source: trunk/lib/idisa_cpp/idisa_ssse3.cpp @ 1570

Last change on this file since 1570 was 1570, checked in by cameron, 8 years ago

bitblock::(load,store) (aligned,unaligned); <1>::add, sub

File size: 153.0 KB
Line 
1#ifndef IDISA_SSSE3_CPP
2#define IDISA_SSSE3_CPP
3#include <stdint.h>
4#include "../config.hpp"
5
6#include "tmmintrin.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 umin(bitblock128_t arg1, bitblock128_t arg2);
31        template <uint64_t val> static IDISA_ALWAYS_INLINE bitblock128_t constant();
32        static IDISA_ALWAYS_INLINE bitblock128_t min(bitblock128_t arg1, bitblock128_t arg2);
33        static IDISA_ALWAYS_INLINE bitblock128_t lomask();
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 load_unaligned(bitblock128_t* arg1);
92        static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t* arg1, bitblock128_t arg2);
93        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
94        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
95        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
96        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(bitblock128_t* arg1);
97        static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t* arg1, bitblock128_t arg2);
98};
99
100//Declaration Part
101IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2);
102IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1);
103IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2);
104IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2);
105IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2);
106IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2);
107template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2);
108template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2);
109template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2);
110template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2);
111template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2);
112template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2);
113template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2);
114template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2);
115template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2);
116template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2);
117template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2);
118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2);
119template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2);
120template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2);
121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2);
122template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2);
123template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2);
124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2);
125template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2);
126template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2);
127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2);
128template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2);
129template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2);
130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2);
131template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2);
132template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2);
133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2);
134template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2);
135template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2);
136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2);
137template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2);
138template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2);
139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2);
140template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2);
141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2);
142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2);
143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2);
144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2);
145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2);
146template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2);
147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2);
148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2);
149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2);
150template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2);
151template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2);
152template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2);
153template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2);
154template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1);
155template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1);
156template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1);
157template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1);
158template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1);
159template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1);
160template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1);
161template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1);
162template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1);
163template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1);
164template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1);
165template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1);
166template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1);
167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1);
168template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1);
169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
170template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
171template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2);
172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2);
173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2);
174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2);
175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2);
176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2);
177template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2);
178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2);
179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1);
180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1);
181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1);
182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1);
183template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1);
184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1);
185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1);
186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1);
187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1);
188template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1);
189template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1);
190template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1);
191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1);
192template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1);
193template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1);
194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::neg(bitblock128_t arg1);
195template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1);
196template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1);
197template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1);
198template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1);
199template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1);
200template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1);
201template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1);
202template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1);
203template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1);
204template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1);
205template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1);
206template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1);
207template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1);
208template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1);
209template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
210template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
211template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
212template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
213template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
214template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
215template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
216template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
217template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2);
218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2);
219template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2);
220template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2);
221template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2);
222template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2);
223template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2);
224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2);
225template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1);
226template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1);
227template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1);
228template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1);
229template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1);
230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1);
231template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1);
232template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
233template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
234template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask();
235template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask();
236template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask();
237template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask();
238template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask();
239template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask();
240template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask();
241template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
242template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
243template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
244template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
245template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
246template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
247template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
248template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
249template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2);
250template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2);
251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2);
252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2);
253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2);
254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2);
255template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2);
256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2);
257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2);
258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2);
259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2);
260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2);
261template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2);
262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2);
263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2);
264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2);
265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2);
266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2);
267template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2);
268template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2);
269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2);
270template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2);
271template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2);
272template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2);
273template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2);
274template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2);
275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2);
276template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2);
277template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2);
278template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2);
279template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2);
280template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2);
281template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1);
282template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1);
283template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1);
284template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1);
285template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1);
286template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1);
287template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1);
288template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask();
289template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask();
290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask();
291template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask();
292template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask();
293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask();
294template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask();
295template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2);
296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2);
297template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2);
298template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2);
299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2);
300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2);
301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2);
302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2);
303template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::abs(bitblock128_t arg1);
304template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1);
305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1);
306template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1);
307template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1);
308template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1);
309template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1);
310template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1);
311template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
312template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
313template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
314template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
315template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
316template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
317template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
318template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
319template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
