| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* -*- c++ -*- */ | ||
| 2 | /* | ||
| 3 | * Copyright 2014 Free Software Foundation, Inc. | ||
| 4 | * | ||
| 5 | * This file is part of VOLK | ||
| 6 | * | ||
| 7 | * SPDX-License-Identifier: LGPL-3.0-or-later | ||
| 8 | */ | ||
| 9 | |||
| 10 | /*! | ||
| 11 | * \page volk_32f_binary_slicer_8i | ||
| 12 | * | ||
| 13 | * \b Overview | ||
| 14 | * | ||
| 15 | * Slices input floats and and returns 1 when the input >= 0 and 0 | ||
| 16 | * when < 0. Results are converted to 8-bit chars. | ||
| 17 | * | ||
| 18 | * <b>Dispatcher Prototype</b> | ||
| 19 | * \code | ||
| 20 | * void volk_32f_binary_slicer_8i(int8_t* cVector, const float* aVector, unsigned int | ||
| 21 | num_points) | ||
| 22 | * \endcode | ||
| 23 | * | ||
| 24 | * \b Inputs | ||
| 25 | * \li aVector: The input vector of floats. | ||
| 26 | * \li num_points: The number of data points. | ||
| 27 | * | ||
| 28 | * \b Outputs | ||
| 29 | * \li cVector: The output vector of 8-bit chars. | ||
| 30 | * | ||
| 31 | * \b Example | ||
| 32 | * Generate bytes of a 7-bit barker code from floats. | ||
| 33 | * \code | ||
| 34 | int N = 7; | ||
| 35 | unsigned int alignment = volk_get_alignment(); | ||
| 36 | float* in = (float*)volk_malloc(sizeof(float)*N, alignment); | ||
| 37 | int8_t* out = (int8_t*)volk_malloc(sizeof(int8_t)*N, alignment); | ||
| 38 | |||
| 39 | in[0] = 0.9f; | ||
| 40 | in[1] = 1.1f; | ||
| 41 | in[2] = 0.4f; | ||
| 42 | in[3] = -0.7f; | ||
| 43 | in[5] = -1.2f; | ||
| 44 | in[6] = 0.2f; | ||
| 45 | in[7] = -0.8f; | ||
| 46 | |||
| 47 | volk_32f_binary_slicer_8i(out, in, N); | ||
| 48 | |||
| 49 | for(unsigned int ii = 0; ii < N; ++ii){ | ||
| 50 | printf("out(%i) = %i\n", ii, out[ii]); | ||
| 51 | } | ||
| 52 | |||
| 53 | volk_free(in); | ||
| 54 | volk_free(out); | ||
| 55 | |||
| 56 | * \endcode | ||
| 57 | */ | ||
| 58 | |||
| 59 | #ifndef INCLUDED_volk_32f_binary_slicer_8i_H | ||
| 60 | #define INCLUDED_volk_32f_binary_slicer_8i_H | ||
| 61 | |||
| 62 | |||
| 63 | #ifdef LV_HAVE_GENERIC | ||
| 64 | |||
| 65 | 2 | static inline void volk_32f_binary_slicer_8i_generic(int8_t* cVector, | |
| 66 | const float* aVector, | ||
| 67 | unsigned int num_points) | ||
| 68 | { | ||
| 69 | 2 | int8_t* cPtr = cVector; | |
| 70 | 2 | const float* aPtr = aVector; | |
| 71 | 2 | unsigned int number = 0; | |
| 72 | |||
| 73 |
2/2✓ Branch 0 taken 262142 times.
✓ Branch 1 taken 2 times.
|
262144 | for (number = 0; number < num_points; number++) { |
| 74 |
2/2✓ Branch 0 taken 131279 times.
✓ Branch 1 taken 130863 times.
|
262142 | if (*aPtr++ >= 0) { |
| 75 | 131279 | *cPtr++ = 1; | |
| 76 | } else { | ||
| 77 | 130863 | *cPtr++ = 0; | |
| 78 | } | ||
| 79 | } | ||
| 80 | 2 | } | |
| 81 | #endif /* LV_HAVE_GENERIC */ | ||
| 82 | |||
| 83 | |||
| 84 | #ifdef LV_HAVE_GENERIC | ||
| 85 | |||
| 86 | 2 | static inline void volk_32f_binary_slicer_8i_generic_branchless(int8_t* cVector, | |
| 87 | const float* aVector, | ||
| 88 | unsigned int num_points) | ||
| 89 | { | ||
| 90 | 2 | int8_t* cPtr = cVector; | |
| 91 | 2 | const float* aPtr = aVector; | |
| 92 | 2 | unsigned int number = 0; | |
| 93 | |||
| 94 |
2/2✓ Branch 0 taken 262142 times.
