Vector Optimized Library of Kernels  2.0
Architecture-tuned implementations of math kernels
volk_16ic_x2_dot_prod_16ic.h
Go to the documentation of this file.
1 /* -*- c++ -*- */
2 /*
3  * Copyright 2016 Free Software Foundation, Inc.
4  *
5  * This file is part of GNU Radio
6  *
7  * GNU Radio is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 3, or (at your option)
10  * any later version.
11  *
12  * GNU Radio is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with GNU Radio; see the file COPYING. If not, write to
19  * the Free Software Foundation, Inc., 51 Franklin Street,
20  * Boston, MA 02110-1301, USA.
21  */
22 
46 #ifndef INCLUDED_volk_16ic_x2_dot_prod_16ic_H
47 #define INCLUDED_volk_16ic_x2_dot_prod_16ic_H
48 
49 #include <volk/volk_common.h>
50 #include <volk/volk_complex.h>
52 
53 
54 #ifdef LV_HAVE_GENERIC
55 
56 static inline void volk_16ic_x2_dot_prod_16ic_generic(lv_16sc_t* result, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
57 {
58  result[0] = lv_cmake((int16_t)0, (int16_t)0);
59  unsigned int n;
60  for (n = 0; n < num_points; n++)
61  {
62  lv_16sc_t tmp = in_a[n] * in_b[n];
63  result[0] = lv_cmake(sat_adds16i(lv_creal(result[0]), lv_creal(tmp)), sat_adds16i(lv_cimag(result[0]), lv_cimag(tmp) ));
64  }
65 }
66 
67 #endif /*LV_HAVE_GENERIC*/
68 
69 
70 #ifdef LV_HAVE_SSE2
71 #include <emmintrin.h>
72 
73 static inline void volk_16ic_x2_dot_prod_16ic_a_sse2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
74 {
75  lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0);
76 
77  const unsigned int sse_iters = num_points / 4;
78  unsigned int number;
79 
80  const lv_16sc_t* _in_a = in_a;
81  const lv_16sc_t* _in_b = in_b;
82  lv_16sc_t* _out = out;
83 
84  if (sse_iters > 0)
85  {
86  __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, realcacc, imagcacc;
87  __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
88 
89  realcacc = _mm_setzero_si128();
90  imagcacc = _mm_setzero_si128();
91 
92  mask_imag = _mm_set_epi8(0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0);
93  mask_real = _mm_set_epi8(0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF);
94 
95  for(number = 0; number < sse_iters; number++)
96  {
97  // a[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
98  a = _mm_load_si128((__m128i*)_in_a); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
99  __VOLK_PREFETCH(_in_a + 8);
100  b = _mm_load_si128((__m128i*)_in_b);
101  __VOLK_PREFETCH(_in_b + 8);
102  c = _mm_mullo_epi16(a, b); // a3.i*b3.i, a3.r*b3.r, ....
103 
104  c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
105  real = _mm_subs_epi16(c, c_sr);
106 
107  b_sl = _mm_slli_si128(b, 2); // b3.r, b2.i ....
108  a_sl = _mm_slli_si128(a, 2); // a3.r, a2.i ....
109 
110  imag1 = _mm_mullo_epi16(a, b_sl); // a3.i*b3.r, ....
111  imag2 = _mm_mullo_epi16(b, a_sl); // b3.i*a3.r, ....
112 
113  imag = _mm_adds_epi16(imag1, imag2); //with saturation arithmetic!
