GCC Code Coverage Report

Directory: ./
File: kernels/volk/volk_32fc_index_min_32u.h
Date: 2023-10-23 23:10:04
Exec Total Coverage
Lines: 173 181 95.6%
Functions: 6 6 100.0%
Branches: 56 62 90.3%
Line Branch Exec Source
1 /* -*- c++ -*- */
2 /*
3 * Copyright 2021 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_32fc_index_min_32u
12 *
13 * \b Overview
14 *
15 * Returns Argmin_i mag(x[i]). Finds and returns the index which contains the
16 * minimum magnitude for complex points in the given vector.
17 *
18 * <b>Dispatcher Prototype</b>
19 * \code
20 * void volk_32fc_index_min_32u(uint32_t* target, lv_32fc_t* source, uint32_t
21 * num_points) \endcode
22 *
23 * \b Inputs
24 * \li source: The complex input vector.
25 * \li num_points: The number of samples.
26 *
27 * \b Outputs
28 * \li target: The index of the point with minimum magnitude.
29 *
30 * \b Example
31 * Calculate the index of the minimum value of \f$x^2 + x\f$ for points around
32 * the unit circle.
33 * \code
34 * int N = 10;
35 * uint32_t alignment = volk_get_alignment();
36 * lv_32fc_t* in = (lv_32fc_t*)volk_malloc(sizeof(lv_32fc_t)*N, alignment);
37 * uint32_t* min = (uint32_t*)volk_malloc(sizeof(uint32_t), alignment);
38 *
39 * for(uint32_t ii = 0; ii < N/2; ++ii){
40 * float real = 2.f * ((float)ii / (float)N) - 1.f;
41 * float imag = std::sqrt(1.f - real * real);
42 * in[ii] = lv_cmake(real, imag);
43 * in[ii] = in[ii] * in[ii] + in[ii];
44 * in[N-ii] = lv_cmake(real, imag);
45 * in[N-ii] = in[N-ii] * in[N-ii] + in[N-ii];
46 * }
47 *
48 * volk_32fc_index_min_32u(min, in, N);
49 *
50 * printf("index of min value = %u\n", *min);
51 *
52 * volk_free(in);
53 * volk_free(min);
54 * \endcode
55 */
56
57 #ifndef INCLUDED_volk_32fc_index_min_32u_a_H
58 #define INCLUDED_volk_32fc_index_min_32u_a_H
59
60 #include <inttypes.h>
61 #include <stdio.h>
62 #include <volk/volk_common.h>
63 #include <volk/volk_complex.h>
64
65 #ifdef LV_HAVE_AVX2
66 #include <immintrin.h>
67 #include <volk/volk_avx2_intrinsics.h>
68
69 2 static inline void volk_32fc_index_min_32u_a_avx2_variant_0(uint32_t* target,
70 const lv_32fc_t* source,
71 uint32_t num_points)
72 {
73 2 const __m256i indices_increment = _mm256_set1_epi32(8);
74 /*
75 * At the start of each loop iteration current_indices holds the indices of
76 * the complex numbers loaded from memory. Explanation for odd order is given
77 * in implementation of vector_32fc_index_min_variant0().
78 */
79 2 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
80
81 2 __m256 min_values = _mm256_set1_ps(FLT_MAX);
82 2 __m256i min_indices = _mm256_setzero_si256();
83
84
2/2
✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
 32768 for (unsigned i = 0; i < num_points / 8u; ++i) {
85 32766 __m256 in0 = _mm256_load_ps((float*)source);
86 32766 __m256 in1 = _mm256_load_ps((float*)(source + 4));
87 32766 vector_32fc_index_min_variant0(
88 in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
89 32766 source += 8;
90 }
91
92 // determine minimum value and index in the result of the vectorized loop
93 __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
94 __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
95 2 _mm256_store_ps(min_values_buffer, min_values);
96 2 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
97
98 2 float min = FLT_MAX;
99 2 uint32_t index = 0;
100
2/2
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 2 times.
 18 for (unsigned i = 0; i < 8; i++) {
101
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 9 times.