✓ Branch 1 taken 2 times.
|
262144 | for (number = 0; number < num_points; number++) { |
| 95 | 262142 | *cPtr++ = (*aPtr++ >= 0); | |
| 96 | } | ||
| 97 | 2 | } | |
| 98 | #endif /* LV_HAVE_GENERIC */ | ||
| 99 | |||
| 100 | |||
| 101 | #ifdef LV_HAVE_AVX2 | ||
| 102 | #include <immintrin.h> | ||
| 103 | |||
| 104 | 2 | static inline void volk_32f_binary_slicer_8i_a_avx2(int8_t* cVector, | |
| 105 | const float* aVector, | ||
| 106 | unsigned int num_points) | ||
| 107 | { | ||
| 108 | 2 | int8_t* cPtr = cVector; | |
| 109 | 2 | const float* aPtr = aVector; | |
| 110 | 2 | unsigned int number = 0; | |
| 111 | 2 | unsigned int n32points = num_points / 32; | |
| 112 | |||
| 113 | 2 | const __m256 zero_val = _mm256_set1_ps(0.0f); | |
| 114 | __m256 a0_val, a1_val, a2_val, a3_val; | ||
| 115 | __m256 res0_f, res1_f, res2_f, res3_f; | ||
| 116 | __m256i res0_i, res1_i, res2_i, res3_i; | ||
| 117 | 2 | __m256i byte_shuffle = _mm256_set_epi8(15, | |
| 118 | 14, | ||
| 119 | 13, | ||
| 120 | 12, | ||
| 121 | 7, | ||
| 122 | 6, | ||
| 123 | 5, | ||
| 124 | 4, | ||
| 125 | 11, | ||
| 126 | 10, | ||
| 127 | 9, | ||
| 128 | 8, | ||
| 129 | 3, | ||
| 130 | 2, | ||
| 131 | 1, | ||
| 132 | 0, | ||
| 133 | 15, | ||
| 134 | 14, | ||
| 135 | 13, | ||
| 136 | 12, | ||
| 137 | 7, | ||
| 138 | 6, | ||
| 139 | 5, | ||
| 140 | 4, | ||
| 141 | 11, | ||
| 142 | 10, | ||
| 143 | 9, | ||
| 144 | 8, | ||
| 145 | 3, | ||
| 146 | 2, | ||
| 147 | 1, | ||
| 148 | 0); | ||
| 149 | |||
| 150 |
2/2✓ Branch 0 taken 8190 times.
✓ Branch 1 taken 2 times.
|
8192 | for (number = 0; number < n32points; number++) { |
| 151 | 8190 | a0_val = _mm256_load_ps(aPtr); | |
| 152 | 8190 | a1_val = _mm256_load_ps(aPtr + 8); | |
| 153 | 8190 | a2_val = _mm256_load_ps(aPtr + 16); | |
| 154 | 8190 | a3_val = _mm256_load_ps(aPtr + 24); | |
| 155 | |||
| 156 | // compare >= 0; return float | ||
| 157 | 8190 | res0_f = _mm256_cmp_ps(a0_val, zero_val, _CMP_GE_OS); | |
| 158 | 8190 | res1_f = _mm256_cmp_ps(a1_val, zero_val, _CMP_GE_OS); | |
| 159 | 8190 | res2_f = _mm256_cmp_ps(a2_val, zero_val, _CMP_GE_OS); | |
| 160 | 8190 | res3_f = _mm256_cmp_ps(a3_val, zero_val, _CMP_GE_OS); | |
| 161 | |||
| 162 | // convert to 32i and >> 31 | ||
| 163 | 16380 | res0_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res0_f), 31); | |
| 164 | 16380 | res1_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res1_f), 31); | |
| 165 | 16380 | res2_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res2_f), 31); | |
| 166 | 16380 | res3_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res3_f), 31); | |
| 167 | |||
| 168 | // pack in to 16-bit results | ||
| 169 | 8190 | res0_i = _mm256_packs_epi32(res0_i, res1_i); | |
| 170 | 8190 | res2_i = _mm256_packs_epi32(res2_i, res3_i); | |
| 171 | // pack in to 8-bit results | ||
