source: trunk/lib/idisa_cpp/idisa_ssse3.cpp @ 1557

Last change on this file since 1557 was 1557, checked in by cameron, 8 years ago

New strategies for ugt, ult

File size: 152.2 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 mergeh(bitblock128_t arg1, bitblock128_t arg2);
63        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendh(bitblock128_t arg1);
64        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendl(bitblock128_t arg1);
65        static IDISA_ALWAYS_INLINE bitblock128_t signextendh(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 bool all(bitblock128_t arg1);
92        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
93        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
94};
95
96//Declaration Part
97IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2);
98IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1);
99IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2);
100IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2);
101IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2);
102IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2);
103template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2);
104template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2);
105template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2);
106template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2);
107template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2);
108template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2);
109template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2);
110template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2);
111template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2);
112template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2);
113template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2);
114template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2);
115template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2);
116template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2);
117template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2);
118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2);
119template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2);
120template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2);
121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2);
122template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2);
123template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2);
124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2);
125template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2);
126template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2);
127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2);
128template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2);
129template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2);
130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2);
131template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2);
132template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2);
133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2);
134template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2);
135template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2);
136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2);
137template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2);
138template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2);
139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2);
140template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2);
141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2);
142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2);
143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2);
144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2);
145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2);
146template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2);
147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2);
148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2);
149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2);
150template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1);
151template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1);
152template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1);
153template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1);
154template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1);
155template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1);
156template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1);
157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1);
158template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1);
159template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1);
160template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1);
161template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1);
162template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1);
163template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1);
164template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1);
165template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
166template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2);
168template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2);
169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2);
170template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2);
171template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2);
172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2);
173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2);
174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2);
175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1);
176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1);
177template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1);
178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1);
179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1);
180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1);
181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1);
182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1);
183template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1);
184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1);
185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1);
186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1);
187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1);
188template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1);
189template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1);
190template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::neg(bitblock128_t arg1);
191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1);
192template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1);
193template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1);
194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1);
195template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1);
196template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1);
197template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1);
198template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1);
199template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1);
200template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1);
201template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1);
202template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1);
203template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1);
204template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1);
205template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
206template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
207template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
208template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
209template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
210template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
211template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
212template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
213template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2);
214template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2);
215template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2);
216template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2);
217template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2);
218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2);
219template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2);
220template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2);
221template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1);
222template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1);
223template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1);
224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1);
225template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1);
226template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1);
227template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1);
228template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
229template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask();
231template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask();
232template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask();
233template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask();
234template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask();
235template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask();
236template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask();
237template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
238template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
239template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
240template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
241template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
242template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
243template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
244template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
245template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2);
246template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2);
247template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2);
248template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2);
249template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2);
250template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2);
251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2);
252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2);
253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2);
254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2);
255template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2);
256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2);
257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2);
258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2);
259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2);
260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2);
261template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2);
262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2);
263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2);
264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2);
265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2);
266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2);
267template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2);
268template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2);
269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2);
270template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2);
271template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2);
272template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2);
273template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2);
274template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2);
275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2);
276template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2);
277template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1);
278template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1);
279template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1);
280template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1);
281template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1);
282template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1);
283template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1);
284template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask();
285template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask();
286template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask();
287template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask();
288template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask();
289template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask();
290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask();
291template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2);
292template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2);
293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2);
294template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2);
295template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2);
296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2);
297template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2);
298template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2);
299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::abs(bitblock128_t arg1);
300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1);
301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1);
302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1);
303template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1);
304template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1);
305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1);
306template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1);
307template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
308template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
309template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
310template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
311template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
312template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
313template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
314template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
315template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
316template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
