• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1/*
2 * Copyright (c) 2016-2020 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24#include <cmath>
25#include <limits>
26
27#ifndef M_PI
28#define M_PI (3.14159265358979323846)
29#endif // M_PI
30
31namespace arm_compute
32{
33/** Exponent polynomial coefficients */
34const std::array<float32x4_t, 8> exp_tab =
35{
36    {
37        vdupq_n_f32(1.f),
38        vdupq_n_f32(0.0416598916054f),
39        vdupq_n_f32(0.500000596046f),
40        vdupq_n_f32(0.0014122662833f),
41        vdupq_n_f32(1.00000011921f),
42        vdupq_n_f32(0.00833693705499f),
43        vdupq_n_f32(0.166665703058f),
44        vdupq_n_f32(0.000195780929062f),
45    }
46};
47
48/** Logarithm polynomial coefficients */
49const std::array<float32x4_t, 8> log_tab =
50{
51    {
52        vdupq_n_f32(-2.29561495781f),
53        vdupq_n_f32(-2.47071170807f),
54        vdupq_n_f32(-5.68692588806f),
55        vdupq_n_f32(-0.165253549814f),
56        vdupq_n_f32(5.17591238022f),
57        vdupq_n_f32(0.844007015228f),
58        vdupq_n_f32(4.58445882797f),
59        vdupq_n_f32(0.0141278216615f),
60    }
61};
62
63/** Sin polynomial coefficients */
64constexpr float te_sin_coeff2 = 0.166666666666f; // 1/(2*3)
65constexpr float te_sin_coeff3 = 0.05f;           // 1/(4*5)
66constexpr float te_sin_coeff4 = 0.023809523810f; // 1/(6*7)
67constexpr float te_sin_coeff5 = 0.013888888889f; // 1/(8*9)
68
69#ifndef DOXYGEN_SKIP_THIS
70inline float32x4_t vfloorq_f32(float32x4_t val)
71{
72    static const float32x4_t CONST_1 = vdupq_n_f32(1.f);
73
74    const int32x4_t   z = vcvtq_s32_f32(val);
75    const float32x4_t r = vcvtq_f32_s32(z);
76
77    return vbslq_f32(vcgtq_f32(r, val), vsubq_f32(r, CONST_1), r);
78}
79
80inline float32x4_t vroundq_rte_f32(float32x4_t val)
81{
82#ifdef __aarch64__
83    return vrndnq_f32(val);
84#else  // __aarch64__
85    static const float32x4_t CONST_HALF_FLOAT = vdupq_n_f32(0.5f);
86    static const float32x4_t CONST_1_FLOAT    = vdupq_n_f32(1.f);
87    static const int32x4_t   CONST_1_INT      = vdupq_n_s32(1);
88    const float32x4_t        floor_val        = vfloorq_f32(val);
89    const float32x4_t        diff             = vsubq_f32(val, floor_val);
90
91    /*
92    * Select the floor value when (diff<0.5 || (diff==0.5 && floor_val%2==0).
93    * This condition is checked by vorrq_u32(vcltq_f32(diff, CONST_HALF_FLOAT) ,vandq_u32(vceqq_f32(diff, CONST_HALF_FLOAT) , vmvnq_u32(vtstq_s32(vandq_s32(vcvtq_s32_f32(floor_val), CONST_1_INT),CONST_1_INT))))
94    */
95
96    return vbslq_f32(vorrq_u32(vcltq_f32(diff, CONST_HALF_FLOAT), vandq_u32(vceqq_f32(diff, CONST_HALF_FLOAT), vmvnq_u32(vtstq_s32(vandq_s32(vcvtq_s32_f32(floor_val), CONST_1_INT), CONST_1_INT)))),
97                     floor_val, vaddq_f32(floor_val, CONST_1_FLOAT));
98#endif // __aarch64__
99}
100
101inline float32x2_t vinvsqrt_f32(float32x2_t x)
102{
103    float32x2_t sqrt_reciprocal = vrsqrte_f32(x);
104    sqrt_reciprocal             = vmul_f32(vrsqrts_f32(vmul_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
105    sqrt_reciprocal             = vmul_f32(vrsqrts_f32(vmul_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
106
107    return sqrt_reciprocal;
108}
109
110inline float32x4_t vinvsqrtq_f32(float32x4_t x)
111{
