * Scale}, 39 {0x1.5cf19b35e3472p+0, -0x1.c993406cd4db6p-2 * Scale}, 40 {0x1.55aac0e956d65p+0, -0x1.aa711d9a7d0f3p-2 * Scale}, 41 {0x1.4eb0022977e01p+0, -0x1.8bf37bacdce9bp-2 * Scale}, 42 {0x1.47fcccda1dd1fp+0, -0x1.6e13b3519946ep-2 * Scale}, 43 {0x1.418ceabab68c1p+0, -0x1.50cb8281e4089p-2 * Scale}, 44 {0x1.3b5c788f1edb3p+0, -0x1.341504a237e2bp-2 * Scale}, 45 {0x1.3567de48e9c9ap+0, -0x1.17eaab624ffbbp-2 * Scale}, 46 {0x1.2fabc80fd19bap+0, -0x1.f88e708f8c853p-3 * Scale}, 47 {0x1.2a25200ce536bp+0, -0x1.c24b6da113914p-3 * Scale}, 48 {0x1.24d108e0152e3p+0, -0x1.8d02ee397cb1dp-3 * Scale}, 49 {0x1.1facd8ab2fbe1p+0, -0x1.58ac1223408b3p-3 * Scale}, 50 {0x1.1ab614a03efdfp+0, -0x1.253e6fd190e89p-3 * Scale}, 51 {0x1.15ea6d03af9ffp+0, -0x1.e5641882c12ffp-4 * Scale}, 52 {0x1.1147b994bb776p+0, -0x1.81fea712926f7p-4 * Scale}, 53 {0x1.0ccbf650593aap+0, -0x1.203e240de64a3p-4 * Scale}, 54 {0x1.0875408477302p+0, -0x1.8029b86a78281p-5 * Scale}, 55 {0x1.0441d42a93328p+0, -0x1.85d713190fb9p-6 * Scale}, 56 {0x1p+0, 0x0p+0 * Scale}, 57 {0x1.f1d006c855e86p-1, 0x1.4c1cc07312997p-5 * Scale}, 58 {0x1.e28c3341aa301p-1, 0x1.5e1848ccec948p-4 * Scale}, 59 {0x1.d4bdf9aa64747p-1, 0x1.04cfcb7f1196fp-3 * Scale}, 60 {0x1.c7b45a24e5803p-1, 0x1.582813d463c21p-3 * Scale}, 61 {0x1.bb5f5eb2ed60ap-1, 0x1.a936fa68760ccp-3 * Scale}, 62 {0x1.afb0bff8fe6b4p-1, 0x1.f81bc31d6cc4ep-3 * Scale}, 63 {0x1.a49badf7ab1f5p-1, 0x1.2279a09fae6b1p-2 * Scale}, 64 {0x1.9a14a111fc4c9p-1, 0x1.47ec0b6df5526p-2 * Scale}, 65 {0x1.901131f5b2fdcp-1, 0x1.6c71762280f1p-2 * Scale}, 66 {0x1.8687f73f6d865p-1, 0x1.90155070798dap-2 * Scale}, 67 {0x1.7d7067eb77986p-1, 0x1.b2e23b1d3068cp-2 * Scale}, 68 {0x1.74c2c1cf97b65p-1, 0x1.d4e21b0daa86ap-2 * Scale}, 69 {0x1.6c77f37cff2a1p-1, 0x1.f61e2a2f67f3fp-2 * Scale},}, 70 .log2_poly = { /* rel err: 1.5 * 2^-30. */ 71 V2 (-0x1.6ff5daa3b3d7cp-2 * Scale), 72 V2 (0x1.ec81d03c01aebp-2 * Scale), 73 V2 (-0x1.71547bb43f101p-1 * Scale), 74 V2 (0x1.7154764a815cbp0 * Scale)}, 75 .exp2f_tab = {0x3ff0000000000000, 0x3fefd9b0d3158574, 0x3fefb5586cf9890f, 76 0x3fef9301d0125b51, 0x3fef72b83c7d517b, 0x3fef54873168b9aa, 77 0x3fef387a6e756238, 0x3fef1e9df51fdee1, 0x3fef06fe0a31b715, 78 0x3feef1a7373aa9cb, 0x3feedea64c123422, 0x3feece086061892d, 79 0x3feebfdad5362a27, 0x3feeb42b569d4f82, 0x3feeab07dd485429, 80 0x3feea47eb03a5585, 0x3feea09e667f3bcd, 0x3fee9f75e8ec5f74, 81 0x3feea11473eb0187, 0x3feea589994cce13, 0x3feeace5422aa0db, 82 0x3feeb737b0cdc5e5, 0x3feec49182a3f090, 