source: trunk/lib/idisa_cpp/idisa_sse4_2.cpp @ 1740

Last change on this file since 1740 was 1740, checked in by huah, 8 years ago

added support for ARM NEON

File size: 148.8 KB
Line 
1#ifndef IDISA_SSE4_2_CPP
2#define IDISA_SSE4_2_CPP
3#include <stdint.h>
4#include "../config.hpp"
5
6#include "smmintrin.h"
7
8typedef __m128i bitblock128_t;
9template <uint32_t fw>
10class simd128
11{
12public:
13        static IDISA_ALWAYS_INLINE bitblock128_t max(bitblock128_t arg1, bitblock128_t arg2);
14        static IDISA_ALWAYS_INLINE bitblock128_t mult(bitblock128_t arg1, bitblock128_t arg2);
15        static IDISA_ALWAYS_INLINE bitblock128_t gt(bitblock128_t arg1, bitblock128_t arg2);
16        static IDISA_ALWAYS_INLINE bitblock128_t umult(bitblock128_t arg1, bitblock128_t arg2);
17        static IDISA_ALWAYS_INLINE bitblock128_t ult(bitblock128_t arg1, bitblock128_t arg2);
18        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
19        static IDISA_ALWAYS_INLINE bitblock128_t ctz(bitblock128_t arg1);
20        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t shift_mask);
21        static IDISA_ALWAYS_INLINE bitblock128_t eq(bitblock128_t arg1, bitblock128_t arg2);
22        static IDISA_ALWAYS_INLINE bitblock128_t popcount(bitblock128_t arg1);
23        static IDISA_ALWAYS_INLINE bitblock128_t neg(bitblock128_t arg1);
24        static IDISA_ALWAYS_INLINE bitblock128_t himask();
25        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
26        static IDISA_ALWAYS_INLINE bitblock128_t ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
27        static IDISA_ALWAYS_INLINE bitblock128_t sub(bitblock128_t arg1, bitblock128_t arg2);
28        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1);
29        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t shift_mask);
30        static IDISA_ALWAYS_INLINE bitblock128_t lomask();
31        static IDISA_ALWAYS_INLINE bitblock128_t umin(bitblock128_t arg1, bitblock128_t arg2);
32        template <uint64_t val> static IDISA_ALWAYS_INLINE bitblock128_t constant();
33        static IDISA_ALWAYS_INLINE bitblock128_t min(bitblock128_t arg1, bitblock128_t arg2);
34        static IDISA_ALWAYS_INLINE bitblock128_t umax(bitblock128_t arg1, bitblock128_t arg2);
35        static IDISA_ALWAYS_INLINE bitblock128_t abs(bitblock128_t arg1);
36        static IDISA_ALWAYS_INLINE bitblock128_t xor_hl(bitblock128_t arg1);
37        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srai(bitblock128_t arg1);
38        static IDISA_ALWAYS_INLINE bitblock128_t lt(bitblock128_t arg1, bitblock128_t arg2);
39        static IDISA_ALWAYS_INLINE bitblock128_t add(bitblock128_t arg1, bitblock128_t arg2);
40        static IDISA_ALWAYS_INLINE bitblock128_t ugt(bitblock128_t arg1, bitblock128_t arg2);
41};
42
43template <uint32_t fw>
44class hsimd128
45{
46public:
47        static IDISA_ALWAYS_INLINE bitblock128_t umin_hl(bitblock128_t arg1, bitblock128_t arg2);
48        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1, bitblock128_t arg2);
49        static IDISA_ALWAYS_INLINE bitblock128_t packss(bitblock128_t arg1, bitblock128_t arg2);
50        static IDISA_ALWAYS_INLINE bitblock128_t packh(bitblock128_t arg1, bitblock128_t arg2);
51        static IDISA_ALWAYS_INLINE uint64_t signmask(bitblock128_t arg1);
52        static IDISA_ALWAYS_INLINE bitblock128_t packl(bitblock128_t arg1, bitblock128_t arg2);
53        static IDISA_ALWAYS_INLINE bitblock128_t min_hl(bitblock128_t arg1, bitblock128_t arg2);
54        static IDISA_ALWAYS_INLINE bitblock128_t packus(bitblock128_t arg1, bitblock128_t arg2);
55};
56
57template <uint32_t fw>
58class esimd128
59{
60public:
61        static IDISA_ALWAYS_INLINE bitblock128_t mergel(bitblock128_t arg1, bitblock128_t arg2);
62        static IDISA_ALWAYS_INLINE bitblock128_t signextendh(bitblock128_t arg1);
63        static IDISA_ALWAYS_INLINE bitblock128_t mergeh(bitblock128_t arg1, bitblock128_t arg2);
64        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendh(bitblock128_t arg1);
65        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendl(bitblock128_t arg1);
66        static IDISA_ALWAYS_INLINE bitblock128_t signextendl(bitblock128_t arg1);
67};
68
69template <uint32_t fw>
70class mvmd128
71{
72public:
73        template <uint64_t msk> static IDISA_ALWAYS_INLINE bitblock128_t shufflei(bitblock128_t arg1);
74        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dsrli(bitblock128_t arg1, bitblock128_t arg2);
75        static IDISA_ALWAYS_INLINE bitblock128_t fill(uint64_t val1);
76        static IDISA_ALWAYS_INLINE bitblock128_t shuffle(bitblock128_t arg1, bitblock128_t arg2);
77        template <uint64_t pos> static IDISA_ALWAYS_INLINE uint64_t extract(bitblock128_t arg1);
78        template <uint64_t pos> static IDISA_ALWAYS_INLINE bitblock128_t splat(bitblock128_t arg1);
79        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
80        static IDISA_ALWAYS_INLINE bitblock128_t fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
81        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
82        static IDISA_ALWAYS_INLINE bitblock128_t fill2(uint64_t val1, uint64_t val2);
83        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dslli(bitblock128_t arg1, bitblock128_t arg2);
84        static IDISA_ALWAYS_INLINE bitblock128_t fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
85        static IDISA_ALWAYS_INLINE bitblock128_t fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
86};
87
88class bitblock128
89{
90public:
91        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
92        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(bitblock128_t* arg1);
93        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
94        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
95        static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t* arg1, bitblock128_t arg2);
96        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
97        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
98        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
99        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
100        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(bitblock128_t* arg1);
101        static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t* arg1, bitblock128_t arg2);
102};
103
104//Declaration Part
105IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2);
106IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1);
107IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2);
108IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2);
109IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2);
110IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2);
111template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2);
112template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2);
113template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2);
114template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2);
115template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2);
116template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2);
117template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2);
118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2);
119template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2);
120template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2);
121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2);
122template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2);
123template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2);
124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2);
125template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2);
126template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2);
127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2);
128template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2);
129template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2);
130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2);
131template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2);
132template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2);
133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2);
134template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2);
135template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2);
136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2);
137template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2);
138template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2);
139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2);
140template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2);
141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2);
142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2);
143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2);
144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2);
145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2);
146template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2);
147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2);
148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2);
149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2);
150template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2);
151template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2);
152template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2);
153template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2);
154template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2);
155template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2);
156template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2);
157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2);
158template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1);
159template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1);
160template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1);
161template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1);
162template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1);
163template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1);
164template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1);
165template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1);
166template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1);
167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1);
168template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1);
169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1);
170template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1);
171template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1);
172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1);
173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2);
176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2);
177template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2);
178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2);
179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2);
180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2);
181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2);
182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2);
183template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1);
184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1);
185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1);
186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1);
187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1);
188template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1);
189template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1);
190template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1);
191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1);
192template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1);
193template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1);
194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1);
195template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1);
196template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1);
197template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1);
198template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1);
199template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1);
200template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1);
201template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1);
202template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1);
203template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1);
204template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1);
205template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1);
206template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1);
207template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1);
208template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1);
209template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1);
210template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1);
211template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1);
212template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
213template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
214template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
215template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
216template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
217template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
219template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
220template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2);
221template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2);
222template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2);
223template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2);
224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2);
225template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2);
226template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2);
227template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2);
228template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1);
229template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1);
230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1);
231template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1);
232template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1);
233template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1);
234template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1);
235template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
236template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
237template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
238template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
239template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
240template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
241template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
242template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
243template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
244template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
245template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2);
246template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2);
247template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2);
248template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2);
249template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2);
250template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2);
251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2);
252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2);
253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask();
254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask();
255template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask();
256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask();
257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask();
258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask();
259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask();
260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2);
261template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2);
262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2);
263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2);
264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2);
265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2);