| 172 | // res0: (after packs_epi32) | ||
| 173 | // a0, a1, a2, a3, b0, b1, b2, b3, a4, a5, a6, a7, b4, b5, b6, b7 | ||
| 174 | // res2: | ||
| 175 | // c0, c1, c2, c3, d0, d1, d2, d3, c4, c5, c6, c7, d4, d5, d6, d7 | ||
| 176 | 8190 | res0_i = _mm256_packs_epi16(res0_i, res2_i); | |
| 177 | // shuffle the lanes | ||
| 178 | // res0: (after packs_epi16) | ||
| 179 | // a0, a1, a2, a3, b0, b1, b2, b3, c0, c1, c2, c3, d0, d1, d2, d3 | ||
| 180 | // a4, a5, a6, a7, b4, b5, b6, b7, c4, c5, c6, c7, d4, d5, d6, d7 | ||
| 181 | // 0, 2, 1, 3 -> 11 01 10 00 (0xd8) | ||
| 182 | 8190 | res0_i = _mm256_permute4x64_epi64(res0_i, 0xd8); | |
| 183 | |||
| 184 | // shuffle bytes within lanes | ||
| 185 | // res0: (after shuffle_epi8) | ||
| 186 | // a0, a1, a2, a3, b0, b1, b2, b3, a4, a5, a6, a7, b4, b5, b6, b7 | ||
| 187 | // c0, c1, c2, c3, d0, d1, d2, d3, c4, c5, c6, c7, d4, d5, d6, d7 | ||
| 188 | 8190 | res0_i = _mm256_shuffle_epi8(res0_i, byte_shuffle); | |
| 189 | |||
| 190 | _mm256_store_si256((__m256i*)cPtr, res0_i); | ||
| 191 | 8190 | aPtr += 32; | |
| 192 | 8190 | cPtr += 32; | |
| 193 | } | ||
| 194 | |||
| 195 |
2/2✓ Branch 0 taken 62 times.
✓ Branch 1 taken 2 times.
|
64 | for (number = n32points * 32; number < num_points; number++) { |
| 196 |
2/2✓ Branch 0 taken 32 times.
✓ Branch 1 taken 30 times.
|
62 | if (*aPtr++ >= 0) { |
| 197 | 32 | *cPtr++ = 1; | |
| 198 | } else { | ||
| 199 | 30 | *cPtr++ = 0; | |
| 200 | } | ||
| 201 | } | ||
| 202 | 2 | } | |
| 203 | #endif | ||
| 204 | |||
| 205 | #ifdef LV_HAVE_AVX2 | ||
| 206 | #include <immintrin.h> | ||
| 207 | |||
| 208 | 2 | static inline void volk_32f_binary_slicer_8i_u_avx2(int8_t* cVector, | |
| 209 | const float* aVector, | ||
| 210 | unsigned int num_points) | ||
| 211 | { | ||
| 212 | 2 | int8_t* cPtr = cVector; | |
| 213 | 2 | const float* aPtr = aVector; | |
| 214 | 2 | unsigned int number = 0; | |
| 215 | 2 | unsigned int n32points = num_points / 32; | |
| 216 | |||
| 217 | 2 | const __m256 zero_val = _mm256_set1_ps(0.0f); | |
| 218 | __m256 a0_val, a1_val, a2_val, a3_val; | ||
| 219 | __m256 res0_f, res1_f, res2_f, res3_f; | ||
| 220 | __m256i res0_i, res1_i, res2_i, res3_i; | ||
| 221 | 2 | __m256i byte_shuffle = _mm256_set_epi8(15, | |
| 222 | 14, | ||
| 223 | 13, | ||
| 224 | 12, | ||
| 225 | 7, | ||
| 226 | 6, | ||
| 227 | 5, | ||
| 228 | 4, | ||
| 229 | 11, | ||
| 230 | 10, | ||
| 231 | 9, | ||
| 232 | 8, | ||
| 233 | 3, | ||
| 234 | 2, | ||
| 235 | 1, | ||
| 236 | 0, | ||
| 237 | 15, | ||
| 238 | 14, | ||
| 239 | 13, | ||
| 240 | 12, | ||
| 241 | 7, | ||
| 242 | 6, | ||
| 243 | 5, | ||
| 244 | 4, | ||
| 245 | 11, | ||
| 246 | 10, | ||
| 247 | 9, | ||
| 248 | 8, | ||
| 249 | 3, | ||
| 250 | 2, | ||
| 251 | 1, | ||
| 252 | 0); | ||
| 253 | |||
| 254 |
2/2✓ Branch 0 taken 8190 times.
