| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* -*- c++ -*- */ | ||
| 2 | /* | ||
| 3 | * Copyright 2014 Free Software Foundation, Inc. | ||
| 4 | * Copyright 2023 Magnus Lundmark <magnuslundmark@gmail.com> | ||
| 5 | * | ||
| 6 | * This file is part of VOLK | ||
| 7 | * | ||
| 8 | * SPDX-License-Identifier: LGPL-3.0-or-later | ||
| 9 | */ | ||
| 10 | |||
| 11 | /*! | ||
| 12 | * \page volk_32f_atan_32f | ||
| 13 | * | ||
| 14 | * \b Overview | ||
| 15 | * | ||
| 16 | * Computes arcsine of input vector and stores results in output vector. | ||
| 17 | * | ||
| 18 | * <b>Dispatcher Prototype</b> | ||
| 19 | * \code | ||
| 20 | * void volk_32f_atan_32f(float* bVector, const float* aVector, unsigned int num_points) | ||
| 21 | * \endcode | ||
| 22 | * | ||
| 23 | * \b Inputs | ||
| 24 | * \li aVector: The input vector of floats. | ||
| 25 | * \li num_points: The number of data points. | ||
| 26 | * | ||
| 27 | * \b Outputs | ||
| 28 | * \li bVector: The vector where results will be stored. | ||
| 29 | * | ||
| 30 | * \b Example | ||
| 31 | * Calculate common angles around the top half of the unit circle. | ||
| 32 | * \code | ||
| 33 | * int N = 10; | ||
| 34 | * unsigned int alignment = volk_get_alignment(); | ||
| 35 | * float* in = (float*)volk_malloc(sizeof(float)*N, alignment); | ||
| 36 | * float* out = (float*)volk_malloc(sizeof(float)*N, alignment); | ||
| 37 | * | ||
| 38 | * in[0] = 0.f; | ||
| 39 | * in[1] = 1.f/std::sqrt(3.f); | ||
| 40 | * in[2] = 1.f; | ||
| 41 | * in[3] = std::sqrt(3.f); | ||
| 42 | * in[4] = in[5] = 1e99; | ||
| 43 | * for(unsigned int ii = 6; ii < N; ++ii){ | ||
| 44 | * in[ii] = - in[N-ii-1]; | ||
| 45 | * } | ||
| 46 | * | ||
| 47 | * volk_32f_atan_32f(out, in, N); | ||
| 48 | * | ||
| 49 | * for(unsigned int ii = 0; ii < N; ++ii){ | ||
| 50 | * printf("atan(%1.3f) = %1.3f\n", in[ii], out[ii]); | ||
| 51 | * } | ||
| 52 | * | ||
| 53 | * volk_free(in); | ||
| 54 | * volk_free(out); | ||
| 55 | * \endcode | ||
| 56 | */ | ||
| 57 | #include <math.h> | ||
| 58 | |||
| 59 | #ifndef INCLUDED_volk_32f_atan_32f_a_H | ||
| 60 | #define INCLUDED_volk_32f_atan_32f_a_H | ||
| 61 | |||
| 62 | #if LV_HAVE_AVX2 && LV_HAVE_FMA | ||
| 63 | #include <immintrin.h> | ||
| 64 | #include <volk/volk_avx2_fma_intrinsics.h> | ||
| 65 | static inline void | ||
| 66 | 2 | volk_32f_atan_32f_a_avx2_fma(float* out, const float* in, unsigned int num_points) | |
| 67 | { | ||
| 68 | 2 | const __m256 one = _mm256_set1_ps(1.f); | |
| 69 | 2 | const __m256 pi_over_2 = _mm256_set1_ps(0x1.921fb6p0f); | |
| 70 | 4 | const __m256 abs_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF)); | |
| 71 | 2 | const __m256 sign_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x80000000)); | |
| 72 | |||
| 73 | 2 | unsigned int number = 0; | |
| 74 | 2 | unsigned int eighth_points = num_points / 8; | |
| 75 |
2/2✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
|
32768 | for (; number < eighth_points; number++) { |
| 76 | 32766 | __m256 x = _mm256_load_ps(in); | |
| 77 | 32766 | __m256 swap_mask = _mm256_cmp_ps(_mm256_and_ps(x, abs_mask), one, _CMP_GT_OS); | |
| 78 | 65532 | __m256 x_star = _mm256_div_ps(_mm256_blendv_ps(x, one, swap_mask), | |
| 79 | _mm256_blendv_ps(one, x, swap_mask)); | ||