 16 if (min_values_buffer[i] < min) {
102 7 min = min_values_buffer[i];
103 7 index = min_indices_buffer[i];
104 }
105 }
106
107 // handle tail not processed by the vectorized loop
108
2/2
✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
 16 for (unsigned i = num_points & (~7u); i < num_points; ++i) {
109 14 const float abs_squared =
110 14 lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
111
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
 14 if (abs_squared < min) {
112 ✗ min = abs_squared;
113 ✗ index = i;
114 }
115 14 ++source;
116 }
117
118 2 *target = index;
119 2 }
120
121 #endif /*LV_HAVE_AVX2*/
122
123 #ifdef LV_HAVE_AVX2
124 #include <immintrin.h>
125 #include <volk/volk_avx2_intrinsics.h>
126
127 2 static inline void volk_32fc_index_min_32u_a_avx2_variant_1(uint32_t* target,
128 const lv_32fc_t* source,
129 uint32_t num_points)
130 {
131 2 const __m256i indices_increment = _mm256_set1_epi32(8);
132 /*
133 * At the start of each loop iteration current_indices holds the indices of
134 * the complex numbers loaded from memory. Explanation for odd order is given
135 * in implementation of vector_32fc_index_min_variant0().
136 */
137 2 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
138
139 2 __m256 min_values = _mm256_set1_ps(FLT_MAX);
140 2 __m256i min_indices = _mm256_setzero_si256();
141
142
2/2
✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
 32768 for (unsigned i = 0; i < num_points / 8u; ++i) {
143 32766 __m256 in0 = _mm256_load_ps((float*)source);
144 32766 __m256 in1 = _mm256_load_ps((float*)(source + 4));
145 32766 vector_32fc_index_min_variant1(
146 in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
147 32766 source += 8;
148 }
149
150 // determine minimum value and index in the result of the vectorized loop
151 __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
152 __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
153 2 _mm256_store_ps(min_values_buffer, min_values);
154 2 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
155
156 2 float min = FLT_MAX;
157 2 uint32_t index = 0;
158
2/2
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 2 times.
 18 for (unsigned i = 0; i < 8; i++) {
159
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 9 times.
 16 if (min_values_buffer[i] < min) {
160 7 min = min_values_buffer[i];
161 7 index = min_indices_buffer[i];
162 }
163 }
164
165 // handle tail not processed by the vectorized loop
166
2/2
✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
 16 for (unsigned i = num_points & (~7u); i < num_points; ++i) {
167 14 const float abs_squared =
168 14 lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
169
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
 14 if (abs_squared < min) {
170 ✗ min = abs_squared;
171 ✗ index = i;
172 }
173 14 ++source;
174 }
175
176 2 *target = index;
177 2 }
178
179 #endif /*LV_HAVE_AVX2*/
180
181 #ifdef LV_HAVE_SSE3
182 #include <pmmintrin.h>
183 #include <xmmintrin.h>
184
185 2 static inline void volk_32fc_index_min_32u_a_sse3(uint32_t* target,
186 const lv_32fc_t* source,
187 uint32_t num_points)
188 {
189 union bit128 holderf;
190 union bit128 holderi;
191 2 float sq_dist = 0.0;
192
193 union bit128 xmm5, xmm4;
194 __m128 xmm1, xmm2, xmm3;
195 __m128i xmm8, xmm11, xmm12, xmm9, xmm10;
196
197 2 xmm5.int_vec = _mm_setzero_si128();
198 2 xmm4.int_vec = _mm_setzero_si128();
199 2 holderf.int_vec = _mm_setzero_si128();
200 2 holderi.int_vec = _mm_setzero_si128();
201
202 2 xmm8 = _mm_setr_epi32(0, 1, 2, 3);
203 2 xmm9 = _mm_setzero_si128();
204 2 xmm10 = _mm_setr_epi32(4, 4, 4, 4);
205 2 xmm3 = _mm_set_ps1(FLT_MAX);
206
207 2 int bound = num_points >> 2;
208
209
2/2
✓ Branch 0 taken 65534 times.
✓ Branch 1 taken 2 times.