320template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
321template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
322template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
323template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
324template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
325template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2);
326template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2);
327template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2);
328template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2);
329template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2);
330template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2);
331template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2);
332template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1);
333template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1);
334template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1);
335template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1);
336template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1);
337template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1);
338template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2);
339template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2);
340template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2);
341template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2);
342template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2);
343template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2);
344template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2);
345template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2);
346template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2);
347template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2);
348template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2);
349template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2);
350template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2);
351template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2);
352template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
353template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
354template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
355template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
356template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
357template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
358template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
359template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2);
360template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2);
361template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2);
362template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2);
363template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2);
364template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2);
365template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2);
366template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2);
367template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2);
368template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2);
369template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2);
370template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2);
371template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2);
372template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2);
373template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
374template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
376template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
377template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
379template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
380template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1);
381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1);
382template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1);
383template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1);
384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1);
385template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1);
386template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1);
387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1);
388template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1);
389template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1);
390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1);
391template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1);
392template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1);
393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1);
394template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1);
395template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1);
396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1);
397template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1);
398template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1);
399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1);
400template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1);
401template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1);
402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1);
403template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1);
404template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1);
405template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1);
406template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1);
407template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1);
408template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1);
409template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1);
410template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1);
411template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
412template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
413template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
414template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
415template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
416template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
417template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
418template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1);
419template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1);
420template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1);
421template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1);
422template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1);
423template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1);
424template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1);
425template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1);
426template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
427template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
428template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
429template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
430template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1);
431template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1);
432template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1);
433template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1);
434template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1);
435template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1);
436template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1);
437template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1);
438template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1);
439template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1);
440template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1);
441template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1);
442template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1);
443template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1);
444template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1);
445template <> 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);
446template <> 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);
447template <> 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);
448template <> 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);
449template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1);
450template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1);
451template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1);
452template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1);
453template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1);
454template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1);
455template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1);
456template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
457template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
458template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
459template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
460template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
461template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
462template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1);
463template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1);
464template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1);
465template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1);
466template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1);
467template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1);
468template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1);
469template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2);
470template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2);
471template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2);
472template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2);
473template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2);
474template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2);
475template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2);
476template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2);
477template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2);
478template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2);
479template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2);
480template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2);
481template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2);
482template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2);
483template <> 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);
484template <> 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);
485template <> 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);
486template <> 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);
487template <> 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);
488
489//Implementation Part
490
491#define shufflemask4(s1, s2, s3, s4) \
492        ((s1<<6) | (s2<<4) | (s3<<2) | s4)
493
494#define shufflemask4_from_shufflemask2(msk) \
495        (msk==3 ? 238 : (msk==2 ? 228 : (msk==1 ? 78 : 68)))
496
497#define shufflemask8_to_shufflemask4(msk) \
498        ((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))
499//The total number of operations is 2
500IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2)
501{
502        return simd_not(simd_or(arg1, arg2));
503}
504
505//The total number of operations is 1
506IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1)
507{
508        return simd_xor(arg1, simd128<32>::constant<-1>());
509}
510
511//The total number of operations is 1
512IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2)
513{
514        return _mm_andnot_si128(arg2, arg1);
515}
516
517//The total number of operations is 1
518IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2)
519{
520        return _mm_or_si128(arg1, arg2);
521}
522
523//The total number of operations is 1
524IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2)
525{
526        return _mm_and_si128(arg1, arg2);
527}
528
529//The total number of operations is 1
530IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2)
531{
532        return _mm_xor_si128(arg1, arg2);
533}
534
535//The total number of operations is 39
536template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2)
537{
538        bitblock128_t high_bit = simd128<1>::constant<(1)>();
539        return simd_xor(simd128<1>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
540}
541
542//The total number of operations is 19
543template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2)
544{
545        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, arg2), simd_or(simd_and(arg2, simd128<128>::srli<1>(simd_or(arg1, simd_not(arg2)))), simd_and(arg1, simd128<128>::srli<1>(simd_or(simd_not(arg1), arg2)))));
546}
547
548//The total number of operations is 9
549template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2)
550{
551        bitblock128_t high_bit = simd128<4>::constant<(8)>();
552        return simd_xor(simd128<4>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
553}
554
555//The total number of operations is 4
556template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2)
557{
558        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg1, arg2);
559}
560
561//The total number of operations is 1
562template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2)
563{
564        return _mm_max_epi16(arg1, arg2);
565}
566
567//The total number of operations is 4
568template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2)
569{
570        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg1, arg2);
571}
572
573//The total number of operations is 18
574template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2)
575{
576        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg1, arg2);
577}
578
579//The total number of operations is 65
580template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2)
581{
582        bitblock128_t hiAns = simd128<(64)>::max(arg1, arg2);
583        bitblock128_t loAns = simd128<(64)>::umax(arg1, arg2);
584        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
585        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
586        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
587}
588
589//The total number of operations is 57
590template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2)
591{
592        bitblock128_t loMask = simd128<(2)>::lomask();
593        bitblock128_t tmpAns1 = simd128<(2)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
594        bitblock128_t tmpAns2 = simd128<(2)>::mult(simd128<(2)>::srli<1>(arg1), simd128<(2)>::srli<1>(arg2));
595        return simd128<1>::ifh(loMask, tmpAns1, simd128<(2)>::slli<1>(tmpAns2));
596}
597
598//The total number of operations is 23
599template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2)
600{
601        bitblock128_t tmp1 = simd128<128>::slli<1>(arg1);
602        bitblock128_t tmp2 = simd128<128>::slli<1>(arg2);
603        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));
604}
605
606//The total number of operations is 31
607template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2)
608{
609        bitblock128_t loMask = simd128<(8)>::lomask();
610        bitblock128_t tmpAns1 = simd128<(8)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