| 80 | 32766 | __m256 result = _m256_arctan_poly_avx2_fma(x_star); | |
| 81 | 32766 | __m256 term = _mm256_and_ps(x_star, sign_mask); | |
| 82 | 32766 | term = _mm256_or_ps(pi_over_2, term); | |
| 83 | 32766 | term = _mm256_sub_ps(term, result); | |
| 84 | 32766 | result = _mm256_blendv_ps(result, term, swap_mask); | |
| 85 | _mm256_store_ps(out, result); | ||
| 86 | 32766 | in += 8; | |
| 87 | 32766 | out += 8; | |
| 88 | } | ||
| 89 | |||
| 90 | 2 | number = eighth_points * 8; | |
| 91 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
|
16 | for (; number < num_points; number++) { |
| 92 | 14 | *out++ = volk_arctan(*in++); | |
| 93 | } | ||
| 94 | 2 | } | |
| 95 | #endif /* LV_HAVE_AVX2 && LV_HAVE_FMA for aligned */ | ||
| 96 | |||
| 97 | #if LV_HAVE_AVX | ||
| 98 | #include <immintrin.h> | ||
| 99 | #include <volk/volk_avx_intrinsics.h> | ||
| 100 | static inline void | ||
| 101 | 2 | volk_32f_atan_32f_a_avx2(float* out, const float* in, unsigned int num_points) | |
| 102 | { | ||
| 103 | 2 | const __m256 one = _mm256_set1_ps(1.f); | |
| 104 | 2 | const __m256 pi_over_2 = _mm256_set1_ps(0x1.921fb6p0f); | |
| 105 | 4 | const __m256 abs_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF)); | |
| 106 | 2 | const __m256 sign_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x80000000)); | |
| 107 | |||
| 108 | 2 | unsigned int number = 0; | |
| 109 | 2 | unsigned int eighth_points = num_points / 8; | |
| 110 |
2/2✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
|
32768 | for (; number < eighth_points; number++) { |
| 111 | 32766 | __m256 x = _mm256_load_ps(in); | |
| 112 | 32766 | __m256 swap_mask = _mm256_cmp_ps(_mm256_and_ps(x, abs_mask), one, _CMP_GT_OS); | |
| 113 | 65532 | __m256 x_star = _mm256_div_ps(_mm256_blendv_ps(x, one, swap_mask), | |
| 114 | _mm256_blendv_ps(one, x, swap_mask)); | ||
| 115 | 32766 | __m256 result = _m256_arctan_poly_avx(x_star); | |
| 116 | 32766 | __m256 term = _mm256_and_ps(x_star, sign_mask); | |
| 117 | 32766 | term = _mm256_or_ps(pi_over_2, term); | |
| 118 | 32766 | term = _mm256_sub_ps(term, result); | |
| 119 | 32766 | result = _mm256_blendv_ps(result, term, swap_mask); | |
| 120 | _mm256_store_ps(out, result); | ||
| 121 | 32766 | in += 8; | |
| 122 | 32766 | out += 8; | |
| 123 | } | ||
| 124 | |||
| 125 | 2 | number = eighth_points * 8; | |
| 126 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
|
16 | for (; number < num_points; number++) { |
| 127 | 14 | *out++ = volk_arctan(*in++); | |
| 128 | } | ||
| 129 | 2 | } | |
| 130 | #endif /* LV_HAVE_AVX for aligned */ | ||
| 131 | |||
| 132 | #ifdef LV_HAVE_SSE4_1 | ||
| 133 | #include <smmintrin.h> | ||
| 134 | #include <volk/volk_sse_intrinsics.h> | ||
| 135 | static inline void | ||
| 136 | 2 | volk_32f_atan_32f_a_sse4_1(float* out, const float* in, unsigned int num_points) | |
| 137 | { | ||
| 138 | 2 | const __m128 one = _mm_set1_ps(1.f); | |
| 139 | 2 | const __m128 pi_over_2 = _mm_set1_ps(0x1.921fb6p0f); | |
| 140 | 4 | const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF)); | |
| 141 | 2 | const __m128 sign_mask = _mm_castsi128_ps(_mm_set1_epi32(0x80000000)); | |
| 142 | |||
| 143 | 2 | unsigned int number = 0; | |
| 144 | 2 | unsigned int quarter_points = num_points / 4; | |
| 145 |
2/2✓ Branch 0 taken 65534 times.