114 
115  realcacc = _mm_adds_epi16(realcacc, real);
116  imagcacc = _mm_adds_epi16(imagcacc, imag);
117 
118  _in_a += 4;
119  _in_b += 4;
120  }
121 
122  realcacc = _mm_and_si128(realcacc, mask_real);
123  imagcacc = _mm_and_si128(imagcacc, mask_imag);
124 
125  a = _mm_or_si128(realcacc, imagcacc);
126 
127  _mm_store_si128((__m128i*)dotProductVector, a); // Store the results back into the dot product vector
128 
129  for (number = 0; number < 4; ++number)
130  {
131  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[number])), sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[number])));
132  }
133  }
134 
135  for (number = 0; number < (num_points % 4); ++number)
136  {
137  lv_16sc_t tmp = (*_in_a++) * (*_in_b++);
138  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(tmp)), sat_adds16i(lv_cimag(dotProduct), lv_cimag(tmp)));
139  }
140 
141  *_out = dotProduct;
142 }
143 
144 #endif /* LV_HAVE_SSE2 */
145 
146 
147 #ifdef LV_HAVE_SSE2
148 #include <emmintrin.h>
149 
150 static inline void volk_16ic_x2_dot_prod_16ic_u_sse2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
151 {
152  lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0);
153 
154  const unsigned int sse_iters = num_points / 4;
155 
156  const lv_16sc_t* _in_a = in_a;
157  const lv_16sc_t* _in_b = in_b;
158  lv_16sc_t* _out = out;
159  unsigned int number;
160 
161  if (sse_iters > 0)
162  {
163  __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, realcacc, imagcacc, result;
164  __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
165 
166  realcacc = _mm_setzero_si128();
167  imagcacc = _mm_setzero_si128();
168 
169  mask_imag = _mm_set_epi8(0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0);
170  mask_real = _mm_set_epi8(0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF);
171 
172  for(number = 0; number < sse_iters; number++)
173  {
174  // a[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
175  a = _mm_loadu_si128((__m128i*)_in_a); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
176  __VOLK_PREFETCH(_in_a + 8);
177  b = _mm_loadu_si128((__m128i*)_in_b);
178  __VOLK_PREFETCH(_in_b + 8);
179  c = _mm_mullo_epi16(a, b); // a3.i*b3.i, a3.r*b3.r, ....
180 
181  c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
182  real = _mm_subs_epi16(c, c_sr);
183 
184  b_sl = _mm_slli_si128(b, 2); // b3.r, b2.i ....
185  a_sl = _mm_slli_si128(a, 2); // a3.r, a2.i ....
186 
187  imag1 = _mm_mullo_epi16(a, b_sl); // a3.i*b3.r, ....
188  imag2 = _mm_mullo_epi16(b, a_sl); // b3.i*a3.r, ....
189 
190  imag = _mm_adds_epi16(imag1, imag2); //with saturation arithmetic!
191 
192  realcacc = _mm_adds_epi16(realcacc, real);
193  imagcacc = _mm_adds_epi16(imagcacc, imag);
194 
195  _in_a += 4;
196  _in_b += 4;
197  }
198 
199  realcacc = _mm_and_si128(realcacc, mask_real);
200  imagcacc = _mm_and_si128(imagcacc, mask_imag);
201 
202  result = _mm_or_si128(realcacc, imagcacc);
203 
204  _mm_storeu_si128((__m128i*)dotProductVector, result); // Store the results back into the dot product vector
205 
206  for (number = 0; number < 4; ++number)
207  {
208  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[number])), sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[number])));
209  }
210  }
211 
212  for (number = 0; number < (num_points % 4); ++number)
213  {
214  lv_16sc_t tmp = (*_in_a++) * (*_in_b++);
215  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(tmp)), sat_adds16i(lv_cimag(dotProduct), lv_cimag(tmp)));
216  }
217 
218  *_out = dotProduct;
219 }
220 #endif /* LV_HAVE_SSE2 */
221 
222 
223 #ifdef LV_HAVE_AVX2
224 #include <immintrin.h>
225 
226 static inline void volk_16ic_x2_dot_prod_16ic_u_axv2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
227 {
228  lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0);
229 
230  const unsigned int avx_iters = num_points / 8;
231 
232  const lv_16sc_t* _in_a = in_a;
233  const lv_16sc_t* _in_b = in_b;
234  lv_16sc_t* _out = out;
235  unsigned int number;
236 
237  if (avx_iters > 0)
238  {
239  __m256i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, realcacc, imagcacc, result;
240  __VOLK_ATTR_ALIGNED(32) lv_16sc_t dotProductVector[8];
241 
242  realcacc = _mm256_setzero_si256();
243  imagcacc = _mm256_setzero_si256();
244 
245  mask_imag = _mm256_set_epi8(0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0);
246  mask_real = _mm256_set_epi8(0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF);
247 
248  for(number = 0; number < avx_iters; number++)
249  {
250  a = _mm256_loadu_si256((__m256i*)_in_a);
251  __VOLK_PREFETCH(_in_a + 16);
252  b = _mm256_loadu_si256((__m256i*)_in_b);
253  __VOLK_PREFETCH(_in_b + 16);
254  c = _mm256_mullo_epi16(a, b);
255 
256  c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
257  real = _mm256_subs_epi16(c, c_sr);
258 
259  b_sl = _mm256_slli_si256(b, 2);
260  a_sl = _mm256_slli_si256(a, 2);
261 
262  imag1 = _mm256_mullo_epi16(a, b_sl);
263  imag2 = _mm256_mullo_epi16(b, a_sl);
264 
265  imag = _mm256_adds_epi16(imag1, imag2); //with saturation arithmetic!