✓ Branch 1 taken 2 times.
|
8192 | for (number = 0; number < n32points; number++) { |
| 255 | 8190 | a0_val = _mm256_loadu_ps(aPtr); | |
| 256 | 8190 | a1_val = _mm256_loadu_ps(aPtr + 8); | |
| 257 | 8190 | a2_val = _mm256_loadu_ps(aPtr + 16); | |
| 258 | 8190 | a3_val = _mm256_loadu_ps(aPtr + 24); | |
| 259 | |||
| 260 | // compare >= 0; return float | ||
| 261 | 8190 | res0_f = _mm256_cmp_ps(a0_val, zero_val, _CMP_GE_OS); | |
| 262 | 8190 | res1_f = _mm256_cmp_ps(a1_val, zero_val, _CMP_GE_OS); | |
| 263 | 8190 | res2_f = _mm256_cmp_ps(a2_val, zero_val, _CMP_GE_OS); | |
| 264 | 8190 | res3_f = _mm256_cmp_ps(a3_val, zero_val, _CMP_GE_OS); | |
| 265 | |||
| 266 | // convert to 32i and >> 31 | ||
| 267 | 16380 | res0_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res0_f), 31); | |
| 268 | 16380 | res1_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res1_f), 31); | |
| 269 | 16380 | res2_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res2_f), 31); | |
| 270 | 16380 | res3_i = _mm256_srli_epi32(_mm256_cvtps_epi32(res3_f), 31); | |
| 271 | |||
| 272 | // pack in to 16-bit results | ||
| 273 | 8190 | res0_i = _mm256_packs_epi32(res0_i, res1_i); | |
| 274 | 8190 | res2_i = _mm256_packs_epi32(res2_i, res3_i); | |
| 275 | // pack in to 8-bit results | ||
| 276 | // res0: (after packs_epi32) | ||
| 277 | // a0, a1, a2, a3, b0, b1, b2, b3, a4, a5, a6, a7, b4, b5, b6, b7 | ||
| 278 | // res2: | ||
| 279 | // c0, c1, c2, c3, d0, d1, d2, d3, c4, c5, c6, c7, d4, d5, d6, d7 | ||
| 280 | 8190 | res0_i = _mm256_packs_epi16(res0_i, res2_i); | |
| 281 | // shuffle the lanes | ||
| 282 | // res0: (after packs_epi16) | ||
| 283 | // a0, a1, a2, a3, b0, b1, b2, b3, c0, c1, c2, c3, d0, d1, d2, d3 | ||
| 284 | // a4, a5, a6, a7, b4, b5, b6, b7, c4, c5, c6, c7, d4, d5, d6, d7 | ||
| 285 | // 0, 2, 1, 3 -> 11 01 10 00 (0xd8) | ||
| 286 | 8190 | res0_i = _mm256_permute4x64_epi64(res0_i, 0xd8); | |
| 287 | |||
| 288 | // shuffle bytes within lanes | ||
| 289 | // res0: (after shuffle_epi8) | ||
| 290 | // a0, a1, a2, a3, b0, b1, b2, b3, a4, a5, a6, a7, b4, b5, b6, b7 | ||
| 291 | // c0, c1, c2, c3, d0, d1, d2, d3, c4, c5, c6, c7, d4, d5, d6, d7 | ||
| 292 | 8190 | res0_i = _mm256_shuffle_epi8(res0_i, byte_shuffle); | |
| 293 | |||
| 294 | _mm256_storeu_si256((__m256i*)cPtr, res0_i); | ||
| 295 | 8190 | aPtr += 32; | |
| 296 | 8190 | cPtr += 32; | |
| 297 | } | ||
| 298 | |||
| 299 |
2/2✓ Branch 0 taken 62 times.
✓ Branch 1 taken 2 times.
|
64 | for (number = n32points * 32; number < num_points; number++) { |
| 300 |
2/2✓ Branch 0 taken 32 times.