266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2);
267template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2);
268template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1);
269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1);
270template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1);
271template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1);
272template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1);
273template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1);
274template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1);
275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2);
276template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2);
277template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2);
278template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2);
279template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2);
280template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2);
281template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2);
282template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2);
283template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1);
284template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1);
285template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1);
286template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1);
287template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1);
288template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1);
289template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1);
290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask();
291template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask();
292template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask();
293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask();
294template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask();
295template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask();
296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask();
297template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2);
298template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2);
299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2);
300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2);
301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2);
302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2);
303template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2);
304template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2);
305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2);
306template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2);
307template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2);
308template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2);
309template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2);
310template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2);
311template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2);
312template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2);
313template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
314template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
315template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
316template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
317template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
318template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
319template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
320template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
321template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
322template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
323template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
324template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
325template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
326template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
327template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2);
328template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2);
329template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2);
330template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2);
331template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2);
332template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2);
333template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2);
334template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1);
335template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1);
336template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1);
337template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1);
338template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1);
339template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1);
340template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2);
341template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2);
342template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2);
343template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2);
344template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2);
345template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2);
346template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2);
347template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2);
348template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2);
349template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2);
350template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2);
351template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2);
352template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2);
353template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2);
354template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
355template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
356template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
357template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
358template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
359template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
360template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
361template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2);
362template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2);
363template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2);
364template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2);
365template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2);
366template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2);
367template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2);
368template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2);
369template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2);
370template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2);
371template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2);
372template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2);
373template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2);
374template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2);
375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
376template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
377template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
379template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
380template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
382template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1);
383template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1);
384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1);
385template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1);
386template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1);
387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1);
388template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1);
389template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1);
390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1);
391template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1);
392template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1);
393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1);
394template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1);
395template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1);
396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1);
397template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1);
398template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1);
399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1);
400template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1);
401template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1);
402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1);
403template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1);
404template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1);
405template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1);
406template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1);
407template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1);
408template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1);
409template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1);
410template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1);
411template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1);
412template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1);
413template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
414template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
415template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
416template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
417template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
418template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
419template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
420template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1);
421template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1);
422template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1);
423template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1);
424template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1);
425template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1);
426template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1);
427template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1);
428template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
429template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
430template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
431template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
432template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1);
433template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1);
434template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1);
435template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1);
436template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1);
437template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1);
438template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1);
439template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1);
440template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1);
441template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1);
442template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1);
443template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1);
444template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1);
445template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1);
446template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1);
447template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
448template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
449template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
450template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
451template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
452template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
453template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
454template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
455template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
456template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
457template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1);
458template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1);
459template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1);
460template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1);
461template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1);
462template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1);
463template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1);
464template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2);
465template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2);
466template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2);
467template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2);
468template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2);
469template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2);
470template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2);
471template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2);
472template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2);
473template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2);
474template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2);
475template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2);
476template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2);
477template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2);
478template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1);
479template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1);
480template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1);
481template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1);
482template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1);
483template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1);
484template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1);
485template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
486template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
487template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
488template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
489template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
490
491//Implementation Part
492
493#define shufflemask4(s1, s2, s3, s4) \
494        ((s1<<6) | (s2<<4) | (s3<<2) | s4)
495
496#define shufflemask4_from_shufflemask2(msk) \
497        (msk==3 ? 238 : (msk==2 ? 228 : (msk==1 ? 78 : 68)))
498
499#define shufflemask8_to_shufflemask4(msk) \
500        ((msk&3) | (((msk>>3)&3)<<2) | (((msk>>6)&3)<<4) | (((msk>>9)&3)<<6) | (((msk>>12)&3)<<8) | (((msk>>15)&3)<<10) | (((msk>>18)&3)<<12) | (((msk>>21)&3)<<14))
501//The total number of operations is 2
502IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2)
503{
504        return simd_not(simd_or(arg1, arg2));
505}
506
507//The total number of operations is 1
508IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1)
509{
510        return simd_xor(arg1, simd128<32>::constant<-1>());
511}
512
513//The total number of operations is 1
514IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2)
515{
516        return _mm_andnot_si128(arg2, arg1);
517}
518
519//The total number of operations is 1
520IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2)
521{
522        return _mm_or_si128(arg1, arg2);
523}
524
525//The total number of operations is 1
526IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2)
527{
528        return _mm_and_si128(arg1, arg2);
529}
530
531//The total number of operations is 1
532IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2)
533{
534        return _mm_xor_si128(arg1, arg2);
535}
536
537//The total number of operations is 1
538template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2)
539{
540        return simd_and(arg1, arg2);
541}
542
543//The total number of operations is 18
544template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2)
545{
546        return simd128<1>::ifh(simd128<2>::lt(arg1, arg2), arg2, arg1);
547}
548
549//The total number of operations is 9
550template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2)