266 
267  realcacc = _mm256_adds_epi16(realcacc, real);
268  imagcacc = _mm256_adds_epi16(imagcacc, imag);
269 
270  _in_a += 8;
271  _in_b += 8;
272  }
273 
274  realcacc = _mm256_and_si256(realcacc, mask_real);
275  imagcacc = _mm256_and_si256(imagcacc, mask_imag);
276 
277  result = _mm256_or_si256(realcacc, imagcacc);
278 
279  _mm256_storeu_si256((__m256i*)dotProductVector, result); // Store the results back into the dot product vector
280  _mm256_zeroupper();
281 
282  for (number = 0; number < 8; ++number)
283  {
284  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[number])), sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[number])));
285  }
286  }
287 
288  for (number = 0; number < (num_points % 8); ++number)
289  {
290  lv_16sc_t tmp = (*_in_a++) * (*_in_b++);
291  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(tmp)), sat_adds16i(lv_cimag(dotProduct), lv_cimag(tmp)));
292  }
293 
294  *_out = dotProduct;
295 }
296 #endif /* LV_HAVE_AVX2 */
297 
298 
299 #ifdef LV_HAVE_AVX2
300 #include <immintrin.h>
301 
302 static inline void volk_16ic_x2_dot_prod_16ic_a_axv2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
303 {
304  lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0);
305 
306  const unsigned int avx_iters = num_points / 8;
307 
308  const lv_16sc_t* _in_a = in_a;
309  const lv_16sc_t* _in_b = in_b;
310  lv_16sc_t* _out = out;
311  unsigned int number;
312 
313  if (avx_iters > 0)
314  {
315  __m256i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, realcacc, imagcacc, result;
316  __VOLK_ATTR_ALIGNED(32) lv_16sc_t dotProductVector[8];
317 
318  realcacc = _mm256_setzero_si256();
319  imagcacc = _mm256_setzero_si256();
320 
321  mask_imag = _mm256_set_epi8(0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0);
322  mask_real = _mm256_set_epi8(0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF, 0, 0, 0xFF, 0xFF);
323 
324  for(number = 0; number < avx_iters; number++)
325  {
326  a = _mm256_load_si256((__m256i*)_in_a);
327  __VOLK_PREFETCH(_in_a + 16);
328  b = _mm256_load_si256((__m256i*)_in_b);
329  __VOLK_PREFETCH(_in_b + 16);
330  c = _mm256_mullo_epi16(a, b);
331 
332  c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
333  real = _mm256_subs_epi16(c, c_sr);
334 
335  b_sl = _mm256_slli_si256(b, 2);
336  a_sl = _mm256_slli_si256(a, 2);
337 
338  imag1 = _mm256_mullo_epi16(a, b_sl);
339  imag2 = _mm256_mullo_epi16(b, a_sl);
340 
341  imag = _mm256_adds_epi16(imag1, imag2); //with saturation arithmetic!