✓ Branch 1 taken 30 times.
|
62 | if (*aPtr++ >= 0) { |
| 301 | 32 | *cPtr++ = 1; | |
| 302 | } else { | ||
| 303 | 30 | *cPtr++ = 0; | |
| 304 | } | ||
| 305 | } | ||
| 306 | 2 | } | |
| 307 | #endif | ||
| 308 | |||
| 309 | |||
| 310 | #ifdef LV_HAVE_SSE2 | ||
| 311 | |||
| 312 | #include <emmintrin.h> | ||
| 313 | |||
| 314 | 2 | static inline void volk_32f_binary_slicer_8i_a_sse2(int8_t* cVector, | |
| 315 | const float* aVector, | ||
| 316 | unsigned int num_points) | ||
| 317 | { | ||
| 318 | 2 | int8_t* cPtr = cVector; | |
| 319 | 2 | const float* aPtr = aVector; | |
| 320 | 2 | unsigned int number = 0; | |
| 321 | |||
| 322 | 2 | unsigned int n16points = num_points / 16; | |
| 323 | __m128 a0_val, a1_val, a2_val, a3_val; | ||
| 324 | __m128 res0_f, res1_f, res2_f, res3_f; | ||
| 325 | __m128i res0_i, res1_i, res2_i, res3_i; | ||
| 326 | __m128 zero_val; | ||
| 327 | 2 | zero_val = _mm_set1_ps(0.0f); | |
| 328 | |||
| 329 |
2/2✓ Branch 0 taken 16382 times.
✓ Branch 1 taken 2 times.
|
16384 | for (number = 0; number < n16points; number++) { |
| 330 | 16382 | a0_val = _mm_load_ps(aPtr); | |
| 331 | 16382 | a1_val = _mm_load_ps(aPtr + 4); | |
| 332 | 16382 | a2_val = _mm_load_ps(aPtr + 8); | |
| 333 | 32764 | a3_val = _mm_load_ps(aPtr + 12); | |
| 334 | |||
| 335 | // compare >= 0; return float | ||
| 336 | 16382 | res0_f = _mm_cmpge_ps(a0_val, zero_val); | |
| 337 | 16382 | res1_f = _mm_cmpge_ps(a1_val, zero_val); | |
| 338 | 16382 | res2_f = _mm_cmpge_ps(a2_val, zero_val); | |
| 339 | 16382 | res3_f = _mm_cmpge_ps(a3_val, zero_val); | |
| 340 | |||
| 341 | // convert to 32i and >> 31 | ||
| 342 | 32764 | res0_i = _mm_srli_epi32(_mm_cvtps_epi32(res0_f), 31); | |
| 343 | 32764 | res1_i = _mm_srli_epi32(_mm_cvtps_epi32(res1_f), 31); | |
| 344 | 32764 | res2_i = _mm_srli_epi32(_mm_cvtps_epi32(res2_f), 31); | |
| 345 | 32764 | res3_i = _mm_srli_epi32(_mm_cvtps_epi32(res3_f), 31); | |
| 346 | |||
| 347 | // pack into 16-bit results | ||
| 348 | 16382 | res0_i = _mm_packs_epi32(res0_i, res1_i); | |
| 349 | 16382 | res2_i = _mm_packs_epi32(res2_i, res3_i); | |
| 350 | |||
| 351 | // pack into 8-bit results | ||
| 352 | 16382 | res0_i = _mm_packs_epi16(res0_i, res2_i); | |
| 353 | |||
| 354 | _mm_store_si128((__m128i*)cPtr, res0_i); | ||
| 355 | |||
| 356 | 16382 | cPtr += 16; | |
| 357 | 16382 | aPtr += 16; | |
| 358 | } | ||
| 359 | |||
| 360 |
2/2✓ Branch 0 taken 30 times.
✓ Branch 1 taken 2 times.
|
32 | for (number = n16points * 16; number < num_points; number++) { |
| 361 |
2/2✓ Branch 0 taken 19 times.