317template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
318template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
319template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
320template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
321template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2);
322template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2);
323template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2);
324template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2);
325template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2);
326template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2);
327template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2);
328template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1);
329template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1);
330template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1);
331template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1);
332template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1);
333template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1);
334template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2);
335template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2);
336template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2);
337template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2);
338template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2);
339template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2);
340template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2);
341template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2);
342template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2);
343template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2);
344template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2);
345template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2);
346template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2);
347template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2);
348template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
349template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
350template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
351template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
352template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
353template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
354template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
355template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2);
356template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2);
357template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2);
358template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2);
359template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2);
360template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2);
361template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2);
362template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2);
363template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2);
364template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2);
365template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2);
366template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2);
367template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2);
368template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2);
369template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
370template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
371template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
372template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
373template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
374template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
376template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1);
377template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1);
378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1);
379template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1);
380template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1);
381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1);
382template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1);
383template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1);
384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1);
385template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1);
386template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1);
387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1);
388template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1);
389template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1);
390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1);
391template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1);
392template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1);
393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1);
394template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1);
395template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1);
396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1);
397template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1);
398template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1);
399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1);
400template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1);
401template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1);
402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1);
403template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1);
404template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1);
405template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1);
406template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1);
407template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
408template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
409template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
410template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
411template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
412template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
413template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
414template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1);
415template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1);
416template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1);
417template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1);
418template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1);
419template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1);
420template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1);
421template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1);
422template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
423template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
424template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
425template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
426template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1);
427template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1);
428template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1);
429template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1);
430template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1);
431template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1);
432template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1);
433template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1);
434template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1);
435template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1);
436template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1);
437template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1);
438template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1);
439template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1);
440template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1);
441template <> 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);
442template <> 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);
443template <> 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);
444template <> 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);
445template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1);
446template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1);
447template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1);
448template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1);
449template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1);
450template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1);
451template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1);
452template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
453template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
454template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
455template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
456template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
457template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
458template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1);
459template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1);
460template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1);
461template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1);
462template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1);
463template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1);
464template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1);
465template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2);
466template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2);
467template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2);
468template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2);
469template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2);
470template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2);
471template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2);
472template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2);
473template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2);
474template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2);
475template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2);
476template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2);
477template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2);
478template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2);
479template <> 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);
480template <> 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);
481template <> 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);
482template <> 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);
483template <> 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);
484
485//Implementation Part
486
487#define shufflemask4(s1, s2, s3, s4) \
488        ((s1<<6) | (s2<<4) | (s3<<2) | s4)
489
490#define shufflemask4_from_shufflemask2(msk) \
491        (msk==3 ? 238 : (msk==2 ? 228 : (msk==1 ? 78 : 68)))
492
493#define shufflemask8_to_shufflemask4(msk) \
494        ((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))
495//The total number of operations is 2
496IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2)
497{
498        return simd_not(simd_or(arg1, arg2));
499}
500
501//The total number of operations is 1
502IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1)
503{
504        return simd_xor(arg1, simd128<32>::constant<-1>());
505}
506
507//The total number of operations is 1
508IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2)
509{
510        return _mm_andnot_si128(arg2, arg1);
511}
512
513//The total number of operations is 1
514IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2)
515{
516        return _mm_or_si128(arg1, arg2);
517}
518
519//The total number of operations is 1
520IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2)
521{
522        return _mm_and_si128(arg1, arg2);
523}
524
525//The total number of operations is 1
526IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2)
527{
528        return _mm_xor_si128(arg1, arg2);
529}
530
531//The total number of operations is 39
532template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2)
533{
534        bitblock128_t high_bit = simd128<1>::constant<(1)>();
535        return simd_xor(simd128<1>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
536}
537
538//The total number of operations is 19
539template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2)
540{
541        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)))));
542}
543
544//The total number of operations is 9
545template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2)
546{
547        bitblock128_t high_bit = simd128<4>::constant<(8)>();
548        return simd_xor(simd128<4>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
549}
550
551//The total number of operations is 4
552template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2)
553{
554        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg1, arg2);
555}
556
557//The total number of operations is 1
558template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2)
559{
560        return _mm_max_epi16(arg1, arg2);
561}
562
563//The total number of operations is 4
564template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2)
565{
566        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg1, arg2);
567}
568
569//The total number of operations is 18
570template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2)
571{
572        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg1, arg2);
573}
574
575//The total number of operations is 65
576template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2)
577{
578        bitblock128_t hiAns = simd128<(64)>::max(arg1, arg2);
579        bitblock128_t loAns = simd128<(64)>::umax(arg1, arg2);
580        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
581        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
582        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
583}
584
585//The total number of operations is 57
586template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2)
587{
588        bitblock128_t loMask = simd128<(2)>::lomask();
589        bitblock128_t tmpAns1 = simd128<(2)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
590        bitblock128_t tmpAns2 = simd128<(2)>::mult(simd128<(2)>::srli<1>(arg1), simd128<(2)>::srli<1>(arg2));
591        return simd128<1>::ifh(loMask, tmpAns1, simd128<(2)>::slli<1>(tmpAns2));
592}
593
594//The total number of operations is 23
595template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2)