✓ Branch 1 taken 2 times.
|
65536 | for (; number < quarter_points; number++) { |
| 146 | 65534 | __m128 x = _mm_load_ps(in); | |
| 147 | 131068 | __m128 swap_mask = _mm_cmpgt_ps(_mm_and_ps(x, abs_mask), one); | |
| 148 | 131068 | __m128 x_star = _mm_div_ps(_mm_blendv_ps(x, one, swap_mask), | |
| 149 | _mm_blendv_ps(one, x, swap_mask)); | ||
| 150 | 65534 | __m128 result = _mm_arctan_poly_sse(x_star); | |
| 151 | 65534 | __m128 term = _mm_and_ps(x_star, sign_mask); | |
| 152 | 65534 | term = _mm_or_ps(pi_over_2, term); | |
| 153 | 65534 | term = _mm_sub_ps(term, result); | |
| 154 | 65534 | result = _mm_blendv_ps(result, term, swap_mask); | |
| 155 | _mm_store_ps(out, result); | ||
| 156 | 65534 | in += 4; | |
| 157 | 65534 | out += 4; | |
| 158 | } | ||
| 159 | |||
| 160 | 2 | number = quarter_points * 4; | |
| 161 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 2 times.
|
8 | for (; number < num_points; number++) { |
| 162 | 6 | *out++ = volk_arctan(*in++); | |
| 163 | } | ||
| 164 | 2 | } | |
| 165 | #endif /* LV_HAVE_SSE4_1 for aligned */ | ||
| 166 | #endif /* INCLUDED_volk_32f_atan_32f_a_H */ | ||
| 167 | |||
| 168 | #ifndef INCLUDED_volk_32f_atan_32f_u_H | ||
| 169 | #define INCLUDED_volk_32f_atan_32f_u_H | ||
| 170 | |||
| 171 | #if LV_HAVE_AVX2 && LV_HAVE_FMA | ||
| 172 | #include <immintrin.h> | ||
| 173 | static inline void | ||
| 174 | 2 | volk_32f_atan_32f_u_avx2_fma(float* out, const float* in, unsigned int num_points) | |
| 175 | { | ||
| 176 | 2 | const __m256 one = _mm256_set1_ps(1.f); | |
| 177 | 2 | const __m256 pi_over_2 = _mm256_set1_ps(0x1.921fb6p0f); | |
| 178 | 4 | const __m256 abs_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF)); | |
| 179 | 2 | const __m256 sign_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x80000000)); | |
| 180 | |||
| 181 | 2 | unsigned int number = 0; | |
| 182 | 2 | unsigned int eighth_points = num_points / 8; | |
| 183 |
2/2✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
|
32768 | for (; number < eighth_points; number++) { |
| 184 | 32766 | __m256 x = _mm256_loadu_ps(in); | |
| 185 | 32766 | __m256 swap_mask = _mm256_cmp_ps(_mm256_and_ps(x, abs_mask), one, _CMP_GT_OS); | |
| 186 | 65532 | __m256 x_star = _mm256_div_ps(_mm256_blendv_ps(x, one, swap_mask), | |
| 187 | _mm256_blendv_ps(one, x, swap_mask)); | ||
| 188 | 32766 | __m256 result = _m256_arctan_poly_avx2_fma(x_star); | |
| 189 | 32766 | __m256 term = _mm256_and_ps(x_star, sign_mask); | |
| 190 | 32766 | term = _mm256_or_ps(pi_over_2, term); | |
| 191 | 32766 | term = _mm256_sub_ps(term, result); | |
| 192 | 32766 | result = _mm256_blendv_ps(result, term, swap_mask); | |
| 193 | _mm256_storeu_ps(out, result); | ||
| 194 | 32766 | in += 8; | |
| 195 | 32766 | out += 8; | |
| 196 | } | ||
| 197 | |||
| 198 | 2 | number = eighth_points * 8; | |
