NEMath.h
  1  /*
  2   * Copyright (c) 2016, 2019 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  #ifndef __ARM_COMPUTE_NEMATH_H__
 25  #define __ARM_COMPUTE_NEMATH_H__
 26  
 27  
 28  #if defined(ARM_MATH_NEON)
 29  /** Calculate floor of a vector.
 30   *
 31   * @param[in] val Input vector value in F32 format.
 32   *
 33   * @return The calculated floor vector.
 34   */
 35  static inline float32x4_t vfloorq_f32(float32x4_t val);
 36  
 37  /** Calculate inverse square root.
 38   *
 39   * @param[in] x Input value.
 40   *
 41   * @return The calculated inverse square root.
 42   */
 43  static inline float32x2_t vinvsqrt_f32(float32x2_t x);
 44  
 45  /** Calculate inverse square root.
 46   *
 47   * @param[in] x Input value.
 48   *
 49   * @return The calculated inverse square root.
 50   */
 51  static inline float32x4_t vinvsqrtq_f32(float32x4_t x);
 52  
 53  /** Calculate reciprocal.
 54   *
 55   * @param[in] x Input value.
 56   *
 57   * @return The calculated reciprocal.
 58   */
 59  static inline float32x2_t vinv_f32(float32x2_t x);
 60  
 61  /** Calculate reciprocal.
 62   *
 63   * @param[in] x Input value.
 64   *
 65   * @return The calculated reciprocal.
 66   */
 67  static inline float32x4_t vinvq_f32(float32x4_t x);
 68  
 69  /** Perform a 7th degree polynomial approximation using Estrin's method.
 70   *
 71   * @param[in] x      Input vector value in F32 format.
 72   * @param[in] coeffs Polynomial coefficients table. (array of flattened float32x4_t vectors)
 73   *
 74   * @return The calculated approximation.
 75   */
 76  static inline float32x4_t vtaylor_polyq_f32(float32x4_t x, const float32_t *coeffs);
 77  
 78  /** Calculate exponential
 79   *
 80   * @param[in] x Input vector value in F32 format.
 81   *
 82   * @return The calculated exponent.
 83   */
 84  static inline float32x4_t vexpq_f32(float32x4_t x);
 85  
 86  /** Calculate logarithm
 87   *
 88   * @param[in] x Input vector value in F32 format.
 89   *
 90   * @return The calculated logarithm.
 91   */
 92  static inline float32x4_t vlogq_f32(float32x4_t x);
 93  
 94  /** Calculate hyperbolic tangent.
 95   *
 96   * tanh(x) = (e^2x - 1)/(e^2x + 1)
 97   *
 98   * @note We clamp x to [-5,5] to avoid overflowing issues.
 99   *
100   * @param[in] val Input vector value in F32 format.
101   *
102   * @return The calculated Hyperbolic Tangent.
103   */
104  static inline float32x4_t vtanhq_f32(float32x4_t val);
105  
106  /** Calculate n power of a number.
107   *
108   * pow(x,n) = e^(n*log(x))
109   *
110   * @param[in] val Input vector value in F32 format.
111   * @param[in] n   Powers to raise the input to.
112   *
113   * @return The calculated power.
114   */
115  static inline float32x4_t vpowq_f32(float32x4_t val, float32x4_t n);
116  
117  #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
118  /** Calculate hyperbolic tangent.
119   *
120   * tanh(x) = (e^2x - 1)/(e^2x + 1)
121   *
122   * @note We clamp x to [-5,5] to avoid overflowing issues.
123   *
124   * @param[in] val Input vector value in F32 format.
125   *
126   * @return The calculated Hyperbolic Tangent.
127   */
128  static inline float16x8_t vtanhq_f16(float16x8_t val);
129  
130  /** Calculate reciprocal.
131   *
132   * @param[in] x Input value.
133   *
134   * @return The calculated reciprocal.
135   */
136  static inline float16x4_t vinv_f16(float16x4_t x);
137  
138  /** Calculate reciprocal.
139   *
140   * @param[in] x Input value.
141   *
142   * @return The calculated reciprocal.
143   */
144  static inline float16x8_t vinvq_f16(float16x8_t x);
145  
146  /** Calculate inverse square root.
147   *
148   * @param[in] x Input value.
149   *
150   * @return The calculated inverse square root.
151   */
152  static inline float16x4_t vinvsqrt_f16(float16x4_t x);
153  
154  /** Calculate inverse square root.
155   *
156   * @param[in] x Input value.
157   *
158   * @return The calculated inverse square root.
159   */
160  static inline float16x8_t vinvsqrtq_f16(float16x8_t x);
161  
162  /** Calculate exponential
163   *
164   * @param[in] x Input vector value in F16 format.
165   *
166   * @return The calculated exponent.
