10#ifndef INCLUDED_volk_8ic_x2_multiply_conjugate_16ic_a_H
11#define INCLUDED_volk_8ic_x2_multiply_conjugate_16ic_a_H
28static inline void volk_8ic_x2_multiply_conjugate_16ic_a_avx2(
lv_16sc_t* cVector,
31 unsigned int num_points)
33 unsigned int number = 0;
34 const unsigned int quarterPoints = num_points / 8;
36 __m256i x, y, realz, imagz;
40 __m256i conjugateSign =
41 _mm256_set_epi16(-1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1);
43 for (; number < quarterPoints; number++) {
49 realz = _mm256_madd_epi16(x, y);
52 y = _mm256_sign_epi16(y, conjugateSign);
55 y = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(y,
_MM_SHUFFLE(2, 3, 0, 1)),
59 imagz = _mm256_madd_epi16(x, y);
63 _mm256_store_si256((__m256i*)c,
64 _mm256_packs_epi32(_mm256_unpacklo_epi32(realz, imagz),
65 _mm256_unpackhi_epi32(realz, imagz)));
72 number = quarterPoints * 8;
73 int16_t* c16Ptr = (int16_t*)&cVector[number];
74 int8_t* a8Ptr = (int8_t*)&aVector[number];
75 int8_t* b8Ptr = (int8_t*)&bVector[number];
76 for (; number < num_points; number++) {
77 float aReal = (float)*a8Ptr++;
78 float aImag = (float)*a8Ptr++;
80 float bReal = (float)*b8Ptr++;
81 float bImag = (float)*b8Ptr++;
85 *c16Ptr++ = (int16_t)(
lv_creal(temp) > SHRT_MAX ? SHRT_MAX :
lv_creal(temp));
102static inline void volk_8ic_x2_multiply_conjugate_16ic_a_sse4_1(
lv_16sc_t* cVector,
105 unsigned int num_points)
107 unsigned int number = 0;
108 const unsigned int quarterPoints = num_points / 4;
116 for (; number < quarterPoints; number++) {
143 number = quarterPoints * 4;
144 int16_t* c16Ptr = (int16_t*)&cVector[number];
145 int8_t* a8Ptr = (int8_t*)&aVector[number];
146 int8_t* b8Ptr = (int8_t*)&bVector[number];
147 for (; number < num_points; number++) {
148 float aReal = (float)*a8Ptr++;
149 float aImag = (float)*a8Ptr++;
151 float bReal = (float)*b8Ptr++;
152 float bImag = (float)*b8Ptr++;
156 *c16Ptr++ = (int16_t)(
lv_creal(temp) > SHRT_MAX ? SHRT_MAX :
lv_creal(temp));
157 *c16Ptr++ = (int16_t)
lv_cimag(temp);
162#ifdef LV_HAVE_GENERIC
174 unsigned int num_points)
176 unsigned int number = 0;
177 int16_t* c16Ptr = (int16_t*)cVector;
178 int8_t* a8Ptr = (int8_t*)aVector;
179 int8_t* b8Ptr = (int8_t*)bVector;
180 for (number = 0; number < num_points; number++) {
181 float aReal = (float)*a8Ptr++;
182 float aImag = (float)*a8Ptr++;
184 float bReal = (float)*b8Ptr++;
185 float bImag = (float)*b8Ptr++;
189 *c16Ptr++ = (int16_t)(
lv_creal(temp) > SHRT_MAX ? SHRT_MAX :
lv_creal(temp));
190 *c16Ptr++ = (int16_t)
lv_cimag(temp);
197#ifndef INCLUDED_volk_8ic_x2_multiply_conjugate_16ic_u_H
198#define INCLUDED_volk_8ic_x2_multiply_conjugate_16ic_u_H
205#include <immintrin.h>
214static inline void volk_8ic_x2_multiply_conjugate_16ic_u_avx2(
lv_16sc_t* cVector,
217 unsigned int num_points)
219 unsigned int number = 0;
220 const unsigned int oneEigthPoints = num_points / 8;
222 __m256i x, y, realz, imagz;