 65536 for (int i = 0; i < bound; ++i) {
210 65534 xmm1 = _mm_load_ps((float*)source);
211 65534 xmm2 = _mm_load_ps((float*)&source[2]);
212
213 65534 source += 4;
214
215 131068 xmm1 = _mm_mul_ps(xmm1, xmm1);
216 65534 xmm2 = _mm_mul_ps(xmm2, xmm2);
217
218 65534 xmm1 = _mm_hadd_ps(xmm1, xmm2);
219
220 65534 xmm3 = _mm_min_ps(xmm1, xmm3);
221
222 65534 xmm4.float_vec = _mm_cmpgt_ps(xmm1, xmm3);
223 65534 xmm5.float_vec = _mm_cmpeq_ps(xmm1, xmm3);
224
225 65534 xmm11 = _mm_and_si128(xmm8, xmm5.int_vec);
226 131068 xmm12 = _mm_and_si128(xmm9, xmm4.int_vec);
227
228 65534 xmm9 = _mm_add_epi32(xmm11, xmm12);
229
230 131068 xmm8 = _mm_add_epi32(xmm8, xmm10);
231 }
232
233
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 if (num_points >> 1 & 1) {
234 2 xmm2 = _mm_load_ps((float*)source);
235
236 2 xmm1 = _mm_movelh_ps(bit128_p(&xmm8)->float_vec, bit128_p(&xmm8)->float_vec);
237 2 xmm8 = bit128_p(&xmm1)->int_vec;
238
239 2 xmm2 = _mm_mul_ps(xmm2, xmm2);
240
241 2 source += 2;
242
243 2 xmm1 = _mm_hadd_ps(xmm2, xmm2);
244
245 4 xmm3 = _mm_min_ps(xmm1, xmm3);
246
247 2 xmm10 = _mm_setr_epi32(2, 2, 2, 2);
248
249 2 xmm4.float_vec = _mm_cmpgt_ps(xmm1, xmm3);
250 2 xmm5.float_vec = _mm_cmpeq_ps(xmm1, xmm3);
251
252 2 xmm11 = _mm_and_si128(xmm8, xmm5.int_vec);
253 4 xmm12 = _mm_and_si128(xmm9, xmm4.int_vec);
254
255 2 xmm9 = _mm_add_epi32(xmm11, xmm12);
256
257 4 xmm8 = _mm_add_epi32(xmm8, xmm10);
258 }
259
260
1/2
✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
 2 if (num_points & 1) {
261 2 sq_dist = lv_creal(source[0]) * lv_creal(source[0]) +
262 2 lv_cimag(source[0]) * lv_cimag(source[0]);
263
264 2 xmm2 = _mm_load1_ps(&sq_dist);
265
266 2 xmm1 = xmm3;
267
268 2 xmm3 = _mm_min_ss(xmm3, xmm2);
269
270 2 xmm4.float_vec = _mm_cmpgt_ps(xmm1, xmm3);
271 2 xmm5.float_vec = _mm_cmpeq_ps(xmm1, xmm3);
272
273 2 xmm8 = _mm_shuffle_epi32(xmm8, 0x00);
274
275 2 xmm11 = _mm_and_si128(xmm8, xmm4.int_vec);
276 4 xmm12 = _mm_and_si128(xmm9, xmm5.int_vec);
277
278 2 xmm9 = _mm_add_epi32(xmm11, xmm12);
279 }
280
281 _mm_store_ps((float*)&(holderf.f), xmm3);
282 _mm_store_si128(&(holderi.int_vec), xmm9);
283
284 2 target[0] = holderi.i[0];
285 2 sq_dist = holderf.f[0];
286
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 1 times.
 2 target[0] = (holderf.f[1] < sq_dist) ? holderi.i[1] : target[0];
287
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 1 times.
 2 sq_dist = (holderf.f[1] < sq_dist) ? holderf.f[1] : sq_dist;
288
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 1 times.
 2 target[0] = (holderf.f[2] < sq_dist) ? holderi.i[2] : target[0];
289
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 1 times.
 2 sq_dist = (holderf.f[2] < sq_dist) ? holderf.f[2] : sq_dist;
290
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 1 times.
 2 target[0] = (holderf.f[3] < sq_dist) ? holderi.i[3] : target[0];
291
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 1 times.