112    float32x4_t sqrt_reciprocal = vrsqrteq_f32(x);
113    sqrt_reciprocal             = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
114    sqrt_reciprocal             = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
115
116    return sqrt_reciprocal;
117}
118
119inline float32x2_t vinv_f32(float32x2_t x)
120{
121    float32x2_t recip = vrecpe_f32(x);
122    recip             = vmul_f32(vrecps_f32(x, recip), recip);
123    recip             = vmul_f32(vrecps_f32(x, recip), recip);
124    return recip;
125}
126
127inline float32x4_t vinvq_f32(float32x4_t x)
128{
129    float32x4_t recip = vrecpeq_f32(x);
130    recip             = vmulq_f32(vrecpsq_f32(x, recip), recip);
131    recip             = vmulq_f32(vrecpsq_f32(x, recip), recip);
132    return recip;
133}
134
135inline float32x4_t vtaylor_polyq_f32(float32x4_t x, const std::array<float32x4_t, 8> &coeffs)
136{
137    float32x4_t A   = vmlaq_f32(coeffs[0], coeffs[4], x);
138    float32x4_t B   = vmlaq_f32(coeffs[2], coeffs[6], x);
139    float32x4_t C   = vmlaq_f32(coeffs[1], coeffs[5], x);
140    float32x4_t D   = vmlaq_f32(coeffs[3], coeffs[7], x);
141    float32x4_t x2  = vmulq_f32(x, x);
142    float32x4_t x4  = vmulq_f32(x2, x2);
143    float32x4_t res = vmlaq_f32(vmlaq_f32(A, B, x2), vmlaq_f32(C, D, x2), x4);
144    return res;
145}
146
147inline float32x4_t vexpq_f32(float32x4_t x)
148{
149    static const float32x4_t CONST_LN2          = vdupq_n_f32(0.6931471805f); // ln(2)
150    static const float32x4_t CONST_INV_LN2      = vdupq_n_f32(1.4426950408f); // 1/ln(2)
151    static const float32x4_t CONST_INF          = vdupq_n_f32(std::numeric_limits<float>::infinity());
152    static const float32x4_t CONST_MAX_INPUT    = vdupq_n_f32(88.7f);
153    static const float32x4_t CONST_0            = vdupq_n_f32(0.f);
154    static const int32x4_t   CONST_NEGATIVE_126 = vdupq_n_s32(-126);
155
156    // Perform range reduction [-log(2),log(2)]
157    int32x4_t   m   = vcvtq_s32_f32(vmulq_f32(x, CONST_INV_LN2));
158    float32x4_t val = vmlsq_f32(x, vcvtq_f32_s32(m), CONST_LN2);
159
160    // Polynomial Approximation
161    float32x4_t poly = vtaylor_polyq_f32(val, exp_tab);
162
163    // Reconstruct
164    poly = vreinterpretq_f32_s32(vqaddq_s32(vreinterpretq_s32_f32(poly), vqshlq_n_s32(m, 23)));
165    poly = vbslq_f32(vcltq_s32(m, CONST_NEGATIVE_126), CONST_0, poly); // Handle underflow
166    poly = vbslq_f32(vcgtq_f32(x, CONST_MAX_INPUT), CONST_INF, poly);  // Handle overflow
167
168    return poly;
169}
170
171inline float32x4_t vlogq_f32(float32x4_t x)
172{
173    static const int32x4_t   CONST_127 = vdupq_n_s32(127);           // 127
174    static const float32x4_t CONST_LN2 = vdupq_n_f32(0.6931471805f); // ln(2)
175
176    // Extract exponent
177    int32x4_t   m   = vsubq_s32(vreinterpretq_s32_u32(vshrq_n_u32(vreinterpretq_u32_f32(x), 23)), CONST_127);
178    float32x4_t val = vreinterpretq_f32_s32(vsubq_s32(vreinterpretq_s32_f32(x), vshlq_n_s32(m, 23)));
179
180    // Polynomial Approximation
181    float32x4_t poly = vtaylor_polyq_f32(val, log_tab);
182
183    // Reconstruct
184    poly = vmlaq_f32(poly, vcvtq_f32_s32(m), CONST_LN2);
185
186    return poly;
187}
188
189inline float32x4_t vtanhq_f32(float32x4_t val)
190{
191    static const float32x4_t CONST_1        = vdupq_n_f32(1.f);