551{
552        bitblock128_t high_bit = simd128<4>::constant<(8)>();
553        return simd_xor(simd128<4>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
554}
555
556//The total number of operations is 1
557template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2)
558{
559        return _mm_max_epi8(arg1, arg2);
560}
561
562//The total number of operations is 1
563template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2)
564{
565        return _mm_max_epi16(arg1, arg2);
566}
567
568//The total number of operations is 1
569template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2)
570{
571        return _mm_max_epi32(arg1, arg2);
572}
573
574//The total number of operations is 4
575template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2)
576{
577        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg1, arg2);
578}
579
580//The total number of operations is 30
581template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2)
582{
583        bitblock128_t hiAns = simd128<(64)>::max(arg1, arg2);
584        bitblock128_t loAns = simd128<(64)>::umax(arg1, arg2);
585        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
586        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
587        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
588}
589
590//The total number of operations is 1
591template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2)
592{
593        return simd_and(arg1, arg2);
594}
595
596//The total number of operations is 23
597template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2)
598{
599        bitblock128_t tmp1 = simd128<128>::slli<1>(arg1);
600        bitblock128_t tmp2 = simd128<128>::slli<1>(arg2);
601        return simd128<1>::ifh(simd128<2>::himask(), simd_or(simd_and(tmp1, simd_and(arg2, simd_or(simd_not(arg1), simd_not(tmp2)))), simd_and(arg1, simd_and(tmp2, simd_or(simd_not(tmp1), simd_not(arg2))))), simd_and(arg1, arg2));
602}
603
604//The total number of operations is 31
605template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2)
606{
607        bitblock128_t loMask = simd128<(8)>::lomask();
608        bitblock128_t tmpAns1 = simd128<(8)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
609        bitblock128_t tmpAns2 = simd128<(8)>::mult(simd128<(8)>::srli<4>(arg1), simd128<(8)>::srli<4>(arg2));
610        return simd128<1>::ifh(loMask, tmpAns1, simd128<(8)>::slli<4>(tmpAns2));
611}
612
613//The total number of operations is 10
614template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2)
615{
616        bitblock128_t loMask = simd128<(16)>::lomask();
617        bitblock128_t tmpAns1 = simd128<(16)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
618        bitblock128_t tmpAns2 = simd128<(16)>::mult(simd128<(16)>::srli<8>(arg1), simd128<(16)>::srli<8>(arg2));
619        return simd128<1>::ifh(loMask, tmpAns1, simd128<(16)>::slli<8>(tmpAns2));
620}
621
622//The total number of operations is 1
623template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2)
624{
625        return _mm_mullo_epi16(arg1, arg2);
626}
627
628//The total number of operations is 1
629template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2)
630{
631        return _mm_mullo_epi32(arg1, arg2);
632}
633
634//The total number of operations is 11
635template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2)
636{
637        bitblock128_t loMask = simd128<64>::lomask();
638        bitblock128_t arg1_low = simd_and(arg1, loMask);
639        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg1);
640        bitblock128_t arg2_low = simd_and(arg2, loMask);
641        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg2);
642        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
643        bitblock128_t tmpAns2 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
644        bitblock128_t tmpAns3 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
645        return simd128<64>::add(tmpAns1, simd128<64>::add(tmpAns2, tmpAns3));
646}
647
648//The total number of operations is 205
649template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2)
650{
651        bitblock128_t loMask = simd128<128>::lomask();
652        bitblock128_t arg1_low = simd_and(arg1, loMask);
653        bitblock128_t arg1_high = simd128<128>::srli<(64)>(arg1);
654        bitblock128_t arg2_low = simd_and(arg2, loMask);
655        bitblock128_t arg2_high = simd128<128>::srli<(64)>(arg2);
656        bitblock128_t tmpAns1 = simd128<(64)>::umult(arg1_low, arg2_low);
657        bitblock128_t tmpAns2 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_low, arg2_high));
658        bitblock128_t tmpAns3 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_high, arg2_low));
659        return simd128<128>::add(tmpAns1, simd128<128>::add(tmpAns2, tmpAns3));
660}
661
662//The total number of operations is 1
663template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2)
664{
665        return simd_andc(arg2, arg1);
666}
667
668//The total number of operations is 15
669template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2)
670{
671        bitblock128_t hiAns = simd128<(1)>::gt(arg1, arg2);
672        bitblock128_t loAns = simd128<(1)>::ugt(arg1, arg2);
673        bitblock128_t mask = simd_and(loAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
674        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
675        return simd_or(simd128<2>::srai<(1)>(hiAns), mask);
676}
677
678//The total number of operations is 10
679template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2)
680{
681        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::gt(simd_and(simd128<(8)>::himask(), arg1), arg2), simd128<(8)>::gt(simd128<(8)>::slli<4>(arg1), simd128<(8)>::slli<4>(arg2)));
682}
683
684//The total number of operations is 1
685template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2)
686{
687        return _mm_cmpgt_epi8(arg1, arg2);
688}
689
690//The total number of operations is 1
691template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2)
692{
693        return _mm_cmpgt_epi16(arg1, arg2);
694}
695
696//The total number of operations is 1
697template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2)
698{
699        return _mm_cmpgt_epi32(arg1, arg2);
700}
701
702//The total number of operations is 1
703template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2)
704{
705        return _mm_cmpgt_epi64(arg1, arg2);
706}
707
708//The total number of operations is 32
709template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2)
710{
711        bitblock128_t hiAns = simd128<(64)>::gt(arg1, arg2);
712        bitblock128_t loAns = simd128<(64)>::ugt(arg1, arg2);
713        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
714        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
715        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
716}
717
718//The total number of operations is 289
719template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2)
720{
721        bitblock128_t loMask = simd128<(2)>::lomask();
722        bitblock128_t tmpAns1 = simd128<(2)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
723        bitblock128_t tmpAns2 = simd128<(2)>::umult(simd_and(loMask, simd128<(4)>::srli<(2)>(arg1)), simd_and(loMask, simd128<(4)>::srli<(2)>(arg2)));
724        return simd_or(tmpAns1, simd128<(4)>::slli<(2)>(tmpAns2));
725}
726
727//The total number of operations is 139
728template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2)
729{
730        bitblock128_t loMask = simd128<(4)>::lomask();
731        bitblock128_t tmpAns1 = simd128<(4)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
732        bitblock128_t tmpAns2 = simd128<(4)>::umult(simd_and(loMask, simd128<(8)>::srli<(4)>(arg1)), simd_and(loMask, simd128<(8)>::srli<(4)>(arg2)));
733        return simd_or(tmpAns1, simd128<(8)>::slli<(4)>(tmpAns2));
734}
735
736//The total number of operations is 64
737template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2)
738{
739        bitblock128_t loMask = simd128<(8)>::lomask();
740        bitblock128_t tmpAns1 = simd128<(8)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
741        bitblock128_t tmpAns2 = simd128<(8)>::umult(simd_and(loMask, simd128<(16)>::srli<(8)>(arg1)), simd_and(loMask, simd128<(16)>::srli<(8)>(arg2)));
742        return simd_or(tmpAns1, simd128<(16)>::slli<(8)>(tmpAns2));
743}
744
745//The total number of operations is 28
746template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2)
747{
748        bitblock128_t loMask = simd128<(16)>::lomask();
749        bitblock128_t tmpAns1 = simd128<(16)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
750        bitblock128_t tmpAns2 = simd128<(16)>::umult(simd_and(loMask, simd128<(32)>::srli<(16)>(arg1)), simd_and(loMask, simd128<(32)>::srli<(16)>(arg2)));
751        return simd_or(tmpAns1, simd128<(32)>::slli<(16)>(tmpAns2));
752}
753
754//The total number of operations is 10
755template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2)
756{
757        bitblock128_t loMask = simd128<(32)>::lomask();
758        bitblock128_t tmpAns1 = simd128<(32)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
759        bitblock128_t tmpAns2 = simd128<(32)>::umult(simd_and(loMask, simd128<(64)>::srli<(32)>(arg1)), simd_and(loMask, simd128<(64)>::srli<(32)>(arg2)));
760        return simd_or(tmpAns1, simd128<(64)>::slli<(32)>(tmpAns2));
761}
762
763//The total number of operations is 1
764template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2)
765{
766        return _mm_mul_epu32(arg1, arg2);
767}
768
769//The total number of operations is 55
770template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2)
771{
772        bitblock128_t loMask1 = simd128<(128)>::lomask();
773        bitblock128_t arg11 = simd_and(arg1, loMask1);
774        bitblock128_t arg22 = simd_and(arg2, loMask1);
775        bitblock128_t loMask2 = simd128<64>::lomask();
776        bitblock128_t arg1_low = simd_and(arg11, loMask2);
777        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg11);
778        bitblock128_t arg2_low = simd_and(arg22, loMask2);
779        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg22);
780        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
781        bitblock128_t tmpAns2 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
782        bitblock128_t tmpAns3 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
783        bitblock128_t tmpAns4 = simd128<(128)>::slli<64>(simd128<(32)>::umult(arg1_high, arg2_high));
784        return simd128<(128)>::add(tmpAns1, simd128<(128)>::add(tmpAns2, simd128<(128)>::add(tmpAns3, tmpAns4)));
785}
786
787//The total number of operations is 1
788template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2)
789{
790        return simd_andc(arg2, arg1);
791}
792
793//The total number of operations is 14
794template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2)
795{
796        bitblock128_t tmpAns = simd128<(1)>::ult(arg1, arg2);
797        bitblock128_t mask = simd_and(tmpAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
798        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
799        return simd_or(simd128<2>::srai<(1)>(tmpAns), mask);
800}
801
802//The total number of operations is 20
803template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2)
804{
805        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::ult(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::ult(simd_andc(arg1, simd128<(8)>::himask()), simd_andc(arg2, simd128<(8)>::himask())));
806}
807
808//The total number of operations is 7
809template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2)
810{
811        bitblock128_t high_bit = simd128<8>::constant<(128)>();
812        return simd128<8>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
813}
814
815//The total number of operations is 7
816template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2)
817{
818        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
819        return simd128<16>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
820}
821
822//The total number of operations is 7
823template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2)
824{
825        bitblock128_t high_bit = simd128<32>::constant<(2147483648ULL)>();
826        return simd128<32>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
827}
828
829//The total number of operations is 7
830template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2)
831{
832        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808ULL)>();
833        return simd128<64>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
834}
835
836//The total number of operations is 35
837template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2)
838{
839        bitblock128_t tmpAns = simd128<(64)>::ult(arg1, arg2);
840        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
841        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
842        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
843}
844
845//The total number of operations is 1
846template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2)
847{
848        return simd_andc(arg1, arg2);
849}
850
851//The total number of operations is 15
852template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2)
853{
854        bitblock128_t hiAns = simd128<(1)>::lt(arg1, arg2);
855        bitblock128_t loAns = simd128<(1)>::ult(arg1, arg2);
856        bitblock128_t mask = simd_and(loAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
857        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
858        return simd_or(simd128<2>::srai<(1)>(hiAns), mask);
859}
860
861//The total number of operations is 18
862template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2)
863{
864        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::lt(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::lt(simd128<(8)>::slli<4>(arg1), simd128<(8)>::slli<4>(arg2)));
865}
866
867//The total number of operations is 5
868template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2)
869{
870        return simd_and(simd_not(simd128<8>::gt(arg1, arg2)), simd_not(simd128<8>::eq(arg1, arg2)));
871}
872
873//The total number of operations is 5
874template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2)
875{
876        return simd_and(simd_not(simd128<16>::gt(arg1, arg2)), simd_not(simd128<16>::eq(arg1, arg2)));
877}
878
879//The total number of operations is 5
880template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2)
881{
882        return simd_and(simd_not(simd128<32>::gt(arg1, arg2)), simd_not(simd128<32>::eq(arg1, arg2)));
883}
884
885//The total number of operations is 5
886template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2)
887{
888        return simd_and(simd_not(simd128<64>::gt(arg1, arg2)), simd_not(simd128<64>::eq(arg1, arg2)));
889}
890
891//The total number of operations is 40
892template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2)
893{
894        bitblock128_t hiAns = simd128<(64)>::lt(arg1, arg2);
895        bitblock128_t loAns = simd128<(64)>::ult(arg1, arg2);
896        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
897        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
898        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
899}
900
901//The total number of operations is 2
902template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1)
903{
904        return simd_and(simd128<32>::srli<sh>(arg1), simd128<2>::constant<((3)>>sh)>());
905}
906
907//The total number of operations is 2
908template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1)
909{
910        return simd_and(simd128<32>::srli<sh>(arg1), simd128<4>::constant<((15)>>sh)>());
911}
912
913//The total number of operations is 2
914template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1)
915{
916        return simd_and(simd128<32>::srli<sh>(arg1), simd128<8>::constant<((255)>>sh)>());
917}
918
919//The total number of operations is 1
920template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1)
921{
922        return _mm_srli_epi16(arg1, (int32_t)(sh));
923}
924
925//The total number of operations is 1
926template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1)
927{
928        return _mm_srli_epi32(arg1, (int32_t)(sh));
929}
930
931//The total number of operations is 1
932template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1)
933{
934        return _mm_srli_epi64(arg1, (int32_t)(sh));
935}
936
937//The total number of operations is 4
938template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1)
939{
