1 #ifndef INCLUDED_volk_16i_max_star_horizontal_16i_a_H
2 #define INCLUDED_volk_16i_max_star_horizontal_16i_a_H
16 static inline void volk_16i_max_star_horizontal_16i_a_ssse3(int16_t* target, int16_t* src0,
unsigned int num_bytes) {
18 const static uint8_t shufmask0[16] = {0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff};
19 const static uint8_t shufmask1[16] = {0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d};
20 const static uint8_t andmask0[16] = {0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
21 const static uint8_t andmask1[16] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02};
25 __m128i xmm0, xmm1, xmm2, xmm3, xmm4;
26 __m128i xmm5, xmm6, xmm7, xmm8;
28 xmm4 = _mm_load_si128((__m128i*)shufmask0);
29 xmm5 = _mm_load_si128((__m128i*)shufmask1);
30 xmm6 = _mm_load_si128((__m128i*)andmask0);
31 xmm7 = _mm_load_si128((__m128i*)andmask1);
33 __m128i *p_target, *p_src0;
35 p_target = (__m128i*)target;
36 p_src0 = (__m128i*)src0;
38 int bound = num_bytes >> 5;
39 int intermediate = (num_bytes >> 4) & 1;
40 int leftovers = (num_bytes >> 1) & 7;
45 for(i = 0; i < bound; ++i) {
47 xmm0 = _mm_load_si128(p_src0);
48 xmm1 = _mm_load_si128(&p_src0[1]);
52 xmm2 = _mm_xor_si128(xmm2, xmm2);
55 xmm3 = _mm_hsub_epi16(xmm0, xmm1);
57 xmm2 = _mm_cmpgt_epi16(xmm2, xmm3);
59 xmm8 = _mm_and_si128(xmm2, xmm6);
60 xmm3 = _mm_and_si128(xmm2, xmm7);
63 xmm8 = _mm_add_epi8(xmm8, xmm4);
64 xmm3 = _mm_add_epi8(xmm3, xmm5);
66 xmm0 = _mm_shuffle_epi8(xmm0, xmm8);
67 xmm1 = _mm_shuffle_epi8(xmm1, xmm3);
70 xmm3 = _mm_add_epi16(xmm0, xmm1);
73 _mm_store_si128(p_target, xmm3);
79 for(i = 0; i < intermediate; ++i) {
81 xmm0 = _mm_load_si128(p_src0);
84 xmm2 = _mm_xor_si128(xmm2, xmm2);
87 xmm3 = _mm_hsub_epi16(xmm0, xmm1);
88 xmm2 = _mm_cmpgt_epi16(xmm2, xmm3);
90 xmm8 = _mm_and_si128(xmm2, xmm6);
92 xmm3 = _mm_add_epi8(xmm8, xmm4);
94 xmm0 = _mm_shuffle_epi8(xmm0, xmm3);
96 _mm_storel_pd((
double*)p_target,
bit128_p(&xmm0)->double_vec);
98 p_target = (__m128i*)((int8_t*)p_target + 8);
102 for(i = (bound << 4) + (intermediate << 3); i < (bound << 4) + (intermediate << 3) + leftovers ; i += 2) {
103 target[i>>1] = ((int16_t)(src0[i] - src0[i + 1]) > 0) ? src0[i] : src0[i + 1];
112 #ifdef LV_HAVE_GENERIC
113 static inline void volk_16i_max_star_horizontal_16i_a_generic(int16_t* target, int16_t* src0,
unsigned int num_bytes) {
117 int bound = num_bytes >> 1;
120 for(i = 0; i < bound; i += 2) {
121 target[i >> 1] = ((int16_t) (src0[i] - src0[i + 1]) > 0) ? src0[i] : src0[i+1];