342 
343  realcacc = _mm256_adds_epi16(realcacc, real);
344  imagcacc = _mm256_adds_epi16(imagcacc, imag);
345 
346  _in_a += 8;
347  _in_b += 8;
348  }
349 
350  realcacc = _mm256_and_si256(realcacc, mask_real);
351  imagcacc = _mm256_and_si256(imagcacc, mask_imag);
352 
353  result = _mm256_or_si256(realcacc, imagcacc);
354 
355  _mm256_store_si256((__m256i*)dotProductVector, result); // Store the results back into the dot product vector
356  _mm256_zeroupper();
357 
358  for (number = 0; number < 8; ++number)
359  {
360  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[number])), sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[number])));
361  }
362  }
363 
364  for (number = 0; number < (num_points % 8); ++number)
365  {
366  lv_16sc_t tmp = (*_in_a++) * (*_in_b++);
367  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(tmp)), sat_adds16i(lv_cimag(dotProduct), lv_cimag(tmp)));
368  }
369 
370  *_out = dotProduct;
371 }
372 #endif /* LV_HAVE_AVX2 */
373 
374 
375 #ifdef LV_HAVE_NEON
376 #include <arm_neon.h>
377 
378 static inline void volk_16ic_x2_dot_prod_16ic_neon(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
379 {
380  unsigned int quarter_points = num_points / 4;
381  unsigned int number;
382 
383  lv_16sc_t* a_ptr = (lv_16sc_t*) in_a;
384  lv_16sc_t* b_ptr = (lv_16sc_t*) in_b;
385  *out = lv_cmake((int16_t)0, (int16_t)0);
386 
387  if (quarter_points > 0)
388  {
389  // for 2-lane vectors, 1st lane holds the real part,
390  // 2nd lane holds the imaginary part
391  int16x4x2_t a_val, b_val, c_val, accumulator;
392  int16x4x2_t tmp_real, tmp_imag;
393  __VOLK_ATTR_ALIGNED(16) lv_16sc_t accum_result[4];
394  accumulator.val[0] = vdup_n_s16(0);
395  accumulator.val[1] = vdup_n_s16(0);
396  lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0);
397 
398  for(number = 0; number < quarter_points; ++number)
399  {
400  a_val = vld2_s16((int16_t*)a_ptr); // a0r|a1r|a2r|a3r || a0i|a1i|a2i|a3i
401  b_val = vld2_s16((int16_t*)b_ptr); // b0r|b1r|b2r|b3r || b0i|b1i|b2i|b3i
402  __VOLK_PREFETCH(a_ptr + 8);
403  __VOLK_PREFETCH(b_ptr + 8);
404 
405  // multiply the real*real and imag*imag to get real result
406  // a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r
407  tmp_real.val[0] = vmul_s16(a_val.val[0], b_val.val[0]);
408  // a0i*b0i|a1i*b1i|a2i*b2i|a3i*b3i
409  tmp_real.val[1] = vmul_s16(a_val.val[1], b_val.val[1]);
410 
411  // Multiply cross terms to get the imaginary result
412  // a0r*b0i|a1r*b1i|a2r*b2i|a3r*b3i
413  tmp_imag.val[0] = vmul_s16(a_val.val[0], b_val.val[1]);
414  // a0i*b0r|a1i*b1r|a2i*b2r|a3i*b3r
415  tmp_imag.val[1] = vmul_s16(a_val.val[1], b_val.val[0]);
416 
417  c_val.val[0] = vqsub_s16(tmp_real.val[0], tmp_real.val[1]);
418  c_val.val[1] = vqadd_s16(tmp_imag.val[0], tmp_imag.val[1]);
419 
420  accumulator.val[0] = vqadd_s16(accumulator.val[0], c_val.val[0]);
421  accumulator.val[1] = vqadd_s16(accumulator.val[1], c_val.val[1]);
