70 unsigned int num_points)
72 const unsigned int num_bytes = num_points * 2;
74 __m128i xmm0, xmm1, xmm2, xmm3, xmm4;
75 __m128i *p_target0, *p_target1, *p_target2, *p_target3, *p_src0, *p_src1, *p_src2,
77 p_target0 = (__m128i*)target0;
78 p_target1 = (__m128i*)target1;
79 p_target2 = (__m128i*)target2;
80 p_target3 = (__m128i*)target3;
82 p_src0 = (__m128i*)src0;
83 p_src1 = (__m128i*)src1;
84 p_src2 = (__m128i*)src2;
85 p_src3 = (__m128i*)src3;
86 p_src4 = (__m128i*)src4;
90 int bound = (num_bytes >> 4);
91 int leftovers = (num_bytes >> 1) & 7;
93 for (;
i < bound; ++
i) {
94 xmm0 = _mm_load_si128(p_src0);
95 xmm1 = _mm_load_si128(p_src1);
96 xmm2 = _mm_load_si128(p_src2);
97 xmm3 = _mm_load_si128(p_src3);
98 xmm4 = _mm_load_si128(p_src4);
103 xmm1 = _mm_add_epi16(xmm0, xmm1);
104 xmm2 = _mm_add_epi16(xmm0, xmm2);
105 xmm3 = _mm_add_epi16(xmm0, xmm3);
106 xmm4 = _mm_add_epi16(xmm0, xmm4);
113 _mm_store_si128(p_target0, xmm1);
114 _mm_store_si128(p_target1, xmm2);
115 _mm_store_si128(p_target2, xmm3);
116 _mm_store_si128(p_target3, xmm4);
124 for (
i = bound * 8;
i < (bound * 8) + leftovers; ++
i) {
125 target0[
i] = src0[
i] + src1[
i];
126 target1[
i] = src0[
i] + src2[
i];
127 target2[
i] = src0[
i] + src3[
i];
128 target3[
i] = src0[
i] + src4[
i];
145 unsigned int num_points)
147 const unsigned int eighth_points = num_points / 8;
148 unsigned int number = 0;
150 int16x8_t src0_vec, src1_vec, src2_vec, src3_vec, src4_vec;
151 int16x8_t target0_vec, target1_vec, target2_vec, target3_vec;
152 for (number = 0; number < eighth_points; ++number) {
153 src0_vec = vld1q_s16(src0);
154 src1_vec = vld1q_s16(src1);
155 src2_vec = vld1q_s16(src2);
156 src3_vec = vld1q_s16(src3);
157 src4_vec = vld1q_s16(src4);
159 target0_vec = vaddq_s16(src0_vec, src1_vec);
160 target1_vec = vaddq_s16(src0_vec, src2_vec);
161 target2_vec = vaddq_s16(src0_vec, src3_vec);
162 target3_vec = vaddq_s16(src0_vec, src4_vec);
164 vst1q_s16(target0, target0_vec);
165 vst1q_s16(target1, target1_vec);
166 vst1q_s16(target2, target2_vec);
167 vst1q_s16(target3, target3_vec);
179 for (number = eighth_points * 8; number < num_points; ++number) {
180 *target0++ = *src0 + *src1++;
181 *target1++ = *src0 + *src2++;
182 *target2++ = *src0 + *src3++;
183 *target3++ = *src0++ + *src4++;
static void volk_16i_x5_add_quad_16i_x4_a_sse2(short *target0, short *target1, short *target2, short *target3, short *src0, short *src1, short *src2, short *src3, short *src4, unsigned int num_points)
Definition volk_16i_x5_add_quad_16i_x4.h:61
static void volk_16i_x5_add_quad_16i_x4_neon(short *target0, short *target1, short *target2, short *target3, short *src0, short *src1, short *src2, short *src3, short *src4, unsigned int num_points)
Definition volk_16i_x5_add_quad_16i_x4.h:136
static void volk_16i_x5_add_quad_16i_x4_generic(short *target0, short *target1, short *target2, short *target3, short *src0, short *src1, short *src2, short *src3, short *src4, unsigned int num_points)
Definition volk_16i_x5_add_quad_16i_x4.h:191