940        return (((sh%8) == 0) ? _mm_srli_si128(arg1, (int32_t)((sh/8))) : ((sh >= 64) ? simd128<64>::srli<(sh-64)>(_mm_srli_si128(arg1, (int32_t)(8))) : simd_or(simd128<64>::srli<sh>(arg1), _mm_srli_si128(simd128<64>::slli<(64-sh)>(arg1), (int32_t)(8)))));
941}
942
943//The total number of operations is 1
944template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1)
945{
946        return simd_not(arg1);
947}
948
949//The total number of operations is 14
950template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1)
951{
952        bitblock128_t tmp = simd_not(arg1);
953        return simd128<1>::ifh(simd128<2>::himask(), simd_and(tmp, simd128<128>::slli<1>(tmp)), simd_and(simd128<128>::srli<1>(arg1), tmp));
954}
955
956//The total number of operations is 14
957template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1)
958{
959        return simd128<4>::popcount(simd_andc(simd128<4>::sub(arg1, simd128<4>::constant<1>()), arg1));
960}
961
962//The total number of operations is 13
963template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1)
964{
965        return simd128<8>::popcount(simd_andc(simd128<8>::sub(arg1, simd128<8>::constant<1>()), arg1));
966}
967
968//The total number of operations is 16
969template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1)
970{
971        return simd128<16>::popcount(simd_andc(simd128<16>::sub(arg1, simd128<16>::constant<1>()), arg1));
972}
973
974//The total number of operations is 19
975template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1)
976{
977        return simd128<32>::popcount(simd_andc(simd128<32>::sub(arg1, simd128<32>::constant<1>()), arg1));
978}
979
980//The total number of operations is 14
981template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1)
982{
983        return simd128<64>::popcount(simd_andc(simd128<64>::sub(arg1, simd128<64>::constant<1>()), arg1));
984}
985
986//The total number of operations is 30
987template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1)
988{
989        return simd128<128>::popcount(simd_andc(simd128<128>::sub(arg1, simd128<128>::constant<1>()), arg1));
990}
991
992//The total number of operations is 7
993template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
994{
995        return simd128<1>::ifh(simd128<128>::himask(), _mm_sll_epi64(arg1, _mm_srli_si128(shift_mask, (int32_t)(8))), _mm_sll_epi64(arg1, simd_andc(shift_mask, simd128<128>::himask())));
996}
997
998//The total number of operations is 11
999template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1000{
1001        return simd_or(_mm_sll_epi64(arg1, shift_mask), simd_or(_mm_slli_si128(_mm_sll_epi64(arg1, simd128<32>::sub(shift_mask, _mm_cvtsi32_si128((int32_t)(64)))), (int32_t)(8)), _mm_slli_si128(_mm_srl_epi64(arg1, simd128<32>::sub(_mm_cvtsi32_si128((int32_t)(64)), shift_mask)), (int32_t)(8))));
1002}
1003
1004//The total number of operations is 1
1005template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1006{
1007        return simd_andc(arg1, arg2);
1008}
1009
1010//The total number of operations is 14
1011template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1012{
1013        bitblock128_t tmpAns = simd128<(1)>::ugt(arg1, arg2);
1014        bitblock128_t mask = simd_and(tmpAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
1015        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
1016        return simd_or(simd128<2>::srai<(1)>(tmpAns), mask);
1017}
1018
1019//The total number of operations is 12
1020template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1021{
1022        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::ugt(simd_and(simd128<(8)>::himask(), arg1), arg2), simd128<(8)>::ugt(simd_andc(arg1, simd128<(8)>::himask()), simd_andc(arg2, simd128<(8)>::himask())));
1023}
1024
1025//The total number of operations is 3
1026template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1027{
1028        bitblock128_t high_bit = simd128<8>::constant<(128)>();
1029        return simd128<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1030}
1031
1032//The total number of operations is 3
1033template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1034{
1035        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1036        return simd128<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1037}
1038
1039//The total number of operations is 3
1040template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1041{
1042        bitblock128_t high_bit = simd128<32>::constant<(2147483648ULL)>();
1043        return simd128<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1044}
1045
1046//The total number of operations is 3
1047template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1048{
1049        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808ULL)>();
1050        return simd128<64>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1051}
1052
1053//The total number of operations is 31
1054template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1055{
1056        bitblock128_t tmpAns = simd128<(64)>::ugt(arg1, arg2);
1057        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
1058        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
1059        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
1060}
1061
1062//The total number of operations is 4
1063template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1)
1064{
1065        return simd_xor(simd128<2>::srli<(1)>(arg1), simd_and(arg1, simd128<2>::lomask()));
1066}
1067
1068//The total number of operations is 4
1069template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1)
1070{
1071        return simd_xor(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1072}
1073
1074//The total number of operations is 4
1075template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1)
1076{
1077        return simd_xor(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1078}
1079
1080//The total number of operations is 3
1081template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1)
1082{
1083        return simd_xor(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1084}
1085
1086//The total number of operations is 3
1087template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1)
1088{
1089        return simd_xor(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1090}
1091
1092//The total number of operations is 3
1093template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1)
1094{
1095        return simd_xor(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1096}
1097
1098//The total number of operations is 6
1099template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1)
1100{
1101        return simd_xor(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1102}
1103
1104//The total number of operations is 0
1105template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1)
1106{
1107        return arg1;
1108}
1109
1110//The total number of operations is 3
1111template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1)
1112{
1113        return simd128<2>::add_hl(simd128<(1)>::popcount(arg1));
1114}
1115
1116//The total number of operations is 7
1117template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1)
1118{
1119        return simd128<4>::add_hl(simd128<(2)>::popcount(arg1));
1120}
1121
1122//The total number of operations is 11
1123template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1)
1124{
1125        return simd128<8>::add_hl(simd128<(4)>::popcount(arg1));
1126}
1127
1128//The total number of operations is 14
1129template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1)
1130{
1131        return simd128<16>::add_hl(simd128<(8)>::popcount(arg1));
1132}
1133
1134//The total number of operations is 17
1135template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1)
1136{
1137        return simd128<32>::add_hl(simd128<(16)>::popcount(arg1));
1138}
1139
1140//The total number of operations is 12
1141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1)
1142{
1143        return _mm_sad_epu8(simd128<8>::popcount(arg1), simd128<8>::constant<0>());
1144}
1145
1146//The total number of operations is 18
1147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1)
1148{
1149        bitblock128_t tmpAns = simd128<(64)>::popcount(arg1);
1150        return simd128<(64)>::add(simd_and(tmpAns, simd128<128>::lomask()), simd128<128>::srli<(64)>(tmpAns));
1151}
1152
1153//The total number of operations is 8
1154template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1)
1155{
1156        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(arg1, simd128<128>::slli<1>(arg1)), arg1);
1157}
1158
1159//The total number of operations is 6
1160template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1)
1161{
1162        return simd128<4>::sub(simd128<4>::constant<0>(), arg1);
1163}
1164
1165//The total number of operations is 1
1166template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1)
1167{
1168        return simd128<8>::sub(simd128<8>::constant<0>(), arg1);
1169}
1170
1171//The total number of operations is 1
1172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1)
1173{
1174        return simd128<16>::sub(simd128<16>::constant<0>(), arg1);
1175}
1176
1177//The total number of operations is 1
1178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1)
1179{
1180        return _mm_sign_epi32(arg1, simd128<32>::constant<-1>());
1181}
1182
1183//The total number of operations is 1
1184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1)
1185{
1186        return simd128<64>::sub(simd128<64>::constant<0>(), arg1);
1187}
1188
1189//The total number of operations is 11
1190template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1)
1191{
1192        return simd128<128>::sub(simd128<128>::constant<0>(), arg1);
1193}
1194
1195//The total number of operations is 2
1196template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1)
1197{
1198        return simd_and(simd128<32>::slli<sh>(arg1), simd128<2>::constant<(((3)<<sh)&(3))>());
1199}
1200
1201//The total number of operations is 2
1202template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1)
1203{
1204        return simd_and(simd128<32>::slli<sh>(arg1), simd128<4>::constant<(((15)<<sh)&(15))>());
1205}
1206
1207//The total number of operations is 2
1208template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1)
1209{
1210        return simd_and(simd128<32>::slli<sh>(arg1), simd128<8>::constant<(((255)<<sh)&(255))>());
1211}
1212
1213//The total number of operations is 1
1214template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1)
1215{
1216        return _mm_slli_epi16(arg1, (int32_t)(sh));
1217}
1218
1219//The total number of operations is 1
1220template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1)
1221{
1222        return _mm_slli_epi32(arg1, (int32_t)(sh));
1223}
1224
1225//The total number of operations is 1
1226template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1)
1227{
1228        return _mm_slli_epi64(arg1, (int32_t)(sh));
1229}
1230
1231//The total number of operations is 4
1232template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1)
1233{
1234        return (((sh%8) == 0) ? _mm_slli_si128(arg1, (int32_t)((sh/8))) : ((sh >= 64) ? simd128<64>::slli<(sh-64)>(_mm_slli_si128(arg1, (int32_t)(8))) : simd_or(simd128<64>::slli<sh>(arg1), _mm_slli_si128(simd128<64>::srli<(64-sh)>(arg1), (int32_t)(8)))));
1235}
1236
1237//The total number of operations is 3
1238template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1239{
1240        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
1241}
1242
1243//The total number of operations is 8
1244template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1245{
1246        return simd128<(1)>::ifh(simd128<1>::ifh(simd128<2>::himask(), arg1, simd128<2>::srli<(1)>(arg1)), arg2, arg3);
1247}
1248
1249//The total number of operations is 13
1250template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1251{
1252        return simd128<1>::ifh(simd128<4>::gt(simd128<4>::constant<0>(), arg1), arg2, arg3);
1253}
1254
1255//The total number of operations is 1
1256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1257{
1258        return _mm_blendv_epi8(arg3, arg2, arg1);
1259}
1260
1261//The total number of operations is 4
1262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1263{
1264        return simd128<1>::ifh(simd128<16>::gt(simd128<16>::constant<0>(), arg1), arg2, arg3);
1265}
1266
1267//The total number of operations is 4
1268template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1269{
1270        return simd128<1>::ifh(simd128<32>::gt(simd128<32>::constant<0>(), arg1), arg2, arg3);
1271}
1272
1273//The total number of operations is 4
1274template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1275{
1276        return simd128<1>::ifh(simd128<64>::gt(simd128<64>::constant<0>(), arg1), arg2, arg3);
1277}
1278
1279//The total number of operations is 11
1280template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1281{
1282        return simd128<(64)>::ifh(simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg1)), arg2, arg3);
1283}
1284
1285//The total number of operations is 1
1286template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2)
1287{
1288        return simd_xor(arg1, arg2);
1289}
1290
1291//The total number of operations is 11
1292template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2)
1293{
1294        bitblock128_t tmp = simd_xor(arg1, arg2);
1295        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2))), tmp);
1296}
1297
1298//The total number of operations is 6
1299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2)
1300{
1301        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::sub(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::sub(arg1, arg2));
1302}
1303
1304//The total number of operations is 1
1305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2)
1306{
1307        return _mm_sub_epi8(arg1, arg2);
1308}
1309
1310//The total number of operations is 1
1311template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2)
1312{
1313        return _mm_sub_epi16(arg1, arg2);
1314}
1315
1316//The total number of operations is 1
1317template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2)
1318{
1319        return _mm_sub_epi32(arg1, arg2);
1320}
1321
1322//The total number of operations is 1
1323template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2)
1324{
1325        return _mm_sub_epi64(arg1, arg2);
1326}
1327
1328//The total number of operations is 11
1329template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2)
1330{
1331        bitblock128_t partial = simd128<(64)>::sub(arg1, arg2);
1332        bitblock128_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
1333        bitblock128_t borrow = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(borrowMask));
1334        return simd128<(64)>::sub(partial, borrow);
1335}
1336
1337//The total number of operations is 3
1338template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1)
1339{
1340        return simd128<16>::sub(arg1, simd_and(simd128<2>::lomask(), simd128<16>::srli<1>(arg1)));
1341}
1342
1343//The total number of operations is 4
1344template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1)
1345{
1346        return simd128<(8)>::add(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1347}
1348
1349//The total number of operations is 4
1350template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1)
1351{
1352        return simd128<(16)>::add(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1353}
1354
1355//The total number of operations is 3
1356template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1)
1357{
1358        return simd128<(32)>::add(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1359}
1360
1361//The total number of operations is 3
1362template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1)
1363{
1364        return simd128<(64)>::add(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1365}
1366
1367//The total number of operations is 3
1368template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1)
1369{
1370        return simd128<64>::add(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1371}
1372
1373//The total number of operations is 16
1374template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1)
1375{
1376        return simd128<128>::add(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1377}
1378
1379//The total number of operations is 7
1380template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1381{
1382        return simd128<1>::ifh(simd128<128>::himask(), _mm_srl_epi64(arg1, _mm_srli_si128(shift_mask, (int32_t)(8))), _mm_srl_epi64(arg1, simd_andc(shift_mask, simd128<128>::himask())));