422 
423  a_ptr += 4;
424  b_ptr += 4;
425  }
426 
427  vst2_s16((int16_t*)accum_result, accumulator);
428  for (number = 0; number < 4; ++number)
429  {
430  dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(accum_result[number])), sat_adds16i(lv_cimag(dotProduct), lv_cimag(accum_result[number])));
431  }
432 
433  *out = dotProduct;
434  }
435 
436  // tail case
437  for(number = quarter_points * 4; number < num_points; ++number)
438  {
439  *out += (*a_ptr++) * (*b_ptr++);
440  }
441 }
442 
443 #endif /* LV_HAVE_NEON */
444 
445 
446 #ifdef LV_HAVE_NEON
447 #include <arm_neon.h>
448 
449 static inline void volk_16ic_x2_dot_prod_16ic_neon_vma(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
450 {
451  unsigned int quarter_points = num_points / 4;
452  unsigned int number;
453 
454  lv_16sc_t* a_ptr = (lv_16sc_t*) in_a;
455  lv_16sc_t* b_ptr = (lv_16sc_t*) in_b;
456  // for 2-lane vectors, 1st lane holds the real part,
457  // 2nd lane holds the imaginary part
458  int16x4x2_t a_val, b_val, accumulator;
459  int16x4x2_t tmp;
460  __VOLK_ATTR_ALIGNED(16) lv_16sc_t accum_result[4];
461  accumulator.val[0] = vdup_n_s16(0);
462  accumulator.val[1] = vdup_n_s16(0);
463 
464  for(number = 0; number < quarter_points; ++number)
465  {
466  a_val = vld2_s16((int16_t*)a_ptr); // a0r|a1r|a2r|a3r || a0i|a1i|a2i|a3i
467  b_val = vld2_s16((int16_t*)b_ptr); // b0r|b1r|b2r|b3r || b0i|b1i|b2i|b3i
468  __VOLK_PREFETCH(a_ptr + 8);
469  __VOLK_PREFETCH(b_ptr + 8);
470 
471  tmp.val[0] = vmul_s16(a_val.val[0], b_val.val[0]);
472  tmp.val[1] = vmul_s16(a_val.val[1], b_val.val[0]);
473 
474  // use multiply accumulate/subtract to get result
475  tmp.val[0] = vmls_s16(tmp.val[0], a_val.val[1], b_val.val[1]);
476  tmp.val[1] = vmla_s16(tmp.val[1], a_val.val[0], b_val.val[1]);
477 
478  accumulator.val[0] = vqadd_s16(accumulator.val[0], tmp.val[0]);
479  accumulator.val[1] = vqadd_s16(accumulator.val[1], tmp.val[1]);
480 
481  a_ptr += 4;
482  b_ptr += 4;
483  }
484 
485  vst2_s16((int16_t*)accum_result, accumulator);
486  *out = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
487 
488  // tail case
489  for(number = quarter_points * 4; number < num_points; ++number)
490  {
491  *out += (*a_ptr++) * (*b_ptr++);
492  }
493 }
494 
495 #endif /* LV_HAVE_NEON */
496 
497 
498 #ifdef LV_HAVE_NEON
499 #include <arm_neon.h>
500 
501 static inline void volk_16ic_x2_dot_prod_16ic_neon_optvma(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
502 {
503  unsigned int quarter_points = num_points / 4;
504  unsigned int number;
505 
506  lv_16sc_t* a_ptr = (lv_16sc_t*) in_a;
507  lv_16sc_t* b_ptr = (lv_16sc_t*) in_b;
508  // for 2-lane vectors, 1st lane holds the real part,
509  // 2nd lane holds the imaginary part
510  int16x4x2_t a_val, b_val, accumulator1, accumulator2;
511 
512  __VOLK_ATTR_ALIGNED(16) lv_16sc_t accum_result[4];
513  accumulator1.val[0] = vdup_n_s16(0);
