58#ifndef INCLUDED_volk_32f_index_min_16u_a_H
59#define INCLUDED_volk_32f_index_min_16u_a_H
72 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
73 const uint32_t eighthPoints = num_points / 8;
75 float* inputPtr = (
float*)source;
77 __m256 indexIncrementValues = _mm256_set1_ps(8);
78 __m256 currentIndexes = _mm256_set_ps(-1, -2, -3, -4, -5, -6, -7, -8);
80 float min = source[0];
82 __m256 minValues = _mm256_set1_ps(min);
83 __m256 minValuesIndex = _mm256_setzero_ps();
84 __m256 compareResults;
90 for (uint32_t number = 0; number < eighthPoints; number++) {
92 currentValues = _mm256_load_ps(inputPtr);
94 currentIndexes = _mm256_add_ps(currentIndexes, indexIncrementValues);
96 compareResults = _mm256_cmp_ps(currentValues, minValues, _CMP_LT_OS);
98 minValuesIndex = _mm256_blendv_ps(minValuesIndex, currentIndexes, compareResults);
99 minValues = _mm256_blendv_ps(minValues, currentValues, compareResults);
103 _mm256_store_ps(minValuesBuffer, minValues);
104 _mm256_store_ps(minIndexesBuffer, minValuesIndex);
106 for (uint32_t number = 0; number < 8; number++) {
107 if (minValuesBuffer[number] < min) {
108 index = minIndexesBuffer[number];
109 min = minValuesBuffer[number];
110 }
else if (minValuesBuffer[number] == min) {
111 if (index > minIndexesBuffer[number])
112 index = minIndexesBuffer[number];
116 for (uint32_t number = eighthPoints * 8; number < num_points; number++) {
117 if (source[number] < min) {
119 min = source[number];
122 target[0] = (uint16_t)index;
128#include <smmintrin.h>
130static inline void volk_32f_index_min_16u_a_sse4_1(uint16_t* target,
134 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
135 const uint32_t quarterPoints = num_points / 4;
137 float* inputPtr = (
float*)source;
139 __m128 indexIncrementValues = _mm_set1_ps(4);
140 __m128 currentIndexes = _mm_set_ps(-1, -2, -3, -4);
142 float min = source[0];
144 __m128 minValues = _mm_set1_ps(min);
145 __m128 minValuesIndex = _mm_setzero_ps();
146 __m128 compareResults;
147 __m128 currentValues;
152 for (uint32_t number = 0; number < quarterPoints; number++) {
154 currentValues = _mm_load_ps(inputPtr);
156 currentIndexes = _mm_add_ps(currentIndexes, indexIncrementValues);
158 compareResults = _mm_cmplt_ps(currentValues, minValues);
160 minValuesIndex = _mm_blendv_ps(minValuesIndex, currentIndexes, compareResults);
161 minValues = _mm_blendv_ps(minValues, currentValues, compareResults);
165 _mm_store_ps(minValuesBuffer, minValues);
166 _mm_store_ps(minIndexesBuffer, minValuesIndex);
168 for (uint32_t number = 0; number < 4; number++) {
169 if (minValuesBuffer[number] < min) {
170 index = minIndexesBuffer[number];
171 min = minValuesBuffer[number];
172 }
else if (minValuesBuffer[number] == min) {
173 if (index > minIndexesBuffer[number])
174 index = minIndexesBuffer[number];
178 for (uint32_t number = quarterPoints * 4; number < num_points; number++) {
179 if (source[number] < min) {
181 min = source[number];
184 target[0] = (uint16_t)index;
192#include <xmmintrin.h>
197 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
198 const uint32_t quarterPoints = num_points / 4;
200 float* inputPtr = (
float*)source;
202 __m128 indexIncrementValues = _mm_set1_ps(4);
203 __m128 currentIndexes = _mm_set_ps(-1, -2, -3, -4);
205 float min = source[0];
207 __m128 minValues = _mm_set1_ps(min);
208 __m128 minValuesIndex = _mm_setzero_ps();
209 __m128 compareResults;
210 __m128 currentValues;
215 for (uint32_t number = 0; number < quarterPoints; number++) {
217 currentValues = _mm_load_ps(inputPtr);
219 currentIndexes = _mm_add_ps(currentIndexes, indexIncrementValues);
221 compareResults = _mm_cmplt_ps(currentValues, minValues);
223 minValuesIndex = _mm_or_ps(_mm_and_ps(compareResults, currentIndexes),
224 _mm_andnot_ps(compareResults, minValuesIndex));
225 minValues = _mm_or_ps(_mm_and_ps(compareResults, currentValues),
226 _mm_andnot_ps(compareResults, minValues));
230 _mm_store_ps(minValuesBuffer, minValues);
231 _mm_store_ps(minIndexesBuffer, minValuesIndex);
233 for (uint32_t number = 0; number < 4; number++) {