| 199 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
|
16 | for (; number < num_points; number++) { |
| 200 | 14 | *out++ = volk_arctan(*in++); | |
| 201 | } | ||
| 202 | 2 | } | |
| 203 | #endif /* LV_HAVE_AVX2 && LV_HAVE_FMA for unaligned */ | ||
| 204 | |||
| 205 | #if LV_HAVE_AVX | ||
| 206 | #include <immintrin.h> | ||
| 207 | static inline void | ||
| 208 | 2 | volk_32f_atan_32f_u_avx2(float* out, const float* in, unsigned int num_points) | |
| 209 | { | ||
| 210 | 2 | const __m256 one = _mm256_set1_ps(1.f); | |
| 211 | 2 | const __m256 pi_over_2 = _mm256_set1_ps(0x1.921fb6p0f); | |
| 212 | 4 | const __m256 abs_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF)); | |
| 213 | 2 | const __m256 sign_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x80000000)); | |
| 214 | |||
| 215 | 2 | unsigned int number = 0; | |
| 216 | 2 | unsigned int eighth_points = num_points / 8; | |
| 217 |
2/2✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
|
32768 | for (; number < eighth_points; number++) { |
| 218 | 32766 | __m256 x = _mm256_loadu_ps(in); | |
| 219 | 32766 | __m256 swap_mask = _mm256_cmp_ps(_mm256_and_ps(x, abs_mask), one, _CMP_GT_OS); | |
| 220 | 65532 | __m256 x_star = _mm256_div_ps(_mm256_blendv_ps(x, one, swap_mask), | |
| 221 | _mm256_blendv_ps(one, x, swap_mask)); | ||
| 222 | 32766 | __m256 result = _m256_arctan_poly_avx(x_star); | |
| 223 | 32766 | __m256 term = _mm256_and_ps(x_star, sign_mask); | |
| 224 | 32766 | term = _mm256_or_ps(pi_over_2, term); | |
| 225 | 32766 | term = _mm256_sub_ps(term, result); | |
| 226 | 32766 | result = _mm256_blendv_ps(result, term, swap_mask); | |
| 227 | _mm256_storeu_ps(out, result); | ||
| 228 | 32766 | in += 8; | |
| 229 | 32766 | out += 8; | |
| 230 | } | ||
| 231 | |||
| 232 | 2 | number = eighth_points * 8; | |
| 233 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
|
16 | for (; number < num_points; number++) { |
| 234 | 14 | *out++ = volk_arctan(*in++); | |
| 235 | } | ||
| 236 | 2 | } | |
| 237 | #endif /* LV_HAVE_AVX for unaligned */ | ||
| 238 | |||
| 239 | #ifdef LV_HAVE_SSE4_1 | ||
| 240 | #include <smmintrin.h> | ||
| 241 | #include <volk/volk_sse_intrinsics.h> | ||
| 242 | static inline void | ||
| 243 | 2 | volk_32f_atan_32f_u_sse4_1(float* out, const float* in, unsigned int num_points) | |
| 244 | { | ||
| 245 | 2 | const __m128 one = _mm_set1_ps(1.f); | |
| 246 | 2 | const __m128 pi_over_2 = _mm_set1_ps(0x1.921fb6p0f); | |
| 247 | 4 | const __m128 abs_mask = _mm_castsi128_ps(_mm_set1_epi32(0x7FFFFFFF)); | |
| 248 | 2 | const __m128 sign_mask = _mm_castsi128_ps(_mm_set1_epi32(0x80000000)); | |
| 249 | |||
| 250 | 2 | unsigned int number = 0; | |
| 251 | 2 | unsigned int quarter_points = num_points / 4; | |
| 252 |
2/2✓ Branch 0 taken 65534 times.