611        bitblock128_t tmpAns2 = simd128<(8)>::mult(simd128<(8)>::srli<4>(arg1), simd128<(8)>::srli<4>(arg2));
612        return simd128<1>::ifh(loMask, tmpAns1, simd128<(8)>::slli<4>(tmpAns2));
613}
614
615//The total number of operations is 10
616template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2)
617{
618        bitblock128_t loMask = simd128<(16)>::lomask();
619        bitblock128_t tmpAns1 = simd128<(16)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
620        bitblock128_t tmpAns2 = simd128<(16)>::mult(simd128<(16)>::srli<8>(arg1), simd128<(16)>::srli<8>(arg2));
621        return simd128<1>::ifh(loMask, tmpAns1, simd128<(16)>::slli<8>(tmpAns2));
622}
623
624//The total number of operations is 1
625template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2)
626{
627        return _mm_mullo_epi16(arg1, arg2);
628}
629
630//The total number of operations is 30
631template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2)
632{
633        bitblock128_t loMask = simd128<(64)>::lomask();
634        bitblock128_t tmpAns1 = simd128<(64)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
635        bitblock128_t tmpAns2 = simd128<(64)>::mult(simd128<(64)>::srli<32>(arg1), simd128<(64)>::srli<32>(arg2));
636        return simd128<1>::ifh(loMask, tmpAns1, simd128<(64)>::slli<32>(tmpAns2));
637}
638
639//The total number of operations is 11
640template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2)
641{
642        bitblock128_t loMask = simd128<64>::lomask();
643        bitblock128_t arg1_low = simd_and(arg1, loMask);
644        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg1);
645        bitblock128_t arg2_low = simd_and(arg2, loMask);
646        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg2);
647        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
648        bitblock128_t tmpAns2 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
649        bitblock128_t tmpAns3 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
650        return simd128<64>::add(tmpAns1, simd128<64>::add(tmpAns2, tmpAns3));
651}
652
653//The total number of operations is 205
654template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2)
655{
656        bitblock128_t loMask = simd128<128>::lomask();
657        bitblock128_t arg1_low = simd_and(arg1, loMask);
658        bitblock128_t arg1_high = simd128<128>::srli<(64)>(arg1);
659        bitblock128_t arg2_low = simd_and(arg2, loMask);
660        bitblock128_t arg2_high = simd128<128>::srli<(64)>(arg2);
661        bitblock128_t tmpAns1 = simd128<(64)>::umult(arg1_low, arg2_low);
662        bitblock128_t tmpAns2 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_low, arg2_high));
663        bitblock128_t tmpAns3 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_high, arg2_low));
664        return simd128<128>::add(tmpAns1, simd128<128>::add(tmpAns2, tmpAns3));
665}
666
667//The total number of operations is 42
668template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2)
669{
670        bitblock128_t high_bit = simd128<1>::constant<(1)>();
671        return simd128<1>::ugt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
672}
673
674//The total number of operations is 18
675template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2)
676{
677        bitblock128_t tmp = simd_not(arg1);
678        bitblock128_t tmpAns = simd_or(simd_and(tmp, arg2), simd_and(simd128<128>::slli<1>(simd_and(arg1, simd_not(arg2))), simd_or(tmp, arg2)));
679        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
680}
681
682//The total number of operations is 10
683template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2)
684{
685        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)));
686}
687
688//The total number of operations is 1
689template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2)
690{
691        return _mm_cmpgt_epi8(arg1, arg2);
692}
693
694//The total number of operations is 1
695template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2)
696{
697        return _mm_cmpgt_epi16(arg1, arg2);
698}
699
700//The total number of operations is 1
701template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2)
702{
703        return _mm_cmpgt_epi32(arg1, arg2);
704}
705
706//The total number of operations is 15
707template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2)
708{
709        bitblock128_t hiAns = simd128<(32)>::gt(arg1, arg2);
710        bitblock128_t loAns = simd128<(32)>::ugt(arg1, arg2);
711        bitblock128_t mask = simd_and(loAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
712        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
713        return simd_or(simd128<64>::srai<(32)>(hiAns), mask);
714}
715
716//The total number of operations is 66
717template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2)
718{
719        bitblock128_t hiAns = simd128<(64)>::gt(arg1, arg2);
720        bitblock128_t loAns = simd128<(64)>::ugt(arg1, arg2);
721        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
722        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
723        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
724}
725
726//The total number of operations is 289
727template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2)
728{
729        bitblock128_t loMask = simd128<(2)>::lomask();
730        bitblock128_t tmpAns1 = simd128<(2)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
731        bitblock128_t tmpAns2 = simd128<(2)>::umult(simd_and(loMask, simd128<(4)>::srli<(2)>(arg1)), simd_and(loMask, simd128<(4)>::srli<(2)>(arg2)));
732        return simd_or(tmpAns1, simd128<(4)>::slli<(2)>(tmpAns2));
733}
734
735//The total number of operations is 139
736template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2)
737{
738        bitblock128_t loMask = simd128<(4)>::lomask();
739        bitblock128_t tmpAns1 = simd128<(4)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
740        bitblock128_t tmpAns2 = simd128<(4)>::umult(simd_and(loMask, simd128<(8)>::srli<(4)>(arg1)), simd_and(loMask, simd128<(8)>::srli<(4)>(arg2)));
741        return simd_or(tmpAns1, simd128<(8)>::slli<(4)>(tmpAns2));
742}
743
744//The total number of operations is 64
745template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2)
746{
747        bitblock128_t loMask = simd128<(8)>::lomask();
748        bitblock128_t tmpAns1 = simd128<(8)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
749        bitblock128_t tmpAns2 = simd128<(8)>::umult(simd_and(loMask, simd128<(16)>::srli<(8)>(arg1)), simd_and(loMask, simd128<(16)>::srli<(8)>(arg2)));
750        return simd_or(tmpAns1, simd128<(16)>::slli<(8)>(tmpAns2));
751}
752
753//The total number of operations is 28
754template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2)
755{
756        bitblock128_t loMask = simd128<(16)>::lomask();
757        bitblock128_t tmpAns1 = simd128<(16)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
758        bitblock128_t tmpAns2 = simd128<(16)>::umult(simd_and(loMask, simd128<(32)>::srli<(16)>(arg1)), simd_and(loMask, simd128<(32)>::srli<(16)>(arg2)));
759        return simd_or(tmpAns1, simd128<(32)>::slli<(16)>(tmpAns2));
760}
761
762//The total number of operations is 10
763template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2)
764{
765        bitblock128_t loMask = simd128<(32)>::lomask();
766        bitblock128_t tmpAns1 = simd128<(32)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
767        bitblock128_t tmpAns2 = simd128<(32)>::umult(simd_and(loMask, simd128<(64)>::srli<(32)>(arg1)), simd_and(loMask, simd128<(64)>::srli<(32)>(arg2)));
768        return simd_or(tmpAns1, simd128<(64)>::slli<(32)>(tmpAns2));
769}
770
771//The total number of operations is 1
772template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2)
773{
774        return _mm_mul_epu32(arg1, arg2);
775}
776
777//The total number of operations is 55
778template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2)
779{
780        bitblock128_t loMask1 = simd128<(128)>::lomask();
781        bitblock128_t arg11 = simd_and(arg1, loMask1);
782        bitblock128_t arg22 = simd_and(arg2, loMask1);
783        bitblock128_t loMask2 = simd128<64>::lomask();
784        bitblock128_t arg1_low = simd_and(arg11, loMask2);
785        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg11);
786        bitblock128_t arg2_low = simd_and(arg22, loMask2);
787        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg22);
788        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
789        bitblock128_t tmpAns2 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
790        bitblock128_t tmpAns3 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
791        bitblock128_t tmpAns4 = simd128<(128)>::slli<64>(simd128<(32)>::umult(arg1_high, arg2_high));
792        return simd128<(128)>::add(tmpAns1, simd128<(128)>::add(tmpAns2, simd128<(128)>::add(tmpAns3, tmpAns4)));
793}
794
795//The total number of operations is 40
796template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2)
797{
798        return simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::ult(arg1, simd_and(simd128<(2)>::himask(), arg2)), simd128<(2)>::ult(simd_andc(arg1, simd128<(2)>::himask()), simd_andc(arg2, simd128<(2)>::himask())));
799}
800
801//The total number of operations is 17
802template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2)
803{
804        bitblock128_t tmp = simd_not(arg1);
805        bitblock128_t tmpAns = simd_or(simd_and(tmp, arg2), simd_and(simd128<128>::slli<1>(simd_and(tmp, arg2)), simd_or(tmp, arg2)));
806        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
807}
808
809//The total number of operations is 20
810template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2)
811{
812        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())));
813}
814
815//The total number of operations is 7
816template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2)
817{
818        bitblock128_t high_bit = simd128<8>::constant<(128)>();
819        return simd128<8>::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<16>::ult(bitblock128_t arg1, bitblock128_t arg2)
824{
825        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
826        return simd128<16>::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<32>::ult(bitblock128_t arg1, bitblock128_t arg2)
831{
832        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
833        return simd128<32>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
834}
835
836//The total number of operations is 18
837template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2)
838{
839        bitblock128_t tmpAns = simd128<(32)>::ult(arg1, arg2);
840        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
841        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
842        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
843}
844
845//The total number of operations is 55
846template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2)
847{
848        return simd_and(simd128<128>::srai<(127)>(simd_or(simd_and(simd_not(arg1), arg2), simd_and(simd_not(simd_xor(arg1, arg2)), simd128<128>::sub(arg1, arg2)))), simd_not(simd128<128>::eq(arg1, arg2)));
849}
850
851//The total number of operations is 42
852template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2)
853{
854        bitblock128_t high_bit = simd128<1>::constant<(1)>();
855        return simd128<1>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
856}
857
858//The total number of operations is 18
859template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2)
860{
861        bitblock128_t tmp = simd_not(arg2);
862        bitblock128_t tmpAns = simd_or(simd_and(arg1, tmp), simd_and(simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2)), simd_or(arg1, tmp)));
863        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
864}
865
866//The total number of operations is 18
867template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2)
868{
869        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)));
870}
871
872//The total number of operations is 5
873template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2)
874{
875        return simd_and(simd_not(simd128<8>::gt(arg1, arg2)), simd_not(simd128<8>::eq(arg1, arg2)));
876}
877
878//The total number of operations is 5
879template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2)
880{
881        return simd_and(simd_not(simd128<16>::gt(arg1, arg2)), simd_not(simd128<16>::eq(arg1, arg2)));
882}
883
884//The total number of operations is 5
885template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2)
886{
887        return simd_and(simd_not(simd128<32>::gt(arg1, arg2)), simd_not(simd128<32>::eq(arg1, arg2)));
888}
889
890//The total number of operations is 20
891template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2)
892{
893        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808UL)>();
894        return simd128<64>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
895}
896
897//The total number of operations is 75
898template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2)
899{
900        bitblock128_t hiAns = simd128<(64)>::lt(arg1, arg2);
901        bitblock128_t loAns = simd128<(64)>::ult(arg1, arg2);
902        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
903        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
904        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
905}
906
907//The total number of operations is 2
908template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1)
909{
910        return simd_and(simd128<32>::srli<sh>(arg1), simd128<2>::constant<((3)>>sh)>());
911}
912
913//The total number of operations is 2
914template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1)
915{
916        return simd_and(simd128<32>::srli<sh>(arg1), simd128<4>::constant<((15)>>sh)>());
917}
918
919//The total number of operations is 2
920template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1)
921{
922        return simd_and(simd128<32>::srli<sh>(arg1), simd128<8>::constant<((255)>>sh)>());
923}
924
925//The total number of operations is 1
926template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1)
927{
928        return _mm_srli_epi16(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<32>::srli(bitblock128_t arg1)
933{
934        return _mm_srli_epi32(arg1, (int32_t)(sh));
935}
936
937//The total number of operations is 1
938template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1)
939{
940        return _mm_srli_epi64(arg1, (int32_t)(sh));
941}
942
943//The total number of operations is 4
944template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1)
945{
946        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)))));
947}
948
949//The total number of operations is 1
950template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1)
951{
952        return simd_not(arg1);
953}
954
955//The total number of operations is 14
956template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1)
957{
958        bitblock128_t tmp = simd_not(arg1);
959        return simd128<1>::ifh(simd128<2>::himask(), simd_and(tmp, simd128<128>::slli<1>(tmp)), simd_and(simd128<128>::srli<1>(arg1), tmp));
960}
961
962//The total number of operations is 14
963template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1)
964{
965        return simd128<4>::popcount(simd_andc(simd128<4>::sub(arg1, simd128<4>::constant<1>()), arg1));
966}
967
968//The total number of operations is 13