514  accumulator1.val[1] = vdup_n_s16(0);
515  accumulator2.val[0] = vdup_n_s16(0);
516  accumulator2.val[1] = vdup_n_s16(0);
517 
518  for(number = 0; number < quarter_points; ++number)
519  {
520  a_val = vld2_s16((int16_t*)a_ptr); // a0r|a1r|a2r|a3r || a0i|a1i|a2i|a3i
521  b_val = vld2_s16((int16_t*)b_ptr); // b0r|b1r|b2r|b3r || b0i|b1i|b2i|b3i
522  __VOLK_PREFETCH(a_ptr + 8);
523  __VOLK_PREFETCH(b_ptr + 8);
524 
525  // use 2 accumulators to remove inter-instruction data dependencies
526  accumulator1.val[0] = vmla_s16(accumulator1.val[0], a_val.val[0], b_val.val[0]);
527  accumulator2.val[0] = vmls_s16(accumulator2.val[0], a_val.val[1], b_val.val[1]);
528  accumulator1.val[1] = vmla_s16(accumulator1.val[1], a_val.val[0], b_val.val[1]);
529  accumulator2.val[1] = vmla_s16(accumulator2.val[1], a_val.val[1], b_val.val[0]);
530 
531  a_ptr += 4;
532  b_ptr += 4;
533  }
534 
535  accumulator1.val[0] = vqadd_s16(accumulator1.val[0], accumulator2.val[0]);
536  accumulator1.val[1] = vqadd_s16(accumulator1.val[1], accumulator2.val[1]);
537 
538  vst2_s16((int16_t*)accum_result, accumulator1);
539  *out = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
540 
541  // tail case
542  for(number = quarter_points * 4; number < num_points; ++number)
543  {
544  *out += (*a_ptr++) * (*b_ptr++);
545  }
546 }
547 
548 #endif /* LV_HAVE_NEON */
549 
550 #endif /*INCLUDED_volk_16ic_x2_dot_prod_16ic_H*/
static int16_t sat_adds16i(int16_t x, int16_t y)
Definition: saturation_arithmetic.h:29
short complex lv_16sc_t
Definition: volk_complex.h:58
static void volk_16ic_x2_dot_prod_16ic_u_sse2(lv_16sc_t *out, const lv_16sc_t *in_a, const lv_16sc_t *in_b, unsigned int num_points)
Definition: volk_16ic_x2_dot_prod_16ic.h:150
#define lv_cmake(r, i)
Definition: volk_complex.h:64
static void volk_16ic_x2_dot_prod_16ic_neon(lv_16sc_t *out, const lv_16sc_t *in_a, const lv_16sc_t *in_b, unsigned int num_points)
Definition: volk_16ic_x2_dot_prod_16ic.h:378
#define __VOLK_PREFETCH(addr)
Definition: volk_common.h:39
static void volk_16ic_x2_dot_prod_16ic_generic(lv_16sc_t *result, const lv_16sc_t *in_a, const lv_16sc_t *in_b, unsigned int num_points)
Definition: volk_16ic_x2_dot_prod_16ic.h:56
#define __VOLK_ATTR_ALIGNED(x)
Definition: volk_common.h:33
static void volk_16ic_x2_dot_prod_16ic_a_sse2(lv_16sc_t *out, const lv_16sc_t *in_a, const lv_16sc_t *in_b, unsigned int num_points)
Definition: volk_16ic_x2_dot_prod_16ic.h:73
static void volk_16ic_x2_dot_prod_16ic_neon_optvma(lv_16sc_t *out, const lv_16sc_t *in_a, const lv_16sc_t *in_b, unsigned int num_points)
Definition: volk_16ic_x2_dot_prod_16ic.h:501
static void volk_16ic_x2_dot_prod_16ic_neon_vma(lv_16sc_t *out, const lv_16sc_t *in_a, const lv_16sc_t *in_b, unsigned int num_points)
Definition: volk_16ic_x2_dot_prod_16ic.h:449
#define lv_creal(x)
Definition: volk_complex.h:83
#define lv_cimag(x)
Definition: volk_complex.h:85