62 unsigned int num_points)
64 const unsigned int num_bytes = num_points * 2;
66 static const uint8_t shufmask0[16] = {
67 0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d,
68 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
70 static const uint8_t shufmask1[16] = {
71 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
72 0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d
74 static const uint8_t andmask0[16] = {
75 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
76 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
78 static const uint8_t andmask1[16] = {
79 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
80 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02
83 __m128i xmm0 = {}, xmm1 = {}, xmm2 = {}, xmm3 = {}, xmm4 = {};
84 __m128i xmm5 = {}, xmm6 = {}, xmm7 = {}, xmm8 = {};
86 xmm4 = _mm_load_si128((__m128i*)shufmask0);
87 xmm5 = _mm_load_si128((__m128i*)shufmask1);
88 xmm6 = _mm_load_si128((__m128i*)andmask0);
89 xmm7 = _mm_load_si128((__m128i*)andmask1);
91 __m128i *p_target, *p_src0;
93 p_target = (__m128i*)target;
94 p_src0 = (__m128i*)src0;
96 int bound = num_bytes >> 5;
97 int intermediate = (num_bytes >> 4) & 1;
98 int leftovers = (num_bytes >> 1) & 7;
102 for (
i = 0;
i < bound; ++
i) {
103 xmm0 = _mm_load_si128(p_src0);
104 xmm1 = _mm_load_si128(&p_src0[1]);
106 xmm2 = _mm_xor_si128(xmm2, xmm2);
109 xmm3 = _mm_hsub_epi16(xmm0, xmm1);
111 xmm2 = _mm_cmpgt_epi16(xmm2, xmm3);
113 xmm8 = _mm_and_si128(xmm2, xmm6);
114 xmm3 = _mm_and_si128(xmm2, xmm7);
117 xmm8 = _mm_add_epi8(xmm8, xmm4);
118 xmm3 = _mm_add_epi8(xmm3, xmm5);
120 xmm0 = _mm_shuffle_epi8(xmm0, xmm8);
121 xmm1 = _mm_shuffle_epi8(xmm1, xmm3);
124 xmm3 = _mm_add_epi16(xmm0, xmm1);
127 _mm_store_si128(p_target, xmm3);
133 xmm0 = _mm_load_si128(p_src0);
135 xmm2 = _mm_xor_si128(xmm2, xmm2);
138 xmm3 = _mm_hsub_epi16(xmm0, xmm1);
139 xmm2 = _mm_cmpgt_epi16(xmm2, xmm3);
141 xmm8 = _mm_and_si128(xmm2, xmm6);
143 xmm3 = _mm_add_epi8(xmm8, xmm4);
145 xmm0 = _mm_shuffle_epi8(xmm0, xmm3);
147 _mm_storel_pd((
double*)p_target,
bit128_p(&xmm0)->double_vec);
149 p_target = (__m128i*)((int8_t*)p_target + 8);
152 for (
i = (bound << 4) + (intermediate << 3);
153 i < (bound << 4) + (intermediate << 3) + leftovers;
155 target[
i >> 1] = ((int16_t)(src0[
i] - src0[
i + 1]) > 0) ? src0[
i] : src0[
i + 1];
166 unsigned int num_points)
168 const unsigned int eighth_points = num_points / 16;
170 int16x8x2_t input_vec;
171 int16x8_t diff, max_vec, zeros;
172 uint16x8_t comp1, comp2;
173 zeros = vdupq_n_s16(0);
174 for (number = 0; number < eighth_points; ++number) {
175 input_vec = vld2q_s16(src0);
177 diff = vsubq_s16(input_vec.val[0], input_vec.val[1]);
178 comp1 = vcgeq_s16(diff, zeros);
179 comp2 = vcltq_s16(diff, zeros);
181 input_vec.val[0] = vandq_s16(input_vec.val[0], (int16x8_t)comp1);
182 input_vec.val[1] = vandq_s16(input_vec.val[1], (int16x8_t)comp2);
184 max_vec = vaddq_s16(input_vec.val[0], input_vec.val[1]);
185 vst1q_s16(target, max_vec);
189 for (number = 0; number < num_points % 16; number += 2) {
190 target[number >> 1] = ((int16_t)(src0[number] - src0[number + 1]) > 0)