969template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1)
970{
971        return simd128<8>::popcount(simd_andc(simd128<8>::sub(arg1, simd128<8>::constant<1>()), arg1));
972}
973
974//The total number of operations is 16
975template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1)
976{
977        return simd128<16>::popcount(simd_andc(simd128<16>::sub(arg1, simd128<16>::constant<1>()), arg1));
978}
979
980//The total number of operations is 19
981template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1)
982{
983        return simd128<32>::popcount(simd_andc(simd128<32>::sub(arg1, simd128<32>::constant<1>()), arg1));
984}
985
986//The total number of operations is 14
987template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1)
988{
989        return simd128<64>::popcount(simd_andc(simd128<64>::sub(arg1, simd128<64>::constant<1>()), arg1));
990}
991
992//The total number of operations is 30
993template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1)
994{
995        return simd128<128>::popcount(simd_andc(simd128<128>::sub(arg1, simd128<128>::constant<1>()), arg1));
996}
997
998//The total number of operations is 7
999template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1000{
1001        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())));
1002}
1003
1004//The total number of operations is 11
1005template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1006{
1007        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))));
1008}
1009
1010//The total number of operations is 40
1011template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1012{
1013        return simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::ugt(simd_and(simd128<(2)>::himask(), arg1), arg2), simd128<(2)>::ugt(simd_andc(arg1, simd128<(2)>::himask()), simd_andc(arg2, simd128<(2)>::himask())));
1014}
1015
1016//The total number of operations is 17
1017template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1018{
1019        bitblock128_t tmp = simd_not(arg2);
1020        bitblock128_t tmpAns = simd_or(simd_and(arg1, tmp), simd_and(simd128<128>::slli<1>(simd_and(arg1, tmp)), simd_or(arg1, tmp)));
1021        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1022}
1023
1024//The total number of operations is 12
1025template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1026{
1027        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())));
1028}
1029
1030//The total number of operations is 3
1031template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1032{
1033        bitblock128_t high_bit = simd128<8>::constant<(128)>();
1034        return simd128<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1035}
1036
1037//The total number of operations is 3
1038template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1039{
1040        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1041        return simd128<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1042}
1043
1044//The total number of operations is 3
1045template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1046{
1047        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1048        return simd128<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1049}
1050
1051//The total number of operations is 14
1052template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1053{
1054        bitblock128_t tmpAns = simd128<(32)>::ugt(arg1, arg2);
1055        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
1056        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
1057        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
1058}
1059
1060//The total number of operations is 51
1061template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1062{
1063        bitblock128_t tmpAns = simd128<(64)>::ugt(arg1, arg2);
1064        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
1065        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
1066        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
1067}
1068
1069//The total number of operations is 4
1070template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1)
1071{
1072        return simd_xor(simd128<2>::srli<(1)>(arg1), simd_and(arg1, simd128<2>::lomask()));
1073}
1074
1075//The total number of operations is 4
1076template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1)
1077{
1078        return simd_xor(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1079}
1080
1081//The total number of operations is 4
1082template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1)
1083{
1084        return simd_xor(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1085}
1086
1087//The total number of operations is 3
1088template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1)
1089{
1090        return simd_xor(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1091}
1092
1093//The total number of operations is 3
1094template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1)
1095{
1096        return simd_xor(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1097}
1098
1099//The total number of operations is 3
1100template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1)
1101{
1102        return simd_xor(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1103}
1104
1105//The total number of operations is 6
1106template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1)
1107{
1108        return simd_xor(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1109}
1110
1111//The total number of operations is 0
1112template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1)
1113{
1114        return arg1;
1115}
1116
1117//The total number of operations is 3
1118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1)
1119{
1120        return simd128<2>::add_hl(simd128<(1)>::popcount(arg1));
1121}
1122
1123//The total number of operations is 7
1124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1)
1125{
1126        return simd128<4>::add_hl(simd128<(2)>::popcount(arg1));
1127}
1128
1129//The total number of operations is 11
1130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1)
1131{
1132        return simd128<8>::add_hl(simd128<(4)>::popcount(arg1));
1133}
1134
1135//The total number of operations is 14
1136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1)
1137{
1138        return simd128<16>::add_hl(simd128<(8)>::popcount(arg1));
1139}
1140
1141//The total number of operations is 17
1142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1)
1143{
1144        return simd128<32>::add_hl(simd128<(16)>::popcount(arg1));
1145}
1146
1147//The total number of operations is 12
1148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1)
1149{
1150        return _mm_sad_epu8(simd128<8>::popcount(arg1), simd128<8>::constant<0>());
1151}
1152
1153//The total number of operations is 18
1154template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1)
1155{
1156        bitblock128_t tmpAns = simd128<(64)>::popcount(arg1);
1157        return simd128<(64)>::add(simd_and(tmpAns, simd128<128>::lomask()), simd128<128>::srli<(64)>(tmpAns));
1158}
1159
1160//The total number of operations is 1
1161template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::neg(bitblock128_t arg1)
1162{
1163        return simd128<1>::sub(simd128<1>::constant<0>(), arg1);
1164}
1165
1166//The total number of operations is 8
1167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1)
1168{
1169        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(arg1, simd128<128>::slli<1>(arg1)), arg1);
1170}
1171
1172//The total number of operations is 6
1173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1)
1174{
1175        return simd128<4>::sub(simd128<4>::constant<0>(), arg1);
1176}
1177
1178//The total number of operations is 1
1179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1)
1180{
1181        return simd128<8>::sub(simd128<8>::constant<0>(), arg1);
1182}
1183
1184//The total number of operations is 1
1185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1)
1186{
1187        return simd128<16>::sub(simd128<16>::constant<0>(), arg1);
1188}
1189
1190//The total number of operations is 1
1191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1)
1192{
1193        return _mm_sign_epi32(arg1, simd128<32>::constant<-1>());
1194}
1195
1196//The total number of operations is 1
1197template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1)
1198{
1199        return simd128<64>::sub(simd128<64>::constant<0>(), arg1);
1200}
1201
1202//The total number of operations is 11
1203template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1)
1204{
1205        return simd128<128>::sub(simd128<128>::constant<0>(), arg1);
1206}
1207
1208//The total number of operations is 2
1209template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1)
1210{
1211        return simd_and(simd128<32>::slli<sh>(arg1), simd128<2>::constant<(((3)<<sh)&(3))>());
1212}
1213
1214//The total number of operations is 2
1215template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1)
1216{
1217        return simd_and(simd128<32>::slli<sh>(arg1), simd128<4>::constant<(((15)<<sh)&(15))>());
1218}
1219
1220//The total number of operations is 2
1221template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1)
1222{
1223        return simd_and(simd128<32>::slli<sh>(arg1), simd128<8>::constant<(((255)<<sh)&(255))>());
1224}
1225
1226//The total number of operations is 1
1227template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1)
1228{
1229        return _mm_slli_epi16(arg1, (int32_t)(sh));
1230}
1231
1232//The total number of operations is 1
1233template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1)
1234{
1235        return _mm_slli_epi32(arg1, (int32_t)(sh));
1236}
1237
1238//The total number of operations is 1
1239template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1)
1240{
1241        return _mm_slli_epi64(arg1, (int32_t)(sh));
1242}
1243
1244//The total number of operations is 4
1245template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1)
1246{
1247        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)))));
1248}
1249
1250//The total number of operations is 3
1251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1252{
1253        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
1254}
1255
1256//The total number of operations is 8
1257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1258{
1259        return simd128<(1)>::ifh(simd128<1>::ifh(simd128<2>::himask(), arg1, simd128<2>::srli<(1)>(arg1)), arg2, arg3);
1260}
1261
1262//The total number of operations is 13
1263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1264{
1265        return simd128<1>::ifh(simd128<4>::gt(simd128<4>::constant<0>(), arg1), arg2, arg3);
1266}
1267
1268//The total number of operations is 4
1269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1270{
1271        return simd128<1>::ifh(simd128<8>::gt(simd128<8>::constant<0>(), arg1), arg2, arg3);
1272}
1273
1274//The total number of operations is 4
1275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1276{
1277        return simd128<1>::ifh(simd128<16>::gt(simd128<16>::constant<0>(), arg1), arg2, arg3);
1278}
1279
1280//The total number of operations is 4
1281template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1282{
1283        return simd128<1>::ifh(simd128<32>::gt(simd128<32>::constant<0>(), arg1), arg2, arg3);
1284}
1285
1286//The total number of operations is 8
1287template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1288{
1289        return simd128<(32)>::ifh(simd128<1>::ifh(simd128<64>::himask(), arg1, simd128<64>::srli<(32)>(arg1)), arg2, arg3);
1290}
1291
1292//The total number of operations is 15
1293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1294{
1295        return simd128<(64)>::ifh(simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg1)), arg2, arg3);
1296}
1297
1298//The total number of operations is 1
1299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2)
1300{
1301        return simd_xor(arg1, arg2);
1302}
1303
1304//The total number of operations is 11
1305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2)
1306{
1307        bitblock128_t tmp = simd_xor(arg1, arg2);
1308        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2))), tmp);
1309}
1310
1311//The total number of operations is 6
1312template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2)
1313{
1314        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::sub(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::sub(arg1, arg2));
1315}
1316
1317//The total number of operations is 1
1318template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2)
1319{
1320        return _mm_sub_epi8(arg1, arg2);
1321}
1322
1323//The total number of operations is 1
1324template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2)
1325{
1326        return _mm_sub_epi16(arg1, arg2);
1327}
1328
1329//The total number of operations is 1
1330template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2)
1331{
1332        return _mm_sub_epi32(arg1, arg2);
1333}
1334
1335//The total number of operations is 1
1336template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2)
1337{
1338        return _mm_sub_epi64(arg1, arg2);
1339}
1340
1341//The total number of operations is 11
1342template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2)
1343{
1344        bitblock128_t partial = simd128<(64)>::sub(arg1, arg2);
1345        bitblock128_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
1346        bitblock128_t borrow = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(borrowMask));
1347        return simd128<(64)>::sub(partial, borrow);
1348}
1349
1350//The total number of operations is 3
1351template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1)
1352{
1353        return simd128<16>::sub(arg1, simd_and(simd128<2>::lomask(), simd128<16>::srli<1>(arg1)));
1354}
1355
1356//The total number of operations is 4
1357template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1)
1358{
1359        return simd128<(8)>::add(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1360}
1361
1362//The total number of operations is 4
1363template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1)
1364{
1365        return simd128<(16)>::add(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1366}
1367
1368//The total number of operations is 3
1369template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1)
1370{
1371        return simd128<(32)>::add(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1372}
1373
1374//The total number of operations is 3
1375template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1)
1376{
1377        return simd128<(64)>::add(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1378}
1379
1380//The total number of operations is 3
1381template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1)
1382{
1383        return simd128<64>::add(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1384}
1385
1386//The total number of operations is 16
1387template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1)
1388{
1389        return simd128<128>::add(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1390}
1391
1392//The total number of operations is 7
1393template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1394{
1395        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())));
1396}
1397
1398//The total number of operations is 11
1399template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1400{
1401        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))));
1402}
1403
1404//The total number of operations is 0
