doc/doxygen-3.5/volk__16i__max__star__horizontal__16i__a_8h_source.html

00001 #ifndef INCLUDED_volk_16i_max_star_horizontal_16i_a_H
00002 #define INCLUDED_volk_16i_max_star_horizontal_16i_a_H
00003
00004 #include <volk/volk_common.h>
00005
00006 #include<inttypes.h>
00007 #include<stdio.h>
00008
00009
00010 #ifdef LV_HAVE_SSSE3
00011
00012 #include<xmmintrin.h>
00013 #include<emmintrin.h>
00014 #include<tmmintrin.h>
00015
00016 static inline  void volk_16i_max_star_horizontal_16i_a_ssse3(int16_t* target, int16_t* src0, unsigned int num_bytes) {
00017
00018   const static uint8_t shufmask0[16] = {0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff};
00019   const static uint8_t shufmask1[16] = {0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d};
00020   const static uint8_t andmask0[16] = {0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
00021   const static uint8_t andmask1[16] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02};
00022
00023
00024
00025   __m128i xmm0, xmm1, xmm2, xmm3, xmm4;
00026   __m128i  xmm5, xmm6, xmm7, xmm8;
00027
00028   xmm4 = _mm_load_si128((__m128i*)shufmask0);
00029   xmm5 = _mm_load_si128((__m128i*)shufmask1);
00030   xmm6 = _mm_load_si128((__m128i*)andmask0);
00031   xmm7 = _mm_load_si128((__m128i*)andmask1);
00032
00033   __m128i *p_target, *p_src0;
00034
00035   p_target = (__m128i*)target;
00036   p_src0 = (__m128i*)src0;
00037
00038   int bound = num_bytes >> 5;
00039   int intermediate = (num_bytes >> 4) & 1;
00040   int leftovers = (num_bytes >> 1) & 7;
00041
00042   int i = 0;
00043
00044
00045   for(i = 0; i < bound; ++i) {
00046
00047     xmm0 = _mm_load_si128(p_src0);
00048     xmm1 = _mm_load_si128(&p_src0[1]);
00049
00050
00051
00052     xmm2 = _mm_xor_si128(xmm2, xmm2);
00053     p_src0 += 2;
00054
00055     xmm3 = _mm_hsub_epi16(xmm0, xmm1);
00056
00057     xmm2 = _mm_cmpgt_epi16(xmm2, xmm3);
00058
00059     xmm8 = _mm_and_si128(xmm2, xmm6);
00060     xmm3 = _mm_and_si128(xmm2, xmm7);
00061
00062
00063     xmm8 = _mm_add_epi8(xmm8, xmm4);
00064     xmm3 = _mm_add_epi8(xmm3, xmm5);
00065
00066     xmm0 = _mm_shuffle_epi8(xmm0, xmm8);
00067     xmm1 = _mm_shuffle_epi8(xmm1, xmm3);
00068
00069
00070     xmm3 = _mm_add_epi16(xmm0, xmm1);
00071
00072
00073     _mm_store_si128(p_target, xmm3);
00074
00075     p_target += 1;
00076
00077   }
00078
00079   for(i = 0; i < intermediate; ++i) {
00080
00081     xmm0 = _mm_load_si128(p_src0);
00082
00083
00084     xmm2 = _mm_xor_si128(xmm2, xmm2);
00085     p_src0 += 1;
00086
00087     xmm3 = _mm_hsub_epi16(xmm0, xmm1);
00088     xmm2 = _mm_cmpgt_epi16(xmm2, xmm3);
00089
00090     xmm8 = _mm_and_si128(xmm2, xmm6);
00091
00092     xmm3 = _mm_add_epi8(xmm8, xmm4);
00093
00094     xmm0 = _mm_shuffle_epi8(xmm0, xmm3);
00095
00096     _mm_storel_pd((double*)p_target, bit128_p(&xmm0)->double_vec);
00097
00098     p_target = (__m128i*)((int8_t*)p_target + 8);
00099
00100   }
00101
00102   for(i = (bound << 4) + (intermediate << 3); i < (bound << 4) + (intermediate << 3) + leftovers ; i += 2) {
00103     target[i>>1] = ((int16_t)(src0[i] - src0[i + 1]) > 0) ? src0[i] : src0[i + 1];
00104   }
00105
00106
00107 }
00108
00109 #endif /*LV_HAVE_SSSE3*/
00110
00111
00112 #ifdef LV_HAVE_GENERIC
00113 static inline void volk_16i_max_star_horizontal_16i_a_generic(int16_t* target, int16_t* src0, unsigned int num_bytes) {
00114
00115         int i = 0;
00116
00117         int bound = num_bytes >> 1;
00118
00119
00120         for(i = 0; i < bound; i += 2) {
00121           target[i >> 1] = ((int16_t) (src0[i] - src0[i + 1]) > 0) ? src0[i] : src0[i+1];
00122         }
00123
00124 }
00125
00126
00127
00128 #endif /*LV_HAVE_GENERIC*/
00129
00130 #endif /*INCLUDED_volk_16i_max_star_horizontal_16i_a_H*/