596{
597        bitblock128_t tmp1 = simd128<128>::slli<1>(arg1);
598        bitblock128_t tmp2 = simd128<128>::slli<1>(arg2);
599        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));
600}
601
602//The total number of operations is 31
603template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2)
604{
605        bitblock128_t loMask = simd128<(8)>::lomask();
606        bitblock128_t tmpAns1 = simd128<(8)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
607        bitblock128_t tmpAns2 = simd128<(8)>::mult(simd128<(8)>::srli<4>(arg1), simd128<(8)>::srli<4>(arg2));
608        return simd128<1>::ifh(loMask, tmpAns1, simd128<(8)>::slli<4>(tmpAns2));
609}
610
611//The total number of operations is 10
612template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2)
613{
614        bitblock128_t loMask = simd128<(16)>::lomask();
615        bitblock128_t tmpAns1 = simd128<(16)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
616        bitblock128_t tmpAns2 = simd128<(16)>::mult(simd128<(16)>::srli<8>(arg1), simd128<(16)>::srli<8>(arg2));
617        return simd128<1>::ifh(loMask, tmpAns1, simd128<(16)>::slli<8>(tmpAns2));
618}
619
620//The total number of operations is 1
621template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2)
622{
623        return _mm_mullo_epi16(arg1, arg2);
624}
625
626//The total number of operations is 30
627template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2)
628{
629        bitblock128_t loMask = simd128<(64)>::lomask();
630        bitblock128_t tmpAns1 = simd128<(64)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
631        bitblock128_t tmpAns2 = simd128<(64)>::mult(simd128<(64)>::srli<32>(arg1), simd128<(64)>::srli<32>(arg2));
632        return simd128<1>::ifh(loMask, tmpAns1, simd128<(64)>::slli<32>(tmpAns2));
633}
634
635//The total number of operations is 11
636template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2)
637{
638        bitblock128_t loMask = simd128<64>::lomask();
639        bitblock128_t arg1_low = simd_and(arg1, loMask);
640        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg1);
641        bitblock128_t arg2_low = simd_and(arg2, loMask);
642        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg2);
643        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
644        bitblock128_t tmpAns2 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
645        bitblock128_t tmpAns3 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
646        return simd128<64>::add(tmpAns1, simd128<64>::add(tmpAns2, tmpAns3));
647}
648
649//The total number of operations is 205
650template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2)
651{
652        bitblock128_t loMask = simd128<128>::lomask();
653        bitblock128_t arg1_low = simd_and(arg1, loMask);
654        bitblock128_t arg1_high = simd128<128>::srli<(64)>(arg1);
655        bitblock128_t arg2_low = simd_and(arg2, loMask);
656        bitblock128_t arg2_high = simd128<128>::srli<(64)>(arg2);
657        bitblock128_t tmpAns1 = simd128<(64)>::umult(arg1_low, arg2_low);
658        bitblock128_t tmpAns2 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_low, arg2_high));
659        bitblock128_t tmpAns3 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_high, arg2_low));
660        return simd128<128>::add(tmpAns1, simd128<128>::add(tmpAns2, tmpAns3));
661}
662
663//The total number of operations is 42
664template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2)
665{
666        bitblock128_t high_bit = simd128<1>::constant<(1)>();
667        return simd128<1>::ugt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
668}
669
670//The total number of operations is 18
671template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2)
672{
673        bitblock128_t tmp = simd_not(arg1);
674        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)));
675        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
676}
677
678//The total number of operations is 10
679template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2)
680{
681        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::gt(simd_and(simd128<(8)>::himask(), arg1), arg2), simd128<(8)>::gt(simd128<(8)>::slli<4>(arg1), simd128<(8)>::slli<4>(arg2)));
682}
683
684//The total number of operations is 1
685template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2)
686{
687        return _mm_cmpgt_epi8(arg1, arg2);
688}
689
690//The total number of operations is 1
691template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2)
692{
693        return _mm_cmpgt_epi16(arg1, arg2);
694}
695
696//The total number of operations is 1
697template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2)
698{
699        return _mm_cmpgt_epi32(arg1, arg2);
700}
701
702//The total number of operations is 15
703template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2)
704{
705        bitblock128_t hiAns = simd128<(32)>::gt(arg1, arg2);
706        bitblock128_t loAns = simd128<(32)>::ugt(arg1, arg2);
707        bitblock128_t mask = simd_and(loAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
708        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
709        return simd_or(simd128<64>::srai<(32)>(hiAns), mask);
710}
711
712//The total number of operations is 66
713template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2)
714{
715        bitblock128_t hiAns = simd128<(64)>::gt(arg1, arg2);
716        bitblock128_t loAns = simd128<(64)>::ugt(arg1, arg2);
717        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
718        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
719        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
720}
721
722//The total number of operations is 289
723template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2)
724{
725        bitblock128_t loMask = simd128<(2)>::lomask();
726        bitblock128_t tmpAns1 = simd128<(2)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
727        bitblock128_t tmpAns2 = simd128<(2)>::umult(simd_and(loMask, simd128<(4)>::srli<(2)>(arg1)), simd_and(loMask, simd128<(4)>::srli<(2)>(arg2)));
728        return simd_or(tmpAns1, simd128<(4)>::slli<(2)>(tmpAns2));
729}
730
731//The total number of operations is 139
732template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2)
733{
734        bitblock128_t loMask = simd128<(4)>::lomask();
735        bitblock128_t tmpAns1 = simd128<(4)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
736        bitblock128_t tmpAns2 = simd128<(4)>::umult(simd_and(loMask, simd128<(8)>::srli<(4)>(arg1)), simd_and(loMask, simd128<(8)>::srli<(4)>(arg2)));
737        return simd_or(tmpAns1, simd128<(8)>::slli<(4)>(tmpAns2));
738}
739
740//The total number of operations is 64
741template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2)
742{
743        bitblock128_t loMask = simd128<(8)>::lomask();
744        bitblock128_t tmpAns1 = simd128<(8)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
745        bitblock128_t tmpAns2 = simd128<(8)>::umult(simd_and(loMask, simd128<(16)>::srli<(8)>(arg1)), simd_and(loMask, simd128<(16)>::srli<(8)>(arg2)));
746        return simd_or(tmpAns1, simd128<(16)>::slli<(8)>(tmpAns2));
747}
748
749//The total number of operations is 28
750template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2)
751{
752        bitblock128_t loMask = simd128<(16)>::lomask();
753        bitblock128_t tmpAns1 = simd128<(16)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
754        bitblock128_t tmpAns2 = simd128<(16)>::umult(simd_and(loMask, simd128<(32)>::srli<(16)>(arg1)), simd_and(loMask, simd128<(32)>::srli<(16)>(arg2)));
755        return simd_or(tmpAns1, simd128<(32)>::slli<(16)>(tmpAns2));
756}
757
758//The total number of operations is 10
759template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2)
760{
761        bitblock128_t loMask = simd128<(32)>::lomask();
762        bitblock128_t tmpAns1 = simd128<(32)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
763        bitblock128_t tmpAns2 = simd128<(32)>::umult(simd_and(loMask, simd128<(64)>::srli<(32)>(arg1)), simd_and(loMask, simd128<(64)>::srli<(32)>(arg2)));
764        return simd_or(tmpAns1, simd128<(64)>::slli<(32)>(tmpAns2));
765}
766
767//The total number of operations is 1
768template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2)
769{
770        return _mm_mul_epu32(arg1, arg2);
771}
772
773//The total number of operations is 55
774template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2)
775{
776        bitblock128_t loMask1 = simd128<(128)>::lomask();
777        bitblock128_t arg11 = simd_and(arg1, loMask1);
778        bitblock128_t arg22 = simd_and(arg2, loMask1);
779        bitblock128_t loMask2 = simd128<64>::lomask();
780        bitblock128_t arg1_low = simd_and(arg11, loMask2);
781        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg11);
782        bitblock128_t arg2_low = simd_and(arg22, loMask2);
783        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg22);
784        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
785        bitblock128_t tmpAns2 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
786        bitblock128_t tmpAns3 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
787        bitblock128_t tmpAns4 = simd128<(128)>::slli<64>(simd128<(32)>::umult(arg1_high, arg2_high));
788        return simd128<(128)>::add(tmpAns1, simd128<(128)>::add(tmpAns2, simd128<(128)>::add(tmpAns3, tmpAns4)));
789}
790
791//The total number of operations is 40
792template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2)
793{
794        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())));
795}
796
797//The total number of operations is 17
798template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2)
799{
800        bitblock128_t tmp = simd_not(arg1);
801        bitblock128_t tmpAns = simd_or(simd_and(tmp, arg2), simd_and(simd128<128>::slli<1>(simd_and(tmp, arg2)), simd_or(tmp, arg2)));
802        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
803}
804
805//The total number of operations is 20
806template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2)
807{
808        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())));
809}
810
811//The total number of operations is 7
812template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2)
813{
814        bitblock128_t high_bit = simd128<8>::constant<(128)>();
815        return simd128<8>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
816}
817
818//The total number of operations is 7
819template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2)
820{
821        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
822        return simd128<16>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
823}
824
825//The total number of operations is 7
826template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2)
827{
828        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
829        return simd128<32>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
830}
831
832//The total number of operations is 18
833template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2)
834{
835        bitblock128_t tmpAns = simd128<(32)>::ult(arg1, arg2);
836        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
837        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
838        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
839}
840
841//The total number of operations is 55
842template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2)
843{
844        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)));
845}
846
847//The total number of operations is 42
848template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2)
849{
850        bitblock128_t high_bit = simd128<1>::constant<(1)>();
851        return simd128<1>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
852}
853
854//The total number of operations is 18
855template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2)
856{
857        bitblock128_t tmp = simd_not(arg2);
858        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)));
859        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
860}
861
862//The total number of operations is 18
863template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2)
864{
865        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)));
866}
867
868//The total number of operations is 5
869template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2)
870{
871        return simd_and(simd_not(simd128<8>::gt(arg1, arg2)), simd_not(simd128<8>::eq(arg1, arg2)));
872}
873
874//The total number of operations is 5
875template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2)
876{
877        return simd_and(simd_not(simd128<16>::gt(arg1, arg2)), simd_not(simd128<16>::eq(arg1, arg2)));
878}
879
880//The total number of operations is 5
881template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2)
882{
883        return simd_and(simd_not(simd128<32>::gt(arg1, arg2)), simd_not(simd128<32>::eq(arg1, arg2)));
884}
885
886//The total number of operations is 20
887template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2)
888{
889        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808UL)>();
890        return simd128<64>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
891}
892
893//The total number of operations is 75
894template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2)
895{
896        bitblock128_t hiAns = simd128<(64)>::lt(arg1, arg2);
897        bitblock128_t loAns = simd128<(64)>::ult(arg1, arg2);
898        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
899        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
900        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
901}
902
903//The total number of operations is 2
904template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1)
905{
906        return simd_and(simd128<32>::srli<sh>(arg1), simd128<2>::constant<((3)>>sh)>());
907}
908
909//The total number of operations is 2
910template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1)
911{
912        return simd_and(simd128<32>::srli<sh>(arg1), simd128<4>::constant<((15)>>sh)>());
913}
914
915//The total number of operations is 2
916template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1)
917{
918        return simd_and(simd128<32>::srli<sh>(arg1), simd128<8>::constant<((255)>>sh)>());
919}
920
921//The total number of operations is 1
922template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1)
923{
924        return _mm_srli_epi16(arg1, (int32_t)(sh));
925}
926
927//The total number of operations is 1
928template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1)
929{
930        return _mm_srli_epi32(arg1, (int32_t)(sh));
931}
932
933//The total number of operations is 1
934template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1)
935{
936        return _mm_srli_epi64(arg1, (int32_t)(sh));
937}