0x3feed503b23e255d, 83 0x3feee89f995ad3ad, 0x3feeff76f2fb5e47, 0x3fef199bdd85529c, 84 0x3fef3720dcef9069, 0x3fef5818dcfba487, 0x3fef7c97337b9b5f, 85 0x3fefa4afa2a490da, 0x3fefd0765b6e4540,}, 86 .exp2f_poly = { /* rel err: 1.69 * 2^-34. */ 87 V2 (0x1.c6af84b912394p-5 / Scale / Scale / Scale), 88 V2 (0x1.ebfce50fac4f3p-3 / Scale / Scale), 89 V2 (0x1.62e42ff0c52d6p-1 / Scale)}}; 90 91 static float32x4_t VPCS_ATTR NOINLINE 92 special_case (float32x4_t x, float32x4_t y, float32x4_t ret, uint32x4_t cmp) 93 { 94 return v_call2_f32 (powf, x, y, ret, cmp); 95 } 96 97 static inline float64x2_t 98 ylogx_core (const struct data *d, float64x2_t iz, float64x2_t k, 99 float64x2_t invc, float64x2_t logc, float64x2_t y) 100 { 101 102 /* log2(x) = log1p(z/c-1)/ln2 + log2(c) + k. */ 103 float64x2_t r = vfmaq_f64 (v_f64 (-1.0), iz, invc); 104 float64x2_t y0 = vaddq_f64 (logc, k); 105 106 /* Polynomial to approximate log1p(r)/ln2. */ 107 float64x2_t logx = vfmaq_f64 (A[1], r, A[0]); 108 logx = vfmaq_f64 (A[2], logx, r); 109 logx = vfmaq_f64 (A[3], logx, r); 110 logx = vfmaq_f64 (y0, logx, r); 111 112 return vmulq_f64 (logx, y); 113 } 114 115 static inline float64x2_t 116 log2_lookup (const struct data *d, uint32_t i) 117 { 118 return vld1q_f64 ( 119 &d->log2_tab[(i >> (23 - V_POWF_LOG2_TABLE_BITS)) & Log2IdxMask].invc); 120 } 121 122 static inline uint64x1_t 123 exp2f_lookup (const struct data *d, uint64_t i) 124 { 125 return vld1_u64 (&d->exp2f_tab[i % (1 << V_EXP2F_TABLE_BITS)]); 126 } 127 128 static inline float32x2_t 129 powf_core (const struct data *d, float64x2_t ylogx) 130 { 131 /* N*x = k + r with r in [-1/2, 1/2]. */ 132 float64x2_t kd = vrndnq_f64 (ylogx); 133 int64x2_t ki = vcvtaq_s64_f64 (ylogx); 134 float64x2_t r = vsubq_f64 (ylogx, kd); 135 136 /* exp2(x) = 2^(k/N) * 2^r ~= s * (C0*r^3 + C1*r^2 + C2*r + 1). */ 137 uint64x2_t t = vcombine_u64 (exp2f_lookup (d, vgetq_lane_s64 (ki, 0)), 138 exp2f_lookup (d, vgetq_lane_s64 (ki, 1))); 139 t = vaddq_u64 ( 140 t, vreinterpretq_u64_s64 (vshlq_n_s64 (ki, 52 - V_EXP2F_TABLE_BITS))); 141 float64x2_t s = vreinterpretq_f64_u64 (t); 142 float64x2_t p = vfmaq_f64 (C[1], r, C[0]); 143 p = vfmaq_f64 (C[2], r, p); 144 p = vfmaq_f64 (s, p, vmulq_f64 (s, r)); 145 return vcvt_f32_f64 (p); 146 } 147 148 float32x4_t VPCS_ATTR NOINLINE V_NAME_F2 (pow) (float32x4_t x, float32x4_t y) 149 { 150 const struct data *d = ptr_barrier (&data); 151 uint32x4_t u = vreinterpretq_u32_f32 (x); 152 