✓ Branch 1 taken 2 times.
|
65536 | for (; number < quarter_points; number++) { |
| 253 | 65534 | __m128 x = _mm_loadu_ps(in); | |
| 254 | 131068 | __m128 swap_mask = _mm_cmpgt_ps(_mm_and_ps(x, abs_mask), one); | |
| 255 | 131068 | __m128 x_star = _mm_div_ps(_mm_blendv_ps(x, one, swap_mask), | |
| 256 | _mm_blendv_ps(one, x, swap_mask)); | ||
| 257 | 65534 | __m128 result = _mm_arctan_poly_sse(x_star); | |
| 258 | 65534 | __m128 term = _mm_and_ps(x_star, sign_mask); | |
| 259 | 65534 | term = _mm_or_ps(pi_over_2, term); | |
| 260 | 65534 | term = _mm_sub_ps(term, result); | |
| 261 | 65534 | result = _mm_blendv_ps(result, term, swap_mask); | |
| 262 | _mm_storeu_ps(out, result); | ||
| 263 | 65534 | in += 4; | |
| 264 | 65534 | out += 4; | |
| 265 | } | ||
| 266 | |||
| 267 | 2 | number = quarter_points * 4; | |
| 268 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 2 times.
|
8 | for (; number < num_points; number++) { |
| 269 | 6 | *out++ = volk_arctan(*in++); | |
| 270 | } | ||
| 271 | 2 | } | |
| 272 | #endif /* LV_HAVE_SSE4_1 for unaligned */ | ||
| 273 | |||
| 274 | #ifdef LV_HAVE_GENERIC | ||
| 275 | static inline void | ||
| 276 | 2 | volk_32f_atan_32f_polynomial(float* out, const float* in, unsigned int num_points) | |
| 277 | { | ||
| 278 | 2 | unsigned int number = 0; | |
| 279 |
2/2✓ Branch 0 taken 262142 times.
✓ Branch 1 taken 2 times.
|
262144 | for (; number < num_points; number++) { |
| 280 | 262142 | *out++ = volk_arctan(*in++); | |
| 281 | } | ||
| 282 | 2 | } | |
| 283 | #endif /* LV_HAVE_GENERIC */ | ||
| 284 | |||
| 285 | #ifdef LV_HAVE_GENERIC | ||
| 286 | static inline void | ||
| 287 | 2 | volk_32f_atan_32f_generic(float* out, const float* in, unsigned int num_points) | |
| 288 | { | ||
| 289 | 2 | unsigned int number = 0; | |
| 290 |
2/2✓ Branch 0 taken 262142 times.
✓ Branch 1 taken 2 times.
|
262144 | for (; number < num_points; number++) { |
| 291 | 262142 | *out++ = atanf(*in++); | |
| 292 | } | ||
| 293 | 2 | } | |
| 294 | #endif /* LV_HAVE_GENERIC */ | ||
| 295 | |||
| 296 | #endif /* INCLUDED_volk_32f_atan_32f_u_H */ | ||
| 297 |