938
939//The total number of operations is 4
940template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1)
941{
942        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)))));
943}
944
945//The total number of operations is 1
946template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1)
947{
948        return simd_not(arg1);
949}
950
951//The total number of operations is 14
952template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1)
953{
954        bitblock128_t tmp = simd_not(arg1);
955        return simd128<1>::ifh(simd128<2>::himask(), simd_and(tmp, simd128<128>::slli<1>(tmp)), simd_and(simd128<128>::srli<1>(arg1), tmp));
956}
957
958//The total number of operations is 14
959template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1)
960{
961        return simd128<4>::popcount(simd_andc(simd128<4>::sub(arg1, simd128<4>::constant<1>()), arg1));
962}
963
964//The total number of operations is 13
965template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1)
966{
967        return simd128<8>::popcount(simd_andc(simd128<8>::sub(arg1, simd128<8>::constant<1>()), arg1));
968}
969
970//The total number of operations is 16
971template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1)
972{
973        return simd128<16>::popcount(simd_andc(simd128<16>::sub(arg1, simd128<16>::constant<1>()), arg1));
974}
975
976//The total number of operations is 19
977template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1)
978{
979        return simd128<32>::popcount(simd_andc(simd128<32>::sub(arg1, simd128<32>::constant<1>()), arg1));
980}
981
982//The total number of operations is 14
983template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1)
984{
985        return simd128<64>::popcount(simd_andc(simd128<64>::sub(arg1, simd128<64>::constant<1>()), arg1));
986}
987
988//The total number of operations is 30
989template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1)
990{
991        return simd128<128>::popcount(simd_andc(simd128<128>::sub(arg1, simd128<128>::constant<1>()), arg1));
992}
993
994//The total number of operations is 7
995template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
996{
997        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())));
998}
999
1000//The total number of operations is 11
1001template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1002{
1003        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))));
1004}
1005
1006//The total number of operations is 40
1007template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1008{
1009        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())));
1010}
1011
1012//The total number of operations is 17
1013template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1014{
1015        bitblock128_t tmp = simd_not(arg2);
1016        bitblock128_t tmpAns = simd_or(simd_and(arg1, tmp), simd_and(simd128<128>::slli<1>(simd_and(arg1, tmp)), simd_or(arg1, tmp)));
1017        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1018}
1019
1020//The total number of operations is 12
1021template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1022{
1023        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())));
1024}
1025
1026//The total number of operations is 3
1027template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1028{
1029        bitblock128_t high_bit = simd128<8>::constant<(128)>();
1030        return simd128<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1031}
1032
1033//The total number of operations is 3
1034template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1035{
1036        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1037        return simd128<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1038}
1039
1040//The total number of operations is 3
1041template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1042{
1043        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1044        return simd128<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1045}
1046
1047//The total number of operations is 14
1048template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1049{
1050        bitblock128_t tmpAns = simd128<(32)>::ugt(arg1, arg2);
1051        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
1052        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
1053        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
1054}
1055
1056//The total number of operations is 51
1057template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1058{
1059        bitblock128_t tmpAns = simd128<(64)>::ugt(arg1, arg2);
1060        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
1061        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
1062        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
1063}
1064
1065//The total number of operations is 4
1066template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1)
1067{
1068        return simd_xor(simd128<2>::srli<(1)>(arg1), simd_and(arg1, simd128<2>::lomask()));
1069}
1070
1071//The total number of operations is 4
1072template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1)
1073{
1074        return simd_xor(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1075}
1076
1077//The total number of operations is 4
1078template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1)
1079{
1080        return simd_xor(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1081}
1082
1083//The total number of operations is 3
1084template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1)
1085{
1086        return simd_xor(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1087}
1088
1089//The total number of operations is 3
1090template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1)
1091{
1092        return simd_xor(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1093}
1094
1095//The total number of operations is 3
1096template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1)
1097{
1098        return simd_xor(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1099}
1100
1101//The total number of operations is 6
1102template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1)
1103{
1104        return simd_xor(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1105}
1106
1107//The total number of operations is 0
1108template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1)
1109{
1110        return arg1;
1111}
1112
1113//The total number of operations is 3
1114template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1)
1115{
1116        return simd128<2>::add_hl(simd128<(1)>::popcount(arg1));
1117}
1118
1119//The total number of operations is 7
1120template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1)
1121{
1122        return simd128<4>::add_hl(simd128<(2)>::popcount(arg1));
1123}
1124
1125//The total number of operations is 11
1126template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1)
1127{
1128        return simd128<8>::add_hl(simd128<(4)>::popcount(arg1));
1129}
1130
1131//The total number of operations is 14
1132template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1)
1133{
1134        return simd128<16>::add_hl(simd128<(8)>::popcount(arg1));
1135}
1136
1137//The total number of operations is 17
1138template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1)
1139{
1140        return simd128<32>::add_hl(simd128<(16)>::popcount(arg1));
1141}
1142
1143//The total number of operations is 12
1144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1)
1145{
1146        return _mm_sad_epu8(simd128<8>::popcount(arg1), simd128<8>::constant<0>());
1147}
1148
1149//The total number of operations is 18
1150template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1)
1151{
1152        bitblock128_t tmpAns = simd128<(64)>::popcount(arg1);
1153        return simd128<(64)>::add(simd_and(tmpAns, simd128<128>::lomask()), simd128<128>::srli<(64)>(tmpAns));
1154}
1155
1156//The total number of operations is 26
1157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::neg(bitblock128_t arg1)
1158{
1159        return simd128<1>::sub(simd128<1>::constant<0>(), arg1);
1160}
1161
1162//The total number of operations is 8
1163template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1)
1164{
1165        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(arg1, simd128<128>::slli<1>(arg1)), arg1);
1166}
1167
1168//The total number of operations is 6
1169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1)
1170{
1171        return simd128<4>::sub(simd128<4>::constant<0>(), arg1);
1172}
1173
1174//The total number of operations is 1
1175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1)
1176{
1177        return simd128<8>::sub(simd128<8>::constant<0>(), arg1);
1178}
1179
1180//The total number of operations is 1
1181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1)
1182{
1183        return simd128<16>::sub(simd128<16>::constant<0>(), arg1);
1184}
1185
1186//The total number of operations is 1
1187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1)
1188{
1189        return _mm_sign_epi32(arg1, simd128<32>::constant<-1>());
1190}
1191
1192//The total number of operations is 1
1193template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1)
1194{
1195        return simd128<64>::sub(simd128<64>::constant<0>(), arg1);
1196}
1197
1198//The total number of operations is 11
1199template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1)
1200{
1201        return simd128<128>::sub(simd128<128>::constant<0>(), arg1);
1202}
1203
1204//The total number of operations is 2
1205template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1)
1206{
1207        return simd_and(simd128<32>::slli<sh>(arg1), simd128<2>::constant<(((3)<<sh)&(3))>());
1208}
1209
1210//The total number of operations is 2
1211template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1)
1212{
1213        return simd_and(simd128<32>::slli<sh>(arg1), simd128<4>::constant<(((15)<<sh)&(15))>());
1214}
1215
1216//The total number of operations is 2
1217template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1)
1218{
1219        return simd_and(simd128<32>::slli<sh>(arg1), simd128<8>::constant<(((255)<<sh)&(255))>());
1220}
1221
1222//The total number of operations is 1
1223template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1)
1224{
1225        return _mm_slli_epi16(arg1, (int32_t)(sh));
1226}
1227
1228//The total number of operations is 1
1229template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1)
1230{
1231        return _mm_slli_epi32(arg1, (int32_t)(sh));
1232}
1233
1234//The total number of operations is 1
1235template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1)
1236{
1237        return _mm_slli_epi64(arg1, (int32_t)(sh));
1238}
1239
1240//The total number of operations is 4
1241template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1)
1242{
1243        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)))));
1244}
1245
1246//The total number of operations is 3
1247template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1248{
1249        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
1250}
1251
1252//The total number of operations is 8
1253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1254{
1255        return simd128<(1)>::ifh(simd128<1>::ifh(simd128<2>::himask(), arg1, simd128<2>::srli<(1)>(arg1)), arg2, arg3);
1256}
1257
1258//The total number of operations is 13
1259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1260{
1261        return simd128<1>::ifh(simd128<4>::gt(simd128<4>::constant<0>(), arg1), arg2, arg3);
1262}
1263
1264//The total number of operations is 4
1265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1266{
1267        return simd128<1>::ifh(simd128<8>::gt(simd128<8>::constant<0>(), arg1), arg2, arg3);
1268}
1269
1270//The total number of operations is 4
1271template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1272{
1273        return simd128<1>::ifh(simd128<16>::gt(simd128<16>::constant<0>(), arg1), arg2, arg3);
1274}
1275
1276//The total number of operations is 4
1277template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1278{
1279        return simd128<1>::ifh(simd128<32>::gt(simd128<32>::constant<0>(), arg1), arg2, arg3);
1280}
1281
1282//The total number of operations is 8
1283template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1284{
1285        return simd128<(32)>::ifh(simd128<1>::ifh(simd128<64>::himask(), arg1, simd128<64>::srli<(32)>(arg1)), arg2, arg3);
1286}
1287
1288//The total number of operations is 15
1289template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1290{
1291        return simd128<(64)>::ifh(simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg1)), arg2, arg3);
1292}
1293
1294//The total number of operations is 26
1295template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2)
1296{
1297        return simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::sub(arg1, simd_and(simd128<(2)>::himask(), arg2)), simd128<(2)>::sub(arg1, arg2));
1298}
1299
1300//The total number of operations is 11
1301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2)
1302{
1303        bitblock128_t tmp = simd_xor(arg1, arg2);
1304        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2))), tmp);
1305}
1306
1307//The total number of operations is 6
1308template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2)
1309{
1310        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::sub(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::sub(arg1, arg2));
1311}
1312
1313//The total number of operations is 1
1314template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2)
1315{
1316        return _mm_sub_epi8(arg1, arg2);
1317}
1318
1319//The total number of operations is 1
1320template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2)
1321{
1322        return _mm_sub_epi16(arg1, arg2);
1323}
1324
1325//The total number of operations is 1
1326template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2)
1327{
1328        return _mm_sub_epi32(arg1, arg2);
1329}
1330
1331//The total number of operations is 1
1332template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2)
1333{
1334        return _mm_sub_epi64(arg1, arg2);
1335}
1336
1337//The total number of operations is 11
1338template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2)
1339{
1340        bitblock128_t partial = simd128<(64)>::sub(arg1, arg2);
1341        bitblock128_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
1342        bitblock128_t borrow = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(borrowMask));
1343        return simd128<(64)>::sub(partial, borrow);
1344}
1345
1346//The total number of operations is 3
1347template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1)
1348{
1349        return simd128<16>::sub(arg1, simd_and(simd128<2>::lomask(), simd128<16>::srli<1>(arg1)));
1350}
1351
1352//The total number of operations is 4
1353template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1)