192    static const float32x4_t CONST_2        = vdupq_n_f32(2.f);
193    static const float32x4_t CONST_MIN_TANH = vdupq_n_f32(-10.f);
194    static const float32x4_t CONST_MAX_TANH = vdupq_n_f32(10.f);
195
196    float32x4_t x     = vminq_f32(vmaxq_f32(val, CONST_MIN_TANH), CONST_MAX_TANH);
197    float32x4_t exp2x = vexpq_f32(vmulq_f32(CONST_2, x));
198    float32x4_t num   = vsubq_f32(exp2x, CONST_1);
199    float32x4_t den   = vaddq_f32(exp2x, CONST_1);
200    float32x4_t tanh  = vmulq_f32(num, vinvq_f32(den));
201    return tanh;
202}
203
204inline float32x4_t vpowq_f32(float32x4_t val, float32x4_t n)
205{
206    return vexpq_f32(vmulq_f32(n, vlogq_f32(val)));
207}
208
209inline float32x4_t vsinq_f32(float32x4_t val)
210{
211    const float32x4_t pi_v   = vdupq_n_f32(M_PI);
212    const float32x4_t pio2_v = vdupq_n_f32(M_PI / 2);
213    const float32x4_t ipi_v  = vdupq_n_f32(1 / M_PI);
214
215    //Find positive or negative
216    const int32x4_t  c_v    = vabsq_s32(vcvtq_s32_f32(vmulq_f32(val, ipi_v)));
217    const uint32x4_t sign_v = vcleq_f32(val, vdupq_n_f32(0));
218    const uint32x4_t odd_v  = vandq_u32(vreinterpretq_u32_s32(c_v), vdupq_n_u32(1));
219
220    uint32x4_t neg_v = veorq_u32(odd_v, sign_v);
221
222    //Modulus a - (n * int(a*(1/n)))
223    float32x4_t      ma    = vsubq_f32(vabsq_f32(val), vmulq_f32(pi_v, vcvtq_f32_s32(c_v)));
224    const uint32x4_t reb_v = vcgeq_f32(ma, pio2_v);
225
226    //Rebase a between 0 and pi/2
227    ma = vbslq_f32(reb_v, vsubq_f32(pi_v, ma), ma);
228
229    //Taylor series
230    const float32x4_t ma2 = vmulq_f32(ma, ma);
231
232    //2nd elem: x^3 / 3!
233    float32x4_t elem = vmulq_f32(vmulq_f32(ma, ma2), vdupq_n_f32(te_sin_coeff2));
234    float32x4_t res  = vsubq_f32(ma, elem);
235
236    //3rd elem: x^5 / 5!
237    elem = vmulq_f32(vmulq_f32(elem, ma2), vdupq_n_f32(te_sin_coeff3));
238    res  = vaddq_f32(res, elem);
239
240    //4th elem: x^7 / 7!float32x2_t vsin_f32(float32x2_t val)
241    elem = vmulq_f32(vmulq_f32(elem, ma2), vdupq_n_f32(te_sin_coeff4));
242    res  = vsubq_f32(res, elem);
243
244    //5th elem: x^9 / 9!
245    elem = vmulq_f32(vmulq_f32(elem, ma2), vdupq_n_f32(te_sin_coeff5));
246    res  = vaddq_f32(res, elem);
247
248    //Change of sign
249    neg_v = vshlq_n_u32(neg_v, 31);
250    res   = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(res), neg_v));
251    return res;
252}
253
254inline float32x2_t vsin_f32(float32x2_t val)
255{
256    const float32x2_t pi_v   = vdup_n_f32(M_PI);
257    const float32x2_t pio2_v = vdup_n_f32(M_PI / 2);
258    const float32x2_t ipi_v  = vdup_n_f32(1 / M_PI);
259
260    //Find positive or negative
261    const int32x2_t  c_v    = vabs_s32(vcvt_s32_f32(vmul_f32(val, ipi_v)));
262    const uint32x2_t sign_v = vcle_f32(val, vdup_n_f32(0));
263    const uint32x2_t odd_v  = vand_u32(vreinterpret_u32_s32(c_v), vdup_n_u32(1));
264
265    uint32x2_t neg_v = veor_u32(odd_v, sign_v);
266
267    //Modulus a - (n * int(a*(1/n)))
268    float32x2_t      ma    = vsub_f32(vabs_f32(val), vmul_f32(pi_v, vcvt_f32_s32(c_v)));
269    const uint32x2_t reb_v = vcge_f32(ma, pio2_v);
270
271    //Rebase a between 0 and pi/2
272    ma = vbsl_f32(reb_v, vsub_f32(pi_v, ma), ma);
273
274    //Taylor series
275    const float32x2_t ma2 = vmul_f32(ma, ma);
276
277    //2nd elem: x^3 / 3!