1405template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask()
1406{
1407        return simd128<2>::constant<(1)>();
1408}
1409
1410//The total number of operations is 0
1411template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask()
1412{
1413        return simd128<4>::constant<(3)>();
1414}
1415
1416//The total number of operations is 0
1417template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask()
1418{
1419        return simd128<8>::constant<(15)>();
1420}
1421
1422//The total number of operations is 0
1423template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask()
1424{
1425        return simd128<16>::constant<(255)>();
1426}
1427
1428//The total number of operations is 0
1429template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask()
1430{
1431        return simd128<32>::constant<(65535)>();
1432}
1433
1434//The total number of operations is 0
1435template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask()
1436{
1437        return _mm_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1));
1438}
1439
1440//The total number of operations is 0
1441template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask()
1442{
1443        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1));
1444}
1445
1446//The total number of operations is 0
1447template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1448{
1449        return simd128<32>::constant<(-1*val)>();
1450}
1451
1452//The total number of operations is 0
1453template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1454{
1455        return ((val < 0) ? simd128<(4)>::constant<((val<<2)|(val^(-4)))>() : simd128<(4)>::constant<((val<<2)|val)>());
1456}
1457
1458//The total number of operations is 0
1459template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1460{
1461        return ((val < 0) ? simd128<(8)>::constant<((val<<4)|(val^(-16)))>() : simd128<(8)>::constant<((val<<4)|val)>());
1462}
1463
1464//The total number of operations is 0
1465template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1466{
1467        return _mm_set1_epi8((int32_t)(val));
1468}
1469
1470//The total number of operations is 0
1471template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1472{
1473        return _mm_set1_epi16((int32_t)(val));
1474}
1475
1476//The total number of operations is 0
1477template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1478{
1479        return _mm_set1_epi32((int32_t)(val));
1480}
1481
1482//The total number of operations is 0
1483template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1484{
1485        return _mm_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val));
1486}
1487
1488//The total number of operations is 0
1489template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1490{
1491        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val));
1492}
1493
1494//The total number of operations is 39
1495template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2)
1496{
1497        bitblock128_t high_bit = simd128<1>::constant<(1)>();
1498        return simd_xor(simd128<1>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1499}
1500
1501//The total number of operations is 19
1502template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2)
1503{
1504        bitblock128_t high_bit = simd128<2>::constant<(2)>();
1505        return simd_xor(simd128<2>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1506}
1507
1508//The total number of operations is 9
1509template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2)
1510{
1511        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1512        return simd_xor(simd128<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1513}
1514
1515//The total number of operations is 4
1516template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2)
1517{
1518        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg2, arg1);
1519}
1520
1521//The total number of operations is 1
1522template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2)
1523{
1524        return _mm_min_epi16(arg1, arg2);
1525}
1526
1527//The total number of operations is 4
1528template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2)
1529{
1530        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg2, arg1);
1531}
1532
1533//The total number of operations is 18
1534template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2)
1535{
1536        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg2, arg1);
1537}
1538
1539//The total number of operations is 65
1540template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2)
1541{
1542        bitblock128_t hiAns = simd128<(64)>::min(arg1, arg2);
1543        bitblock128_t loAns = simd128<(64)>::umin(arg1, arg2);
1544        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
1545        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
1546        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
1547}
1548
1549//The total number of operations is 36
1550template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2)
1551{
1552        return simd_or(simd_and(simd128<(2)>::himask(), simd128<(2)>::umin(arg1, arg2)), simd128<(2)>::umin(simd_and(simd128<(2)>::lomask(), arg1), simd_and(simd128<(2)>::lomask(), arg2)));
1553}
1554
1555//The total number of operations is 16
1556template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2)
1557{
1558        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)));
1559}
1560
1561//The total number of operations is 6
1562template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2)
1563{
1564        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)));
1565}
1566
1567//The total number of operations is 1
1568template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2)
1569{
1570        return _mm_min_epu8(arg1, arg2);
1571}
1572
1573//The total number of operations is 4
1574template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2)
1575{
1576        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1577        return simd_xor(simd128<16>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1578}
1579
1580//The total number of operations is 7
1581template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2)
1582{
1583        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1584        return simd_xor(simd128<32>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1585}
1586
1587//The total number of operations is 20
1588template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2)
1589{
1590        bitblock128_t tmpAns = simd128<(32)>::umin(arg1, arg2);
1591        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1592        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1593        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1594}
1595
1596//The total number of operations is 47
1597template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2)
1598{
1599        bitblock128_t tmpAns = simd128<(64)>::umin(arg1, arg2);
1600        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1601        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1602        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1603}
1604
1605//The total number of operations is 36
1606template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2)
1607{
1608        return simd_or(simd_and(simd128<(2)>::himask(), simd128<(2)>::umax(arg1, arg2)), simd128<(2)>::umax(simd_and(simd128<(2)>::lomask(), arg1), simd_and(simd128<(2)>::lomask(), arg2)));
1609}
1610
1611//The total number of operations is 16
1612template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2)
1613{
1614        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)));
1615}
1616
1617//The total number of operations is 6
1618template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2)
1619{
1620        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)));
1621}
1622
1623//The total number of operations is 1
1624template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2)
1625{
1626        return _mm_max_epu8(arg1, arg2);
1627}
1628
1629//The total number of operations is 4
1630template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2)
1631{
1632        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1633        return simd_xor(simd128<16>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1634}
1635
1636//The total number of operations is 7
1637template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2)
1638{
1639        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1640        return simd_xor(simd128<32>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1641}
1642
1643//The total number of operations is 20
1644template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2)
1645{
1646        bitblock128_t tmpAns = simd128<(32)>::umax(arg1, arg2);
1647        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1648        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1649        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1650}
1651
1652//The total number of operations is 47
1653template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2)
1654{
1655        bitblock128_t tmpAns = simd128<(64)>::umax(arg1, arg2);
1656        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1657        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1658        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1659}
1660
1661//The total number of operations is 37
1662template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2)
1663{
1664        return simd_or(simd_and(simd128<(2)>::himask(), simd128<(2)>::eq(simd_and(simd128<(2)>::himask(), arg1), simd_and(simd128<(2)>::himask(), arg2))), simd_and(simd128<(2)>::lomask(), simd128<(2)>::eq(simd_and(simd128<(2)>::lomask(), arg1), simd_and(simd128<(2)>::lomask(), arg2))));
1665}
1666
1667//The total number of operations is 15
1668template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2)
1669{
1670        bitblock128_t tmp = simd_xor(arg1, arg2);
1671        bitblock128_t tmpAns = simd_and(simd_not(simd128<128>::slli<1>(tmp)), simd_not(tmp));
1672        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1673}
1674
1675//The total number of operations is 9
1676template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2)
1677{
1678        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))));
1679}
1680
1681//The total number of operations is 1
1682template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2)
1683{
1684        return _mm_cmpeq_epi8(arg1, arg2);
1685}
1686
1687//The total number of operations is 1
1688template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2)
1689{
1690        return _mm_cmpeq_epi16(arg1, arg2);
1691}
1692
1693//The total number of operations is 1
1694template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2)
1695{
1696        return _mm_cmpeq_epi32(arg1, arg2);
1697}
1698
1699//The total number of operations is 5
1700template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2)
1701{
1702        bitblock128_t tmpAns = simd128<(32)>::eq(arg1, arg2);
1703        bitblock128_t loMask = simd_and(tmpAns, simd128<64>::srli<(32)>(tmpAns));
1704        bitblock128_t hiMask = simd128<64>::slli<(32)>(loMask);
1705        return simd_or(loMask, hiMask);
1706}
1707
1708//The total number of operations is 15
1709template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2)
1710{
1711        bitblock128_t tmpAns = simd128<(64)>::eq(arg1, arg2);
1712        bitblock128_t loMask = simd_and(tmpAns, simd128<128>::srli<(64)>(tmpAns));
1713        bitblock128_t hiMask = simd128<128>::slli<(64)>(loMask);
1714        return simd_or(loMask, hiMask);
1715}
1716
1717//The total number of operations is 4
1718template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1)
1719{
1720        return ((sh == 0) ? arg1 : simd_or(simd_and(simd128<2>::himask(), arg1), simd128<2>::srli<1>(arg1)));
1721}
1722
1723//The total number of operations is 10
1724template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1)
1725{
1726        bitblock128_t tmp = simd128<4>::srli<((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh))>(arg1);
1727        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)));
1728}
1729
1730//The total number of operations is 5
1731template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1)
1732{
1733        bitblock128_t tmp = simd128<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
1734        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)));
1735}
1736
1737//The total number of operations is 1
1738template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1)
1739{
1740        return _mm_srai_epi16(arg1, (int32_t)(sh));
1741}
1742
1743//The total number of operations is 1
1744template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1)
1745{
1746        return _mm_srai_epi32(arg1, (int32_t)(sh));
1747}
1748
1749//The total number of operations is 5
1750template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1)
1751{
1752        bitblock128_t tmp = simd128<64>::srli<((sh >= 64) ? (63) : ((sh < 0) ? 0 : sh))>(arg1);
1753        return simd_or(tmp, simd128<64>::sub(simd128<64>::constant<0>(), simd_and(simd128<64>::slli<((64-((sh >= 64) ? (63) : ((sh < 0) ? 0 : sh)))-1)>(simd128<64>::constant<1>()), tmp)));
1754}
1755
1756//The total number of operations is 21
1757template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1)
1758{
1759        bitblock128_t tmp = simd128<128>::srli<((sh >= 128) ? (127) : ((sh < 0) ? 0 : sh))>(arg1);
1760        return simd_or(tmp, simd128<128>::sub(simd128<128>::constant<0>(), simd_and(simd128<128>::slli<((128-((sh >= 128) ? (127) : ((sh < 0) ? 0 : sh)))-1)>(simd128<128>::constant<1>()), tmp)));
1761}
1762
1763//The total number of operations is 0
1764template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask()
1765{
1766        return simd128<2>::constant<(2)>();
1767}
1768
1769//The total number of operations is 0
1770template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask()
1771{
1772        return simd128<4>::constant<(12)>();
1773}
1774
1775//The total number of operations is 0
1776template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask()
1777{
1778        return simd128<8>::constant<(240)>();
1779}
1780
1781//The total number of operations is 0
1782template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask()
1783{
1784        return simd128<16>::constant<(65280)>();
1785}
1786
1787//The total number of operations is 0
1788template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask()
1789{
1790        return simd128<32>::constant<-65536>();
1791}
1792
1793//The total number of operations is 0
1794template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask()
1795{
1796        return _mm_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0));
1797}
1798
1799//The total number of operations is 0
1800template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask()
1801{
1802        return _mm_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0));
1803}
1804
1805//The total number of operations is 1
1806template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2)
1807{
1808        return simd_xor(arg1, arg2);
1809}
1810
1811//The total number of operations is 10
1812template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2)
1813{
1814        bitblock128_t tmp = simd_xor(arg1, arg2);
1815        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(arg1, arg2))), tmp);
1816}
1817
1818//The total number of operations is 6
1819template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2)