✓ Branch 1 taken 11 times.
|
30 | if (*aPtr++ >= 0) { |
| 362 | 19 | *cPtr++ = 1; | |
| 363 | } else { | ||
| 364 | 11 | *cPtr++ = 0; | |
| 365 | } | ||
| 366 | } | ||
| 367 | 2 | } | |
| 368 | #endif /* LV_HAVE_SSE2 */ | ||
| 369 | |||
| 370 | |||
| 371 | #ifdef LV_HAVE_SSE2 | ||
| 372 | #include <emmintrin.h> | ||
| 373 | |||
| 374 | 2 | static inline void volk_32f_binary_slicer_8i_u_sse2(int8_t* cVector, | |
| 375 | const float* aVector, | ||
| 376 | unsigned int num_points) | ||
| 377 | { | ||
| 378 | 2 | int8_t* cPtr = cVector; | |
| 379 | 2 | const float* aPtr = aVector; | |
| 380 | 2 | unsigned int number = 0; | |
| 381 | |||
| 382 | 2 | unsigned int n16points = num_points / 16; | |
| 383 | __m128 a0_val, a1_val, a2_val, a3_val; | ||
| 384 | __m128 res0_f, res1_f, res2_f, res3_f; | ||
| 385 | __m128i res0_i, res1_i, res2_i, res3_i; | ||
| 386 | __m128 zero_val; | ||
| 387 | 2 | zero_val = _mm_set1_ps(0.0f); | |
| 388 | |||
| 389 |
2/2✓ Branch 0 taken 16382 times.
✓ Branch 1 taken 2 times.
|
16384 | for (number = 0; number < n16points; number++) { |
| 390 | 16382 | a0_val = _mm_loadu_ps(aPtr); | |
| 391 | 16382 | a1_val = _mm_loadu_ps(aPtr + 4); | |
| 392 | 16382 | a2_val = _mm_loadu_ps(aPtr + 8); | |
| 393 | 32764 | a3_val = _mm_loadu_ps(aPtr + 12); | |
| 394 | |||
| 395 | // compare >= 0; return float | ||
| 396 | 16382 | res0_f = _mm_cmpge_ps(a0_val, zero_val); | |
| 397 | 16382 | res1_f = _mm_cmpge_ps(a1_val, zero_val); | |
| 398 | 16382 | res2_f = _mm_cmpge_ps(a2_val, zero_val); | |
| 399 | 16382 | res3_f = _mm_cmpge_ps(a3_val, zero_val); | |
| 400 | |||
| 401 | // convert to 32i and >> 31 | ||
| 402 | 32764 | res0_i = _mm_srli_epi32(_mm_cvtps_epi32(res0_f), 31); | |
| 403 | 32764 | res1_i = _mm_srli_epi32(_mm_cvtps_epi32(res1_f), 31); | |
| 404 | 32764 | res2_i = _mm_srli_epi32(_mm_cvtps_epi32(res2_f), 31); | |
| 405 | 32764 | res3_i = _mm_srli_epi32(_mm_cvtps_epi32(res3_f), 31); | |
| 406 | |||
| 407 | // pack into 16-bit results | ||
| 408 | 16382 | res0_i = _mm_packs_epi32(res0_i, res1_i); | |
| 409 | 16382 | res2_i = _mm_packs_epi32(res2_i, res3_i); | |
| 410 | |||
| 411 | // pack into 8-bit results | ||
| 412 | 16382 | res0_i = _mm_packs_epi16(res0_i, res2_i); | |
| 413 | |||
| 414 | _mm_storeu_si128((__m128i*)cPtr, res0_i); | ||
| 415 | |||
| 416 | 16382 | cPtr += 16; | |
| 417 | 16382 | aPtr += 16; | |
| 418 | } | ||
| 419 | |||
| 420 |
2/2✓ Branch 0 taken 30 times.
✓ Branch 1 taken 2 times.
|
32 | for (number = n16points * 16; number < num_points; number++) { |
| 421 |
2/2✓ Branch 0 taken 19 times.
✓ Branch 1 taken 11 times.