 2 sq_dist = (holderf.f[3] < sq_dist) ? holderf.f[3] : sq_dist;
292 2 }
293
294 #endif /*LV_HAVE_SSE3*/
295
296 #ifdef LV_HAVE_GENERIC
297 2 static inline void volk_32fc_index_min_32u_generic(uint32_t* target,
298 const lv_32fc_t* source,
299 uint32_t num_points)
300 {
301 2 float sq_dist = 0.0;
302 2 float min = FLT_MAX;
303 2 uint32_t index = 0;
304
305
2/2
✓ Branch 0 taken 262142 times.
✓ Branch 1 taken 2 times.
 262144 for (uint32_t i = 0; i < num_points; ++i) {
306 262142 sq_dist = lv_creal(source[i]) * lv_creal(source[i]) +
307 262142 lv_cimag(source[i]) * lv_cimag(source[i]);
308
309
2/2
✓ Branch 0 taken 34 times.
✓ Branch 1 taken 262108 times.
 262142 if (sq_dist < min) {
310 34 index = i;
311 34 min = sq_dist;
312 }
313 }
314 2 target[0] = index;
315 2 }
316
317 #endif /*LV_HAVE_GENERIC*/
318
319 #endif /*INCLUDED_volk_32fc_index_min_32u_a_H*/
320
321 #ifndef INCLUDED_volk_32fc_index_min_32u_u_H
322 #define INCLUDED_volk_32fc_index_min_32u_u_H
323
324 #include <inttypes.h>
325 #include <stdio.h>
326 #include <volk/volk_common.h>
327 #include <volk/volk_complex.h>
328
329 #ifdef LV_HAVE_AVX2
330 #include <immintrin.h>
331 #include <volk/volk_avx2_intrinsics.h>
332
333 2 static inline void volk_32fc_index_min_32u_u_avx2_variant_0(uint32_t* target,
334 const lv_32fc_t* source,
335 uint32_t num_points)
336 {
337 2 const __m256i indices_increment = _mm256_set1_epi32(8);
338 /*
339 * At the start of each loop iteration current_indices holds the indices of
340 * the complex numbers loaded from memory. Explanation for odd order is given
341 * in implementation of vector_32fc_index_min_variant0().
342 */
343 2 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
344
345 2 __m256 min_values = _mm256_set1_ps(FLT_MAX);
346 2 __m256i min_indices = _mm256_setzero_si256();
347
348
2/2
✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
 32768 for (unsigned i = 0; i < num_points / 8u; ++i) {
349 32766 __m256 in0 = _mm256_loadu_ps((float*)source);
350 32766 __m256 in1 = _mm256_loadu_ps((float*)(source + 4));
351 32766 vector_32fc_index_min_variant0(
352 in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
353 32766 source += 8;
354 }
355
356 // determine minimum value and index in the result of the vectorized loop
357 __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
358 __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
359 2 _mm256_store_ps(min_values_buffer, min_values);
360 2 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
361
362 2 float min = FLT_MAX;
363 2 uint32_t index = 0;
364
2/2
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 2 times.
 18 for (unsigned i = 0; i < 8; i++) {
365
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 9 times.
 16 if (min_values_buffer[i] < min) {
366 7 min = min_values_buffer[i];
367 7 index = min_indices_buffer[i];
368 }
369 }
370
371 // handle tail not processed by the vectorized loop
372
2/2
✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
 16 for (unsigned i = num_points & (~7u); i < num_points; ++i) {
373 14 const float abs_squared =
374 14 lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
375
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
 14 if (abs_squared < min) {
376 ✗ min = abs_squared;
377 ✗ index = i;
378 }
379 14 ++source;
380 }
381
382 2 *target = index;
383 2 }
384
385 #endif /*LV_HAVE_AVX2*/
386
387 #ifdef LV_HAVE_AVX2
388 #include <immintrin.h>
389 #include <volk/volk_avx2_intrinsics.h>
390
391 2 static inline void volk_32fc_index_min_32u_u_avx2_variant_1(uint32_t* target,
392 const lv_32fc_t* source,
393 uint32_t num_points)
394 {
395 2 const __m256i indices_increment = _mm256_set1_epi32(8);
396 /*
397 * At the start of each loop iteration current_indices holds the indices of
398 * the complex numbers loaded from memory. Explanation for odd order is given
399 * in implementation of vector_32fc_index_min_variant0().