1820{
1821        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::add(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::add(arg1, arg2));
1822}
1823
1824//The total number of operations is 1
1825template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2)
1826{
1827        return _mm_add_epi8(arg1, arg2);
1828}
1829
1830//The total number of operations is 1
1831template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2)
1832{
1833        return _mm_add_epi16(arg1, arg2);
1834}
1835
1836//The total number of operations is 1
1837template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2)
1838{
1839        return _mm_add_epi32(arg1, arg2);
1840}
1841
1842//The total number of operations is 1
1843template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2)
1844{
1845        return _mm_add_epi64(arg1, arg2);
1846}
1847
1848//The total number of operations is 11
1849template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2)
1850{
1851        bitblock128_t partial = simd128<(64)>::add(arg1, arg2);
1852        bitblock128_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
1853        bitblock128_t carry = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(carryMask));
1854        return simd128<(64)>::add(partial, carry);
1855}
1856
1857//The total number of operations is 46
1858template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::abs(bitblock128_t arg1)
1859{
1860        bitblock128_t gtMask = simd128<1>::gt(arg1, simd128<1>::constant<0>());
1861        return simd128<1>::ifh(gtMask, arg1, simd128<1>::sub(gtMask, arg1));
1862}
1863
1864//The total number of operations is 9
1865template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1)
1866{
1867        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, simd128<128>::slli<1>(simd_not(arg1))), arg1);
1868}
1869
1870//The total number of operations is 19
1871template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1)
1872{
1873        bitblock128_t gtMask = simd128<4>::gt(arg1, simd128<4>::constant<0>());
1874        return simd128<1>::ifh(gtMask, arg1, simd128<4>::sub(gtMask, arg1));
1875}
1876
1877//The total number of operations is 1
1878template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1)
1879{
1880        return _mm_abs_epi8(arg1);
1881}
1882
1883//The total number of operations is 1
1884template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1)
1885{
1886        return _mm_abs_epi16(arg1);
1887}
1888
1889//The total number of operations is 1
1890template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1)
1891{
1892        return _mm_abs_epi32(arg1);
1893}
1894
1895//The total number of operations is 13
1896template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1)
1897{
1898        bitblock128_t eqMask = simd128<64>::eq(simd128<1>::ifh(simd128<64>::himask(), simd128<(32)>::abs(arg1), arg1), arg1);
1899        return simd128<1>::ifh(eqMask, arg1, simd128<64>::sub(eqMask, arg1));
1900}
1901
1902//The total number of operations is 45
1903template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1)
1904{
1905        bitblock128_t eqMask = simd128<128>::eq(simd128<1>::ifh(simd128<128>::himask(), simd128<(64)>::abs(arg1), arg1), arg1);
1906        return simd128<1>::ifh(eqMask, arg1, simd128<128>::sub(eqMask, arg1));
1907}
1908
1909//The total number of operations is 128
1910template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1911{
1912        return simd128<(1)>::umin(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1913}
1914
1915//The total number of operations is 80
1916template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1917{
1918        return simd128<(2)>::umin(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1919}
1920
1921//The total number of operations is 42
1922template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1923{
1924        return simd128<(4)>::umin(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1925}
1926
1927//The total number of operations is 7
1928template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1929{
1930        return simd128<(8)>::umin(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1931}
1932
1933//The total number of operations is 10
1934template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1935{
1936        return simd128<(16)>::umin(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1937}
1938
1939//The total number of operations is 13
1940template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1941{
1942        return simd128<(32)>::umin(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1943}
1944
1945//The total number of operations is 34
1946template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1947{
1948        return simd128<(64)>::umin(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1949}
1950
1951//The total number of operations is 93
1952template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1953{
1954        return simd128<(1)>::add(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1955}
1956
1957//The total number of operations is 74
1958template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1959{
1960        return simd128<(2)>::add(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1961}
1962
1963//The total number of operations is 42
1964template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1965{
1966        return simd128<(4)>::add(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1967}
1968
1969//The total number of operations is 7
1970template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1971{
1972        return simd128<(8)>::add(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1973}
1974
1975//The total number of operations is 1
1976template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1977{
1978        return _mm_hadd_epi16(arg2, arg1);
1979}
1980
1981//The total number of operations is 1
1982template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1983{
1984        return _mm_hadd_epi32(arg2, arg1);
1985}
1986
1987//The total number of operations is 15
1988template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1989{
1990        return simd128<(64)>::add(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1991}
1992
1993//The total number of operations is 132
1994template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2)
1995{
1996        bitblock128_t hiBound = simd128<2>::srli<1>(simd128<2>::lomask());
1997        bitblock128_t loBound = simd_not(hiBound);
1998        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)));
1999}
2000
2001//The total number of operations is 86
2002template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2)
2003{
2004        bitblock128_t hiBound = simd128<4>::srli<1>(simd128<4>::lomask());
2005        bitblock128_t loBound = simd_not(hiBound);
2006        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)));
2007}
2008
2009//The total number of operations is 36
2010template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2)
2011{
2012        bitblock128_t hiBound = simd128<8>::srli<1>(simd128<8>::lomask());
2013        bitblock128_t loBound = simd_not(hiBound);
2014        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)));
2015}
2016
2017//The total number of operations is 1
2018template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2)
2019{
2020        return _mm_packs_epi16(arg2, arg1);
2021}
2022
2023//The total number of operations is 1
2024template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2)
2025{
2026        return _mm_packs_epi32(arg2, arg1);
2027}
2028
2029//The total number of operations is 77
2030template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2)
2031{
2032        bitblock128_t hiBound = simd128<64>::srli<1>(simd128<64>::lomask());
2033        bitblock128_t loBound = simd_not(hiBound);
2034        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)));
2035}
2036
2037//The total number of operations is 288
2038template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2)
2039{
2040        bitblock128_t hiBound = simd128<128>::srli<1>(simd128<128>::lomask());
2041        bitblock128_t loBound = simd_not(hiBound);
2042        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)));
2043}
2044
2045//The total number of operations is 24
2046template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1)
2047{
2048        uint64_t tmpAns1 = hsimd128<(8)>::signmask(esimd128<4>::mergeh(arg1, simd128<4>::constant<0>()));
2049        uint64_t tmpAns2 = hsimd128<(8)>::signmask(esimd128<4>::mergel(arg1, simd128<4>::constant<0>()));
2050        return ((tmpAns1<<(16))+tmpAns2);
2051}
2052
2053//The total number of operations is 1
2054template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1)
2055{
2056        return _mm_movemask_epi8(arg1);
2057}
2058
2059//The total number of operations is 4
2060template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1)
2061{
2062        return hsimd128<(8)>::signmask(hsimd128<16>::packh(simd128<16>::constant<0>(), arg1));
2063}
2064
2065//The total number of operations is 7
2066template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1)
2067{
2068        return hsimd128<(16)>::signmask(hsimd128<32>::packh(simd128<32>::constant<0>(), arg1));
2069}
2070
2071//The total number of operations is 10
2072template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1)
2073{
2074        return hsimd128<(32)>::signmask(hsimd128<64>::packh(simd128<64>::constant<0>(), arg1));
2075}
2076
2077//The total number of operations is 17
2078template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1)
2079{
2080        return hsimd128<(64)>::signmask(hsimd128<128>::packh(simd128<128>::constant<0>(), arg1));
2081}
2082
2083//The total number of operations is 45
2084template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2)
2085{
2086        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));
2087}
2088
2089//The total number of operations is 31
2090template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2)
2091{
2092        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));
2093}
2094
2095//The total number of operations is 17
2096template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2)
2097{
2098        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));
2099}
2100
2101//The total number of operations is 3
2102template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2)
2103{
2104        return hsimd128<16>::packus(simd_and(arg1, simd128<16>::lomask()), simd_and(arg2, simd128<16>::lomask()));
2105}
2106
2107//The total number of operations is 3
2108template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2)
2109{
2110        return _mm_hsub_epi16(simd_and(arg2, simd128<32>::lomask()), simd_and(arg1, simd128<32>::lomask()));
2111}
2112
2113//The total number of operations is 3
2114template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2)
2115{
2116        return _mm_hsub_epi32(simd_and(arg2, simd128<64>::lomask()), simd_and(arg1, simd128<64>::lomask()));
2117}
2118
2119//The total number of operations is 7
2120template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2)
2121{
2122        return simd128<1>::ifh(simd128<128>::himask(), simd128<128>::slli<(64)>(arg1), arg2);
2123}
2124
2125//The total number of operations is 47
2126template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2)
2127{
2128        return hsimd128<2>::packl(simd128<64>::srli<(1)>(arg1), simd128<64>::srli<(1)>(arg2));
2129}
2130
2131//The total number of operations is 33
2132template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2)
2133{
2134        return hsimd128<4>::packl(simd128<64>::srli<(2)>(arg1), simd128<64>::srli<(2)>(arg2));
2135}
2136
2137//The total number of operations is 19
2138template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2)
2139{
2140        return hsimd128<8>::packl(simd128<64>::srli<(4)>(arg1), simd128<64>::srli<(4)>(arg2));
2141}
2142
2143//The total number of operations is 3
2144template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2)
2145{
2146        return hsimd128<16>::packus(simd128<16>::srli<(8)>(arg1), simd128<16>::srli<(8)>(arg2));
2147}
2148
2149//The total number of operations is 3
2150template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2)
2151{
2152        return _mm_hsub_epi16(simd128<32>::srli<(16)>(arg2), simd128<32>::srli<(16)>(arg1));
2153}
2154
2155//The total number of operations is 3
2156template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2)
2157{
2158        return _mm_hsub_epi32(simd128<64>::srli<(32)>(arg2), simd128<64>::srli<(32)>(arg1));
2159}
2160
2161//The total number of operations is 7
2162template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2)
2163{
2164        return simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg2));
2165}
2166
2167//The total number of operations is 131
2168template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2169{
2170        return simd128<(1)>::min(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
2171}
2172
2173//The total number of operations is 83
2174template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2175{
2176        return simd128<(2)>::min(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
2177}
2178
2179//The total number of operations is 45
2180template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2181{
2182        return simd128<(4)>::min(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
2183}
2184
2185//The total number of operations is 10
2186template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2187{
2188        return simd128<(8)>::min(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2189}
2190
2191//The total number of operations is 7
2192template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2193{
2194        return simd128<(16)>::min(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2195}
2196
2197//The total number of operations is 10
2198template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2199{
2200        return simd128<(32)>::min(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2201}
2202
2203//The total number of operations is 32
2204template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2205{
2206        return simd128<(64)>::min(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2207}
2208
2209//The total number of operations is 99
2210template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2)
2211{
2212        bitblock128_t arg11 = simd128<2>::ifh(arg1, simd128<2>::constant<0>(), arg1);
2213        bitblock128_t arg12 = simd_and(simd128<2>::lomask(), arg11);
2214        bitblock128_t arg21 = simd128<2>::ifh(arg2, simd128<2>::constant<0>(), arg2);
2215        bitblock128_t arg22 = simd_and(simd128<2>::lomask(), arg21);
2216        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()));
2217}
2218
2219//The total number of operations is 83
2220template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2)
2221{
2222        bitblock128_t arg11 = simd128<4>::ifh(arg1, simd128<4>::constant<0>(), arg1);
2223        bitblock128_t arg12 = simd_and(simd128<4>::lomask(), arg11);
2224        bitblock128_t arg21 = simd128<4>::ifh(arg2, simd128<4>::constant<0>(), arg2);
2225        bitblock128_t arg22 = simd_and(simd128<4>::lomask(), arg21);
2226        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()));
2227}
2228
2229//The total number of operations is 35
2230template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2)
2231{
2232        bitblock128_t arg11 = simd128<8>::ifh(arg1, simd128<8>::constant<0>(), arg1);
2233        bitblock128_t arg12 = simd_and(simd128<8>::lomask(), arg11);