234 if (minValuesBuffer[number] < min) {
235 index = minIndexesBuffer[number];
236 min = minValuesBuffer[number];
237 }
else if (minValuesBuffer[number] == min) {
238 if (index > minIndexesBuffer[number])
239 index = minIndexesBuffer[number];
243 for (uint32_t number = quarterPoints * 4; number < num_points; number++) {
244 if (source[number] < min) {
246 min = source[number];
249 target[0] = (uint16_t)index;
255#ifdef LV_HAVE_GENERIC
260 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
262 float min = source[0];
265 for (uint32_t
i = 1;
i < num_points; ++
i) {
266 if (source[
i] < min) {
280#ifndef INCLUDED_volk_32f_index_min_16u_u_H
281#define INCLUDED_volk_32f_index_min_16u_u_H
289#include <immintrin.h>
294 num_points = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
295 const uint32_t eighthPoints = num_points / 8;
297 float* inputPtr = (
float*)source;
299 __m256 indexIncrementValues = _mm256_set1_ps(8);
300 __m256 currentIndexes = _mm256_set_ps(-1, -2, -3, -4, -5, -6, -7, -8);
302 float min = source[0];
304 __m256 minValues = _mm256_set1_ps(min);
305 __m256 minValuesIndex = _mm256_setzero_ps();
306 __m256 compareResults;
307 __m256 currentValues;
312 for (uint32_t number = 0; number < eighthPoints; number++) {
314 currentValues = _mm256_loadu_ps(inputPtr);
316 currentIndexes = _mm256_add_ps(currentIndexes, indexIncrementValues);
318 compareResults = _mm256_cmp_ps(currentValues, minValues, _CMP_LT_OS);
320 minValuesIndex = _mm256_blendv_ps(minValuesIndex, currentIndexes, compareResults);
321 minValues = _mm256_blendv_ps(minValues, currentValues, compareResults);
325 _mm256_storeu_ps(minValuesBuffer, minValues);
326 _mm256_storeu_ps(minIndexesBuffer, minValuesIndex);
328 for (uint32_t number = 0; number < 8; number++) {
329 if (minValuesBuffer[number] < min) {
330 index = minIndexesBuffer[number];
331 min = minValuesBuffer[number];
332 }
else if (minValuesBuffer[number] == min) {
333 if (index > minIndexesBuffer[number])
334 index = minIndexesBuffer[number];
338 for (uint32_t number = eighthPoints * 8; number < num_points; number++) {
339 if (source[number] < min) {
341 min = source[number];
344 target[0] = (uint16_t)index;
351#include <riscv_vector.h>
354volk_32f_index_min_16u_rvv(uint16_t* target,
const float* src0, uint32_t num_points)
356 vfloat32m8_t vmin = __riscv_vfmv_v_f_f32m8(FLT_MAX, __riscv_vsetvlmax_e32m8());
357 vuint16m4_t vmini = __riscv_vmv_v_x_u16m4(0, __riscv_vsetvlmax_e16m4());
358 vuint16m4_t vidx = __riscv_vid_v_u16m4(__riscv_vsetvlmax_e16m4());
359 size_t n = (num_points > USHRT_MAX) ? USHRT_MAX : num_points;
360 for (
size_t vl; n > 0; n -= vl, src0 += vl) {
361 vl = __riscv_vsetvl_e32m8(n);
362 vfloat32m8_t v = __riscv_vle32_v_f32m8(src0, vl);
363 vbool4_t m = __riscv_vmflt(v, vmin, vl);
364 vmin = __riscv_vfmin_tu(vmin, vmin, v, vl);
365 vmini = __riscv_vmerge_tu(vmini, vmini, vidx, m, vl);
366 vidx = __riscv_vadd(vidx, vl, __riscv_vsetvlmax_e16m4());
368 size_t vl = __riscv_vsetvlmax_e32m8();
369 float min = __riscv_vfmv_f(__riscv_vfredmin(
RISCV_SHRINK8(vfmin, f, 32, vmin),
370 __riscv_vfmv_v_f_f32m1(FLT_MAX, 1),
371 __riscv_vsetvlmax_e32m1()));
372 vbool4_t m = __riscv_vmfeq(vmin, min, vl);
373 *target = __riscv_vmv_x(__riscv_vslidedown(vmini, __riscv_vfirst(m, vl), vl));
static void volk_32f_index_min_16u_a_avx(uint16_t *target, const float *source, uint32_t num_points)
Definition volk_32f_index_min_16u.h:70
static void volk_32f_index_min_16u_generic(uint16_t *target, const float *source, uint32_t num_points)
Definition volk_32f_index_min_16u.h:258
static void volk_32f_index_min_16u_a_sse(uint16_t *target, const float *source, uint32_t num_points)
Definition volk_32f_index_min_16u.h:195
static void volk_32f_index_min_16u_u_avx(uint16_t *target, const float *source, uint32_t num_points)
Definition volk_32f_index_min_16u.h:292
#define __VOLK_ATTR_ALIGNED(x)
Definition volk_common.h:62
for i
Definition volk_config_fixed.tmpl.h:13
#define RISCV_SHRINK8(op, T, S, v)
Definition volk_rvv_intrinsics.h:33