1383}
1384
1385//The total number of operations is 11
1386template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1387{
1388        return simd_or(_mm_srl_epi64(arg1, shift_mask), simd_or(_mm_srli_si128(_mm_srl_epi64(arg1, simd128<32>::sub(shift_mask, _mm_cvtsi32_si128((int32_t)(64)))), (int32_t)(8)), _mm_srli_si128(_mm_sll_epi64(arg1, simd128<32>::sub(_mm_cvtsi32_si128((int32_t)(64)), shift_mask)), (int32_t)(8))));
1389}
1390
1391//The total number of operations is 0
1392template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1393{
1394        return simd128<32>::constant<(-1*val)>();
1395}
1396
1397//The total number of operations is 0
1398template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1399{
1400        return ((val < 0) ? simd128<(4)>::constant<((val<<2)|(val^(-4)))>() : simd128<(4)>::constant<((val<<2)|val)>());
1401}
1402
1403//The total number of operations is 0
1404template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1405{
1406        return ((val < 0) ? simd128<(8)>::constant<((val<<4)|(val^(-16)))>() : simd128<(8)>::constant<((val<<4)|val)>());
1407}
1408
1409//The total number of operations is 0
1410template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1411{
1412        return _mm_set1_epi8((int32_t)(val));
1413}
1414
1415//The total number of operations is 0
1416template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1417{
1418        return _mm_set1_epi16((int32_t)(val));
1419}
1420
1421//The total number of operations is 0
1422template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1423{
1424        return _mm_set1_epi32((int32_t)(val));
1425}
1426
1427//The total number of operations is 0
1428template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1429{
1430        return _mm_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val));
1431}
1432
1433//The total number of operations is 0
1434template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1435{
1436        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val));
1437}
1438
1439//The total number of operations is 1
1440template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2)
1441{
1442        return simd_or(arg1, arg2);
1443}
1444
1445//The total number of operations is 18
1446template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2)
1447{
1448        return simd128<1>::ifh(simd128<2>::lt(arg1, arg2), arg1, arg2);
1449}
1450
1451//The total number of operations is 9
1452template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2)
1453{
1454        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1455        return simd_xor(simd128<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1456}
1457
1458//The total number of operations is 1
1459template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2)
1460{
1461        return _mm_min_epi8(arg1, arg2);
1462}
1463
1464//The total number of operations is 1
1465template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2)
1466{
1467        return _mm_min_epi16(arg1, arg2);
1468}
1469
1470//The total number of operations is 1
1471template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2)
1472{
1473        return _mm_min_epi32(arg1, arg2);
1474}
1475
1476//The total number of operations is 4
1477template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2)
1478{
1479        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg2, arg1);
1480}
1481
1482//The total number of operations is 30
1483template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2)
1484{
1485        bitblock128_t hiAns = simd128<(64)>::min(arg1, arg2);
1486        bitblock128_t loAns = simd128<(64)>::umin(arg1, arg2);
1487        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
1488        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
1489        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
1490}
1491
1492//The total number of operations is 0
1493template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask()
1494{
1495        return simd128<2>::constant<(1)>();
1496}
1497
1498//The total number of operations is 0
1499template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask()
1500{
1501        return simd128<4>::constant<(3)>();
1502}
1503
1504//The total number of operations is 0
1505template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask()
1506{
1507        return simd128<8>::constant<(15)>();
1508}
1509
1510//The total number of operations is 0
1511template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask()
1512{
1513        return simd128<16>::constant<(255)>();
1514}
1515
1516//The total number of operations is 0
1517template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask()
1518{
1519        return simd128<32>::constant<(65535)>();
1520}
1521
1522//The total number of operations is 0
1523template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask()
1524{
1525        return _mm_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1));
1526}
1527
1528//The total number of operations is 0
1529template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask()
1530{
1531        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1));
1532}
1533
1534//The total number of operations is 1
1535template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2)
1536{
1537        return simd_and(arg1, arg2);
1538}
1539
1540//The total number of operations is 16
1541template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2)
1542{
1543        return simd_or(simd_and(simd128<(4)>::himask(), simd128<(4)>::umin(arg1, arg2)), simd128<(4)>::umin(simd_and(simd128<(4)>::lomask(), arg1), simd_and(simd128<(4)>::lomask(), arg2)));
1544}
1545
1546//The total number of operations is 6
1547template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2)
1548{
1549        return simd_or(simd_and(simd128<(8)>::himask(), simd128<(8)>::umin(arg1, arg2)), simd128<(8)>::umin(simd_and(simd128<(8)>::lomask(), arg1), simd_and(simd128<(8)>::lomask(), arg2)));
1550}
1551
1552//The total number of operations is 1
1553template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2)
1554{
1555        return _mm_min_epu8(arg1, arg2);
1556}
1557
1558//The total number of operations is 1
1559template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2)
1560{
1561        return _mm_min_epu16(arg1, arg2);
1562}
1563
1564//The total number of operations is 1
1565template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2)
1566{
1567        return _mm_min_epu32(arg1, arg2);
1568}
1569
1570//The total number of operations is 7
1571template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2)
1572{
1573        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808ULL)>();
1574        return simd_xor(simd128<64>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1575}
1576
1577//The total number of operations is 26
1578template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2)
1579{
1580        bitblock128_t tmpAns = simd128<(64)>::umin(arg1, arg2);
1581        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1582        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1583        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1584}
1585
1586//The total number of operations is 9
1587template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1)
1588{
1589        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, simd128<128>::slli<1>(simd_not(arg1))), arg1);
1590}
1591
1592//The total number of operations is 19
1593template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1)
1594{
1595        bitblock128_t gtMask = simd128<4>::gt(arg1, simd128<4>::constant<0>());
1596        return simd128<1>::ifh(gtMask, arg1, simd128<4>::sub(gtMask, arg1));
1597}
1598
1599//The total number of operations is 1
1600template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1)
1601{
1602        return _mm_abs_epi8(arg1);
1603}
1604
1605//The total number of operations is 1
1606template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1)
1607{
1608        return _mm_abs_epi16(arg1);
1609}
1610
1611//The total number of operations is 1
1612template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1)
1613{
1614        return _mm_abs_epi32(arg1);
1615}
1616
1617//The total number of operations is 5
1618template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1)
1619{
1620        bitblock128_t gtMask = simd128<64>::gt(arg1, simd128<64>::constant<0>());
1621        return simd128<1>::ifh(gtMask, arg1, simd128<64>::sub(gtMask, arg1));
1622}
1623
1624//The total number of operations is 33
1625template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1)
1626{
1627        bitblock128_t eqMask = simd128<128>::eq(simd128<1>::ifh(simd128<128>::himask(), simd128<(64)>::abs(arg1), arg1), arg1);
1628        return simd128<1>::ifh(eqMask, arg1, simd128<128>::sub(eqMask, arg1));
1629}
1630
1631//The total number of operations is 2
1632template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2)
1633{
1634        return simd_not(simd_xor(arg1, arg2));
1635}
1636
1637//The total number of operations is 8
1638template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2)
1639{
1640        bitblock128_t tmpAns = simd128<(1)>::eq(arg1, arg2);
1641        bitblock128_t loMask = simd_and(tmpAns, simd128<2>::srli<(1)>(tmpAns));
1642        bitblock128_t hiMask = simd128<2>::slli<(1)>(loMask);
1643        return simd_or(loMask, hiMask);
1644}
1645
1646//The total number of operations is 9
1647template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2)
1648{
1649        return simd_or(simd_and(simd128<(8)>::himask(), simd128<(8)>::eq(simd_and(simd128<(8)>::himask(), arg1), simd_and(simd128<(8)>::himask(), arg2))), simd_and(simd128<(8)>::lomask(), simd128<(8)>::eq(simd_and(simd128<(8)>::lomask(), arg1), simd_and(simd128<(8)>::lomask(), arg2))));
1650}
1651
1652//The total number of operations is 1
1653template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2)
1654{
1655        return _mm_cmpeq_epi8(arg1, arg2);
1656}
1657
1658//The total number of operations is 1
1659template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2)
1660{
1661        return _mm_cmpeq_epi16(arg1, arg2);
1662}
1663
1664//The total number of operations is 1
1665template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2)
1666{
1667        return _mm_cmpeq_epi32(arg1, arg2);
1668}
1669
1670//The total number of operations is 1
1671template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2)
1672{
1673        return _mm_cmpeq_epi64(arg1, arg2);
1674}
1675
1676//The total number of operations is 11
1677template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2)
1678{
1679        bitblock128_t tmpAns = simd128<(64)>::eq(arg1, arg2);
1680        bitblock128_t loMask = simd_and(tmpAns, simd128<128>::srli<(64)>(tmpAns));
1681        bitblock128_t hiMask = simd128<128>::slli<(64)>(loMask);
1682        return simd_or(loMask, hiMask);
1683}
1684
1685//The total number of operations is 4
1686template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1)
1687{
1688        return ((sh == 0) ? arg1 : simd_or(simd_and(simd128<2>::himask(), arg1), simd128<2>::srli<1>(arg1)));
1689}
1690
1691//The total number of operations is 10
1692template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1)
1693{
1694        bitblock128_t tmp = simd128<4>::srli<((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh))>(arg1);
1695        return simd_or(tmp, simd128<4>::sub(simd128<4>::constant<0>(), simd_and(simd128<4>::constant<(1<<((4-((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh)))-1))>(), tmp)));
1696}
1697
1698//The total number of operations is 5
1699template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1)
1700{
1701        bitblock128_t tmp = simd128<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
1702        return simd_or(tmp, simd128<8>::sub(simd128<8>::constant<0>(), simd_and(simd128<8>::constant<(1<<((8-((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh)))-1))>(), tmp)));
1703}
1704
1705//The total number of operations is 1
1706template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1)
1707{
1708        return _mm_srai_epi16(arg1, (int32_t)(sh));
1709}
1710
1711//The total number of operations is 1
1712template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1)
1713{
1714        return _mm_srai_epi32(arg1, (int32_t)(sh));
1715}
1716
1717//The total number of operations is 5
1718template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1)
1719{
1720        return simd_or(simd_and(simd128<64>::himask(), simd128<(32)>::srai<((sh < (32)) ? sh : (32))>(arg1)), ((sh <= (32)) ? simd128<64>::srli<sh>(arg1) : simd128<(32)>::srai<(sh-(32))>(simd128<64>::srli<(32)>(arg1))));
1721}
1722
1723//The total number of operations is 16
1724template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1)
1725{
1726        return simd_or(simd_and(simd128<128>::himask(), simd128<(64)>::srai<((sh < (64)) ? sh : (64))>(arg1)), ((sh <= (64)) ? simd128<128>::srli<sh>(arg1) : simd128<(64)>::srai<(sh-(64))>(simd128<128>::srli<(64)>(arg1))));
1727}
1728
1729//The total number of operations is 0
1730template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask()
1731{
1732        return simd128<2>::constant<(2)>();
1733}
1734
1735//The total number of operations is 0
1736template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask()
1737{
1738        return simd128<4>::constant<(12)>();
1739}
1740
1741//The total number of operations is 0
1742template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask()
1743{
1744        return simd128<8>::constant<(240)>();
1745}
1746
1747//The total number of operations is 0
1748template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask()
1749{
1750        return simd128<16>::constant<(65280)>();
1751}
1752
1753//The total number of operations is 0
1754template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask()
1755{
1756        return simd128<32>::constant<-65536>();
1757}
1758
1759//The total number of operations is 0
1760template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask()
1761{
1762        return _mm_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0));
1763}
1764
1765//The total number of operations is 0
1766template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask()
1767{
1768        return _mm_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0));
1769}
1770
1771//The total number of operations is 1
1772template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2)
1773{
1774        return simd_xor(arg1, arg2);
1775}
1776
1777//The total number of operations is 10
1778template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2)
1779{
1780        bitblock128_t tmp = simd_xor(arg1, arg2);
1781        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(arg1, arg2))), tmp);
1782}
1783
1784//The total number of operations is 6
1785template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2)
1786{
1787        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::add(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::add(arg1, arg2));
1788}
1789
1790//The total number of operations is 1
1791template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2)
1792{
1793        return _mm_add_epi8(arg1, arg2);
1794}
1795
1796//The total number of operations is 1
1797template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2)
1798{
1799        return _mm_add_epi16(arg1, arg2);
1800}
1801
1802//The total number of operations is 1
1803template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2)
1804{
1805        return _mm_add_epi32(arg1, arg2);
1806}
1807
1808//The total number of operations is 1
1809template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2)
1810{
1811        return _mm_add_epi64(arg1, arg2);
1812}
1813
1814//The total number of operations is 11
1815template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2)
1816{
1817        bitblock128_t partial = simd128<(64)>::add(arg1, arg2);
1818        bitblock128_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
1819        bitblock128_t carry = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(carryMask));
1820        return simd128<(64)>::add(partial, carry);
1821}
1822
1823//The total number of operations is 1
1824template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2)
1825{
1826        return simd_or(arg1, arg2);
1827}
1828
1829//The total number of operations is 16
1830template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2)
1831{