uint32x4_t cmp = vcgeq_u32 (vsubq_u32 (u, Min), Thresh); 153 uint32x4_t tmp = vsubq_u32 (u, Off); 154 uint32x4_t top = vbicq_u32 (tmp, MantissaMask); 155 float32x4_t iz = vreinterpretq_f32_u32 (vsubq_u32 (u, top)); 156 int32x4_t k = vshrq_n_s32 (vreinterpretq_s32_u32 (top), 157 23 - V_EXP2F_TABLE_BITS); /* arithmetic shift. */ 158 159 /* Use double precision for each lane: split input vectors into lo and hi 160 halves and promote. */ 161 float64x2_t tab0 = log2_lookup (d, vgetq_lane_u32 (tmp, 0)), 162 tab1 = log2_lookup (d, vgetq_lane_u32 (tmp, 1)), 163 tab2 = log2_lookup (d, vgetq_lane_u32 (tmp, 2)), 164 tab3 = log2_lookup (d, vgetq_lane_u32 (tmp, 3)); 165 166 float64x2_t iz_lo = vcvt_f64_f32 (vget_low_f32 (iz)), 167 iz_hi = vcvt_high_f64_f32 (iz); 168 169 float64x2_t k_lo = vcvtq_f64_s64 (vmovl_s32 (vget_low_s32 (k))), 170 k_hi = vcvtq_f64_s64 (vmovl_high_s32 (k)); 171 172 float64x2_t invc_lo = vzip1q_f64 (tab0, tab1), 173 invc_hi = vzip1q_f64 (tab2, tab3), 174 logc_lo = vzip2q_f64 (tab0, tab1), 175 logc_hi = vzip2q_f64 (tab2, tab3); 176 177 float64x2_t y_lo = vcvt_f64_f32 (vget_low_f32 (y)), 178 y_hi = vcvt_high_f64_f32 (y); 179 180 float64x2_t ylogx_lo = ylogx_core (d, iz_lo, k_lo, invc_lo, logc_lo, y_lo); 181 float64x2_t ylogx_hi = ylogx_core (d, iz_hi, k_hi, invc_hi, logc_hi, y_hi); 182 183 uint32x4_t ylogx_top = vuzp2q_u32 (vreinterpretq_u32_f64 (ylogx_lo), 184 vreinterpretq_u32_f64 (ylogx_hi)); 185 186 cmp = vorrq_u32 ( 187 cmp, vcgeq_u32 (vandq_u32 (vshrq_n_u32 (ylogx_top, 15), v_u32 (0xffff)), 188 vdupq_n_u32 (asuint64 (126.0 * (1 << V_EXP2F_TABLE_BITS)) 189 >> 47))); 190 191 float32x2_t p_lo = powf_core (d, ylogx_lo); 192 float32x2_t p_hi = powf_core (d, ylogx_hi); 193 194 if (unlikely (v_any_u32 (cmp))) 195 return special_case (x, y, vcombine_f32 (p_lo, p_hi), cmp); 196 return vcombine_f32 (p_lo, p_hi); 197 } 198 199 HALF_WIDTH_ALIAS_F2 (pow) 200 201 TEST_SIG (V, F, 2, pow) 202 TEST_ULP (V_NAME_F2 (pow), 2.1) 203 TEST_DISABLE_FENV (V_NAME_F2 (pow)) 204 TEST_INTERVAL2 (V_NAME_F2 (pow), 0x1p-1, 0x1p1, 0x1p-7, 0x1p7, 50000) 205 TEST_INTERVAL2 (V_NAME_F2 (pow), 0x1p-1, 0x1p1, -0x1p-7, -0x1p7, 50000) 206 TEST_INTERVAL2 (V_NAME_F2 (pow), 0x1p-70, 0x1p70, 0x1p-1, 0x1p1, 50000) 207 TEST_INTERVAL2 (V_NAME_F2 (pow), 0x1p-70, 0x1p70, -0x1p-1, -0x1p1, 50000) 208 TEST_INTERVAL2 (V_NAME_F2 (pow), 0x1.ep-1, 0x1.1p0, 0x1p8, 0x1p14, 50000) 209 TEST_INTERVAL2 (V_NAME_F2 (pow), 0x1.ep-1, 0x1.1p0, -0x1p8, -0x1p14, 50000) 210