400 */
401 2 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
402
403 2 __m256 min_values = _mm256_set1_ps(FLT_MAX);
404 2 __m256i min_indices = _mm256_setzero_si256();
405
406
2/2
✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
 32768 for (unsigned i = 0; i < num_points / 8u; ++i) {
407 32766 __m256 in0 = _mm256_loadu_ps((float*)source);
408 32766 __m256 in1 = _mm256_loadu_ps((float*)(source + 4));
409 32766 vector_32fc_index_min_variant1(
410 in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
411 32766 source += 8;
412 }
413
414 // determine minimum value and index in the result of the vectorized loop
415 __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
416 __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
417 2 _mm256_store_ps(min_values_buffer, min_values);
418 2 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
419
420 2 float min = FLT_MAX;
421 2 uint32_t index = 0;
422
2/2
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 2 times.
 18 for (unsigned i = 0; i < 8; i++) {
423
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 9 times.
 16 if (min_values_buffer[i] < min) {
424 7 min = min_values_buffer[i];
425 7 index = min_indices_buffer[i];
426 }
427 }
428
429 // handle tail not processed by the vectorized loop
430
2/2
✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
 16 for (unsigned i = num_points & (~7u); i < num_points; ++i) {
431 14 const float abs_squared =
432 14 lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
433
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
 14 if (abs_squared < min) {
434 ✗ min = abs_squared;
435 ✗ index = i;
436 }
437 14 ++source;
438 }
439
440 2 *target = index;
441 2 }
442
443 #endif /*LV_HAVE_AVX2*/
444
445 #ifdef LV_HAVE_NEON
446 #include <arm_neon.h>
447 #include <volk/volk_neon_intrinsics.h>
448
449 static inline void volk_32fc_index_min_32u_neon(uint32_t* target,
450 const lv_32fc_t* source,
451 uint32_t num_points)
452 {
453 const uint32_t quarter_points = num_points / 4;
454 const lv_32fc_t* sourcePtr = source;
455
456 uint32_t indices[4] = { 0, 1, 2, 3 };
457 const uint32x4_t vec_indices_incr = vdupq_n_u32(4);
458 uint32x4_t vec_indices = vld1q_u32(indices);
459 uint32x4_t vec_min_indices = vec_indices;
460
461 if (num_points) {
462 float min = FLT_MAX;
463 uint32_t index = 0;
464
465 float32x4_t vec_min = vdupq_n_f32(FLT_MAX);
466
467 for (uint32_t number = 0; number < quarter_points; number++) {
468 // Load complex and compute magnitude squared
469 const float32x4_t vec_mag2 =
470 _vmagnitudesquaredq_f32(vld2q_f32((float*)sourcePtr));
471 __VOLK_PREFETCH(sourcePtr += 4);
472 // a < b?
473 const uint32x4_t lt_mask = vcltq_f32(vec_mag2, vec_min);
474 vec_min = vbslq_f32(lt_mask, vec_mag2, vec_min);
475 vec_min_indices = vbslq_u32(lt_mask, vec_indices, vec_min_indices);
476 vec_indices = vaddq_u32(vec_indices, vec_indices_incr);
477 }
478 uint32_t tmp_min_indices[4];
479 float tmp_min[4];
480 vst1q_u32(tmp_min_indices, vec_min_indices);
481 vst1q_f32(tmp_min, vec_min);
482
483 for (int i = 0; i < 4; i++) {
484 if (tmp_min[i] < min) {
485 min = tmp_min[i];
486 index = tmp_min_indices[i];
487 }
488 }
489
490 // Deal with the rest
491 for (uint32_t number = quarter_points * 4; number < num_points; number++) {
492 const float re = lv_creal(*sourcePtr);
493 const float im = lv_cimag(*sourcePtr);
494 const float sq_dist = re * re + im * im;
495 if (sq_dist < min) {
496 min = sq_dist;
497 index = number;
498 }
499 sourcePtr++;
500 }
501 *target = index;
502 }
503 }
504
505 #endif /*LV_HAVE_NEON*/
506
507 #endif /*INCLUDED_volk_32fc_index_min_32u_u_H*/
508