1832        return simd_or(simd_and(simd128<(4)>::himask(), simd128<(4)>::umax(arg1, arg2)), simd128<(4)>::umax(simd_and(simd128<(4)>::lomask(), arg1), simd_and(simd128<(4)>::lomask(), arg2)));
1833}
1834
1835//The total number of operations is 6
1836template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2)
1837{
1838        return simd_or(simd_and(simd128<(8)>::himask(), simd128<(8)>::umax(arg1, arg2)), simd128<(8)>::umax(simd_and(simd128<(8)>::lomask(), arg1), simd_and(simd128<(8)>::lomask(), arg2)));
1839}
1840
1841//The total number of operations is 1
1842template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2)
1843{
1844        return _mm_max_epu8(arg1, arg2);
1845}
1846
1847//The total number of operations is 1
1848template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2)
1849{
1850        return _mm_max_epu16(arg1, arg2);
1851}
1852
1853//The total number of operations is 1
1854template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2)
1855{
1856        return _mm_max_epu32(arg1, arg2);
1857}
1858
1859//The total number of operations is 7
1860template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2)
1861{
1862        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808ULL)>();
1863        return simd_xor(simd128<64>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1864}
1865
1866//The total number of operations is 26
1867template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2)
1868{
1869        bitblock128_t tmpAns = simd128<(64)>::umax(arg1, arg2);
1870        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1871        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1872        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1873}
1874
1875//The total number of operations is 93
1876template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1877{
1878        return simd128<(1)>::umin(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1879}
1880
1881//The total number of operations is 80
1882template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1883{
1884        return simd128<(2)>::umin(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1885}
1886
1887//The total number of operations is 42
1888template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1889{
1890        return simd128<(4)>::umin(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1891}
1892
1893//The total number of operations is 7
1894template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1895{
1896        return simd128<(8)>::umin(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1897}
1898
1899//The total number of operations is 7
1900template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1901{
1902        return simd128<(16)>::umin(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1903}
1904
1905//The total number of operations is 7
1906template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1907{
1908        return simd128<(32)>::umin(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1909}
1910
1911//The total number of operations is 21
1912template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1913{
1914        return simd128<(64)>::umin(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1915}
1916
1917//The total number of operations is 93
1918template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1919{
1920        return simd128<(1)>::add(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1921}
1922
1923//The total number of operations is 74
1924template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1925{
1926        return simd128<(2)>::add(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1927}
1928
1929//The total number of operations is 42
1930template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1931{
1932        return simd128<(4)>::add(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1933}
1934
1935//The total number of operations is 7
1936template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1937{
1938        return simd128<(8)>::add(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1939}
1940
1941//The total number of operations is 1
1942template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1943{
1944        return _mm_hadd_epi16(arg2, arg1);
1945}
1946
1947//The total number of operations is 1
1948template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1949{
1950        return _mm_hadd_epi32(arg2, arg1);
1951}
1952
1953//The total number of operations is 15
1954template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1955{
1956        return simd128<(64)>::add(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1957}
1958
1959//The total number of operations is 120
1960template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2)
1961{
1962        bitblock128_t hiBound = simd128<2>::srli<1>(simd128<2>::lomask());
1963        bitblock128_t loBound = simd_not(hiBound);
1964        return hsimd128<2>::packl(simd128<1>::ifh(simd128<2>::gt(arg1, hiBound), hiBound, simd128<1>::ifh(simd128<2>::gt(arg1, loBound), arg1, loBound)), simd128<1>::ifh(simd128<2>::gt(arg2, hiBound), hiBound, simd128<1>::ifh(simd128<2>::gt(arg2, loBound), arg2, loBound)));
1965}
1966
1967//The total number of operations is 86
1968template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2)
1969{
1970        bitblock128_t hiBound = simd128<4>::srli<1>(simd128<4>::lomask());
1971        bitblock128_t loBound = simd_not(hiBound);
1972        return hsimd128<4>::packl(simd128<1>::ifh(simd128<4>::gt(arg1, hiBound), hiBound, simd128<1>::ifh(simd128<4>::gt(arg1, loBound), arg1, loBound)), simd128<1>::ifh(simd128<4>::gt(arg2, hiBound), hiBound, simd128<1>::ifh(simd128<4>::gt(arg2, loBound), arg2, loBound)));
1973}
1974
1975//The total number of operations is 36
1976template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2)
1977{
1978        bitblock128_t hiBound = simd128<8>::srli<1>(simd128<8>::lomask());
1979        bitblock128_t loBound = simd_not(hiBound);
1980        return hsimd128<8>::packl(simd128<1>::ifh(simd128<8>::gt(arg1, hiBound), hiBound, simd128<1>::ifh(simd128<8>::gt(arg1, loBound), arg1, loBound)), simd128<1>::ifh(simd128<8>::gt(arg2, hiBound), hiBound, simd128<1>::ifh(simd128<8>::gt(arg2, loBound), arg2, loBound)));
1981}
1982
1983//The total number of operations is 1
1984template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2)
1985{
1986        return _mm_packs_epi16(arg2, arg1);
1987}
1988
1989//The total number of operations is 1
1990template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2)
1991{
1992        return _mm_packs_epi32(arg2, arg1);
1993}
1994
1995//The total number of operations is 21
1996template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2)
1997{
1998        bitblock128_t hiBound = simd128<64>::srli<1>(simd128<64>::lomask());
1999        bitblock128_t loBound = simd_not(hiBound);
2000        return hsimd128<64>::packl(simd128<1>::ifh(simd128<64>::gt(arg1, hiBound), hiBound, simd128<1>::ifh(simd128<64>::gt(arg1, loBound), arg1, loBound)), simd128<1>::ifh(simd128<64>::gt(arg2, hiBound), hiBound, simd128<1>::ifh(simd128<64>::gt(arg2, loBound), arg2, loBound)));
2001}
2002
2003//The total number of operations is 152
2004template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2)
2005{
2006        bitblock128_t hiBound = simd128<128>::srli<1>(simd128<128>::lomask());
2007        bitblock128_t loBound = simd_not(hiBound);
2008        return hsimd128<128>::packl(simd128<1>::ifh(simd128<128>::gt(arg1, hiBound), hiBound, simd128<1>::ifh(simd128<128>::gt(arg1, loBound), arg1, loBound)), simd128<1>::ifh(simd128<128>::gt(arg2, hiBound), hiBound, simd128<1>::ifh(simd128<128>::gt(arg2, loBound), arg2, loBound)));
2009}
2010
2011//The total number of operations is 24
2012template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1)
2013{
2014        uint64_t tmpAns1 = hsimd128<(8)>::signmask(esimd128<4>::mergeh(arg1, simd128<4>::constant<0>()));
2015        uint64_t tmpAns2 = hsimd128<(8)>::signmask(esimd128<4>::mergel(arg1, simd128<4>::constant<0>()));
2016        return ((tmpAns1<<(16))+tmpAns2);
2017}
2018
2019//The total number of operations is 1
2020template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1)
2021{
2022        return _mm_movemask_epi8(arg1);
2023}
2024
2025//The total number of operations is 4
2026template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1)
2027{
2028        return hsimd128<(8)>::signmask(hsimd128<16>::packh(simd128<16>::constant<0>(), arg1));
2029}
2030
2031//The total number of operations is 4
2032template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1)
2033{
2034        return (((((mvmd128<32>::extract<3>(arg1)>>28)&8)|((mvmd128<32>::extract<2>(arg1)>>29)&4))|((mvmd128<32>::extract<1>(arg1)>>30)&2))|(mvmd128<32>::extract<0>(arg1)>>31));
2035}
2036
2037//The total number of operations is 4
2038template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1)
2039{
2040        return (((mvmd128<64>::extract<1>(arg1)>>62)&2)|(mvmd128<64>::extract<0>(arg1)>>63));
2041}
2042
2043//The total number of operations is 11
2044template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1)
2045{
2046        return hsimd128<(64)>::signmask(hsimd128<128>::packh(simd128<128>::constant<0>(), arg1));
2047}
2048
2049//The total number of operations is 45
2050template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2)
2051{
2052        return hsimd128<(4)>::packl(simd128<1>::ifh(simd128<2>::himask(), simd128<128>::srli<(1)>(arg1), arg1), simd128<1>::ifh(simd128<2>::himask(), simd128<128>::srli<(1)>(arg2), arg2));
2053}
2054
2055//The total number of operations is 31
2056template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2)
2057{
2058        return hsimd128<(8)>::packl(simd128<1>::ifh(simd128<4>::himask(), simd128<128>::srli<(2)>(arg1), arg1), simd128<1>::ifh(simd128<4>::himask(), simd128<128>::srli<(2)>(arg2), arg2));
2059}
2060
2061//The total number of operations is 17
2062template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2)
2063{
2064        return hsimd128<(16)>::packl(simd128<1>::ifh(simd128<8>::himask(), simd128<128>::srli<(4)>(arg1), arg1), simd128<1>::ifh(simd128<8>::himask(), simd128<128>::srli<(4)>(arg2), arg2));
2065}
2066
2067//The total number of operations is 3
2068template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2)
2069{
2070        return hsimd128<16>::packus(simd_and(arg1, simd128<16>::lomask()), simd_and(arg2, simd128<16>::lomask()));
2071}
2072
2073//The total number of operations is 3
2074template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2)
2075{
2076        return hsimd128<32>::packus(simd_and(arg1, simd128<32>::lomask()), simd_and(arg2, simd128<32>::lomask()));
2077}
2078
2079//The total number of operations is 3
2080template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2)
2081{
2082        return _mm_hsub_epi32(simd_and(arg2, simd128<64>::lomask()), simd_and(arg1, simd128<64>::lomask()));
2083}
2084
2085//The total number of operations is 7
2086template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2)
2087{
2088        return simd128<1>::ifh(simd128<128>::himask(), simd128<128>::slli<(64)>(arg1), arg2);
2089}
2090
2091//The total number of operations is 47
2092template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2)
2093{
2094        return hsimd128<2>::packl(simd128<64>::srli<(1)>(arg1), simd128<64>::srli<(1)>(arg2));
2095}
2096
2097//The total number of operations is 33
2098template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2)
2099{
2100        return hsimd128<4>::packl(simd128<64>::srli<(2)>(arg1), simd128<64>::srli<(2)>(arg2));
2101}
2102
2103//The total number of operations is 19
2104template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2)
2105{
2106        return hsimd128<8>::packl(simd128<64>::srli<(4)>(arg1), simd128<64>::srli<(4)>(arg2));
2107}
2108
2109//The total number of operations is 3
2110template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2)
2111{
2112        return hsimd128<16>::packus(simd128<16>::srli<(8)>(arg1), simd128<16>::srli<(8)>(arg2));
2113}
2114
2115//The total number of operations is 3
2116template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2)
2117{
2118        return _mm_hsub_epi16(simd128<32>::srli<(16)>(arg2), simd128<32>::srli<(16)>(arg1));
2119}
2120
2121//The total number of operations is 3
2122template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2)
2123{
2124        return _mm_hsub_epi32(simd128<64>::srli<(32)>(arg2), simd128<64>::srli<(32)>(arg1));
2125}
2126
2127//The total number of operations is 7
2128template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2)
2129{
2130        return simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg2));
2131}
2132
2133//The total number of operations is 93
2134template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2135{
2136        return simd128<(1)>::min(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
2137}
2138
2139//The total number of operations is 82
2140template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2141{
2142        return simd128<(2)>::min(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
2143}
2144
2145//The total number of operations is 45
2146template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2147{
2148        return simd128<(4)>::min(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
2149}
2150
2151//The total number of operations is 7
2152template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2153{
2154        return simd128<(8)>::min(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2155}
2156
2157//The total number of operations is 7
2158template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2159{
2160        return simd128<(16)>::min(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2161}
2162
2163//The total number of operations is 7
2164template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2165{
2166        return simd128<(32)>::min(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2167}
2168
2169//The total number of operations is 18
2170template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2171{
2172        return simd128<(64)>::min(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2173}
2174
2175//The total number of operations is 85
2176template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2)
2177{
2178        bitblock128_t arg11 = simd128<2>::ifh(arg1, simd128<2>::constant<0>(), arg1);
2179        bitblock128_t arg12 = simd_and(simd128<2>::lomask(), arg11);
2180        bitblock128_t arg21 = simd128<2>::ifh(arg2, simd128<2>::constant<0>(), arg2);
2181        bitblock128_t arg22 = simd_and(simd128<2>::lomask(), arg21);
2182        return hsimd128<2>::packl(simd128<1>::ifh(simd128<2>::eq(arg12, arg11), arg12, simd128<2>::lomask()), simd128<1>::ifh(simd128<2>::eq(arg22, arg21), arg22, simd128<2>::lomask()));
2183}
2184
2185//The total number of operations is 83
2186template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2)
2187{
2188        bitblock128_t arg11 = simd128<4>::ifh(arg1, simd128<4>::constant<0>(), arg1);
2189        bitblock128_t arg12 = simd_and(simd128<4>::lomask(), arg11);
2190        bitblock128_t arg21 = simd128<4>::ifh(arg2, simd128<4>::constant<0>(), arg2);
2191        bitblock128_t arg22 = simd_and(simd128<4>::lomask(), arg21);
2192        return hsimd128<4>::packl(simd128<1>::ifh(simd128<4>::eq(arg12, arg11), arg12, simd128<4>::lomask()), simd128<1>::ifh(simd128<4>::eq(arg22, arg21), arg22, simd128<4>::lomask()));
2193}
2194
2195//The total number of operations is 29
2196template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2)
2197{
2198        bitblock128_t arg11 = simd128<8>::ifh(arg1, simd128<8>::constant<0>(), arg1);
2199        bitblock128_t arg12 = simd_and(simd128<8>::lomask(), arg11);
2200        bitblock128_t arg21 = simd128<8>::ifh(arg2, simd128<8>::constant<0>(), arg2);
2201        bitblock128_t arg22 = simd_and(simd128<8>::lomask(), arg21);
2202        return hsimd128<8>::packl(simd128<1>::ifh(simd128<8>::eq(arg12, arg11), arg12, simd128<8>::lomask()), simd128<1>::ifh(simd128<8>::eq(arg22, arg21), arg22, simd128<8>::lomask()));
2203}
2204
2205//The total number of operations is 1
2206template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2)
2207{
2208        return _mm_packus_epi16(arg2, arg1);
2209}
2210
2211//The total number of operations is 1
2212template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2)
2213{
2214        return _mm_packus_epi32(arg2, arg1);
2215}
2216
2217//The total number of operations is 12
2218template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2)
2219{
2220        bitblock128_t hiPart = hsimd128<64>::packh(arg1, arg2);
