55 #ifndef INCLUDED_volk_16i_x4_quad_max_star_16i_a_H 56 #define INCLUDED_volk_16i_x4_quad_max_star_16i_a_H 67 short* src2,
short* src3,
unsigned int num_points)
69 const unsigned int num_bytes = num_points*2;
73 int bound = (num_bytes >> 4);
74 int bound_copy = bound;
75 int leftovers = (num_bytes >> 1) & 7;
77 __m128i *p_target, *p_src0, *p_src1, *p_src2, *p_src3;
78 p_target = (__m128i*) target;
79 p_src0 = (__m128i*)src0;
80 p_src1 = (__m128i*)src1;
81 p_src2 = (__m128i*)src2;
82 p_src3 = (__m128i*)src3;
84 __m128i xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7, xmm8;
86 while(bound_copy > 0) {
87 xmm1 = _mm_load_si128(p_src0);
88 xmm2 = _mm_load_si128(p_src1);
89 xmm3 = _mm_load_si128(p_src2);
90 xmm4 = _mm_load_si128(p_src3);
92 xmm5 = _mm_setzero_si128();
93 xmm6 = _mm_setzero_si128();
97 xmm1 = _mm_sub_epi16(xmm2, xmm1);
99 xmm3 = _mm_sub_epi16(xmm4, xmm3);
101 xmm5 = _mm_cmpgt_epi16(xmm1, xmm5);
102 xmm6 = _mm_cmpgt_epi16(xmm3, xmm6);
104 xmm2 = _mm_and_si128(xmm5, xmm2);
105 xmm4 = _mm_and_si128(xmm6, xmm4);
106 xmm5 = _mm_andnot_si128(xmm5, xmm7);
107 xmm6 = _mm_andnot_si128(xmm6, xmm8);
109 xmm5 = _mm_add_epi16(xmm2, xmm5);
110 xmm6 = _mm_add_epi16(xmm4, xmm6);
112 xmm1 = _mm_xor_si128(xmm1, xmm1);
114 xmm5 = _mm_sub_epi16(xmm6, xmm5);
118 xmm1 = _mm_cmpgt_epi16(xmm5, xmm1);
121 xmm6 = _mm_and_si128(xmm1, xmm6);
123 xmm1 = _mm_andnot_si128(xmm1, xmm2);
126 xmm1 = _mm_add_epi16(xmm6, xmm1);
129 _mm_store_si128(p_target, xmm1);
195 for(i = bound * 8; i < (bound * 8) + leftovers; ++
i) {
196 temp0 = ((short)(src0[i] - src1[i]) > 0) ? src0[i] : src1[i];
197 temp1 = ((short)(src2[i] - src3[i])>0) ? src2[i] : src3[i];
198 target[
i] = ((short)(temp0 - temp1)>0) ? temp0 : temp1;
207 #include <arm_neon.h> 211 short* src2,
short* src3,
unsigned int num_points)
213 const unsigned int eighth_points = num_points / 8;
216 int16x8_t src0_vec, src1_vec, src2_vec, src3_vec;
217 int16x8_t diff12, diff34;
218 int16x8_t comp0, comp1, comp2, comp3;
219 int16x8_t result1_vec, result2_vec;
221 zeros = vdupq_n_s16(0);
222 for(i=0; i < eighth_points; ++
i) {
223 src0_vec = vld1q_s16(src0);
224 src1_vec = vld1q_s16(src1);
225 src2_vec = vld1q_s16(src2);
226 src3_vec = vld1q_s16(src3);
227 diff12 = vsubq_s16(src0_vec, src1_vec);
228 diff34 = vsubq_s16(src2_vec, src3_vec);
229 comp0 = (int16x8_t)vcgeq_s16(diff12, zeros);
230 comp1 = (int16x8_t)vcltq_s16(diff12, zeros);
231 comp2 = (int16x8_t)vcgeq_s16(diff34, zeros);
232 comp3 = (int16x8_t)vcltq_s16(diff34, zeros);
233 comp0 = vandq_s16(src0_vec, comp0);
234 comp1 = vandq_s16(src1_vec, comp1);
235 comp2 = vandq_s16(src2_vec, comp2);
236 comp3 = vandq_s16(src3_vec, comp3);
238 result1_vec = vaddq_s16(comp0, comp1);
239 result2_vec = vaddq_s16(comp2, comp3);
241 diff12 = vsubq_s16(result1_vec, result2_vec);
242 comp0 = (int16x8_t)vcgeq_s16(diff12, zeros);
243 comp1 = (int16x8_t)vcltq_s16(diff12, zeros);
244 comp0 = vandq_s16(result1_vec, comp0);
245 comp1 = vandq_s16(result2_vec, comp1);
246 result1_vec = vaddq_s16(comp0, comp1);
247 vst1q_s16(target, result1_vec);
257 for(i=eighth_points*8; i < num_points; ++
i) {
258 temp0 = ((short)(*src0 - *src1) > 0) ? *src0 : *src1;
259 temp1 = ((short)(*src2 - *src3) > 0) ? *src2 : *src3;
260 *target++ = ((short)(temp0 - temp1)>0) ? temp0 : temp1;
270 #ifdef LV_HAVE_GENERIC 273 short* src2,
short* src3,
unsigned int num_points)
275 const unsigned int num_bytes = num_points*2;
279 int bound = num_bytes >> 1;
283 for(i = 0; i < bound; ++
i) {
284 temp0 = ((short)(src0[i] - src1[i]) > 0) ? src0[i] : src1[i];
285 temp1 = ((short)(src2[i] - src3[i])>0) ? src2[i] : src3[i];
286 target[
i] = ((short)(temp0 - temp1)>0) ? temp0 : temp1;
for i
Definition: volk_config_fixed.tmpl.h:25
static void volk_16i_x4_quad_max_star_16i_generic(short *target, short *src0, short *src1, short *src2, short *src3, unsigned int num_points)
Definition: volk_16i_x4_quad_max_star_16i.h:272
static void volk_16i_x4_quad_max_star_16i_a_sse2(short *target, short *src0, short *src1, short *src2, short *src3, unsigned int num_points)
Definition: volk_16i_x4_quad_max_star_16i.h:66
static void volk_16i_x4_quad_max_star_16i_neon(short *target, short *src0, short *src1, short *src2, short *src3, unsigned int num_points)
Definition: volk_16i_x4_quad_max_star_16i.h:210