167   */
168  static inline float16x8_t vexpq_f16(float16x8_t x);
169  
170  /** Calculate n power of a number.
171   *
172   * pow(x,n) = e^(n*log(x))
173   *
174   * @param[in] val Input vector value in F16 format.
175   * @param[in] n   Powers to raise the input to.
176   *
177   * @return The calculated power.
178   */
179  static inline float16x8_t vpowq_f16(float16x8_t val, float16x8_t n);
180  #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
181  
182  /** Exponent polynomial coefficients */
183  extern const float32_t exp_tab[4*8];
184  
185  
186  /** Logarithm polynomial coefficients */
187  extern const float32_t log_tab[4*8];
188  
189  #ifndef DOXYGEN_SKIP_THIS
190  inline float32x4_t vfloorq_f32(float32x4_t val)
191  {
192      static const float32_t CONST_1[4] = {1.f,1.f,1.f,1.f};
193  
194      const int32x4_t   z = vcvtq_s32_f32(val);
195      const float32x4_t r = vcvtq_f32_s32(z);
196  
197      return vbslq_f32(vcgtq_f32(r, val), vsubq_f32(r, vld1q_f32(CONST_1)), r);
198  }
199  
200  inline float32x2_t vinvsqrt_f32(float32x2_t x)
201  {
202      float32x2_t sqrt_reciprocal = vrsqrte_f32(x);
203      sqrt_reciprocal             = vmul_f32(vrsqrts_f32(vmul_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
204      sqrt_reciprocal             = vmul_f32(vrsqrts_f32(vmul_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
205  
206      return sqrt_reciprocal;
207  }
208  
209  inline float32x4_t vinvsqrtq_f32(float32x4_t x)
210  {
211      float32x4_t sqrt_reciprocal = vrsqrteq_f32(x);
212      sqrt_reciprocal             = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
213      sqrt_reciprocal             = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
214  
215      return sqrt_reciprocal;
216  }
217  
218  inline float32x2_t vinv_f32(float32x2_t x)
219  {
220      float32x2_t recip = vrecpe_f32(x);
221      recip             = vmul_f32(vrecps_f32(x, recip), recip);
222      recip             = vmul_f32(vrecps_f32(x, recip), recip);
223      return recip;
224  }
225  
226  inline float32x4_t vinvq_f32(float32x4_t x)
227  {
228      float32x4_t recip = vrecpeq_f32(x);
229      recip             = vmulq_f32(vrecpsq_f32(x, recip), recip);
230      recip             = vmulq_f32(vrecpsq_f32(x, recip), recip);
231      return recip;
232  }
233  
234  inline float32x4_t vtaylor_polyq_f32(float32x4_t x, const float32_t *coeffs)
235  {
236      float32x4_t A   = vmlaq_f32(vld1q_f32(&coeffs[4*0]), vld1q_f32(&coeffs[4*4]), x);
237      float32x4_t B   = vmlaq_f32(vld1q_f32(&coeffs[4*2]), vld1q_f32(&coeffs[4*6]), x);
238      float32x4_t C   = vmlaq_f32(vld1q_f32(&coeffs[4*1]), vld1q_f32(&coeffs[4*5]), x);
239      float32x4_t D   = vmlaq_f32(vld1q_f32(&coeffs[4*3]), vld1q_f32(&coeffs[4*7]), x);
240      float32x4_t x2  = vmulq_f32(x, x);
241      float32x4_t x4  = vmulq_f32(x2, x2);
242      float32x4_t res = vmlaq_f32(vmlaq_f32(A, B, x2), vmlaq_f32(C, D, x2), x4);
243      return res;
244  }
245  
246  inline float32x4_t vexpq_f32(float32x4_t x)
247  {
248      static const float32_t CONST_LN2[4]          = {0.6931471805f,0.6931471805f,0.6931471805f,0.6931471805f}; // ln(2)
249      static const float32_t CONST_INV_LN2[4]      = {1.4426950408f,1.4426950408f,1.4426950408f,1.4426950408f}; // 1/ln(2)
250      static const float32_t CONST_0[4]            = {0.f,0.f,0.f,0.f};
251      static const int32_t   CONST_NEGATIVE_126[4] = {-126,-126,-126,-126};
252  
253      // Perform range reduction [-log(2),log(2)]
254      int32x4_t   m   = vcvtq_s32_f32(vmulq_f32(x, vld1q_f32(CONST_INV_LN2)));
