63#ifndef INCLUDED_volk_32fc_index_max_16u_a_H
64#define INCLUDED_volk_32fc_index_max_16u_a_H
76static inline void volk_32fc_index_max_16u_a_avx2_variant_0(uint16_t* target,
80 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
82 const __m256i indices_increment = _mm256_set1_epi32(8);
88 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
90 __m256 max_values = _mm256_setzero_ps();
91 __m256i max_indices = _mm256_setzero_si256();
93 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
94 __m256 in0 = _mm256_load_ps((
float*)src0);
95 __m256 in1 = _mm256_load_ps((
float*)(src0 + 4));
97 in0, in1, &max_values, &max_indices, ¤t_indices, indices_increment);
104 _mm256_store_ps(max_values_buffer, max_values);
105 _mm256_store_si256((__m256i*)max_indices_buffer, max_indices);
109 for (
unsigned i = 0;
i < 8;
i++) {
110 if (max_values_buffer[
i] > max) {
111 max = max_values_buffer[
i];
112 index = max_indices_buffer[
i];
117 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
118 const float abs_squared =
120 if (abs_squared > max) {
133#include <immintrin.h>
136static inline void volk_32fc_index_max_16u_a_avx2_variant_1(uint16_t* target,
140 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
142 const __m256i indices_increment = _mm256_set1_epi32(8);
148 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
150 __m256 max_values = _mm256_setzero_ps();
151 __m256i max_indices = _mm256_setzero_si256();
153 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
154 __m256 in0 = _mm256_load_ps((
float*)src0);
155 __m256 in1 = _mm256_load_ps((
float*)(src0 + 4));
157 in0, in1, &max_values, &max_indices, ¤t_indices, indices_increment);
164 _mm256_store_ps(max_values_buffer, max_values);
165 _mm256_store_si256((__m256i*)max_indices_buffer, max_indices);
169 for (
unsigned i = 0;
i < 8;
i++) {
170 if (max_values_buffer[
i] > max) {
171 max = max_values_buffer[
i];
172 index = max_indices_buffer[
i];
177 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
178 const float abs_squared =
180 if (abs_squared > max) {
193#include <pmmintrin.h>
194#include <xmmintrin.h>
200 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
201 const uint32_t num_bytes = num_points * 8;
208 __m128 xmm1, xmm2, xmm3;
209 __m128i xmm8, xmm11, xmm12, xmm9, xmm10;
211 xmm5.
int_vec = _mm_setzero_si128();
212 xmm4.
int_vec = _mm_setzero_si128();
213 holderf.
int_vec = _mm_setzero_si128();
214 holderi.
int_vec = _mm_setzero_si128();
216 int bound = num_bytes >> 5;
219 xmm8 = _mm_setr_epi32(0, 1, 2, 3);
220 xmm9 = _mm_setzero_si128();
221 xmm10 = _mm_setr_epi32(4, 4, 4, 4);
222 xmm3 = _mm_setzero_ps();
224 for (;
i < bound; ++
i) {
225 xmm1 = _mm_load_ps((
float*)src0);
226 xmm2 = _mm_load_ps((
float*)&src0[2]);
230 xmm1 = _mm_mul_ps(xmm1, xmm1);
231 xmm2 = _mm_mul_ps(xmm2, xmm2);
233 xmm1 = _mm_hadd_ps(xmm1, xmm2);
235 xmm3 = _mm_max_ps(xmm1, xmm3);
237 xmm4.
float_vec = _mm_cmplt_ps(xmm1, xmm3);
238 xmm5.
float_vec = _mm_cmpeq_ps(xmm1, xmm3);
240 xmm11 = _mm_and_si128(xmm8, xmm5.
int_vec);
241 xmm12 = _mm_and_si128(xmm9, xmm4.
int_vec);
243 xmm9 = _mm_add_epi32(xmm11, xmm12);
245 xmm8 = _mm_add_epi32(xmm8, xmm10);
248 if (num_bytes >> 4 & 1) {
249 xmm2 = _mm_load_ps((
float*)src0);
254 xmm2 = _mm_mul_ps(xmm2, xmm2);
258 xmm1 = _mm_hadd_ps(xmm2, xmm2);
260 xmm3 = _mm_max_ps(xmm1, xmm3);
262 xmm10 = _mm_setr_epi32(2, 2, 2, 2);
264 xmm4.
float_vec = _mm_cmplt_ps(xmm1, xmm3);
265 xmm5.
float_vec = _mm_cmpeq_ps(xmm1, xmm3);
267 xmm11 = _mm_and_si128(xmm8, xmm5.
int_vec);
268 xmm12 = _mm_and_si128(xmm9, xmm4.
int_vec);
270 xmm9 = _mm_add_epi32(xmm11, xmm12);
272 xmm8 = _mm_add_epi32(xmm8, xmm10);
275 if (num_bytes >> 3 & 1) {
279 xmm2 = _mm_load1_ps(&sq_dist);
283 xmm3 = _mm_max_ss(xmm3, xmm2);
285 xmm4.
