70 #ifndef INCLUDED_volk_32fc_index_min_32u_a_H
71 #define INCLUDED_volk_32fc_index_min_32u_a_H
79 #include <immintrin.h>
82 static inline void volk_32fc_index_min_32u_a_avx2_variant_0(uint32_t* target,
86 const __m256i indices_increment = _mm256_set1_epi32(8);
92 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
94 __m256 min_values = _mm256_set1_ps(FLT_MAX);
95 __m256i min_indices = _mm256_setzero_si256();
97 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
98 __m256 in0 = _mm256_load_ps((
float*)source);
99 __m256 in1 = _mm256_load_ps((
float*)(source + 4));
101 in0, in1, &min_values, &min_indices, ¤t_indices, indices_increment);
108 _mm256_store_ps(min_values_buffer, min_values);
109 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
113 for (
unsigned i = 0;
i < 8;
i++) {
114 if (min_values_buffer[
i] < min) {
115 min = min_values_buffer[
i];
116 index = min_indices_buffer[
i];
121 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
122 const float abs_squared =
124 if (abs_squared < min) {
137 #include <immintrin.h>
140 static inline void volk_32fc_index_min_32u_a_avx2_variant_1(uint32_t* target,
144 const __m256i indices_increment = _mm256_set1_epi32(8);
150 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
152 __m256 min_values = _mm256_set1_ps(FLT_MAX);
153 __m256i min_indices = _mm256_setzero_si256();
155 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
156 __m256 in0 = _mm256_load_ps((
float*)source);
157 __m256 in1 = _mm256_load_ps((
float*)(source + 4));
159 in0, in1, &min_values, &min_indices, ¤t_indices, indices_increment);
166 _mm256_store_ps(min_values_buffer, min_values);
167 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
171 for (
unsigned i = 0;
i < 8;
i++) {
172 if (min_values_buffer[
i] < min) {
173 min = min_values_buffer[
i];
174 index = min_indices_buffer[
i];
179 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
180 const float abs_squared =
182 if (abs_squared < min) {
195 #include <pmmintrin.h>
196 #include <xmmintrin.h>
207 __m128 xmm1, xmm2, xmm3;
208 __m128i xmm8, xmm11, xmm12, xmm9, xmm10;
210 xmm5.
int_vec = _mm_setzero_si128();
211 xmm4.
int_vec = _mm_setzero_si128();
212 holderf.
int_vec = _mm_setzero_si128();
213 holderi.
int_vec = _mm_setzero_si128();
215 xmm8 = _mm_setr_epi32(0, 1, 2, 3);
216 xmm9 = _mm_setzero_si128();
217 xmm10 = _mm_setr_epi32(4, 4, 4, 4);
218 xmm3 = _mm_set_ps1(FLT_MAX);
220 int bound = num_points >> 2;
222 for (
int i = 0;
i < bound; ++
i) {
223 xmm1 = _mm_load_ps((
float*)source);
224 xmm2 = _mm_load_ps((
float*)&source[2]);
228 xmm1 = _mm_mul_ps(xmm1, xmm1);
229 xmm2 = _mm_mul_ps(xmm2, xmm2);
231 xmm1 = _mm_hadd_ps(xmm1, xmm2);
233 xmm3 = _mm_min_ps(xmm1, xmm3);
235 xmm4.
float_vec = _mm_cmpgt_ps(xmm1, xmm3);
236 xmm5.
float_vec = _mm_cmpeq_ps(xmm1, xmm3);
238 xmm11 = _mm_and_si128(xmm8, xmm5.
int_vec);
239 xmm12 = _mm_and_si128(xmm9, xmm4.
int_vec);
241 xmm9 = _mm_add_epi32(xmm11, xmm12);
243 xmm8 = _mm_add_epi32(xmm8, xmm10);
246 if (num_points >> 1 & 1) {
247 xmm2 = _mm_load_ps((
float*)source);
252 xmm2 = _mm_mul_ps(xmm2, xmm2);
256 xmm1 = _mm_hadd_ps(xmm2, xmm2);
258 xmm3 = _mm_min_ps(xmm1, xmm3);
260 xmm10 = _mm_setr_epi32(2, 2, 2, 2);
262 xmm4.