1354{
1355        return simd128<(8)>::add(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1356}
1357
1358//The total number of operations is 4
1359template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1)
1360{
1361        return simd128<(16)>::add(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1362}
1363
1364//The total number of operations is 3
1365template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1)
1366{
1367        return simd128<(32)>::add(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1368}
1369
1370//The total number of operations is 3
1371template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1)
1372{
1373        return simd128<(64)>::add(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1374}
1375
1376//The total number of operations is 3
1377template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1)
1378{
1379        return simd128<64>::add(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1380}
1381
1382//The total number of operations is 16
1383template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1)
1384{
1385        return simd128<128>::add(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1386}
1387
1388//The total number of operations is 7
1389template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1390{
1391        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())));
1392}
1393
1394//The total number of operations is 11
1395template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1396{
1397        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))));
1398}
1399
1400//The total number of operations is 0
1401template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask()
1402{
1403        return simd128<2>::constant<(1)>();
1404}
1405
1406//The total number of operations is 0
1407template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask()
1408{
1409        return simd128<4>::constant<(3)>();
1410}
1411
1412//The total number of operations is 0
1413template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask()
1414{
1415        return simd128<8>::constant<(15)>();
1416}
1417
1418//The total number of operations is 0
1419template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask()
1420{
1421        return simd128<16>::constant<(255)>();
1422}
1423
1424//The total number of operations is 0
1425template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask()
1426{
1427        return simd128<32>::constant<(65535)>();
1428}
1429
1430//The total number of operations is 0
1431template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask()
1432{
1433        return _mm_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1));
1434}
1435
1436//The total number of operations is 0
1437template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask()
1438{
1439        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1));
1440}
1441
1442//The total number of operations is 0
1443template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1444{
1445        return simd128<32>::constant<(-1*val)>();
1446}
1447
1448//The total number of operations is 0
1449template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1450{
1451        return ((val < 0) ? simd128<(4)>::constant<((val<<2)|(val^(-4)))>() : simd128<(4)>::constant<((val<<2)|val)>());
1452}
1453
1454//The total number of operations is 0
1455template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1456{
1457        return ((val < 0) ? simd128<(8)>::constant<((val<<4)|(val^(-16)))>() : simd128<(8)>::constant<((val<<4)|val)>());
1458}
1459
1460//The total number of operations is 0
1461template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1462{
1463        return _mm_set1_epi8((int32_t)(val));
1464}
1465
1466//The total number of operations is 0
1467template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1468{
1469        return _mm_set1_epi16((int32_t)(val));
1470}
1471
1472//The total number of operations is 0
1473template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1474{
1475        return _mm_set1_epi32((int32_t)(val));
1476}
1477
1478//The total number of operations is 0
1479template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1480{
1481        return _mm_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val));
1482}
1483
1484//The total number of operations is 0
1485template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1486{
1487        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val));
1488}
1489
1490//The total number of operations is 39
1491template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2)
1492{
1493        bitblock128_t high_bit = simd128<1>::constant<(1)>();
1494        return simd_xor(simd128<1>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1495}
1496
1497//The total number of operations is 19
1498template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2)
1499{
1500        bitblock128_t high_bit = simd128<2>::constant<(2)>();
1501        return simd_xor(simd128<2>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1502}
1503
1504//The total number of operations is 9
1505template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2)
1506{
1507        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1508        return simd_xor(simd128<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1509}
1510
1511//The total number of operations is 4
1512template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2)
1513{
1514        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg2, arg1);
1515}
1516
1517//The total number of operations is 1
1518template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2)
1519{
1520        return _mm_min_epi16(arg1, arg2);
1521}
1522
1523//The total number of operations is 4
1524template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2)
1525{
1526        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg2, arg1);
1527}
1528
1529//The total number of operations is 18
1530template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2)
1531{
1532        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg2, arg1);
1533}
1534
1535//The total number of operations is 65
1536template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2)
1537{
1538        bitblock128_t hiAns = simd128<(64)>::min(arg1, arg2);
1539        bitblock128_t loAns = simd128<(64)>::umin(arg1, arg2);
1540        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
1541        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
1542        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
1543}
1544
1545//The total number of operations is 36
1546template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2)
1547{
1548        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)));
1549}
1550
1551//The total number of operations is 16
1552template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2)
1553{
1554        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)));
1555}
1556
1557//The total number of operations is 6
1558template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2)
1559{
1560        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)));
1561}
1562
1563//The total number of operations is 1
1564template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2)
1565{
1566        return _mm_min_epu8(arg1, arg2);
1567}
1568
1569//The total number of operations is 4
1570template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2)
1571{
1572        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1573        return simd_xor(simd128<16>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1574}
1575
1576//The total number of operations is 7
1577template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2)
1578{
1579        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1580        return simd_xor(simd128<32>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1581}
1582
1583//The total number of operations is 20
1584template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2)
1585{
1586        bitblock128_t tmpAns = simd128<(32)>::umin(arg1, arg2);
1587        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1588        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1589        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1590}
1591
1592//The total number of operations is 47
1593template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2)
1594{
1595        bitblock128_t tmpAns = simd128<(64)>::umin(arg1, arg2);
1596        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1597        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1598        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1599}
1600
1601//The total number of operations is 36
1602template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2)
1603{
1604        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)));
1605}
1606
1607//The total number of operations is 16
1608template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2)
1609{
1610        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)));
1611}
1612
1613//The total number of operations is 6
1614template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2)
1615{
1616        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)));
1617}
1618
1619//The total number of operations is 1
1620template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2)
1621{
1622        return _mm_max_epu8(arg1, arg2);
1623}
1624
1625//The total number of operations is 4
1626template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2)
1627{
1628        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1629        return simd_xor(simd128<16>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1630}
1631
1632//The total number of operations is 7
1633template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2)
1634{
1635        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1636        return simd_xor(simd128<32>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1637}
1638
1639//The total number of operations is 20
1640template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2)
1641{
1642        bitblock128_t tmpAns = simd128<(32)>::umax(arg1, arg2);
1643        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1644        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1645        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1646}
1647
1648//The total number of operations is 47
1649template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2)
1650{
1651        bitblock128_t tmpAns = simd128<(64)>::umax(arg1, arg2);
1652        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1653        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1654        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1655}
1656
1657//The total number of operations is 37
1658template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2)
1659{
1660        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))));
1661}
1662
1663//The total number of operations is 15
1664template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2)
1665{
1666        bitblock128_t tmp = simd_xor(arg1, arg2);
1667        bitblock128_t tmpAns = simd_and(simd_not(simd128<128>::slli<1>(tmp)), simd_not(tmp));
1668        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1669}
1670
1671//The total number of operations is 9
1672template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2)
1673{
1674        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))));
1675}
1676
1677//The total number of operations is 1
1678template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2)
1679{
1680        return _mm_cmpeq_epi8(arg1, arg2);
1681}
1682
1683//The total number of operations is 1
1684template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2)
1685{
1686        return _mm_cmpeq_epi16(arg1, arg2);
1687}
1688
1689//The total number of operations is 1
1690template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2)
1691{
1692        return _mm_cmpeq_epi32(arg1, arg2);
1693}
1694
1695//The total number of operations is 5
1696template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2)
1697{
1698        bitblock128_t tmpAns = simd128<(32)>::eq(arg1, arg2);
1699        bitblock128_t loMask = simd_and(tmpAns, simd128<64>::srli<(32)>(tmpAns));
1700        bitblock128_t hiMask = simd128<64>::slli<(32)>(loMask);
1701        return simd_or(loMask, hiMask);
1702}
1703
1704//The total number of operations is 15
1705template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2)
1706{
1707        bitblock128_t tmpAns = simd128<(64)>::eq(arg1, arg2);
1708        bitblock128_t loMask = simd_and(tmpAns, simd128<128>::srli<(64)>(tmpAns));
1709        bitblock128_t hiMask = simd128<128>::slli<(64)>(loMask);
1710        return simd_or(loMask, hiMask);
1711}
1712
1713//The total number of operations is 4
1714template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1)
1715{
1716        return ((sh == 0) ? arg1 : simd_or(simd_and(simd128<2>::himask(), arg1), simd128<2>::srli<1>(arg1)));
1717}
1718
1719//The total number of operations is 10
1720template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1)
1721{
1722        bitblock128_t tmp = simd128<4>::srli<((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh))>(arg1);
1723        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)));
1724}
1725
1726//The total number of operations is 5
1727template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1)
1728{
1729        bitblock128_t tmp = simd128<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
1730        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)));
1731}
1732
1733//The total number of operations is 1
1734template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1)
1735{
1736        return _mm_srai_epi16(arg1, (int32_t)(sh));
1737}
1738
1739//The total number of operations is 1
1740template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1)
1741{
1742        return _mm_srai_epi32(arg1, (int32_t)(sh));
1743}
1744
1745//The total number of operations is 5
1746template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1)
1747{
1748        bitblock128_t tmp = simd128<64>::srli<((sh >= 64) ? (63) : ((sh < 0) ? 0 : sh))>(arg1);
1749        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)));
1750}
1751
1752//The total number of operations is 21
1753template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1)
1754{
1755        bitblock128_t tmp = simd128<128>::srli<((sh >= 128) ? (127) : ((sh < 0) ? 0 : sh))>(arg1);
1756        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)));
1757}
1758
1759//The total number of operations is 0
1760template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask()
1761{
1762        return simd128<2>::constant<(2)>();
1763}
1764
1765//The total number of operations is 0
1766template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask()
1767{
1768        return simd128<4>::constant<(12)>();