278    float32x2_t elem = vmul_f32(vmul_f32(ma, ma2), vdup_n_f32(te_sin_coeff2));
279    float32x2_t res  = vsub_f32(ma, elem);
280
281    //3rd elem: x^5 / 5!
282    elem = vmul_f32(vmul_f32(elem, ma2), vdup_n_f32(te_sin_coeff3));
283    res  = vadd_f32(res, elem);
284
285    //4th elem: x^7 / 7!float32x2_t vsin_f32(float32x2_t val)
286    elem = vmul_f32(vmul_f32(elem, ma2), vdup_n_f32(te_sin_coeff4));
287    res  = vsub_f32(res, elem);
288
289    //5th elem: x^9 / 9!
290    elem = vmul_f32(vmul_f32(elem, ma2), vdup_n_f32(te_sin_coeff5));
291    res  = vadd_f32(res, elem);
292
293    //Change of sign
294    neg_v = vshl_n_u32(neg_v, 31);
295    res   = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(res), neg_v));
296    return res;
297}
298
299#endif /* DOXYGEN_SKIP_THIS */
300
301inline int32x4_t rounding_divide_by_pow2(int32x4_t x, int32x4_t exponent)
302{
303    const int32x4_t shift_vec  = vnegq_s32(exponent);
304    const int32x4_t fixup      = vshrq_n_s32(vandq_s32(x, shift_vec), 31);
305    const int32x4_t fixed_up_x = vqaddq_s32(x, fixup);
306    return vrshlq_s32(fixed_up_x, shift_vec);
307}
308
309inline int32x4_t rounding_divide_by_pow2(int32x4_t x, int exponent)
310{
311    const int32x4_t shift_vec  = vdupq_n_s32(-exponent);
312    const int32x4_t fixup      = vshrq_n_s32(vandq_s32(x, shift_vec), 31);
313    const int32x4_t fixed_up_x = vqaddq_s32(x, fixup);
314    return vrshlq_s32(fixed_up_x, shift_vec);
315}
316
317inline int32_t rounding_divide_by_pow2(int32_t x, int exponent)
318{
319    const int32_t mask      = (1 << exponent) - 1;
320    const int32_t threshold = (mask >> 1) + (x < 0 ? 1 : 0);
321    return (x >> exponent) + ((x & mask) > threshold ? 1 : 0);
322}
323
324inline float32x4x4_t convert_uint8x16_to_float32x4x4(const uint8x16_t &in)
325{
326    float32x4x4_t out;
327
328    const auto tmp1 = vmovl_u8(vget_low_u8(in));
329    out.val[0]      = vcvtq_f32_u32(vmovl_u16(vget_low_u16(tmp1)));
330    out.val[1]      = vcvtq_f32_u32(vmovl_u16(vget_high_u16(tmp1)));
331
332    const auto tmp2 = vmovl_u8(vget_high_u8(in));
333    out.val[2]      = vcvtq_f32_u32(vmovl_u16(vget_low_u16(tmp2)));
334    out.val[3]      = vcvtq_f32_u32(vmovl_u16(vget_high_u16(tmp2)));
335    return out;
336}
337
338inline float32x4x4_t convert_int8x16_to_float32x4x4(const int8x16_t &in)
339{
340    float32x4x4_t out;
341
342    const auto tmp1 = vmovl_s8(vget_low_s8(in));
343    out.val[0]      = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp1)));
344    out.val[1]      = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp1)));
345
346    const auto tmp2 = vmovl_s8(vget_high_s8(in));
347    out.val[2]      = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp2)));
348    out.val[3]      = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp2)));
349    return out;
350}
351
352template <>
353inline float32x4x4_t convert_to_float32x4x4(const uint8x16_t &in)
354{
355    return convert_uint8x16_to_float32x4x4(in);
356}
357
358template <>
359inline float32x4x4_t convert_to_float32x4x4(const int8x16_t &in)
360{
361    return convert_int8x16_to_float32x4x4(in);
362}
363
364inline void convert_float32x4x3_to_uint8x8x3(const float32x4x3_t &in1, const float32x4x3_t &in2, uint8x8x3_t &out)
365{
366    out.val[0] = vqmovn_u16(vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in1.val[0])),
367                                         vqmovn_u32(vcvtq_u32_f32(in2.val[0]))));