float_vec = _mm_cmpgt_ps(xmm1, xmm3);
263 xmm5.
float_vec = _mm_cmpeq_ps(xmm1, xmm3);
265 xmm11 = _mm_and_si128(xmm8, xmm5.
int_vec);
266 xmm12 = _mm_and_si128(xmm9, xmm4.
int_vec);
268 xmm9 = _mm_add_epi32(xmm11, xmm12);
270 xmm8 = _mm_add_epi32(xmm8, xmm10);
273 if (num_points & 1) {
277 xmm2 = _mm_load1_ps(&sq_dist);
281 xmm3 = _mm_min_ss(xmm3, xmm2);
283 xmm4.
float_vec = _mm_cmpgt_ps(xmm1, xmm3);
284 xmm5.
float_vec = _mm_cmpeq_ps(xmm1, xmm3);
286 xmm8 = _mm_shuffle_epi32(xmm8, 0x00);
288 xmm11 = _mm_and_si128(xmm8, xmm4.
int_vec);
289 xmm12 = _mm_and_si128(xmm9, xmm5.
int_vec);
291 xmm9 = _mm_add_epi32(xmm11, xmm12);
294 _mm_store_ps((
float*)&(holderf.
f), xmm3);
295 _mm_store_si128(&(holderi.
int_vec), xmm9);
297 target[0] = holderi.
i[0];
298 sq_dist = holderf.
f[0];
299 target[0] = (holderf.
f[1] < sq_dist) ? holderi.
i[1] : target[0];
300 sq_dist = (holderf.
f[1] < sq_dist) ? holderf.
f[1] : sq_dist;
301 target[0] = (holderf.
f[2] < sq_dist) ? holderi.
i[2] : target[0];
302 sq_dist = (holderf.
f[2] < sq_dist) ? holderf.
f[2] : sq_dist;
303 target[0] = (holderf.
f[3] < sq_dist) ? holderi.
i[3] : target[0];
304 sq_dist = (holderf.
f[3] < sq_dist) ? holderf.
f[3] : sq_dist;
309 #ifdef LV_HAVE_GENERIC
318 for (uint32_t
i = 0;
i < num_points; ++
i) {
334 #ifndef INCLUDED_volk_32fc_index_min_32u_u_H
335 #define INCLUDED_volk_32fc_index_min_32u_u_H
337 #include <inttypes.h>
343 #include <immintrin.h>
346 static inline void volk_32fc_index_min_32u_u_avx2_variant_0(uint32_t* target,
350 const __m256i indices_increment = _mm256_set1_epi32(8);
356 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
358 __m256 min_values = _mm256_set1_ps(FLT_MAX);
359 __m256i min_indices = _mm256_setzero_si256();
361 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
362 __m256 in0 = _mm256_loadu_ps((
float*)source);
363 __m256 in1 = _mm256_loadu_ps((
float*)(source + 4));
365 in0, in1, &min_values, &min_indices, ¤t_indices, indices_increment);
372 _mm256_store_ps(min_values_buffer, min_values);
373 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
377 for (
unsigned i = 0;
i < 8;
i++) {
378 if (min_values_buffer[
i] < min) {
379 min = min_values_buffer[
i];
380 index = min_indices_buffer[
i];
385 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
386 const float abs_squared =
388 if (abs_squared < min) {
401 #include <immintrin.h>
404 static inline void volk_32fc_index_min_32u_u_avx2_variant_1(uint32_t* target,
408 const __m256i indices_increment = _mm256_set1_epi32(8);
414 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
416 __m256 min_values = _mm256_set1_ps(FLT_MAX);
417 __m256i min_indices = _mm256_setzero_si256();
419 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
420 __m256 in0 = _mm256_loadu_ps((
float*)source);
421 __m256 in1 = _mm256_loadu_ps((
float*)(source + 4));
423 in0, in1, &min_values, &min_indices, ¤t_indices, indices_increment);
430 _mm256_store_ps(min_values_buffer, min_values);