1769}
1770
1771//The total number of operations is 0
1772template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask()
1773{
1774        return simd128<8>::constant<(240)>();
1775}
1776
1777//The total number of operations is 0
1778template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask()
1779{
1780        return simd128<16>::constant<(65280)>();
1781}
1782
1783//The total number of operations is 0
1784template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask()
1785{
1786        return simd128<32>::constant<-65536>();
1787}
1788
1789//The total number of operations is 0
1790template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask()
1791{
1792        return _mm_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0));
1793}
1794
1795//The total number of operations is 0
1796template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask()
1797{
1798        return _mm_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0));
1799}
1800
1801//The total number of operations is 24
1802template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2)
1803{
1804        return simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::add(arg1, simd_and(simd128<(2)>::himask(), arg2)), simd128<(2)>::add(arg1, arg2));
1805}
1806
1807//The total number of operations is 10
1808template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2)
1809{
1810        bitblock128_t tmp = simd_xor(arg1, arg2);
1811        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(arg1, arg2))), tmp);
1812}
1813
1814//The total number of operations is 6
1815template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2)
1816{
1817        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::add(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::add(arg1, arg2));
1818}
1819
1820//The total number of operations is 1
1821template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2)
1822{
1823        return _mm_add_epi8(arg1, arg2);
1824}
1825
1826//The total number of operations is 1
1827template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2)
1828{
1829        return _mm_add_epi16(arg1, arg2);
1830}
1831
1832//The total number of operations is 1
1833template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2)
1834{
1835        return _mm_add_epi32(arg1, arg2);
1836}
1837
1838//The total number of operations is 1
1839template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2)
1840{
1841        return _mm_add_epi64(arg1, arg2);
1842}
1843
1844//The total number of operations is 11
1845template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2)
1846{
1847        bitblock128_t partial = simd128<(64)>::add(arg1, arg2);
1848        bitblock128_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
1849        bitblock128_t carry = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(carryMask));
1850        return simd128<(64)>::add(partial, carry);
1851}
1852
1853//The total number of operations is 71
1854template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::abs(bitblock128_t arg1)
1855{
1856        bitblock128_t gtMask = simd128<1>::gt(arg1, simd128<1>::constant<0>());
1857        return simd128<1>::ifh(gtMask, arg1, simd128<1>::sub(gtMask, arg1));
1858}
1859
1860//The total number of operations is 9
1861template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1)
1862{
1863        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, simd128<128>::slli<1>(simd_not(arg1))), arg1);
1864}
1865
1866//The total number of operations is 19
1867template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1)
1868{
1869        bitblock128_t gtMask = simd128<4>::gt(arg1, simd128<4>::constant<0>());
1870        return simd128<1>::ifh(gtMask, arg1, simd128<4>::sub(gtMask, arg1));
1871}
1872
1873//The total number of operations is 1
1874template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1)
1875{
1876        return _mm_abs_epi8(arg1);
1877}
1878
1879//The total number of operations is 1
1880template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1)
1881{
1882        return _mm_abs_epi16(arg1);
1883}
1884
1885//The total number of operations is 1
1886template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1)
1887{
1888        return _mm_abs_epi32(arg1);
1889}
1890
1891//The total number of operations is 13
1892template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1)
1893{
1894        bitblock128_t eqMask = simd128<64>::eq(simd128<1>::ifh(simd128<64>::himask(), simd128<(32)>::abs(arg1), arg1), arg1);
1895        return simd128<1>::ifh(eqMask, arg1, simd128<64>::sub(eqMask, arg1));
1896}
1897
1898//The total number of operations is 45
1899template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1)
1900{
1901        bitblock128_t eqMask = simd128<128>::eq(simd128<1>::ifh(simd128<128>::himask(), simd128<(64)>::abs(arg1), arg1), arg1);
1902        return simd128<1>::ifh(eqMask, arg1, simd128<128>::sub(eqMask, arg1));
1903}
1904
1905//The total number of operations is 128
1906template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1907{
1908        return simd128<(1)>::umin(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1909}
1910
1911//The total number of operations is 80
1912template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1913{
1914        return simd128<(2)>::umin(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1915}
1916
1917//The total number of operations is 42
1918template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1919{
1920        return simd128<(4)>::umin(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1921}
1922
1923//The total number of operations is 7
1924template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1925{
1926        return simd128<(8)>::umin(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1927}
1928
1929//The total number of operations is 10
1930template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1931{
1932        return simd128<(16)>::umin(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1933}
1934
1935//The total number of operations is 13
1936template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1937{
1938        return simd128<(32)>::umin(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1939}
1940
1941//The total number of operations is 34
1942template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1943{
1944        return simd128<(64)>::umin(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1945}
1946
1947//The total number of operations is 116
1948template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1949{
1950        return simd128<(1)>::add(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1951}
1952
1953//The total number of operations is 74
1954template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1955{
1956        return simd128<(2)>::add(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1957}
1958
1959//The total number of operations is 42
1960template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1961{
1962        return simd128<(4)>::add(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1963}
1964
1965//The total number of operations is 7
1966template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1967{
1968        return simd128<(8)>::add(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1969}
1970
1971//The total number of operations is 1
1972template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1973{
1974        return _mm_hadd_epi16(arg2, arg1);
1975}
1976
1977//The total number of operations is 1
1978template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1979{
1980        return _mm_hadd_epi32(arg2, arg1);
1981}
1982
1983//The total number of operations is 15
1984template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1985{
1986        return simd128<(64)>::add(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1987}
1988
1989//The total number of operations is 132
1990template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2)
1991{
1992        bitblock128_t hiBound = simd128<2>::srli<1>(simd128<2>::lomask());
1993        bitblock128_t loBound = simd_not(hiBound);
1994        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)));
1995}
1996
1997//The total number of operations is 86
1998template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2)
1999{
2000        bitblock128_t hiBound = simd128<4>::srli<1>(simd128<4>::lomask());
2001        bitblock128_t loBound = simd_not(hiBound);
2002        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)));
2003}
2004
2005//The total number of operations is 36
2006template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2)
2007{
2008        bitblock128_t hiBound = simd128<8>::srli<1>(simd128<8>::lomask());
2009        bitblock128_t loBound = simd_not(hiBound);
2010        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)));
2011}
2012
2013//The total number of operations is 1
2014template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2)
2015{
2016        return _mm_packs_epi16(arg2, arg1);
2017}
2018
2019//The total number of operations is 1
2020template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2)
2021{
2022        return _mm_packs_epi32(arg2, arg1);
2023}
2024
2025//The total number of operations is 77
2026template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2)
2027{
2028        bitblock128_t hiBound = simd128<64>::srli<1>(simd128<64>::lomask());
2029        bitblock128_t loBound = simd_not(hiBound);
2030        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)));
2031}
2032
2033//The total number of operations is 288
2034template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2)
2035{
2036        bitblock128_t hiBound = simd128<128>::srli<1>(simd128<128>::lomask());
2037        bitblock128_t loBound = simd_not(hiBound);
2038        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)));
2039}
2040
2041//The total number of operations is 24
2042template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1)
2043{
2044        uint64_t tmpAns1 = hsimd128<(8)>::signmask(esimd128<4>::mergeh(arg1, simd128<4>::constant<0>()));
2045        uint64_t tmpAns2 = hsimd128<(8)>::signmask(esimd128<4>::mergel(arg1, simd128<4>::constant<0>()));
2046        return ((tmpAns1<<(16))+tmpAns2);
2047}
2048
2049//The total number of operations is 1
2050template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1)
2051{
2052        return _mm_movemask_epi8(arg1);
2053}
2054
2055//The total number of operations is 4
2056template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1)
2057{
2058        return hsimd128<(8)>::signmask(hsimd128<16>::packh(simd128<16>::constant<0>(), arg1));
2059}
2060
2061//The total number of operations is 7
2062template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1)
2063{
2064        return hsimd128<(16)>::signmask(hsimd128<32>::packh(simd128<32>::constant<0>(), arg1));
2065}
2066
2067//The total number of operations is 10
2068template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1)
2069{
2070        return hsimd128<(32)>::signmask(hsimd128<64>::packh(simd128<64>::constant<0>(), arg1));
2071}
2072
2073//The total number of operations is 17
2074template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1)
2075{
2076        return hsimd128<(64)>::signmask(hsimd128<128>::packh(simd128<128>::constant<0>(), arg1));
2077}
2078
2079//The total number of operations is 45
2080template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2)
2081{
2082        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));
2083}
2084
2085//The total number of operations is 31
2086template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2)
2087{
2088        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));
2089}
2090
2091//The total number of operations is 17
2092template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2)
2093{
2094        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));
2095}
2096
2097//The total number of operations is 3
2098template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2)
2099{
2100        return hsimd128<16>::packus(simd_and(arg1, simd128<16>::lomask()), simd_and(arg2, simd128<16>::lomask()));
2101}
2102
2103//The total number of operations is 3
2104template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2)
2105{
2106        return _mm_hsub_epi16(simd_and(arg2, simd128<32>::lomask()), simd_and(arg1, simd128<32>::lomask()));
2107}
2108
2109//The total number of operations is 3
2110template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2)
2111{
2112        return _mm_hsub_epi32(simd_and(arg2, simd128<64>::lomask()), simd_and(arg1, simd128<64>::lomask()));
2113}
2114
2115//The total number of operations is 7
2116template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2)
2117{
2118        return simd128<1>::ifh(simd128<128>::himask(), simd128<128>::slli<(64)>(arg1), arg2);
2119}
2120
2121//The total number of operations is 47
2122template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2)
2123{
2124        return hsimd128<2>::packl(simd128<64>::srli<(1)>(arg1), simd128<64>::srli<(1)>(arg2));
2125}
2126
2127//The total number of operations is 33
2128template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2)
2129{
2130        return hsimd128<4>::packl(simd128<64>::srli<(2)>(arg1), simd128<64>::srli<(2)>(arg2));
2131}
2132
2133//The total number of operations is 19
2134template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2)
2135{
2136        return hsimd128<8>::packl(simd128<64>::srli<(4)>(arg1), simd128<64>::srli<(4)>(arg2));
2137}
2138
2139//The total number of operations is 3
2140template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2)
2141{
2142        return hsimd128<16>::packus(simd128<16>::srli<(8)>(arg1), simd128<16>::srli<(8)>(arg2));
2143}
2144
2145//The total number of operations is 3
2146template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2)
2147{
2148        return _mm_hsub_epi16(simd128<32>::srli<(16)>(arg2), simd128<32>::srli<(16)>(arg1));
2149}
2150
2151//The total number of operations is 3
2152template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2)
2153{
2154        return _mm_hsub_epi32(simd128<64>::srli<(32)>(arg2), simd128<64>::srli<(32)>(arg1));
2155}
2156
2157//The total number of operations is 7
2158template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2)
2159{
2160        return simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg2));
2161}
2162
2163//The total number of operations is 131
2164template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2165{
2166        return simd128<(1)>::min(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
2167}
2168
2169//The total number of operations is 83
2170template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2171{
2172        return simd128<(2)>::min(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
2173}
2174
2175//The total number of operations is 45
2176template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2177{