2221        return simd128<(32)>::ifh(hiPart, simd128<(32)>::constant<0>(), simd_or(simd128<(32)>::gt(hiPart, simd128<(32)>::constant<0>()), hsimd128<64>::packl(arg1, arg2)));
2222}
2223
2224//The total number of operations is 20
2225template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2)
2226{
2227        bitblock128_t hiPart = hsimd128<128>::packh(arg1, arg2);
2228        return simd128<(64)>::ifh(hiPart, simd128<(64)>::constant<0>(), simd_or(simd128<(64)>::gt(hiPart, simd128<(64)>::constant<0>()), hsimd128<128>::packl(arg1, arg2)));
2229}
2230
2231//The total number of operations is 31
2232template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2233{
2234        return esimd128<(2)>::mergel(simd128<1>::ifh(simd128<(2)>::himask(), arg1, simd128<(2)>::srli<1>(arg2)), simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::slli<1>(arg1), arg2));
2235}
2236
2237//The total number of operations is 21
2238template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2239{
2240        return esimd128<(4)>::mergel(simd128<1>::ifh(simd128<(4)>::himask(), arg1, simd128<(4)>::srli<2>(arg2)), simd128<1>::ifh(simd128<(4)>::himask(), simd128<(4)>::slli<2>(arg1), arg2));
2241}
2242
2243//The total number of operations is 11
2244template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2245{
2246        return esimd128<(8)>::mergel(simd128<1>::ifh(simd128<(8)>::himask(), arg1, simd128<(8)>::srli<4>(arg2)), simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::slli<4>(arg1), arg2));
2247}
2248
2249//The total number of operations is 1
2250template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2251{
2252        return _mm_unpacklo_epi8(arg2, arg1);
2253}
2254
2255//The total number of operations is 1
2256template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2257{
2258        return _mm_unpacklo_epi16(arg2, arg1);
2259}
2260
2261//The total number of operations is 1
2262template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2263{
2264        return _mm_unpacklo_epi32(arg2, arg1);
2265}
2266
2267//The total number of operations is 1
2268template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2269{
2270        return _mm_unpacklo_epi64(arg2, arg1);
2271}
2272
2273//The total number of operations is 31
2274template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2275{
2276        return esimd128<(2)>::mergeh(simd128<1>::ifh(simd128<(2)>::himask(), arg1, simd128<(2)>::srli<1>(arg2)), simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::slli<1>(arg1), arg2));
2277}
2278
2279//The total number of operations is 21
2280template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2281{
2282        return esimd128<(4)>::mergeh(simd128<1>::ifh(simd128<(4)>::himask(), arg1, simd128<(4)>::srli<2>(arg2)), simd128<1>::ifh(simd128<(4)>::himask(), simd128<(4)>::slli<2>(arg1), arg2));
2283}
2284
2285//The total number of operations is 11
2286template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2287{
2288        return esimd128<(8)>::mergeh(simd128<1>::ifh(simd128<(8)>::himask(), arg1, simd128<(8)>::srli<4>(arg2)), simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::slli<4>(arg1), arg2));
2289}
2290
2291//The total number of operations is 1
2292template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2293{
2294        return _mm_unpackhi_epi8(arg2, arg1);
2295}
2296
2297//The total number of operations is 1
2298template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2299{
2300        return _mm_unpackhi_epi16(arg2, arg1);
2301}
2302
2303//The total number of operations is 1
2304template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2305{
2306        return _mm_unpackhi_epi32(arg2, arg1);
2307}
2308
2309//The total number of operations is 1
2310template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2311{
2312        return _mm_unpackhi_epi64(arg2, arg1);
2313}
2314
2315//The total number of operations is 24
2316template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1)
2317{
2318        return esimd128<(2)>::mergeh(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2319}
2320
2321//The total number of operations is 14
2322template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1)
2323{
2324        return esimd128<(4)>::mergeh(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2325}
2326
2327//The total number of operations is 4
2328template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1)
2329{
2330        return esimd128<(8)>::mergeh(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2331}
2332
2333//The total number of operations is 3
2334template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1)
2335{
2336        return esimd128<(16)>::mergeh(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2337}
2338
2339//The total number of operations is 3
2340template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1)
2341{
2342        return esimd128<(32)>::mergeh(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2343}
2344
2345//The total number of operations is 3
2346template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1)
2347{
2348        return esimd128<(64)>::mergeh(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2349}
2350
2351//The total number of operations is 4
2352template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1)
2353{
2354        return simd128<(128)>::srli<64>(arg1);
2355}
2356
2357//The total number of operations is 24
2358template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1)
2359{
2360        return esimd128<(2)>::mergel(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2361}
2362
2363//The total number of operations is 14
2364template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1)
2365{
2366        return esimd128<(4)>::mergel(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2367}
2368
2369//The total number of operations is 4
2370template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1)
2371{
2372        return esimd128<(8)>::mergel(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2373}
2374
2375//The total number of operations is 1
2376template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1)
2377{
2378        return _mm_cvtepu8_epi16(arg1);
2379}
2380
2381//The total number of operations is 1
2382template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1)
2383{
2384        return _mm_cvtepu16_epi32(arg1);
2385}
2386
2387//The total number of operations is 1
2388template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1)
2389{
2390        return _mm_cvtepu32_epi64(arg1);
2391}
2392
2393//The total number of operations is 1
2394template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1)
2395{
2396        return simd_and(simd128<(128)>::lomask(), arg1);
2397}
2398
2399//The total number of operations is 31
2400template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1)
2401{
2402        return esimd128<(2)>::mergeh(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2403}
2404
2405//The total number of operations is 33
2406template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1)
2407{
2408        return esimd128<(4)>::mergeh(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2409}
2410
2411//The total number of operations is 13
2412template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1)
2413{
2414        return esimd128<(8)>::mergeh(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2415}
2416
2417//The total number of operations is 4
2418template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1)
2419{
2420        return esimd128<(16)>::mergeh(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2421}
2422
2423//The total number of operations is 4
2424template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1)
2425{
2426        return esimd128<(32)>::mergeh(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2427}
2428
2429//The total number of operations is 5
2430template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1)
2431{
2432        return esimd128<32>::signextendl(simd128<128>::srli<(64)>(arg1));
2433}
2434
2435//The total number of operations is 16
2436template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1)
2437{
2438        return simd128<(128)>::srai<64>(arg1);
2439}
2440
2441//The total number of operations is 31
2442template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1)
2443{
2444        return esimd128<(2)>::mergel(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2445}
2446
2447//The total number of operations is 33
2448template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1)
2449{
2450        return esimd128<(4)>::mergel(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2451}
2452
2453//The total number of operations is 13
2454template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1)
2455{
2456        return esimd128<(8)>::mergel(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2457}
2458
2459//The total number of operations is 1
2460template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1)
2461{
2462        return _mm_cvtepi8_epi16(arg1);
2463}
2464
2465//The total number of operations is 1
2466template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1)
2467{
2468        return _mm_cvtepi16_epi32(arg1);
2469}
2470
2471//The total number of operations is 1
2472template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1)
2473{
2474        return _mm_cvtepi32_epi64(arg1);
2475}
2476
2477//The total number of operations is 20
2478template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1)
2479{
2480        return simd128<(128)>::srai<64>(simd128<(128)>::slli<64>(arg1));
2481}
2482
2483//The total number of operations is 17
2484template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1)
2485{
2486        bitblock128_t tmphi = _mm_shufflehi_epi16(arg1, (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2487        bitblock128_t tmpAns = _mm_shufflelo_epi16(tmphi, (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2488        bitblock128_t tmplh = _mm_shufflehi_epi16(simd128<128>::slli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2489        bitblock128_t tmphl = _mm_shufflelo_epi16(simd128<128>::srli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2490        uint32_t a1 = ((((msk>>21)&4) == 0) ? 0 : (131071));
2491        uint32_t a2 = ((((msk>>18)&4) == 0) ? 0 : (131071));
2492        uint32_t a3 = ((((msk>>15)&4) == 0) ? 0 : (131071));
2493        uint32_t a4 = ((((msk>>12)&4) == 0) ? 0 : (131071));
2494        uint32_t a5 = ((((msk>>9)&4) == 0) ? (131071) : 0);
2495        uint32_t a6 = ((((msk>>6)&4) == 0) ? (131071) : 0);
2496        uint32_t a7 = ((((msk>>3)&4) == 0) ? (131071) : 0);
2497        uint32_t a8 = (((msk&4) == 0) ? (131071) : 0);
2498        return simd128<1>::ifh(mvmd128<16>::fill8(a1, a2, a3, a4, a5, a6, a7, a8), tmpAns, simd_or(tmplh, tmphl));
2499}
2500
2501//The total number of operations is 1
2502template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1)
2503{
2504        return _mm_shuffle_epi32(arg1, (int32_t)(msk));
2505}
2506
2507//The total number of operations is 1
2508template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1)
2509{
2510        return mvmd128<32>::shufflei<shufflemask4_from_shufflemask2(msk)>(arg1);
2511}
2512
2513//The total number of operations is 9
2514template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2515{
2516        return simd_or(mvmd128<2>::srli<sh>(arg1), mvmd128<2>::slli<((64)-sh)>(arg2));
2517}
2518
2519//The total number of operations is 9
2520template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2521{
2522        return simd_or(mvmd128<4>::srli<sh>(arg1), mvmd128<4>::slli<((32)-sh)>(arg2));
2523}
2524
2525//The total number of operations is 9
2526template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2527{
2528        return simd_or(mvmd128<8>::srli<sh>(arg1), mvmd128<8>::slli<((16)-sh)>(arg2));
2529}
2530
2531//The total number of operations is 9
2532template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2533{
2534        return simd_or(mvmd128<16>::srli<sh>(arg1), mvmd128<16>::slli<((8)-sh)>(arg2));
2535}
2536
2537//The total number of operations is 9
2538template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2539{
2540        return simd_or(mvmd128<32>::srli<sh>(arg1), mvmd128<32>::slli<((4)-sh)>(arg2));
2541}
2542
2543//The total number of operations is 9
2544template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2545{
2546        return simd_or(mvmd128<64>::srli<sh>(arg1), mvmd128<64>::slli<((2)-sh)>(arg2));
2547}
2548
2549//The total number of operations is 9
2550template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2551{
2552        return simd_or(mvmd128<128>::srli<sh>(arg1), mvmd128<128>::slli<((1)-sh)>(arg2));
2553}
2554
2555//The total number of operations is 1
2556template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1)
2557{
2558        return mvmd128<32>::fill((-1*val1));
2559}
2560
2561//The total number of operations is 1
2562template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1)
2563{
2564        return mvmd128<(4)>::fill(((val1<<2)|val1));
2565}
2566
2567//The total number of operations is 1
2568template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1)
2569{
2570        return mvmd128<(8)>::fill(((val1<<4)|val1));
2571}
2572
2573//The total number of operations is 1
2574template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1)
2575{
2576        return _mm_set1_epi8((int32_t)(val1));
2577}
2578
2579//The total number of operations is 1
2580template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1)
2581{
2582        return _mm_set1_epi16((int32_t)(val1));
2583}
2584
2585//The total number of operations is 1
2586template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1)
2587{
2588        return _mm_set1_epi32((int32_t)(val1));
2589}
2590
2591//The total number of operations is 1
2592template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1)
2593{
2594        return _mm_set_epi32((int32_t)((val1>>32)), (int32_t)(val1), (int32_t)((val1>>32)), (int32_t)(val1));
2595}
2596
2597//The total number of operations is 1
2598template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1)
2599{
2600        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val1>>32)), (int32_t)(val1));
2601}
2602
2603//The total number of operations is 1
2604template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2605{
2606        return _mm_shuffle_epi8(arg1, arg2);
2607}
2608
2609//The total number of operations is 10
2610template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2611{
2612        bitblock128_t tmp1 = simd_and(simd128<16>::constant<(7)>(), arg2);
2613        bitblock128_t msk1 = simd128<16>::add(tmp1, tmp1);
2614        bitblock128_t msk2 = simd128<16>::add(msk1, simd128<16>::constant<1>());
2615        bitblock128_t msk = simd_or(msk1, simd128<16>::slli<(8)>(msk2));
2616        return simd128<16>::ifh(arg2, simd128<16>::constant<0>(), mvmd128<(8)>::shuffle(arg1, msk));
2617}
2618
2619//The total number of operations is 19
2620template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2621{
2622        bitblock128_t tmp1 = simd_and(simd128<32>::constant<(3)>(), arg2);
2623        bitblock128_t msk1 = simd128<32>::add(tmp1, tmp1);
2624        bitblock128_t msk2 = simd128<32>::add(msk1, simd128<32>::constant<1>());
2625        bitblock128_t msk = simd_or(msk1, simd128<32>::slli<(16)>(msk2));
2626        return simd128<32>::ifh(arg2, simd128<32>::constant<0>(), mvmd128<(16)>::shuffle(arg1, msk));
2627}
2628
2629//The total number of operations is 28
2630template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2631{
2632        bitblock128_t tmp1 = simd_and(simd128<64>::constant<(1)>(), arg2);
2633        bitblock128_t msk1 = simd128<64>::add(tmp1, tmp1);
2634        bitblock128_t msk2 = simd128<64>::add(msk1, simd128<64>::constant<1>());
2635        bitblock128_t msk = simd_or(msk1, simd128<64>::slli<(32)>(msk2));
2636        return simd128<64>::ifh(arg2, simd128<64>::constant<0>(), mvmd128<(32)>::shuffle(arg1, msk));
2637}
2638
2639//The total number of operations is 1
2640template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1)
2641{
2642        return (((pos%2) == 0) ? (mvmd128<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd128<(2)>::extract<(pos/2)>(arg1)>>1));
2643}
2644
2645//The total number of operations is 1
2646template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1)
2647{
2648        return (((pos%2) == 0) ? (mvmd128<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd128<(4)>::extract<(pos/2)>(arg1)>>2));
2649}
2650
2651//The total number of operations is 1
2652template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1)