2234        bitblock128_t arg21 = simd128<8>::ifh(arg2, simd128<8>::constant<0>(), arg2);
2235        bitblock128_t arg22 = simd_and(simd128<8>::lomask(), arg21);
2236        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()));
2237}
2238
2239//The total number of operations is 1
2240template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2)
2241{
2242        return _mm_packus_epi16(arg2, arg1);
2243}
2244
2245//The total number of operations is 12
2246template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2)
2247{
2248        bitblock128_t hiPart = hsimd128<32>::packh(arg1, arg2);
2249        return simd128<(16)>::ifh(hiPart, simd128<(16)>::constant<0>(), simd_or(simd128<(16)>::gt(hiPart, simd128<(16)>::constant<0>()), hsimd128<32>::packl(arg1, arg2)));
2250}
2251
2252//The total number of operations is 12
2253template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2)
2254{
2255        bitblock128_t hiPart = hsimd128<64>::packh(arg1, arg2);
2256        return simd128<(32)>::ifh(hiPart, simd128<(32)>::constant<0>(), simd_or(simd128<(32)>::gt(hiPart, simd128<(32)>::constant<0>()), hsimd128<64>::packl(arg1, arg2)));
2257}
2258
2259//The total number of operations is 38
2260template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2)
2261{
2262        bitblock128_t hiPart = hsimd128<128>::packh(arg1, arg2);
2263        return simd128<(64)>::ifh(hiPart, simd128<(64)>::constant<0>(), simd_or(simd128<(64)>::gt(hiPart, simd128<(64)>::constant<0>()), hsimd128<128>::packl(arg1, arg2)));
2264}
2265
2266//The total number of operations is 31
2267template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2268{
2269        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));
2270}
2271
2272//The total number of operations is 21
2273template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2274{
2275        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));
2276}
2277
2278//The total number of operations is 11
2279template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2280{
2281        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));
2282}
2283
2284//The total number of operations is 1
2285template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2286{
2287        return _mm_unpacklo_epi8(arg2, arg1);
2288}
2289
2290//The total number of operations is 1
2291template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2292{
2293        return _mm_unpacklo_epi16(arg2, arg1);
2294}
2295
2296//The total number of operations is 1
2297template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2298{
2299        return _mm_unpacklo_epi32(arg2, arg1);
2300}
2301
2302//The total number of operations is 1
2303template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2304{
2305        return _mm_unpacklo_epi64(arg2, arg1);
2306}
2307
2308//The total number of operations is 31
2309template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2310{
2311        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));
2312}
2313
2314//The total number of operations is 21
2315template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2316{
2317        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));
2318}
2319
2320//The total number of operations is 11
2321template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2322{
2323        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));
2324}
2325
2326//The total number of operations is 1
2327template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2328{
2329        return _mm_unpackhi_epi8(arg2, arg1);
2330}
2331
2332//The total number of operations is 1
2333template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2334{
2335        return _mm_unpackhi_epi16(arg2, arg1);
2336}
2337
2338//The total number of operations is 1
2339template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2340{
2341        return _mm_unpackhi_epi32(arg2, arg1);
2342}
2343
2344//The total number of operations is 1
2345template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2346{
2347        return _mm_unpackhi_epi64(arg2, arg1);
2348}
2349
2350//The total number of operations is 24
2351template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1)
2352{
2353        return esimd128<(2)>::mergeh(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2354}
2355
2356//The total number of operations is 14
2357template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1)
2358{
2359        return esimd128<(4)>::mergeh(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2360}
2361
2362//The total number of operations is 4
2363template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1)
2364{
2365        return esimd128<(8)>::mergeh(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2366}
2367
2368//The total number of operations is 3
2369template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1)
2370{
2371        return esimd128<(16)>::mergeh(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2372}
2373
2374//The total number of operations is 3
2375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1)
2376{
2377        return esimd128<(32)>::mergeh(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2378}
2379
2380//The total number of operations is 3
2381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1)
2382{
2383        return esimd128<(64)>::mergeh(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2384}
2385
2386//The total number of operations is 4
2387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1)
2388{
2389        return simd128<(128)>::srli<64>(arg1);
2390}
2391
2392//The total number of operations is 24
2393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1)
2394{
2395        return esimd128<(2)>::mergel(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2396}
2397
2398//The total number of operations is 14
2399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1)
2400{
2401        return esimd128<(4)>::mergel(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2402}
2403
2404//The total number of operations is 4
2405template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1)
2406{
2407        return esimd128<(8)>::mergel(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2408}
2409
2410//The total number of operations is 3
2411template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1)
2412{
2413        return esimd128<(16)>::mergel(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2414}
2415
2416//The total number of operations is 3
2417template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1)
2418{
2419        return esimd128<(32)>::mergel(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2420}
2421
2422//The total number of operations is 3
2423template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1)
2424{
2425        return esimd128<(64)>::mergel(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2426}
2427
2428//The total number of operations is 1
2429template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1)
2430{
2431        return simd_and(simd128<(128)>::lomask(), arg1);
2432}
2433
2434//The total number of operations is 31
2435template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1)
2436{
2437        return esimd128<(2)>::mergeh(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2438}
2439
2440//The total number of operations is 33
2441template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1)
2442{
2443        return esimd128<(4)>::mergeh(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2444}
2445
2446//The total number of operations is 13
2447template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1)
2448{
2449        return esimd128<(8)>::mergeh(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2450}
2451
2452//The total number of operations is 4
2453template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1)
2454{
2455        return esimd128<(16)>::mergeh(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2456}
2457
2458//The total number of operations is 4
2459template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1)
2460{
2461        return esimd128<(32)>::mergeh(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2462}
2463
2464//The total number of operations is 12
2465template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1)
2466{
2467        return esimd128<(64)>::mergeh(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2468}
2469
2470//The total number of operations is 21
2471template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1)
2472{
2473        return simd128<(128)>::srai<64>(arg1);
2474}
2475
2476//The total number of operations is 31
2477template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1)
2478{
2479        return esimd128<(2)>::mergel(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2480}
2481
2482//The total number of operations is 33
2483template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1)
2484{
2485        return esimd128<(4)>::mergel(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2486}
2487
2488//The total number of operations is 13
2489template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1)
2490{
2491        return esimd128<(8)>::mergel(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2492}
2493
2494//The total number of operations is 4
2495template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1)
2496{
2497        return esimd128<(16)>::mergel(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2498}
2499
2500//The total number of operations is 4
2501template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1)
2502{
2503        return esimd128<(32)>::mergel(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2504}
2505
2506//The total number of operations is 12
2507template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1)
2508{
2509        return esimd128<(64)>::mergel(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2510}
2511
2512//The total number of operations is 25
2513template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1)
2514{
2515        return simd128<(128)>::srai<64>(simd128<(128)>::slli<64>(arg1));
2516}
2517
2518//The total number of operations is 17
2519template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1)
2520{
2521        bitblock128_t tmphi = _mm_shufflehi_epi16(arg1, (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2522        bitblock128_t tmpAns = _mm_shufflelo_epi16(tmphi, (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2523        bitblock128_t tmplh = _mm_shufflehi_epi16(simd128<128>::slli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2524        bitblock128_t tmphl = _mm_shufflelo_epi16(simd128<128>::srli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2525        uint32_t a1 = ((((msk>>21)&4) == 0) ? 0 : (131071));
2526        uint32_t a2 = ((((msk>>18)&4) == 0) ? 0 : (131071));
2527        uint32_t a3 = ((((msk>>15)&4) == 0) ? 0 : (131071));
2528        uint32_t a4 = ((((msk>>12)&4) == 0) ? 0 : (131071));
2529        uint32_t a5 = ((((msk>>9)&4) == 0) ? (131071) : 0);
2530        uint32_t a6 = ((((msk>>6)&4) == 0) ? (131071) : 0);
2531        uint32_t a7 = ((((msk>>3)&4) == 0) ? (131071) : 0);
2532        uint32_t a8 = (((msk&4) == 0) ? (131071) : 0);
2533        return simd128<1>::ifh(mvmd128<16>::fill8(a1, a2, a3, a4, a5, a6, a7, a8), tmpAns, simd_or(tmplh, tmphl));
2534}
2535
2536//The total number of operations is 1
2537template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1)
2538{
2539        return _mm_shuffle_epi32(arg1, (int32_t)(msk));
2540}
2541
2542//The total number of operations is 1
2543template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1)
2544{
2545        return mvmd128<32>::shufflei<shufflemask4_from_shufflemask2(msk)>(arg1);
2546}
2547
2548//The total number of operations is 9
2549template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2550{
2551        return simd_or(mvmd128<2>::srli<sh>(arg1), mvmd128<2>::slli<((64)-sh)>(arg2));
2552}
2553
2554//The total number of operations is 9
2555template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2556{
2557        return simd_or(mvmd128<4>::srli<sh>(arg1), mvmd128<4>::slli<((32)-sh)>(arg2));
2558}
2559
2560//The total number of operations is 9
2561template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2562{
2563        return simd_or(mvmd128<8>::srli<sh>(arg1), mvmd128<8>::slli<((16)-sh)>(arg2));
2564}
2565
2566//The total number of operations is 9
2567template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2568{
2569        return simd_or(mvmd128<16>::srli<sh>(arg1), mvmd128<16>::slli<((8)-sh)>(arg2));
2570}
2571
2572//The total number of operations is 9
2573template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2574{
2575        return simd_or(mvmd128<32>::srli<sh>(arg1), mvmd128<32>::slli<((4)-sh)>(arg2));
2576}
2577
2578//The total number of operations is 9
2579template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2580{
2581        return simd_or(mvmd128<64>::srli<sh>(arg1), mvmd128<64>::slli<((2)-sh)>(arg2));
2582}
2583
2584//The total number of operations is 9
2585template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2586{
2587        return simd_or(mvmd128<128>::srli<sh>(arg1), mvmd128<128>::slli<((1)-sh)>(arg2));
2588}
2589
2590//The total number of operations is 1
2591template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1)
2592{
2593        return mvmd128<32>::fill((-1*val1));
2594}
2595
2596//The total number of operations is 1
2597template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1)
2598{
2599        return mvmd128<(4)>::fill(((val1<<2)|val1));
2600}
2601
2602//The total number of operations is 1
2603template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1)
2604{
2605        return mvmd128<(8)>::fill(((val1<<4)|val1));
2606}
2607
2608//The total number of operations is 1
2609template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1)
2610{
2611        return _mm_set1_epi8((int32_t)(val1));
2612}
2613
2614//The total number of operations is 1
2615template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1)
2616{
2617        return _mm_set1_epi16((int32_t)(val1));
2618}
2619
2620//The total number of operations is 1
2621template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1)
2622{
2623        return _mm_set1_epi32((int32_t)(val1));
2624}
2625
2626//The total number of operations is 1
2627template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1)
2628{
2629        return _mm_set_epi32((int32_t)((val1>>32)), (int32_t)(val1), (int32_t)((val1>>32)), (int32_t)(val1));
2630}
2631
2632//The total number of operations is 1
2633template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1)
2634{
2635        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val1>>32)), (int32_t)(val1));
2636}
2637
2638//The total number of operations is 1
2639template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2640{
2641        return _mm_shuffle_epi8(arg1, arg2);
2642}
2643
2644//The total number of operations is 10
2645template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2646{
2647        bitblock128_t tmp1 = simd_and(simd128<16>::constant<(7)>(), arg2);
2648        bitblock128_t msk1 = simd128<16>::add(tmp1, tmp1);
2649        bitblock128_t msk2 = simd128<16>::add(msk1, simd128<16>::constant<1>());
2650        bitblock128_t msk = simd_or(msk1, simd128<16>::slli<(8)>(msk2));
2651        return simd128<16>::ifh(arg2, simd128<16>::constant<0>(), mvmd128<(8)>::shuffle(arg1, msk));
2652}
2653
2654//The total number of operations is 19
2655template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2656{