2178        return simd128<(4)>::min(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
2179}
2180
2181//The total number of operations is 10
2182template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2183{
2184        return simd128<(8)>::min(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2185}
2186
2187//The total number of operations is 7
2188template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2189{
2190        return simd128<(16)>::min(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2191}
2192
2193//The total number of operations is 10
2194template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2195{
2196        return simd128<(32)>::min(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2197}
2198
2199//The total number of operations is 32
2200template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2201{
2202        return simd128<(64)>::min(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2203}
2204
2205//The total number of operations is 99
2206template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2)
2207{
2208        bitblock128_t arg11 = simd128<2>::ifh(arg1, simd128<2>::constant<0>(), arg1);
2209        bitblock128_t arg12 = simd_and(simd128<2>::lomask(), arg11);
2210        bitblock128_t arg21 = simd128<2>::ifh(arg2, simd128<2>::constant<0>(), arg2);
2211        bitblock128_t arg22 = simd_and(simd128<2>::lomask(), arg21);
2212        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()));
2213}
2214
2215//The total number of operations is 83
2216template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2)
2217{
2218        bitblock128_t arg11 = simd128<4>::ifh(arg1, simd128<4>::constant<0>(), arg1);
2219        bitblock128_t arg12 = simd_and(simd128<4>::lomask(), arg11);
2220        bitblock128_t arg21 = simd128<4>::ifh(arg2, simd128<4>::constant<0>(), arg2);
2221        bitblock128_t arg22 = simd_and(simd128<4>::lomask(), arg21);
2222        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()));
2223}
2224
2225//The total number of operations is 35
2226template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2)
2227{
2228        bitblock128_t arg11 = simd128<8>::ifh(arg1, simd128<8>::constant<0>(), arg1);
2229        bitblock128_t arg12 = simd_and(simd128<8>::lomask(), arg11);
2230        bitblock128_t arg21 = simd128<8>::ifh(arg2, simd128<8>::constant<0>(), arg2);
2231        bitblock128_t arg22 = simd_and(simd128<8>::lomask(), arg21);
2232        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()));
2233}
2234
2235//The total number of operations is 1
2236template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2)
2237{
2238        return _mm_packus_epi16(arg2, arg1);
2239}
2240
2241//The total number of operations is 12
2242template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2)
2243{
2244        bitblock128_t hiPart = hsimd128<32>::packh(arg1, arg2);
2245        return simd128<(16)>::ifh(hiPart, simd128<(16)>::constant<0>(), simd_or(simd128<(16)>::gt(hiPart, simd128<(16)>::constant<0>()), hsimd128<32>::packl(arg1, arg2)));
2246}
2247
2248//The total number of operations is 12
2249template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2)
2250{
2251        bitblock128_t hiPart = hsimd128<64>::packh(arg1, arg2);
2252        return simd128<(32)>::ifh(hiPart, simd128<(32)>::constant<0>(), simd_or(simd128<(32)>::gt(hiPart, simd128<(32)>::constant<0>()), hsimd128<64>::packl(arg1, arg2)));
2253}
2254
2255//The total number of operations is 38
2256template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2)
2257{
2258        bitblock128_t hiPart = hsimd128<128>::packh(arg1, arg2);
2259        return simd128<(64)>::ifh(hiPart, simd128<(64)>::constant<0>(), simd_or(simd128<(64)>::gt(hiPart, simd128<(64)>::constant<0>()), hsimd128<128>::packl(arg1, arg2)));
2260}
2261
2262//The total number of operations is 31
2263template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2264{
2265        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));
2266}
2267
2268//The total number of operations is 21
2269template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2270{
2271        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));
2272}
2273
2274//The total number of operations is 11
2275template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2276{
2277        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));
2278}
2279
2280//The total number of operations is 1
2281template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2282{
2283        return _mm_unpacklo_epi8(arg2, arg1);
2284}
2285
2286//The total number of operations is 1
2287template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2288{
2289        return _mm_unpacklo_epi16(arg2, arg1);
2290}
2291
2292//The total number of operations is 1
2293template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2294{
2295        return _mm_unpacklo_epi32(arg2, arg1);
2296}
2297
2298//The total number of operations is 1
2299template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2300{
2301        return _mm_unpacklo_epi64(arg2, arg1);
2302}
2303
2304//The total number of operations is 31
2305template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2306{
2307        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));
2308}
2309
2310//The total number of operations is 21
2311template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2312{
2313        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));
2314}
2315
2316//The total number of operations is 11
2317template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2318{
2319        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));
2320}
2321
2322//The total number of operations is 1
2323template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2324{
2325        return _mm_unpackhi_epi8(arg2, arg1);
2326}
2327
2328//The total number of operations is 1
2329template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2330{
2331        return _mm_unpackhi_epi16(arg2, arg1);
2332}
2333
2334//The total number of operations is 1
2335template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2336{
2337        return _mm_unpackhi_epi32(arg2, arg1);
2338}
2339
2340//The total number of operations is 1
2341template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2342{
2343        return _mm_unpackhi_epi64(arg2, arg1);
2344}
2345
2346//The total number of operations is 24
2347template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1)
2348{
2349        return esimd128<(2)>::mergeh(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2350}
2351
2352//The total number of operations is 14
2353template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1)
2354{
2355        return esimd128<(4)>::mergeh(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2356}
2357
2358//The total number of operations is 4
2359template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1)
2360{
2361        return esimd128<(8)>::mergeh(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2362}
2363
2364//The total number of operations is 3
2365template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1)
2366{
2367        return esimd128<(16)>::mergeh(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2368}
2369
2370//The total number of operations is 3
2371template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1)
2372{
2373        return esimd128<(32)>::mergeh(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2374}
2375
2376//The total number of operations is 3
2377template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1)
2378{
2379        return esimd128<(64)>::mergeh(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2380}
2381
2382//The total number of operations is 4
2383template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1)
2384{
2385        return simd128<(128)>::srli<64>(arg1);
2386}
2387
2388//The total number of operations is 24
2389template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1)
2390{
2391        return esimd128<(2)>::mergel(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2392}
2393
2394//The total number of operations is 14
2395template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1)
2396{
2397        return esimd128<(4)>::mergel(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2398}
2399
2400//The total number of operations is 4
2401template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1)
2402{
2403        return esimd128<(8)>::mergel(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2404}
2405
2406//The total number of operations is 3
2407template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1)
2408{
2409        return esimd128<(16)>::mergel(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2410}
2411
2412//The total number of operations is 3
2413template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1)
2414{
2415        return esimd128<(32)>::mergel(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2416}
2417
2418//The total number of operations is 3
2419template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1)
2420{
2421        return esimd128<(64)>::mergel(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2422}
2423
2424//The total number of operations is 1
2425template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1)
2426{
2427        return simd_and(simd128<(128)>::lomask(), arg1);
2428}
2429
2430//The total number of operations is 31
2431template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1)
2432{
2433        return esimd128<(2)>::mergeh(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2434}
2435
2436//The total number of operations is 33
2437template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1)
2438{
2439        return esimd128<(4)>::mergeh(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2440}
2441
2442//The total number of operations is 13
2443template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1)
2444{
2445        return esimd128<(8)>::mergeh(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2446}
2447
2448//The total number of operations is 4
2449template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1)
2450{
2451        return esimd128<(16)>::mergeh(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2452}
2453
2454//The total number of operations is 4
2455template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1)
2456{
2457        return esimd128<(32)>::mergeh(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2458}
2459
2460//The total number of operations is 12
2461template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1)
2462{
2463        return esimd128<(64)>::mergeh(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2464}
2465
2466//The total number of operations is 21
2467template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1)
2468{
2469        return simd128<(128)>::srai<64>(arg1);
2470}
2471
2472//The total number of operations is 31
2473template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1)
2474{
2475        return esimd128<(2)>::mergel(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2476}
2477
2478//The total number of operations is 33
2479template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1)
2480{
2481        return esimd128<(4)>::mergel(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2482}
2483
2484//The total number of operations is 13
2485template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1)
2486{
2487        return esimd128<(8)>::mergel(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2488}
2489
2490//The total number of operations is 4
2491template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1)
2492{
2493        return esimd128<(16)>::mergel(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2494}
2495
2496//The total number of operations is 4
2497template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1)
2498{
2499        return esimd128<(32)>::mergel(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2500}
2501
2502//The total number of operations is 12
2503template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1)
2504{
2505        return esimd128<(64)>::mergel(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2506}
2507
2508//The total number of operations is 25
2509template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1)
2510{
2511        return simd128<(128)>::srai<64>(simd128<(128)>::slli<64>(arg1));
2512}
2513
2514//The total number of operations is 17
2515template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1)
2516{
2517        bitblock128_t tmphi = _mm_shufflehi_epi16(arg1, (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2518        bitblock128_t tmpAns = _mm_shufflelo_epi16(tmphi, (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2519        bitblock128_t tmplh = _mm_shufflehi_epi16(simd128<128>::slli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2520        bitblock128_t tmphl = _mm_shufflelo_epi16(simd128<128>::srli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2521        uint32_t a1 = ((((msk>>21)&4) == 0) ? 0 : (131071));
2522        uint32_t a2 = ((((msk>>18)&4) == 0) ? 0 : (131071));
2523        uint32_t a3 = ((((msk>>15)&4) == 0) ? 0 : (131071));
2524        uint32_t a4 = ((((msk>>12)&4) == 0) ? 0 : (131071));
2525        uint32_t a5 = ((((msk>>9)&4) == 0) ? (131071) : 0);
2526        uint32_t a6 = ((((msk>>6)&4) == 0) ? (131071) : 0);
2527        uint32_t a7 = ((((msk>>3)&4) == 0) ? (131071) : 0);
2528        uint32_t a8 = (((msk&4) == 0) ? (131071) : 0);
2529        return simd128<1>::ifh(mvmd128<16>::fill8(a1, a2, a3, a4, a5, a6, a7, a8), tmpAns, simd_or(tmplh, tmphl));
2530}
2531
2532//The total number of operations is 1
2533template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1)
2534{
2535        return _mm_shuffle_epi32(arg1, (int32_t)(msk));
2536}
2537
2538//The total number of operations is 1
2539template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1)
2540{
2541        return mvmd128<32>::shufflei<shufflemask4_from_shufflemask2(msk)>(arg1);
2542}
2543
2544//The total number of operations is 9
2545template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2546{
2547        return simd_or(mvmd128<2>::srli<sh>(arg1), mvmd128<2>::slli<((64)-sh)>(arg2));
2548}
2549
2550//The total number of operations is 9
2551template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2552{
2553        return simd_or(mvmd128<4>::srli<sh>(arg1), mvmd128<4>::slli<((32)-sh)>(arg2));
2554}
2555
2556//The total number of operations is 9
2557template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2558{
2559        return simd_or(mvmd128<8>::srli<sh>(arg1), mvmd128<8>::slli<((16)-sh)>(arg2));
2560}
2561
2562//The total number of operations is 9
2563template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2564{
2565        return simd_or(mvmd128<16>::srli<sh>(arg1), mvmd128<16>::slli<((8)-sh)>(arg2));
2566}
2567
2568//The total number of operations is 9
