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
23 /*!
24 * \page volk_16ic_x2_dot_prod_16ic
25 *
26 * \b Overview
27 *
28 * Multiplies two input complex vectors (16-bit integer each component) and accumulates them,
29 * storing the result. Results are saturated so never go beyond the limits of the data type.
30 *
31 * <b>Dispatcher Prototype</b>
32 * \code
33 * void volk_16ic_x2_dot_prod_16ic(lv_16sc_t* result, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points);
34 * \endcode
35 *
36 * \b Inputs
37 * \li in_a: One of the vectors to be multiplied and accumulated.
38 * \li in_b: The other vector to be multiplied and accumulated.
39 * \li num_points: Number of complex values to be multiplied together, accumulated and stored into \p result
40 *
41 * \b Outputs
42 * \li result: Value of the accumulated result.
43 *
44 */
45
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>
51 #include <volk/saturation_arithmetic.h>
52
53
54 #ifdef LV_HAVE_GENERIC
55
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)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
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)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
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)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
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)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
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)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
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)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
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)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
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)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*/
551