431 _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
435 for (
unsigned i = 0;
i < 8;
i++) {
436 if (min_values_buffer[
i] < min) {
437 min = min_values_buffer[
i];
438 index = min_indices_buffer[
i];
443 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
444 const float abs_squared =
446 if (abs_squared < min) {
459 #include <arm_neon.h>
466 const uint32_t quarter_points = num_points / 4;
469 uint32_t indices[4] = { 0, 1, 2, 3 };
470 const uint32x4_t vec_indices_incr = vdupq_n_u32(4);
471 uint32x4_t vec_indices = vld1q_u32(indices);
472 uint32x4_t vec_min_indices = vec_indices;
478 float32x4_t vec_min = vdupq_n_f32(FLT_MAX);
480 for (uint32_t number = 0; number < quarter_points; number++) {
482 const float32x4_t vec_mag2 =
486 const uint32x4_t lt_mask = vcltq_f32(vec_mag2, vec_min);
487 vec_min = vbslq_f32(lt_mask, vec_mag2, vec_min);
488 vec_min_indices = vbslq_u32(lt_mask, vec_indices, vec_min_indices);
489 vec_indices = vaddq_u32(vec_indices, vec_indices_incr);
491 uint32_t tmp_min_indices[4];
493 vst1q_u32(tmp_min_indices, vec_min_indices);
494 vst1q_f32(tmp_min, vec_min);
496 for (
int i = 0;
i < 4;
i++) {
497 if (tmp_min[
i] < min) {
499 index = tmp_min_indices[
i];
504 for (uint32_t number = quarter_points * 4; number < num_points; number++) {
505 const float re =
lv_creal(*sourcePtr);
506 const float im =
lv_cimag(*sourcePtr);
507 if ((re * re + im * im) < min) {
Definition: volk_common.h:111
float f[4]
Definition: volk_common.h:115
__m128i int_vec
Definition: volk_common.h:123
uint32_t i[4]
Definition: volk_common.h:114
__m128 float_vec
Definition: volk_common.h:119
static void volk_32fc_index_min_32u_generic(uint32_t *target, const lv_32fc_t *source, uint32_t num_points)
Definition: volk_32fc_index_min_32u.h:310
static void volk_32fc_index_min_32u_a_sse3(uint32_t *target, const lv_32fc_t *source, uint32_t num_points)
Definition: volk_32fc_index_min_32u.h:198
static void volk_32fc_index_min_32u_neon(uint32_t *target, const lv_32fc_t *source, uint32_t num_points)
Definition: volk_32fc_index_min_32u.h:462
static void vector_32fc_index_min_variant0(__m256 in0, __m256 in1, __m256 *min_values, __m256i *min_indices, __m256i *current_indices, __m256i indices_increment)
Definition: volk_avx2_intrinsics.h:251
static void vector_32fc_index_min_variant1(__m256 in0, __m256 in1, __m256 *min_values, __m256i *min_indices, __m256i *current_indices, __m256i indices_increment)
Definition: volk_avx2_intrinsics.h:313
#define bit128_p(x)
Definition: volk_common.h:142
#define __VOLK_PREFETCH(addr)
Definition: volk_common.h:62
#define __VOLK_ATTR_ALIGNED(x)
Definition: volk_common.h:56
#define lv_cimag(x)
Definition: volk_complex.h:89
#define lv_creal(x)
Definition: volk_complex.h:87
float complex lv_32fc_t
Definition: volk_complex.h:65
for i
Definition: volk_config_fixed.tmpl.h:25
static float32x4_t _vmagnitudesquaredq_f32(float32x4x2_t cmplxValue)
Definition: volk_neon_intrinsics.h:87