255      float32x4_t val = vmlsq_f32(x, vcvtq_f32_s32(m), vld1q_f32(CONST_LN2));
256  
257      // Polynomial Approximation
258      float32x4_t poly = vtaylor_polyq_f32(val, exp_tab);
259  
260      // Reconstruct
261      poly = vreinterpretq_f32_s32(vqaddq_s32(vreinterpretq_s32_f32(poly), vqshlq_n_s32(m, 23)));
262      poly = vbslq_f32(vcltq_s32(m, vld1q_s32(CONST_NEGATIVE_126)), vld1q_f32(CONST_0), poly);
263  
264      return poly;
265  }
266  
267  inline float32x4_t vlogq_f32(float32x4_t x)
268  {
269      static const int32_t   CONST_127[4] = {127,127,127,127};           // 127
270      static const float32_t CONST_LN2[4] = {0.6931471805f,0.6931471805f,0.6931471805f,0.6931471805f}; // ln(2)
271  
272      // Extract exponent
273      int32x4_t   m   = vsubq_s32(vreinterpretq_s32_u32(vshrq_n_u32(vreinterpretq_u32_f32(x), 23)), vld1q_s32(CONST_127));
274      float32x4_t val = vreinterpretq_f32_s32(vsubq_s32(vreinterpretq_s32_f32(x), vshlq_n_s32(m, 23)));
275  
276      // Polynomial Approximation
277      float32x4_t poly = vtaylor_polyq_f32(val, log_tab);
278  
279      // Reconstruct
280      poly = vmlaq_f32(poly, vcvtq_f32_s32(m), vld1q_f32(CONST_LN2));
281  
282      return poly;
283  }
284  
285  inline float32x4_t vtanhq_f32(float32x4_t val)
286  {
287      static const float32_t CONST_1[4]        = {1.f,1.f,1.f,1.f};
288      static const float32_t CONST_2[4]        = {2.f,2.f,2.f,2.f};
289      static const float32_t CONST_MIN_TANH[4] = {-10.f,-10.f,-10.f,-10.f};
290      static const float32_t CONST_MAX_TANH[4] = {10.f,10.f,10.f,10.f};
291  
292      float32x4_t x     = vminq_f32(vmaxq_f32(val, vld1q_f32(CONST_MIN_TANH)), vld1q_f32(CONST_MAX_TANH));
293      float32x4_t exp2x = vexpq_f32(vmulq_f32(vld1q_f32(CONST_2), x));
294      float32x4_t num   = vsubq_f32(exp2x, vld1q_f32(CONST_1));
295      float32x4_t den   = vaddq_f32(exp2x, vld1q_f32(CONST_1));
296      float32x4_t tanh  = vmulq_f32(num, vinvq_f32(den));
297      return tanh;
298  }
299  
300  inline float32x4_t vpowq_f32(float32x4_t val, float32x4_t n)
301  {
302      return vexpq_f32(vmulq_f32(n, vlogq_f32(val)));
303  }
304  #endif /* DOXYGEN_SKIP_THIS */
305  
306  #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
307  /** Exponent polynomial coefficients */
308  /** Logarithm polynomial coefficients */
309  #ifndef DOXYGEN_SKIP_THIS
310  inline float16x8_t vfloorq_f16(float16x8_t val)
311  {
312      static const float16_t CONST_1[8] = {1.f,1.f,1.f,1.f,1.f,1.f,1.f,1.f};
313  
314      const int16x8_t   z = vcvtq_s16_f16(val);
315      const float16x8_t r = vcvtq_f16_s16(z);
316  
317      return vbslq_f16(vcgtq_f16(r, val), vsubq_f16(r, vld1q_f16(CONST_1)), r);
318  }
319  inline float16x4_t vinvsqrt_f16(float16x4_t x)
320  {
321      float16x4_t sqrt_reciprocal = vrsqrte_f16(x);
322      sqrt_reciprocal             = vmul_f16(vrsqrts_f16(vmul_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
323      sqrt_reciprocal             = vmul_f16(vrsqrts_f16(vmul_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
324      return sqrt_reciprocal;
325  }
326  
327  inline float16x8_t vinvsqrtq_f16(float16x8_t x)
328  {
329      float16x8_t sqrt_reciprocal = vrsqrteq_f16(x);
330      sqrt_reciprocal             = vmulq_f16(vrsqrtsq_f16(vmulq_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
331      sqrt_reciprocal             = vmulq_f16(vrsqrtsq_f16(vmulq_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
332      return sqrt_reciprocal;
333  }
334  
335  inline float16x4_t vinv_f16(float16x4_t x)
336  {
337      float16x4_t recip = vrecpe_f16(x);
338      recip             = vmul_f16(vrecps_f16(x, recip), recip);