float_vec = _mm_cmplt_ps(xmm1, xmm3);
286 xmm5.
float_vec = _mm_cmpeq_ps(xmm1, xmm3);
288 xmm8 = _mm_shuffle_epi32(xmm8, 0x00);
290 xmm11 = _mm_and_si128(xmm8, xmm4.
int_vec);
291 xmm12 = _mm_and_si128(xmm9, xmm5.
int_vec);
293 xmm9 = _mm_add_epi32(xmm11, xmm12);
296 _mm_store_ps((
float*)&(holderf.
f), xmm3);
297 _mm_store_si128(&(holderi.
int_vec), xmm9);
299 target[0] = holderi.
i[0];
300 sq_dist = holderf.
f[0];
301 target[0] = (holderf.
f[1] > sq_dist) ? holderi.
i[1] : target[0];
302 sq_dist = (holderf.
f[1] > sq_dist) ? holderf.
f[1] : sq_dist;
303 target[0] = (holderf.
f[2] > sq_dist) ? holderi.
i[2] : target[0];
304 sq_dist = (holderf.
f[2] > sq_dist) ? holderf.
f[2] : sq_dist;
305 target[0] = (holderf.
f[3] > sq_dist) ? holderi.
i[3] : target[0];
306 sq_dist = (holderf.
f[3] > sq_dist) ? holderf.
f[3] : sq_dist;
311#ifdef LV_HAVE_GENERIC
316 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
318 const uint32_t num_bytes = num_points * 8;
326 for (;
i < (num_bytes >> 3); ++
i) {
342#ifndef INCLUDED_volk_32fc_index_max_16u_u_H
343#define INCLUDED_volk_32fc_index_max_16u_u_H
352#include <immintrin.h>
355static inline void volk_32fc_index_max_16u_u_avx2_variant_0(uint16_t* target,
359 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
361 const __m256i indices_increment = _mm256_set1_epi32(8);
367 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
369 __m256 max_values = _mm256_setzero_ps();
370 __m256i max_indices = _mm256_setzero_si256();
372 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
373 __m256 in0 = _mm256_loadu_ps((
float*)src0);
374 __m256 in1 = _mm256_loadu_ps((
float*)(src0 + 4));
376 in0, in1, &max_values, &max_indices, ¤t_indices, indices_increment);
383 _mm256_store_ps(max_values_buffer, max_values);
384 _mm256_store_si256((__m256i*)max_indices_buffer, max_indices);
388 for (
unsigned i = 0;
i < 8;
i++) {
389 if (max_values_buffer[
i] > max) {
390 max = max_values_buffer[
i];
391 index = max_indices_buffer[
i];
396 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
397 const float abs_squared =
399 if (abs_squared > max) {
412#include <immintrin.h>
415static inline void volk_32fc_index_max_16u_u_avx2_variant_1(uint16_t* target,
419 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
421 const __m256i indices_increment = _mm256_set1_epi32(8);
427 __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
429 __m256 max_values = _mm256_setzero_ps();
430 __m256i max_indices = _mm256_setzero_si256();
432 for (
unsigned i = 0;
i < num_points / 8u; ++
i) {
433 __m256 in0 = _mm256_loadu_ps((
float*)src0);
434 __m256 in1 = _mm256_loadu_ps((
float*)(src0 + 4));
436 in0, in1, &max_values, &max_indices, ¤t_indices, indices_increment);
443 _mm256_store_ps(max_values_buffer, max_values);
444 _mm256_store_si256((__m256i*)max_indices_buffer, max_indices);
448 for (
unsigned i = 0;
i < 8;
i++) {
449 if (max_values_buffer[
i] > max) {
450 max = max_values_buffer[
i];
451 index = max_indices_buffer[
i];
456 for (
unsigned i = num_points & (~7u);
i < num_points; ++
i) {
457 const float abs_squared =
459 if (abs_squared > max) {
473#include <riscv_vector.h>
476volk_32fc_index_max_16u_rvv(uint16_t* target,
const lv_32fc_t* src0, uint32_t num_points)
478 vfloat32m4_t vmax = __riscv_vfmv_v_f_f32m4(0, __riscv_vsetvlmax_e32m4());
479 vuint16m2_t vmaxi = __riscv_vmv_v_x_u16m2(0, __riscv_vsetvlmax_e16m2());
480 vuint16m2_t vidx = __riscv_vid_v_u16m2(__riscv_vsetvlmax_e16m2());
