80 #ifndef INCLUDE_VOLK_VOLK_NEON_INTRINSICS_H_ 81 #define INCLUDE_VOLK_VOLK_NEON_INTRINSICS_H_ 89 float32x4_t iValue, qValue, result;
90 iValue = vmulq_f32(cmplxValue.val[0], cmplxValue.val[0]);
91 qValue = vmulq_f32(cmplxValue.val[1], cmplxValue.val[1]);
92 result = vaddq_f32(iValue, qValue);
99 float32x4_t sqrt_reciprocal = vrsqrteq_f32(x);
100 sqrt_reciprocal = vmulq_f32(
101 vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
102 sqrt_reciprocal = vmulq_f32(
103 vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
105 return sqrt_reciprocal;
112 float32x4_t recip = vrecpeq_f32(x);
113 recip = vmulq_f32(vrecpsq_f32(x, recip), recip);
114 recip = vmulq_f32(vrecpsq_f32(x, recip), recip);
122 float32x4x2_t tmp_real;
123 float32x4x2_t tmp_imag;
128 tmp_real.val[0] = vmulq_f32(a_val.val[0], b_val.val[0]);
130 tmp_real.val[1] = vmulq_f32(a_val.val[1], b_val.val[1]);
133 tmp_imag.val[0] = vmulq_f32(a_val.val[0], b_val.val[1]);
135 tmp_imag.val[1] = vmulq_f32(a_val.val[1], b_val.val[0]);
137 c_val.val[0] = vsubq_f32(tmp_real.val[0], tmp_real.val[1]);
138 c_val.val[1] = vaddq_f32(tmp_imag.val[0], tmp_imag.val[1]);
145 float32x4_t cA = vmlaq_f32(coeffs[0], coeffs[4], x);
146 float32x4_t cB = vmlaq_f32(coeffs[2], coeffs[6], x);
147 float32x4_t cC = vmlaq_f32(coeffs[1], coeffs[5], x);
148 float32x4_t cD = vmlaq_f32(coeffs[3], coeffs[7], x);
149 float32x4_t x2 = vmulq_f32(x, x);
150 float32x4_t x4 = vmulq_f32(x2, x2);
151 float32x4_t res = vmlaq_f32(vmlaq_f32(cA, cB, x2), vmlaq_f32(cC, cD, x2), x4);
159 const float32x4_t log_tab[8] = {
160 vdupq_n_f32(-2.29561495781f), vdupq_n_f32(-2.47071170807f),
161 vdupq_n_f32(-5.68692588806f), vdupq_n_f32(-0.165253549814f),
162 vdupq_n_f32(5.17591238022f), vdupq_n_f32(0.844007015228f),
163 vdupq_n_f32(4.58445882797f), vdupq_n_f32(0.0141278216615f),
166 const int32x4_t CONST_127 = vdupq_n_s32(127);
167 const float32x4_t CONST_LN2 = vdupq_n_f32(0.6931471805f);
170 int32x4_t m = vsubq_s32(
171 vreinterpretq_s32_u32(vshrq_n_u32(vreinterpretq_u32_f32(x), 23)), CONST_127);
173 vreinterpretq_f32_s32(vsubq_s32(vreinterpretq_s32_f32(x), vshlq_n_s32(m, 23)));
179 poly = vmlaq_f32(poly, vcvtq_f32_s32(m), CONST_LN2);
189 const float32x4_t c_minus_cephes_DP1 = vdupq_n_f32(-0.78515625);
190 const float32x4_t c_minus_cephes_DP2 = vdupq_n_f32(-2.4187564849853515625e-4);
191 const float32x4_t c_minus_cephes_DP3 = vdupq_n_f32(-3.77489497744594108e-8);
192 const float32x4_t c_sincof_p0 = vdupq_n_f32(-1.9515295891e-4);
193 const float32x4_t c_sincof_p1 = vdupq_n_f32(8.3321608736e-3);
194 const float32x4_t c_sincof_p2 = vdupq_n_f32(-1.6666654611e-1);
195 const float32x4_t c_coscof_p0 = vdupq_n_f32(2.443315711809948e-005);
196 const float32x4_t c_coscof_p1 = vdupq_n_f32(-1.388731625493765e-003);