339      recip             = vmul_f16(vrecps_f16(x, recip), recip);
340      return recip;
341  }
342  
343  inline float16x8_t vinvq_f16(float16x8_t x)
344  {
345      float16x8_t recip = vrecpeq_f16(x);
346      recip             = vmulq_f16(vrecpsq_f16(x, recip), recip);
347      recip             = vmulq_f16(vrecpsq_f16(x, recip), recip);
348      return recip;
349  }
350  
351  inline float16x8_t vtanhq_f16(float16x8_t val)
352  {
353      const float16_t CONST_1[8]        = {1.f,1.f,1.f,1.f,1.f,1.f,1.f,1.f};
354      const float16_t CONST_2[8]        = {2.f,2.f,2.f,2.f,2.f,2.f,2.f,2.f};
355      const float16_t CONST_MIN_TANH[8] = {-10.f,-10.f,-10.f,-10.f,-10.f,-10.f,-10.f,-10.f};
356      const float16_t CONST_MAX_TANH[8] = {10.f,10.f,10.f,10.f,10.f,10.f,10.f,10.f};
357  
358      const float16x8_t x     = vminq_f16(vmaxq_f16(val, vld1q_f16(CONST_MIN_TANH)), vld1q_f16(CONST_MAX_TANH));
359      const float16x8_t exp2x = vexpq_f16(vmulq_f16(vld1q_f16(CONST_2), x));
360      const float16x8_t num   = vsubq_f16(exp2x, vld1q_f16(CONST_1));
361      const float16x8_t den   = vaddq_f16(exp2x, vld1q_f16(CONST_1));
362      const float16x8_t tanh  = vmulq_f16(num, vinvq_f16(den));
363      return tanh;
364  }
365  
366  inline float16x8_t vtaylor_polyq_f16(float16x8_t x, const float16_t *coeffs)
367  {
368      const float16x8_t A   = vaddq_f16(vld1q_f16(&coeffs[8*0]), vmulq_f16(vld1q_f16(&coeffs[8*4]), x));
369      const float16x8_t B   = vaddq_f16(vld1q_f16(&coeffs[8*2]), vmulq_f16(vld1q_f16(&coeffs[8*6]), x));
370      const float16x8_t C   = vaddq_f16(vld1q_f16(&coeffs[8*1]), vmulq_f16(vld1q_f16(&coeffs[8*5]), x));
371      const float16x8_t D   = vaddq_f16(vld1q_f16(&coeffs[8*3]), vmulq_f16(vld1q_f16(&coeffs[8*7]), x));
372      const float16x8_t x2  = vmulq_f16(x, x);
373      const float16x8_t x4  = vmulq_f16(x2, x2);
374      const float16x8_t res = vaddq_f16(vaddq_f16(A, vmulq_f16(B, x2)), vmulq_f16(vaddq_f16(C, vmulq_f16(D, x2)), x4));
375      return res;
376  }
377  
378  inline float16x8_t vexpq_f16(float16x8_t x)
379  {
380      // TODO (COMPMID-1535) : Revisit FP16 approximations
381      const float32x4_t x_high = vcvt_f32_f16(vget_high_f16(x));
382      const float32x4_t x_low  = vcvt_f32_f16(vget_low_f16(x));
383  
384      const float16x8_t res = vcvt_high_f16_f32(vcvt_f16_f32(vexpq_f32(x_low)), vexpq_f32(x_high));
385      return res;
386  }
387  
388  inline float16x8_t vlogq_f16(float16x8_t x)
389  {
390      // TODO (COMPMID-1535) : Revisit FP16 approximations
391      const float32x4_t x_high = vcvt_f32_f16(vget_high_f16(x));
392      const float32x4_t x_low  = vcvt_f32_f16(vget_low_f16(x));
393  
394      const float16x8_t res = vcvt_high_f16_f32(vcvt_f16_f32(vlogq_f32(x_low)), vlogq_f32(x_high));
395      return res;
396  }
397  
398  inline float16x8_t vpowq_f16(float16x8_t val, float16x8_t n)
399  {
400      // TODO (giaiod01) - COMPMID-1535
401      float32x4_t n0_f32   = vcvt_f32_f16(vget_low_f16(n));
402      float32x4_t n1_f32   = vcvt_f32_f16(vget_high_f16(n));
403      float32x4_t val0_f32 = vcvt_f32_f16(vget_low_f16(val));
404      float32x4_t val1_f32 = vcvt_f32_f16(vget_high_f16(val));
405  
406      float32x4_t res0_f32 = vexpq_f32(vmulq_f32(n0_f32, vlogq_f32(val0_f32)));
407      float32x4_t res1_f32 = vexpq_f32(vmulq_f32(n1_f32, vlogq_f32(val1_f32)));
408  
409      return vcombine_f16(vcvt_f16_f32(res0_f32), vcvt_f16_f32(res1_f32));
410  }
411  #endif /* DOXYGEN_SKIP_THIS */
412  #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
413  #endif
414  #endif /* __ARM_COMPUTE_NEMATH_H__ */