481 size_t n = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
482 for (
size_t vl; n > 0; n -= vl, src0 += vl) {
483 vl = __riscv_vsetvl_e32m4(n);
484 vuint64m8_t vc = __riscv_vle64_v_u64m8((
const uint64_t*)src0, vl);
485 vfloat32m4_t vr = __riscv_vreinterpret_f32m4(__riscv_vnsrl(vc, 0, vl));
486 vfloat32m4_t vi = __riscv_vreinterpret_f32m4(__riscv_vnsrl(vc, 32, vl));
487 vfloat32m4_t v = __riscv_vfmacc(__riscv_vfmul(vr, vr, vl), vi, vi, vl);
488 vbool8_t m = __riscv_vmflt(vmax, v, vl);
489 vmax = __riscv_vfmax_tu(vmax, vmax, v, vl);
490 vmaxi = __riscv_vmerge_tu(vmaxi, vmaxi, vidx, m, vl);
491 vidx = __riscv_vadd(vidx, vl, __riscv_vsetvlmax_e16m4());
493 size_t vl = __riscv_vsetvlmax_e32m4();
494 float max = __riscv_vfmv_f(__riscv_vfredmax(
RISCV_SHRINK4(vfmax,
f, 32, vmax),
495 __riscv_vfmv_v_f_f32m1(0, 1),
496 __riscv_vsetvlmax_e32m1()));
497 vbool8_t m = __riscv_vmfeq(vmax, max, vl);
498 *target = __riscv_vmv_x(__riscv_vslidedown(vmaxi, __riscv_vfirst(m, vl), vl));
504#include <riscv_vector.h>
506static inline void volk_32fc_index_max_16u_rvvseg(uint16_t* target,
510 vfloat32m4_t vmax = __riscv_vfmv_v_f_f32m4(0, __riscv_vsetvlmax_e32m4());
511 vuint16m2_t vmaxi = __riscv_vmv_v_x_u16m2(0, __riscv_vsetvlmax_e16m2());
512 vuint16m2_t vidx = __riscv_vid_v_u16m2(__riscv_vsetvlmax_e16m2());
513 size_t n = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
514 for (
size_t vl; n > 0; n -= vl, src0 += vl) {
515 vl = __riscv_vsetvl_e32m4(n);
516 vfloat32m4x2_t vc = __riscv_vlseg2e32_v_f32m4x2((
const float*)src0, vl);
517 vfloat32m4_t vr = __riscv_vget_f32m4(vc, 0), vi = __riscv_vget_f32m4(vc, 1);
518 vfloat32m4_t v = __riscv_vfmacc(__riscv_vfmul(vr, vr, vl), vi, vi, vl);
519 vbool8_t m = __riscv_vmflt(vmax, v, vl);
520 vmax = __riscv_vfmax_tu(vmax, vmax, v, vl);
521 vmaxi = __riscv_vmerge_tu(vmaxi, vmaxi, vidx, m, vl);
522 vidx = __riscv_vadd(vidx, vl, __riscv_vsetvlmax_e16m4());
524 size_t vl = __riscv_vsetvlmax_e32m4();
525 float max = __riscv_vfmv_f(__riscv_vfredmax(
RISCV_SHRINK4(vfmax,
f, 32, vmax),
526 __riscv_vfmv_v_f_f32m1(0, 1),
527 __riscv_vsetvlmax_e32m1()));
528 vbool8_t m = __riscv_vmfeq(vmax, max, vl);
529 *target = __riscv_vmv_x(__riscv_vslidedown(vmaxi, __riscv_vfirst(m, vl), vl));
Definition volk_common.h:116
float f[4]
Definition volk_common.h:120
__m128i int_vec
Definition volk_common.h:128
uint32_t i[4]
Definition volk_common.h:119
__m128 float_vec
Definition volk_common.h:124
static void volk_32fc_index_max_16u_generic(uint16_t *target, const lv_32fc_t *src0, uint32_t num_points)
Definition volk_32fc_index_max_16u.h:312
static void volk_32fc_index_max_16u_a_sse3(uint16_t *target, const lv_32fc_t *src0, uint32_t num_points)
Definition volk_32fc_index_max_16u.h:196
static void vector_32fc_index_max_variant1(__m256 in0, __m256 in1, __m256 *max_values, __m256i *max_indices, __m256i *current_indices, __m256i indices_increment)
Definition volk_avx2_intrinsics.h:203
static void vector_32fc_index_max_variant0(__m256 in0, __m256 in1, __m256 *max_values, __m256i *max_indices, __m256i *current_indices, __m256i indices_increment)
Definition volk_avx2_intrinsics.h:141
#define bit128_p(x)
Definition volk_common.h:147
#define __VOLK_ATTR_ALIGNED(x)
Definition volk_common.h:62
#define lv_cimag(x)
Definition volk_complex.h:98
#define lv_creal(x)
Definition volk_complex.h:96
float complex lv_32fc_t
Definition volk_complex.h:74
for i
Definition volk_config_fixed.tmpl.h:13
#define RISCV_SHRINK4(op, T, S, v)
Definition volk_rvv_intrinsics.h:24