197 const float32x4_t c_coscof_p2 = vdupq_n_f32(4.166664568298827e-002);
198 const float32x4_t c_cephes_FOPI = vdupq_n_f32(1.27323954473516);
200 const float32x4_t CONST_1 = vdupq_n_f32(1.f);
201 const float32x4_t CONST_1_2 = vdupq_n_f32(0.5f);
202 const float32x4_t CONST_0 = vdupq_n_f32(0.f);
203 const uint32x4_t CONST_2 = vdupq_n_u32(2);
204 const uint32x4_t CONST_4 = vdupq_n_u32(4);
208 uint32x4_t sign_mask_sin, sign_mask_cos;
209 sign_mask_sin = vcltq_f32(x, CONST_0);
212 float32x4_t y = vmulq_f32(x, c_cephes_FOPI);
215 emm2 = vcvtq_u32_f32(y);
217 emm2 = vaddq_u32(emm2, vdupq_n_u32(1));
218 emm2 = vandq_u32(emm2, vdupq_n_u32(~1));
219 y = vcvtq_f32_u32(emm2);
225 const uint32x4_t poly_mask = vtstq_u32(emm2, CONST_2);
228 x = vmlaq_f32(x, y, c_minus_cephes_DP1);
229 x = vmlaq_f32(x, y, c_minus_cephes_DP2);
230 x = vmlaq_f32(x, y, c_minus_cephes_DP3);
232 sign_mask_sin = veorq_u32(sign_mask_sin, vtstq_u32(emm2, CONST_4));
233 sign_mask_cos = vtstq_u32(vsubq_u32(emm2, CONST_2), CONST_4);
238 float32x4_t z = vmulq_f32(x, x);
240 y1 = vmlaq_f32(c_coscof_p1, z, c_coscof_p0);
241 y1 = vmlaq_f32(c_coscof_p2, z, y1);
242 y1 = vmulq_f32(y1, z);
243 y1 = vmulq_f32(y1, z);
244 y1 = vmlsq_f32(y1, z, CONST_1_2);
245 y1 = vaddq_f32(y1, CONST_1);
247 y2 = vmlaq_f32(c_sincof_p1, z, c_sincof_p0);
248 y2 = vmlaq_f32(c_sincof_p2, z, y2);
249 y2 = vmulq_f32(y2, z);
250 y2 = vmlaq_f32(x, x, y2);
253 const float32x4_t ys = vbslq_f32(poly_mask, y1, y2);
254 const float32x4_t yc = vbslq_f32(poly_mask, y2, y1);
256 float32x4x2_t sincos;
257 sincos.val[0] = vbslq_f32(sign_mask_sin, vnegq_f32(ys), ys);
258 sincos.val[1] = vbslq_f32(sign_mask_cos, yc, vnegq_f32(yc));
266 return sincos.val[0];
272 return sincos.val[1];
278 return vmulq_f32(sincos.val[0],
_vinvq_f32(sincos.val[1]));
static float32x4_t _vsinq_f32(float32x4_t x)
Definition: volk_neon_intrinsics.h:263
val
Definition: volk_arch_defs.py:66
static float32x4_t _vinvsqrtq_f32(float32x4_t x)
Definition: volk_neon_intrinsics.h:97
static float32x4x2_t _vmultiply_complexq_f32(float32x4x2_t a_val, float32x4x2_t b_val)
Definition: volk_neon_intrinsics.h:119
static float32x4_t _vtaylor_polyq_f32(float32x4_t x, const float32x4_t coeffs[8])
Definition: volk_neon_intrinsics.h:143
static float32x4x2_t _vsincosq_f32(float32x4_t x)
Definition: volk_neon_intrinsics.h:187
static float32x4_t _vmagnitudesquaredq_f32(float32x4x2_t cmplxValue)
Definition: volk_neon_intrinsics.h:87
static float32x4_t _vtanq_f32(float32x4_t x)
Definition: volk_neon_intrinsics.h:275
static float32x4_t _vcosq_f32(float32x4_t x)
Definition: volk_neon_intrinsics.h:269
static float32x4_t _vinvq_f32(float32x4_t x)
Definition: volk_neon_intrinsics.h:109
static float32x4_t _vlogq_f32(float32x4_t x)
Definition: volk_neon_intrinsics.h:157