2657        bitblock128_t tmp1 = simd_and(simd128<32>::constant<(3)>(), arg2);
2658        bitblock128_t msk1 = simd128<32>::add(tmp1, tmp1);
2659        bitblock128_t msk2 = simd128<32>::add(msk1, simd128<32>::constant<1>());
2660        bitblock128_t msk = simd_or(msk1, simd128<32>::slli<(16)>(msk2));
2661        return simd128<32>::ifh(arg2, simd128<32>::constant<0>(), mvmd128<(16)>::shuffle(arg1, msk));
2662}
2663
2664//The total number of operations is 32
2665template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2666{
2667        bitblock128_t tmp1 = simd_and(simd128<64>::constant<(1)>(), arg2);
2668        bitblock128_t msk1 = simd128<64>::add(tmp1, tmp1);
2669        bitblock128_t msk2 = simd128<64>::add(msk1, simd128<64>::constant<1>());
2670        bitblock128_t msk = simd_or(msk1, simd128<64>::slli<(32)>(msk2));
2671        return simd128<64>::ifh(arg2, simd128<64>::constant<0>(), mvmd128<(32)>::shuffle(arg1, msk));
2672}
2673
2674//The total number of operations is 1
2675template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1)
2676{
2677        return (((pos%2) == 0) ? (mvmd128<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd128<(2)>::extract<(pos/2)>(arg1)>>1));
2678}
2679
2680//The total number of operations is 1
2681template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1)
2682{
2683        return (((pos%2) == 0) ? (mvmd128<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd128<(4)>::extract<(pos/2)>(arg1)>>2));
2684}
2685
2686//The total number of operations is 1
2687template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1)
2688{
2689        return (((pos%2) == 0) ? (mvmd128<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd128<(8)>::extract<(pos/2)>(arg1)>>4));
2690}
2691
2692//The total number of operations is 1
2693template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1)
2694{
2695        return (((pos%2) == 0) ? (mvmd128<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd128<(16)>::extract<(pos/2)>(arg1)>>8));
2696}
2697
2698//The total number of operations is 1
2699template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1)
2700{
2701        return (65535&_mm_extract_epi16(arg1, (int32_t)(pos)));
2702}
2703
2704//The total number of operations is 2
2705template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1)
2706{
2707        return ((((uint64_t)mvmd128<(16)>::extract<((2*pos)+1)>(arg1))<<(16))|mvmd128<(16)>::extract<(2*pos)>(arg1));
2708}
2709
2710//The total number of operations is 4
2711template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1)
2712{
2713        return ((((uint64_t)mvmd128<(32)>::extract<((2*pos)+1)>(arg1))<<(32))|mvmd128<(32)>::extract<(2*pos)>(arg1));
2714}
2715
2716//The total number of operations is 16
2717template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1)
2718{
2719        return simd128<128>::sub(simd128<128>::constant<0>(), simd_and(simd128<128>::constant<1>(), simd128<128>::srli<pos>(arg1)));
2720}
2721
2722//The total number of operations is 13
2723template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1)
2724{
2725        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(4)>::slli<2>(arg1) : simd128<(4)>::srli<2>(arg1));
2726        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(4)>::lomask(), arg1) : simd_and(simd128<(4)>::himask(), arg1));
2727        return mvmd128<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2728}
2729
2730//The total number of operations is 9
2731template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1)
2732{
2733        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(8)>::slli<4>(arg1) : simd128<(8)>::srli<4>(arg1));
2734        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(8)>::lomask(), arg1) : simd_and(simd128<(8)>::himask(), arg1));
2735        return mvmd128<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2736}
2737
2738//The total number of operations is 5
2739template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1)
2740{
2741        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(16)>::slli<8>(arg1) : simd128<(16)>::srli<8>(arg1));
2742        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(16)>::lomask(), arg1) : simd_and(simd128<(16)>::himask(), arg1));
2743        return mvmd128<(16)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2744}
2745
2746//The total number of operations is 2
2747template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1)
2748{
2749        return mvmd128<16>::fill(_mm_extract_epi16(arg1, (int32_t)(pos)));
2750}
2751
2752//The total number of operations is 1
2753template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1)
2754{
2755        return mvmd128<32>::shufflei<shufflemask4(pos, pos, pos, pos)>(arg1);
2756}
2757
2758//The total number of operations is 5
2759template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1)
2760{
2761        return simd128<1>::ifh(simd128<64>::himask(), mvmd128<(32)>::splat<((2*pos)+1)>(arg1), mvmd128<(32)>::splat<(2*pos)>(arg1));
2762}
2763
2764//The total number of operations is 13
2765template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1)
2766{
2767        return simd128<1>::ifh(simd128<128>::himask(), mvmd128<(64)>::splat<((2*pos)+1)>(arg1), mvmd128<(64)>::splat<(2*pos)>(arg1));
2768}
2769
2770//The total number of operations is 15
2771template <> 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)
2772{
2773        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))));
2774}
2775
2776//The total number of operations is 7
2777template <> 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)
2778{
2779        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))));
2780}
2781
2782//The total number of operations is 3
2783template <> 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)
2784{
2785        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))));
2786}
2787
2788//The total number of operations is 1
2789template <> 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)
2790{
2791        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));
2792}
2793
2794//The total number of operations is 4
2795template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1)
2796{
2797        return simd128<128>::slli<(sh*2)>(arg1);
2798}
2799
2800//The total number of operations is 4
2801template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1)
2802{
2803        return simd128<128>::slli<(sh*4)>(arg1);
2804}
2805
2806//The total number of operations is 4
2807template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1)
2808{
2809        return simd128<128>::slli<(sh*8)>(arg1);
2810}
2811
2812//The total number of operations is 4
2813template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1)
2814{
2815        return simd128<128>::slli<(sh*16)>(arg1);
2816}
2817
2818//The total number of operations is 4
2819template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1)
2820{
2821        return simd128<128>::slli<(sh*32)>(arg1);
2822}
2823
2824//The total number of operations is 4
2825template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1)
2826{
2827        return simd128<128>::slli<(sh*64)>(arg1);
2828}
2829
2830//The total number of operations is 4
2831template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1)
2832{
2833        return simd128<128>::slli<(sh*128)>(arg1);
2834}
2835
2836//The total number of operations is 5
2837template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2838{
2839        return simd128<1>::ifh(simd128<(4)>::himask(), mvmd128<1>::fill2(val1, val2), mvmd128<1>::fill2(val3, val4));
2840}
2841
2842//The total number of operations is 5
2843template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2844{
2845        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<2>::fill2(val1, val2), mvmd128<2>::fill2(val3, val4));
2846}
2847
2848//The total number of operations is 5
2849template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2850{
2851        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<4>::fill2(val1, val2), mvmd128<4>::fill2(val3, val4));
2852}
2853
2854//The total number of operations is 5
2855template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2856{
2857        return simd128<1>::ifh(simd128<(32)>::himask(), mvmd128<8>::fill2(val1, val2), mvmd128<8>::fill2(val3, val4));
2858}
2859
2860//The total number of operations is 3
2861template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2862{
2863        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))));
2864}
2865
2866//The total number of operations is 1
2867template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2868{
2869        return _mm_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
2870}
2871
2872//The total number of operations is 4
2873template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1)
2874{
2875        return simd128<128>::srli<(sh*2)>(arg1);
2876}
2877
2878//The total number of operations is 4
2879template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1)
2880{
2881        return simd128<128>::srli<(sh*4)>(arg1);
2882}
2883
2884//The total number of operations is 4
2885template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1)
2886{
2887        return simd128<128>::srli<(sh*8)>(arg1);
2888}
2889
2890//The total number of operations is 4
2891template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1)
2892{
2893        return simd128<128>::srli<(sh*16)>(arg1);
2894}
2895
2896//The total number of operations is 4
2897template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1)
2898{
2899        return simd128<128>::srli<(sh*32)>(arg1);
2900}
2901
2902//The total number of operations is 4
2903template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1)
2904{
2905        return simd128<128>::srli<(sh*64)>(arg1);
2906}
2907
2908//The total number of operations is 4
2909template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1)
2910{
2911        return simd128<128>::srli<(sh*128)>(arg1);
2912}
2913
2914//The total number of operations is 1
2915template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2)
2916{
2917        return mvmd128<(2)>::fill(((val1<<1)|(val2&(1))));
2918}
2919
2920//The total number of operations is 1
2921template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2)
2922{
2923        return mvmd128<(4)>::fill(((val1<<2)|(val2&(3))));
2924}
2925
2926//The total number of operations is 1
2927template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2)
2928{
2929        return mvmd128<(8)>::fill(((val1<<4)|(val2&(15))));
2930}
2931
2932//The total number of operations is 1
2933template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2)
2934{
2935        return mvmd128<(16)>::fill(((val1<<8)|(val2&(255))));
2936}
2937
2938//The total number of operations is 1
2939template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2)
2940{
2941        return mvmd128<(32)>::fill(((val1<<16)|(val2&(65535))));
2942}
2943
2944//The total number of operations is 5
2945template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2)
2946{
2947        return simd128<1>::ifh(simd128<(64)>::himask(), mvmd128<32>::fill(val1), mvmd128<32>::fill(val2));
2948}
2949
2950//The total number of operations is 5
2951template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2)
2952{
2953        return simd128<1>::ifh(simd128<(128)>::himask(), mvmd128<64>::fill(val1), mvmd128<64>::fill(val2));
2954}
2955
2956//The total number of operations is 9
2957template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2958{
2959        return simd_or(mvmd128<2>::slli<sh>(arg1), mvmd128<2>::srli<((64)-sh)>(arg2));
2960}
2961
2962//The total number of operations is 9
2963template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2964{
2965        return simd_or(mvmd128<4>::slli<sh>(arg1), mvmd128<4>::srli<((32)-sh)>(arg2));
2966}
2967
2968//The total number of operations is 9
2969template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2970{
2971        return simd_or(mvmd128<8>::slli<sh>(arg1), mvmd128<8>::srli<((16)-sh)>(arg2));
2972}
2973
2974//The total number of operations is 9
2975template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2976{
2977        return simd_or(mvmd128<16>::slli<sh>(arg1), mvmd128<16>::srli<((8)-sh)>(arg2));
2978}
2979
2980//The total number of operations is 9
2981template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2982{
2983        return simd_or(mvmd128<32>::slli<sh>(arg1), mvmd128<32>::srli<((4)-sh)>(arg2));
2984}
2985
2986//The total number of operations is 9
2987template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2988{
2989        return simd_or(mvmd128<64>::slli<sh>(arg1), mvmd128<64>::srli<((2)-sh)>(arg2));
2990}
2991
2992//The total number of operations is 9
2993template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2994{
2995        return simd_or(mvmd128<128>::slli<sh>(arg1), mvmd128<128>::srli<((1)-sh)>(arg2));
2996}
2997
2998//The total number of operations is 13
2999template <> 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)
3000{
3001        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<1>::fill4(val1, val2, val3, val4), mvmd128<1>::fill4(val5, val6, val7, val8));
3002}
3003
3004//The total number of operations is 13
3005template <> 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)
3006{
3007        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<2>::fill4(val1, val2, val3, val4), mvmd128<2>::fill4(val5, val6, val7, val8));
3008}
3009
3010//The total number of operations is 7
3011template <> 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)
3012{
3013        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))));
3014}
3015
3016//The total number of operations is 3
3017template <> 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)
3018{
3019        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))));
3020}
3021
3022//The total number of operations is 1
3023template <> 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)
3024{
3025        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));
3026}
3027
3028//The total number of operations is 1
3029IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(bitblock128_t* arg1)
3030{
3031        return _mm_loadu_si128((bitblock128_t*)(arg1));
3032}
3033
3034//The total number of operations is 1
3035IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t* arg1, bitblock128_t arg2)
3036{
3037        _mm_store_si128((bitblock128_t*)(arg1), arg2);
3038}
3039
3040//The total number of operations is 22
3041IDISA_ALWAYS_INLINE uint64_t bitblock128::popcount(bitblock128_t arg1)
3042{
3043        return mvmd128<64>::extract<0>(simd128<128>::popcount(arg1));
3044}
3045
3046//The total number of operations is 2
3047IDISA_ALWAYS_INLINE bool bitblock128::all(bitblock128_t arg1)
3048{
3049        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<-1>())) == 65535;
3050}
3051
3052//The total number of operations is 2
3053IDISA_ALWAYS_INLINE bool bitblock128::any(bitblock128_t arg1)
3054{
3055        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<0>())) != 65535;
3056}
3057
3058//The total number of operations is 1
3059IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(bitblock128_t* arg1)
3060{
3061        return _mm_load_si128((bitblock128_t*)(arg1));
3062}
3063
3064//The total number of operations is 1
3065IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t* arg1, bitblock128_t arg2)
3066{
3067        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
3068}
3069
3070#endif
Note: See TracBrowser for help on using the repository browser.