368    out.val[1] = vqmovn_u16(vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in1.val[1])),
369                                         vqmovn_u32(vcvtq_u32_f32(in2.val[1]))));
370    out.val[2] = vqmovn_u16(vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in1.val[2])),
371                                         vqmovn_u32(vcvtq_u32_f32(in2.val[2]))));
372}
373
374inline void convert_float32x4x4_to_uint8x16(const float32x4x4_t &in, uint8x16_t &out)
375{
376    const auto low = vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in.val[0])),
377                                  vqmovn_u32(vcvtq_u32_f32(in.val[1])));
378    const auto high = vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in.val[2])),
379                                   vqmovn_u32(vcvtq_u32_f32(in.val[3])));
380    out = vcombine_u8(vqmovn_u16(low), vqmovn_u16(high));
381}
382
383inline void convert_float32x4x4_to_int8x16(const float32x4x4_t &in, int8x16_t &out)
384{
385    const auto low = vcombine_s16(vqmovn_s32(vcvtq_s32_f32(in.val[0])),
386                                  vqmovn_s32(vcvtq_s32_f32(in.val[1])));
387    const auto high = vcombine_s16(vqmovn_s32(vcvtq_s32_f32(in.val[2])),
388                                   vqmovn_s32(vcvtq_s32_f32(in.val[3])));
389    out = vcombine_s8(vqmovn_s16(low), vqmovn_s16(high));
390}
391
392#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
393/** Exponent polynomial coefficients */
394/** Logarithm polynomial coefficients */
395#ifndef DOXYGEN_SKIP_THIS
396inline float16x8_t vfloorq_f16(float16x8_t val)
397{
398    static const float16x8_t CONST_1 = vdupq_n_f16(1.f);
399
400    const int16x8_t   z = vcvtq_s16_f16(val);
401    const float16x8_t r = vcvtq_f16_s16(z);
402
403    return vbslq_f16(vcgtq_f16(r, val), vsubq_f16(r, CONST_1), r);
404}
405
406inline float16x8_t vroundq_rte_f16(float16x8_t val)
407{
408    return vrndnq_f16(val);
409}
410
411inline float16x4_t vinvsqrt_f16(float16x4_t x)
412{
413    float16x4_t sqrt_reciprocal = vrsqrte_f16(x);
414    sqrt_reciprocal             = vmul_f16(vrsqrts_f16(vmul_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
415    sqrt_reciprocal             = vmul_f16(vrsqrts_f16(vmul_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
416    return sqrt_reciprocal;
417}
418
419inline float16x8_t vinvsqrtq_f16(float16x8_t x)
420{
421    float16x8_t sqrt_reciprocal = vrsqrteq_f16(x);
422    sqrt_reciprocal             = vmulq_f16(vrsqrtsq_f16(vmulq_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
423    sqrt_reciprocal             = vmulq_f16(vrsqrtsq_f16(vmulq_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
424    return sqrt_reciprocal;
425}
426
427inline float16x4_t vinv_f16(float16x4_t x)
428{
429    float16x4_t recip = vrecpe_f16(x);
430    recip             = vmul_f16(vrecps_f16(x, recip), recip);
431    recip             = vmul_f16(vrecps_f16(x, recip), recip);
432    return recip;
433}
434
435inline float16x8_t vinvq_f16(float16x8_t x)
436{
437    float16x8_t recip = vrecpeq_f16(x);
438    recip             = vmulq_f16(vrecpsq_f16(x, recip), recip);
439    recip             = vmulq_f16(vrecpsq_f16(x, recip), recip);
440    return recip;
441}
442
443inline float16x8_t vtanhq_f16(float16x8_t val)
444{
445    const float16x8_t CONST_1        = vdupq_n_f16(1.f);
446    const float16x8_t CONST_2        = vdupq_n_f16(2.f);
447    const float16x8_t CONST_MIN_TANH = vdupq_n_f16(-10.f);