|
30 | if (*aPtr++ >= 0) { |
| 422 | 19 | *cPtr++ = 1; | |
| 423 | } else { | ||
| 424 | 11 | *cPtr++ = 0; | |
| 425 | } | ||
| 426 | } | ||
| 427 | 2 | } | |
| 428 | #endif /* LV_HAVE_SSE2 */ | ||
| 429 | |||
| 430 | |||
| 431 | #ifdef LV_HAVE_NEON | ||
| 432 | #include <arm_neon.h> | ||
| 433 | |||
| 434 | static inline void volk_32f_binary_slicer_8i_neon(int8_t* cVector, | ||
| 435 | const float* aVector, | ||
| 436 | unsigned int num_points) | ||
| 437 | { | ||
| 438 | int8_t* cPtr = cVector; | ||
| 439 | const float* aPtr = aVector; | ||
| 440 | unsigned int number = 0; | ||
| 441 | unsigned int n16points = num_points / 16; | ||
| 442 | |||
| 443 | float32x4x2_t input_val0, input_val1; | ||
| 444 | float32x4_t zero_val; | ||
| 445 | uint32x4x2_t res0_u32, res1_u32; | ||
| 446 | uint16x4x2_t res0_u16x4, res1_u16x4; | ||
| 447 | uint16x8x2_t res_u16x8; | ||
| 448 | uint8x8x2_t res_u8; | ||
| 449 | uint8x8_t one; | ||
| 450 | |||
| 451 | zero_val = vdupq_n_f32(0.0); | ||
| 452 | one = vdup_n_u8(0x01); | ||
| 453 | |||
| 454 | // TODO: this is a good candidate for asm because the vcombines | ||
| 455 | // can be eliminated simply by picking dst registers that are | ||
| 456 | // adjacent. | ||
| 457 | for (number = 0; number < n16points; number++) { | ||
| 458 | input_val0 = vld2q_f32(aPtr); | ||
| 459 | input_val1 = vld2q_f32(aPtr + 8); | ||
| 460 | |||
| 461 | // test against 0; return uint32 | ||
| 462 | res0_u32.val[0] = vcgeq_f32(input_val0.val[0], zero_val); | ||
| 463 | res0_u32.val[1] = vcgeq_f32(input_val0.val[1], zero_val); | ||
| 464 | res1_u32.val[0] = vcgeq_f32(input_val1.val[0], zero_val); | ||
| 465 | res1_u32.val[1] = vcgeq_f32(input_val1.val[1], zero_val); | ||
| 466 | |||
| 467 | // narrow uint32 -> uint16 followed by combine to 8-element vectors | ||
| 468 | res0_u16x4.val[0] = vmovn_u32(res0_u32.val[0]); | ||
| 469 | res0_u16x4.val[1] = vmovn_u32(res0_u32.val[1]); | ||
| 470 | res1_u16x4.val[0] = vmovn_u32(res1_u32.val[0]); | ||
| 471 | res1_u16x4.val[1] = vmovn_u32(res1_u32.val[1]); | ||
| 472 | |||
| 473 | res_u16x8.val[0] = vcombine_u16(res0_u16x4.val[0], res1_u16x4.val[0]); | ||
| 474 | res_u16x8.val[1] = vcombine_u16(res0_u16x4.val[1], res1_u16x4.val[1]); | ||
| 475 | |||
| 476 | // narrow uint16x8 -> uint8x8 | ||
| 477 | res_u8.val[0] = vmovn_u16(res_u16x8.val[0]); | ||
| 478 | res_u8.val[1] = vmovn_u16(res_u16x8.val[1]); | ||
| 479 | // we *could* load twice as much data and do another vcombine here | ||
| 480 | // to get a uint8x16x2 vector, still only do 2 vandqs and a single store | ||
| 481 | // but that turns out to be ~16% slower than this version on zc702 | ||
| 482 | // it's possible register contention in GCC scheduler slows it down | ||
| 483 | // and a hand-written asm with quad-word u8 registers is much faster. | ||
| 484 | |||
| 485 | res_u8.val[0] = vand_u8(one, res_u8.val[0]); | ||
| 486 | res_u8.val[1] = vand_u8(one, res_u8.val[1]); | ||
| 487 | |||
| 488 | vst2_u8((unsigned char*)cPtr, res_u8); | ||
| 489 | cPtr += 16; | ||
| 490 | aPtr += 16; | ||
| 491 | } | ||
| 492 | |||
| 493 | for (number = n16points * 16; number < num_points; number++) { | ||
| 494 | if (*aPtr++ >= 0) { | ||
| 495 | *cPtr++ = 1; | ||
| 496 | } else { | ||
| 497 | *cPtr++ = 0; | ||
| 498 | } | ||
| 499 | } | ||
| 500 | } | ||
| 501 | #endif /* LV_HAVE_NEON */ | ||
| 502 | |||
| 503 | |||
| 504 | #endif /* INCLUDED_volk_32f_binary_slicer_8i_H */ | ||
| 505 |