226 __m256i conjugateSign =
227 _mm256_set_epi16(-1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1);
229 for (; number < oneEigthPoints; number++) {
235 realz = _mm256_madd_epi16(x, y);
238 y = _mm256_sign_epi16(y, conjugateSign);
241 y = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(y,
_MM_SHUFFLE(2, 3, 0, 1)),
245 imagz = _mm256_madd_epi16(x, y);
249 _mm256_storeu_si256((__m256i*)c,
250 _mm256_packs_epi32(_mm256_unpacklo_epi32(realz, imagz),
251 _mm256_unpackhi_epi32(realz, imagz)));
258 number = oneEigthPoints * 8;
259 int16_t* c16Ptr = (int16_t*)&cVector[number];
260 int8_t* a8Ptr = (int8_t*)&aVector[number];
261 int8_t* b8Ptr = (int8_t*)&bVector[number];
262 for (; number < num_points; number++) {
263 float aReal = (float)*a8Ptr++;
264 float aImag = (float)*a8Ptr++;
266 float bReal = (float)*b8Ptr++;
267 float bImag = (float)*b8Ptr++;
271 *c16Ptr++ = (int16_t)(
lv_creal(temp) > SHRT_MAX ? SHRT_MAX :
lv_creal(temp));
272 *c16Ptr++ = (int16_t)
lv_cimag(temp);
FORCE_INLINE __m128i _mm_packs_epi32(__m128i a, __m128i b)
Definition: sse2neon.h:4780
FORCE_INLINE void _mm_store_si128(__m128i *p, __m128i a)
Definition: sse2neon.h:5452
FORCE_INLINE __m128i _mm_unpacklo_epi32(__m128i a, __m128i b)
Definition: sse2neon.h:5809
FORCE_INLINE __m128i _mm_unpackhi_epi32(__m128i a, __m128i b)
Definition: sse2neon.h:5729
FORCE_INLINE __m128i _mm_loadu_si128(const __m128i *p)
Definition: sse2neon.h:4344
FORCE_INLINE __m128i _mm_madd_epi16(__m128i a, __m128i b)
Definition: sse2neon.h:4361
FORCE_INLINE __m128i _mm_load_si128(const __m128i *p)
Definition: sse2neon.h:4260
#define _mm_shufflelo_epi16(a, imm)
Definition: sse2neon.h:5157
FORCE_INLINE __m128i _mm_loadl_epi64(__m128i const *p)
Definition: sse2neon.h:4294
FORCE_INLINE __m128i _mm_cvtepi8_epi16(__m128i a)
Definition: sse2neon.h:6853
FORCE_INLINE __m128i _mm_set_epi16(short i7, short i6, short i5, short i4, short i3, short i2, short i1, short i0)
Definition: sse2neon.h:4825
#define _mm_shufflehi_epi16(a, imm)
Definition: sse2neon.h:5142
FORCE_INLINE __m128i _mm_sign_epi16(__m128i _a, __m128i _b)
Definition: sse2neon.h:6488
#define _MM_SHUFFLE(fp3, fp2, fp1, fp0)
Definition: sse2neon.h:306
int64x2_t __m128i
Definition: sse2neon.h:375
static void volk_8ic_x2_multiply_conjugate_16ic_generic(lv_16sc_t *cVector, const lv_8sc_t *aVector, const lv_8sc_t *bVector, unsigned int num_points)
Multiplys the one complex vector with the complex conjugate of the second complex vector and stores t...
Definition: volk_8ic_x2_multiply_conjugate_16ic.h:171
#define lv_cimag(x)
Definition: volk_complex.h:98
#define lv_cmake(r, i)
Definition: volk_complex.h:77
char complex lv_8sc_t
Provide typedefs and operators for all complex types in C and C++.
Definition: volk_complex.h:70
#define lv_creal(x)
Definition: volk_complex.h:96
float complex lv_32fc_t
Definition: volk_complex.h:74
short complex lv_16sc_t
Definition: volk_complex.h:71