2653{
2654        return (((pos%2) == 0) ? (mvmd128<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd128<(8)>::extract<(pos/2)>(arg1)>>4));
2655}
2656
2657//The total number of operations is 1
2658template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1)
2659{
2660        return (((pos%2) == 0) ? (mvmd128<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd128<(16)>::extract<(pos/2)>(arg1)>>8));
2661}
2662
2663//The total number of operations is 1
2664template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1)
2665{
2666        return (65535&_mm_extract_epi16(arg1, (int32_t)(pos)));
2667}
2668
2669//The total number of operations is 1
2670template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1)
2671{
2672        return (((uint64_t)((4294967296ULL)-1))&_mm_extract_epi32(arg1, (int32_t)(pos)));
2673}
2674
2675//The total number of operations is 2
2676template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1)
2677{
2678        return ((((uint64_t)mvmd128<(32)>::extract<((2*pos)+1)>(arg1))<<(32))|mvmd128<(32)>::extract<(2*pos)>(arg1));
2679}
2680
2681//The total number of operations is 14
2682template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1)
2683{
2684        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(2)>::slli<1>(arg1) : simd128<(2)>::srli<1>(arg1));
2685        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(2)>::lomask(), arg1) : simd_and(simd128<(2)>::himask(), arg1));
2686        return mvmd128<(2)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2687}
2688
2689//The total number of operations is 10
2690template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1)
2691{
2692        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(4)>::slli<2>(arg1) : simd128<(4)>::srli<2>(arg1));
2693        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(4)>::lomask(), arg1) : simd_and(simd128<(4)>::himask(), arg1));
2694        return mvmd128<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2695}
2696
2697//The total number of operations is 6
2698template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1)
2699{
2700        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(8)>::slli<4>(arg1) : simd128<(8)>::srli<4>(arg1));
2701        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(8)>::lomask(), arg1) : simd_and(simd128<(8)>::himask(), arg1));
2702        return mvmd128<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2703}
2704
2705//The total number of operations is 2
2706template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1)
2707{
2708        return mvmd128<8>::fill(_mm_extract_epi8(arg1, (int32_t)(pos)));
2709}
2710
2711//The total number of operations is 2
2712template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1)
2713{
2714        return mvmd128<16>::fill(_mm_extract_epi16(arg1, (int32_t)(pos)));
2715}
2716
2717//The total number of operations is 1
2718template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1)
2719{
2720        return mvmd128<32>::shufflei<shufflemask4(pos, pos, pos, pos)>(arg1);
2721}
2722
2723//The total number of operations is 5
2724template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1)
2725{
2726        return simd128<1>::ifh(simd128<64>::himask(), mvmd128<(32)>::splat<((2*pos)+1)>(arg1), mvmd128<(32)>::splat<(2*pos)>(arg1));
2727}
2728
2729//The total number of operations is 13
2730template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1)
2731{
2732        return simd128<1>::ifh(simd128<128>::himask(), mvmd128<(64)>::splat<((2*pos)+1)>(arg1), mvmd128<(64)>::splat<(2*pos)>(arg1));
2733}
2734
2735//The total number of operations is 15
2736template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2737{
2738        return simd_or(mvmd128<(2)>::fill16((val1<<1), (val3<<1), (val5<<1), (val7<<1), (val9<<1), (val11<<1), (val13<<1), (val15<<1), (val1<<1), (val3<<1), (val5<<1), (val7<<1), (val9<<1), (val11<<1), (val13<<1), (val15<<1)), mvmd128<(2)>::fill16((val2&(1)), (val4&(1)), (val6&(1)), (val8&(1)), (val10&(1)), (val12&(1)), (val14&(1)), (val16&(1)), (val2&(1)), (val4&(1)), (val6&(1)), (val8&(1)), (val10&(1)), (val12&(1)), (val14&(1)), (val16&(1))));
2739}
2740
2741//The total number of operations is 7
2742template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2743{
2744        return simd_or(mvmd128<(4)>::fill16((val1<<2), (val3<<2), (val5<<2), (val7<<2), (val9<<2), (val11<<2), (val13<<2), (val15<<2), (val1<<2), (val3<<2), (val5<<2), (val7<<2), (val9<<2), (val11<<2), (val13<<2), (val15<<2)), mvmd128<(4)>::fill16((val2&(3)), (val4&(3)), (val6&(3)), (val8&(3)), (val10&(3)), (val12&(3)), (val14&(3)), (val16&(3)), (val2&(3)), (val4&(3)), (val6&(3)), (val8&(3)), (val10&(3)), (val12&(3)), (val14&(3)), (val16&(3))));
2745}
2746
2747//The total number of operations is 3
2748template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2749{
2750        return simd_or(mvmd128<(8)>::fill16((val1<<4), (val3<<4), (val5<<4), (val7<<4), (val9<<4), (val11<<4), (val13<<4), (val15<<4), (val1<<4), (val3<<4), (val5<<4), (val7<<4), (val9<<4), (val11<<4), (val13<<4), (val15<<4)), mvmd128<(8)>::fill16((val2&(15)), (val4&(15)), (val6&(15)), (val8&(15)), (val10&(15)), (val12&(15)), (val14&(15)), (val16&(15)), (val2&(15)), (val4&(15)), (val6&(15)), (val8&(15)), (val10&(15)), (val12&(15)), (val14&(15)), (val16&(15))));
2751}
2752
2753//The total number of operations is 1
2754template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2755{
2756        return _mm_set_epi8((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4), (int32_t)(val5), (int32_t)(val6), (int32_t)(val7), (int32_t)(val8), (int32_t)(val9), (int32_t)(val10), (int32_t)(val11), (int32_t)(val12), (int32_t)(val13), (int32_t)(val14), (int32_t)(val15), (int32_t)(val16));
2757}
2758
2759//The total number of operations is 5
2760template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2761{
2762        return simd128<1>::ifh(simd128<(4)>::himask(), mvmd128<1>::fill2(val1, val2), mvmd128<1>::fill2(val3, val4));
2763}
2764
2765//The total number of operations is 5
2766template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2767{
2768        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<2>::fill2(val1, val2), mvmd128<2>::fill2(val3, val4));
2769}
2770
2771//The total number of operations is 5
2772template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2773{
2774        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<4>::fill2(val1, val2), mvmd128<4>::fill2(val3, val4));
2775}
2776
2777//The total number of operations is 5
2778template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2779{
2780        return simd128<1>::ifh(simd128<(32)>::himask(), mvmd128<8>::fill2(val1, val2), mvmd128<8>::fill2(val3, val4));
2781}
2782
2783//The total number of operations is 3
2784template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2785{
2786        return simd_or(mvmd128<(32)>::fill4((val1<<16), (val3<<16), (val1<<16), (val3<<16)), mvmd128<(32)>::fill4((val2&(65535)), (val4&(65535)), (val2&(65535)), (val4&(65535))));
2787}
2788
2789//The total number of operations is 1
2790template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2791{
2792        return _mm_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
2793}
2794
2795//The total number of operations is 4
2796template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1)
2797{
2798        return simd128<128>::srli<(sh*2)>(arg1);
2799}
2800
2801//The total number of operations is 4
2802template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1)
2803{
2804        return simd128<128>::srli<(sh*4)>(arg1);
2805}
2806
2807//The total number of operations is 4
2808template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1)
2809{
2810        return simd128<128>::srli<(sh*8)>(arg1);
2811}
2812
2813//The total number of operations is 4
2814template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1)
2815{
2816        return simd128<128>::srli<(sh*16)>(arg1);
2817}
2818
2819//The total number of operations is 4
2820template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1)
2821{
2822        return simd128<128>::srli<(sh*32)>(arg1);
2823}
2824
2825//The total number of operations is 4
2826template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1)
2827{
2828        return simd128<128>::srli<(sh*64)>(arg1);
2829}
2830
2831//The total number of operations is 4
2832template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1)
2833{
2834        return simd128<128>::srli<(sh*128)>(arg1);
2835}
2836
2837//The total number of operations is 1
2838template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2)
2839{
2840        return mvmd128<(2)>::fill(((val1<<1)|(val2&(1))));
2841}
2842
2843//The total number of operations is 1
2844template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2)
2845{
2846        return mvmd128<(4)>::fill(((val1<<2)|(val2&(3))));
2847}
2848
2849//The total number of operations is 1
2850template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2)
2851{
2852        return mvmd128<(8)>::fill(((val1<<4)|(val2&(15))));
2853}
2854
2855//The total number of operations is 1
2856template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2)
2857{
2858        return mvmd128<(16)>::fill(((val1<<8)|(val2&(255))));
2859}
2860
2861//The total number of operations is 1
2862template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2)
2863{
2864        return mvmd128<(32)>::fill(((val1<<16)|(val2&(65535))));
2865}
2866
2867//The total number of operations is 5
2868template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2)
2869{
2870        return simd128<1>::ifh(simd128<(64)>::himask(), mvmd128<32>::fill(val1), mvmd128<32>::fill(val2));
2871}
2872
2873//The total number of operations is 5
2874template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2)
2875{
2876        return simd128<1>::ifh(simd128<(128)>::himask(), mvmd128<64>::fill(val1), mvmd128<64>::fill(val2));
2877}
2878
2879//The total number of operations is 9
2880template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2881{
2882        return simd_or(mvmd128<2>::slli<sh>(arg1), mvmd128<2>::srli<((64)-sh)>(arg2));
2883}
2884
2885//The total number of operations is 9
2886template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2887{
2888        return simd_or(mvmd128<4>::slli<sh>(arg1), mvmd128<4>::srli<((32)-sh)>(arg2));
2889}
2890
2891//The total number of operations is 9
2892template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2893{
2894        return simd_or(mvmd128<8>::slli<sh>(arg1), mvmd128<8>::srli<((16)-sh)>(arg2));
2895}
2896
2897//The total number of operations is 9
2898template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2899{
2900        return simd_or(mvmd128<16>::slli<sh>(arg1), mvmd128<16>::srli<((8)-sh)>(arg2));
2901}
2902
2903//The total number of operations is 9
2904template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2905{
2906        return simd_or(mvmd128<32>::slli<sh>(arg1), mvmd128<32>::srli<((4)-sh)>(arg2));
2907}
2908
2909//The total number of operations is 9
2910template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2911{
2912        return simd_or(mvmd128<64>::slli<sh>(arg1), mvmd128<64>::srli<((2)-sh)>(arg2));
2913}
2914
2915//The total number of operations is 9
2916template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2917{
2918        return simd_or(mvmd128<128>::slli<sh>(arg1), mvmd128<128>::srli<((1)-sh)>(arg2));
2919}
2920
2921//The total number of operations is 4
2922template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1)
2923{
2924        return simd128<128>::slli<(sh*2)>(arg1);
2925}
2926
2927//The total number of operations is 4
2928template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1)
2929{
2930        return mvmd128<(2)>::slli<(sh*2)>(arg1);
2931}
2932
2933//The total number of operations is 4
2934template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1)
2935{
2936        return simd128<128>::slli<(sh*8)>(arg1);
2937}
2938
2939//The total number of operations is 4
2940template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1)
2941{
2942        return mvmd128<(8)>::slli<(sh*2)>(arg1);
2943}
2944
2945//The total number of operations is 4
2946template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1)
2947{
2948        return simd128<128>::slli<(sh*32)>(arg1);
2949}
2950
2951//The total number of operations is 4
2952template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1)
2953{
2954        return mvmd128<(32)>::slli<(sh*2)>(arg1);
2955}
2956
2957//The total number of operations is 4
2958template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1)
2959{
2960        return simd128<128>::slli<(sh*128)>(arg1);
2961}
2962
2963//The total number of operations is 13
2964template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2965{
2966        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<1>::fill4(val1, val2, val3, val4), mvmd128<1>::fill4(val5, val6, val7, val8));
2967}
2968
2969//The total number of operations is 13
2970template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2971{
2972        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<2>::fill4(val1, val2, val3, val4), mvmd128<2>::fill4(val5, val6, val7, val8));
2973}
2974
2975//The total number of operations is 7
2976template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2977{
2978        return simd_or(mvmd128<(8)>::fill8((val1<<4), (val3<<4), (val5<<4), (val7<<4), (val1<<4), (val3<<4), (val5<<4), (val7<<4)), mvmd128<(8)>::fill8((val2&(15)), (val4&(15)), (val6&(15)), (val8&(15)), (val2&(15)), (val4&(15)), (val6&(15)), (val8&(15))));
2979}
2980
2981//The total number of operations is 3
2982template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2983{
2984        return simd_or(mvmd128<(16)>::fill8((val1<<8), (val3<<8), (val5<<8), (val7<<8), (val1<<8), (val3<<8), (val5<<8), (val7<<8)), mvmd128<(16)>::fill8((val2&(255)), (val4&(255)), (val6&(255)), (val8&(255)), (val2&(255)), (val4&(255)), (val6&(255)), (val8&(255))));
2985}
2986
2987//The total number of operations is 1
2988template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2989{
2990        return _mm_set_epi16((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4), (int32_t)(val5), (int32_t)(val6), (int32_t)(val7), (int32_t)(val8));
2991}
2992
2993//The total number of operations is 11
2994IDISA_ALWAYS_INLINE bitblock128_t bitblock128::sll(bitblock128_t arg1, bitblock128_t arg2)
2995{
2996        return simd128<128>::sll(arg1, arg2);
2997}
2998
2999//The total number of operations is 1
3000IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(bitblock128_t* arg1)
3001{
3002        return _mm_loadu_si128((bitblock128_t*)(arg1));
3003}
3004
3005//The total number of operations is 4
3006template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srli(bitblock128_t arg1)
3007{
3008        return simd128<128>::srli<sh>(arg1);
3009}
3010
3011//The total number of operations is 1
3012IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t* arg1, bitblock128_t arg2)
3013{
3014        _mm_store_si128((bitblock128_t*)(arg1), arg2);
3015}
3016
3017//The total number of operations is 20
3018IDISA_ALWAYS_INLINE uint64_t bitblock128::popcount(bitblock128_t arg1)
3019{
3020        return mvmd128<64>::extract<0>(simd128<128>::popcount(arg1));
3021}
3022
3023//The total number of operations is 2
3024IDISA_ALWAYS_INLINE bool bitblock128::all(bitblock128_t arg1)
3025{
3026        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<-1>())) == 65535;
3027}
3028
3029//The total number of operations is 11
3030IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srl(bitblock128_t arg1, bitblock128_t arg2)
3031{
3032        return simd128<128>::srl(arg1, arg2);
3033}
3034
3035//The total number of operations is 4
3036template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::slli(bitblock128_t arg1)
3037{
3038        return simd128<128>::slli<sh>(arg1);
3039}
3040
3041//The total number of operations is 2
3042IDISA_ALWAYS_INLINE bool bitblock128::any(bitblock128_t arg1)
3043{
3044        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<0>())) != 65535;
3045}
3046
3047//The total number of operations is 1
3048IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(bitblock128_t* arg1)
3049{
3050        return _mm_load_si128((bitblock128_t*)(arg1));
3051}
3052
3053//The total number of operations is 1
3054IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t* arg1, bitblock128_t arg2)
3055{
3056        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
3057}
3058
3059#endif
Note: See TracBrowser for help on using the repository browser.