448    const float16x8_t CONST_MAX_TANH = vdupq_n_f16(10.f);
449
450    const float16x8_t x     = vminq_f16(vmaxq_f16(val, CONST_MIN_TANH), CONST_MAX_TANH);
451    const float16x8_t exp2x = vexpq_f16(vmulq_f16(CONST_2, x));
452    const float16x8_t num   = vsubq_f16(exp2x, CONST_1);
453    const float16x8_t den   = vaddq_f16(exp2x, CONST_1);
454    const float16x8_t tanh  = vmulq_f16(num, vinvq_f16(den));
455    return tanh;
456}
457
458inline float16x8_t vtaylor_polyq_f16(float16x8_t x, const std::array<float16x8_t, 8> &coeffs)
459{
460    const float16x8_t A   = vaddq_f16(coeffs[0], vmulq_f16(coeffs[4], x));
461    const float16x8_t B   = vaddq_f16(coeffs[2], vmulq_f16(coeffs[6], x));
462    const float16x8_t C   = vaddq_f16(coeffs[1], vmulq_f16(coeffs[5], x));
463    const float16x8_t D   = vaddq_f16(coeffs[3], vmulq_f16(coeffs[7], x));
464    const float16x8_t x2  = vmulq_f16(x, x);
465    const float16x8_t x4  = vmulq_f16(x2, x2);
466    const float16x8_t res = vaddq_f16(vaddq_f16(A, vmulq_f16(B, x2)), vmulq_f16(vaddq_f16(C, vmulq_f16(D, x2)), x4));
467    return res;
468}
469
470inline float16x8_t vexpq_f16(float16x8_t x)
471{
472    // TODO (COMPMID-1535) : Revisit FP16 approximations
473    const float32x4_t x_high = vcvt_f32_f16(vget_high_f16(x));
474    const float32x4_t x_low  = vcvt_f32_f16(vget_low_f16(x));
475
476    const float16x8_t res = vcombine_f16(vcvt_f16_f32(vexpq_f32(x_low)), vcvt_f16_f32(vexpq_f32(x_high)));
477    return res;
478}
479
480inline float16x8_t vlogq_f16(float16x8_t x)
481{
482    // TODO (COMPMID-1535) : Revisit FP16 approximations
483    const float32x4_t x_high = vcvt_f32_f16(vget_high_f16(x));
484    const float32x4_t x_low  = vcvt_f32_f16(vget_low_f16(x));
485
486    const float16x8_t res = vcombine_f16(vcvt_f16_f32(vlogq_f32(x_low)), vcvt_f16_f32(vlogq_f32(x_high)));
487    return res;
488}
489
490inline float16x8_t vpowq_f16(float16x8_t val, float16x8_t n)
491{
492    // TODO (giaiod01) - COMPMID-1535
493    float32x4_t n0_f32   = vcvt_f32_f16(vget_low_f16(n));
494    float32x4_t n1_f32   = vcvt_f32_f16(vget_high_f16(n));
495    float32x4_t val0_f32 = vcvt_f32_f16(vget_low_f16(val));
496    float32x4_t val1_f32 = vcvt_f32_f16(vget_high_f16(val));
497
498    float32x4_t res0_f32 = vexpq_f32(vmulq_f32(n0_f32, vlogq_f32(val0_f32)));
499    float32x4_t res1_f32 = vexpq_f32(vmulq_f32(n1_f32, vlogq_f32(val1_f32)));
500
501    return vcombine_f16(vcvt_f16_f32(res0_f32), vcvt_f16_f32(res1_f32));
502}
503
504inline float16x8_t vsinq_f16(float16x8_t val)
505{
506    const float32x4_t val_high = vcvt_f32_f16(vget_high_f16(val));
507    const float32x4_t val_low  = vcvt_f32_f16(vget_low_f16(val));
508
509    const float32x4_t res_high = vsinq_f32(val_high);
510    const float32x4_t res_low  = vsinq_f32(val_low);
511
512    return vcombine_f16(vcvt_f16_f32(res_low), vcvt_f16_f32(res_high));
513}
514
515inline float16x4_t vsin_f16(float16x4_t val)
516{
517    const float32x4_t val_f32  = vcvt_f32_f16(val);
518    const float32x2_t val_high = vget_high_f32(val_f32);
519    const float32x2_t val_low  = vget_low_f32(val_f32);
520
521    const float32x2_t res_high = vsin_f32(val_high);
522    const float32x2_t res_low  = vsin_f32(val_low);
523
524    return vcvt_f16_f32(vcombine_f32(res_low, res_high));
525}
526
527#endif /* DOXYGEN_SKIP_THIS */
528#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
529} // namespace arm_compute
530