2569template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2570{
2571        return simd_or(mvmd128<32>::srli<sh>(arg1), mvmd128<32>::slli<((4)-sh)>(arg2));
2572}
2573
2574//The total number of operations is 9
2575template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2576{
2577        return simd_or(mvmd128<64>::srli<sh>(arg1), mvmd128<64>::slli<((2)-sh)>(arg2));
2578}
2579
2580//The total number of operations is 9
2581template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2582{
2583        return simd_or(mvmd128<128>::srli<sh>(arg1), mvmd128<128>::slli<((1)-sh)>(arg2));
2584}
2585
2586//The total number of operations is 1
2587template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1)
2588{
2589        return mvmd128<32>::fill((-1*val1));
2590}
2591
2592//The total number of operations is 1
2593template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1)
2594{
2595        return mvmd128<(4)>::fill(((val1<<2)|val1));
2596}
2597
2598//The total number of operations is 1
2599template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1)
2600{
2601        return mvmd128<(8)>::fill(((val1<<4)|val1));
2602}
2603
2604//The total number of operations is 1
2605template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1)
2606{
2607        return _mm_set1_epi8((int32_t)(val1));
2608}
2609
2610//The total number of operations is 1
2611template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1)
2612{
2613        return _mm_set1_epi16((int32_t)(val1));
2614}
2615
2616//The total number of operations is 1
2617template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1)
2618{
2619        return _mm_set1_epi32((int32_t)(val1));
2620}
2621
2622//The total number of operations is 1
2623template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1)
2624{
2625        return _mm_set_epi32((int32_t)((val1>>32)), (int32_t)(val1), (int32_t)((val1>>32)), (int32_t)(val1));
2626}
2627
2628//The total number of operations is 1
2629template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1)
2630{
2631        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val1>>32)), (int32_t)(val1));
2632}
2633
2634//The total number of operations is 1
2635template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2636{
2637        return _mm_shuffle_epi8(arg1, arg2);
2638}
2639
2640//The total number of operations is 10
2641template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2642{
2643        bitblock128_t tmp1 = simd_and(simd128<16>::constant<(7)>(), arg2);
2644        bitblock128_t msk1 = simd128<16>::add(tmp1, tmp1);
2645        bitblock128_t msk2 = simd128<16>::add(msk1, simd128<16>::constant<1>());
2646        bitblock128_t msk = simd_or(msk1, simd128<16>::slli<(8)>(msk2));
2647        return simd128<16>::ifh(arg2, simd128<16>::constant<0>(), mvmd128<(8)>::shuffle(arg1, msk));
2648}
2649
2650//The total number of operations is 19
2651template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2652{
2653        bitblock128_t tmp1 = simd_and(simd128<32>::constant<(3)>(), arg2);
2654        bitblock128_t msk1 = simd128<32>::add(tmp1, tmp1);
2655        bitblock128_t msk2 = simd128<32>::add(msk1, simd128<32>::constant<1>());
2656        bitblock128_t msk = simd_or(msk1, simd128<32>::slli<(16)>(msk2));
2657        return simd128<32>::ifh(arg2, simd128<32>::constant<0>(), mvmd128<(16)>::shuffle(arg1, msk));
2658}
2659
2660//The total number of operations is 32
2661template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2662{
2663        bitblock128_t tmp1 = simd_and(simd128<64>::constant<(1)>(), arg2);
2664        bitblock128_t msk1 = simd128<64>::add(tmp1, tmp1);
2665        bitblock128_t msk2 = simd128<64>::add(msk1, simd128<64>::constant<1>());
2666        bitblock128_t msk = simd_or(msk1, simd128<64>::slli<(32)>(msk2));
2667        return simd128<64>::ifh(arg2, simd128<64>::constant<0>(), mvmd128<(32)>::shuffle(arg1, msk));
2668}
2669
2670//The total number of operations is 1
2671template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1)
2672{
2673        return (((pos%2) == 0) ? (mvmd128<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd128<(2)>::extract<(pos/2)>(arg1)>>1));
2674}
2675
2676//The total number of operations is 1
2677template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1)
2678{
2679        return (((pos%2) == 0) ? (mvmd128<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd128<(4)>::extract<(pos/2)>(arg1)>>2));
2680}
2681
2682//The total number of operations is 1
2683template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1)
2684{
2685        return (((pos%2) == 0) ? (mvmd128<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd128<(8)>::extract<(pos/2)>(arg1)>>4));
2686}
2687
2688//The total number of operations is 1
2689template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1)
2690{
2691        return (((pos%2) == 0) ? (mvmd128<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd128<(16)>::extract<(pos/2)>(arg1)>>8));
2692}
2693
2694//The total number of operations is 1
2695template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1)
2696{
2697        return (65535&_mm_extract_epi16(arg1, (int32_t)(pos)));
2698}
2699
2700//The total number of operations is 2
2701template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1)
2702{
2703        return ((((uint64_t)mvmd128<(16)>::extract<((2*pos)+1)>(arg1))<<(16))|mvmd128<(16)>::extract<(2*pos)>(arg1));
2704}
2705
2706//The total number of operations is 4
2707template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1)
2708{
2709        return ((((uint64_t)mvmd128<(32)>::extract<((2*pos)+1)>(arg1))<<(32))|mvmd128<(32)>::extract<(2*pos)>(arg1));
2710}
2711
2712//The total number of operations is 16
2713template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1)
2714{
2715        return simd128<128>::sub(simd128<128>::constant<0>(), simd_and(simd128<128>::constant<1>(), simd128<128>::srli<pos>(arg1)));
2716}
2717
2718//The total number of operations is 13
2719template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1)
2720{
2721        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(4)>::slli<2>(arg1) : simd128<(4)>::srli<2>(arg1));
2722        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(4)>::lomask(), arg1) : simd_and(simd128<(4)>::himask(), arg1));
2723        return mvmd128<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2724}
2725
2726//The total number of operations is 9
2727template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1)
2728{
2729        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(8)>::slli<4>(arg1) : simd128<(8)>::srli<4>(arg1));
2730        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(8)>::lomask(), arg1) : simd_and(simd128<(8)>::himask(), arg1));
2731        return mvmd128<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2732}
2733
2734//The total number of operations is 5
2735template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1)
2736{
2737        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(16)>::slli<8>(arg1) : simd128<(16)>::srli<8>(arg1));
2738        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(16)>::lomask(), arg1) : simd_and(simd128<(16)>::himask(), arg1));
2739        return mvmd128<(16)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2740}
2741
2742//The total number of operations is 2
2743template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1)
2744{
2745        return mvmd128<16>::fill(_mm_extract_epi16(arg1, (int32_t)(pos)));
2746}
2747
2748//The total number of operations is 1
2749template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1)
2750{
2751        return mvmd128<32>::shufflei<shufflemask4(pos, pos, pos, pos)>(arg1);
2752}
2753
2754//The total number of operations is 5
2755template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1)
2756{
2757        return simd128<1>::ifh(simd128<64>::himask(), mvmd128<(32)>::splat<((2*pos)+1)>(arg1), mvmd128<(32)>::splat<(2*pos)>(arg1));
2758}
2759
2760//The total number of operations is 13
2761template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1)
2762{
2763        return simd128<1>::ifh(simd128<128>::himask(), mvmd128<(64)>::splat<((2*pos)+1)>(arg1), mvmd128<(64)>::splat<(2*pos)>(arg1));
2764}
2765
2766//The total number of operations is 15
2767template <> 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)
2768{
2769        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))));
2770}
2771
2772//The total number of operations is 7
2773template <> 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)
2774{
2775        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))));
2776}
2777
2778//The total number of operations is 3
2779template <> 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)
2780{
2781        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))));
2782}
2783
2784//The total number of operations is 1
2785template <> 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)
2786{
2787        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));
2788}
2789
2790//The total number of operations is 4
2791template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1)
2792{
2793        return simd128<128>::slli<(sh*2)>(arg1);
2794}
2795
2796//The total number of operations is 4
2797template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1)
2798{
2799        return simd128<128>::slli<(sh*4)>(arg1);
2800}
2801
2802//The total number of operations is 4
2803template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1)
2804{
2805        return simd128<128>::slli<(sh*8)>(arg1);
2806}
2807
2808//The total number of operations is 4
2809template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1)
2810{
2811        return simd128<128>::slli<(sh*16)>(arg1);
2812}
2813
2814//The total number of operations is 4
2815template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1)
2816{
2817        return simd128<128>::slli<(sh*32)>(arg1);
2818}
2819
2820//The total number of operations is 4
2821template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1)
2822{
2823        return simd128<128>::slli<(sh*64)>(arg1);
2824}
2825
2826//The total number of operations is 4
2827template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1)
2828{
2829        return simd128<128>::slli<(sh*128)>(arg1);
2830}
2831
2832//The total number of operations is 5
2833template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2834{
2835        return simd128<1>::ifh(simd128<(4)>::himask(), mvmd128<1>::fill2(val1, val2), mvmd128<1>::fill2(val3, val4));
2836}
2837
2838//The total number of operations is 5
2839template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2840{
2841        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<2>::fill2(val1, val2), mvmd128<2>::fill2(val3, val4));
2842}
2843
2844//The total number of operations is 5
2845template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2846{
2847        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<4>::fill2(val1, val2), mvmd128<4>::fill2(val3, val4));
2848}
2849
2850//The total number of operations is 5
2851template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2852{
2853        return simd128<1>::ifh(simd128<(32)>::himask(), mvmd128<8>::fill2(val1, val2), mvmd128<8>::fill2(val3, val4));
2854}
2855
2856//The total number of operations is 3
2857template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2858{
2859        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))));
2860}
2861
2862//The total number of operations is 1
2863template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2864{
2865        return _mm_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
2866}
2867
2868//The total number of operations is 4
2869template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1)
2870{
2871        return simd128<128>::srli<(sh*2)>(arg1);
2872}
2873
2874//The total number of operations is 4
2875template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1)
2876{
2877        return simd128<128>::srli<(sh*4)>(arg1);
2878}
2879
2880//The total number of operations is 4
2881template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1)
2882{
2883        return simd128<128>::srli<(sh*8)>(arg1);
2884}
2885
2886//The total number of operations is 4
2887template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1)
2888{
2889        return simd128<128>::srli<(sh*16)>(arg1);
2890}
2891
2892//The total number of operations is 4
2893template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1)
2894{
2895        return simd128<128>::srli<(sh*32)>(arg1);
2896}
2897
2898//The total number of operations is 4
2899template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1)
2900{
2901        return simd128<128>::srli<(sh*64)>(arg1);
2902}
2903
2904//The total number of operations is 4
2905template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1)
2906{
2907        return simd128<128>::srli<(sh*128)>(arg1);
2908}
2909
2910//The total number of operations is 1
2911template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2)
2912{
2913        return mvmd128<(2)>::fill(((val1<<1)|(val2&(1))));
2914}
2915
2916//The total number of operations is 1
2917template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2)
2918{
2919        return mvmd128<(4)>::fill(((val1<<2)|(val2&(3))));
2920}
2921
2922//The total number of operations is 1
2923template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2)
2924{
2925        return mvmd128<(8)>::fill(((val1<<4)|(val2&(15))));
2926}
2927
2928//The total number of operations is 1
2929template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2)
2930{
2931        return mvmd128<(16)>::fill(((val1<<8)|(val2&(255))));
2932}
2933
2934//The total number of operations is 1
2935template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2)
2936{
2937        return mvmd128<(32)>::fill(((val1<<16)|(val2&(65535))));
2938}
2939
2940//The total number of operations is 5
2941template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2)
2942{
2943        return simd128<1>::ifh(simd128<(64)>::himask(), mvmd128<32>::fill(val1), mvmd128<32>::fill(val2));
2944}
2945
2946//The total number of operations is 5
2947template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2)
2948{
2949        return simd128<1>::ifh(simd128<(128)>::himask(), mvmd128<64>::fill(val1), mvmd128<64>::fill(val2));
2950}
2951
2952//The total number of operations is 9
2953template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2954{
2955        return simd_or(mvmd128<2>::slli<sh>(arg1), mvmd128<2>::srli<((64)-sh)>(arg2));
2956}
2957
2958//The total number of operations is 9
2959template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2960{
2961        return simd_or(mvmd128<4>::slli<sh>(arg1), mvmd128<4>::srli<((32)-sh)>(arg2));
2962}
2963
2964//The total number of operations is 9
2965template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2966{
2967        return simd_or(mvmd128<8>::slli<sh>(arg1), mvmd128<8>::srli<((16)-sh)>(arg2));
2968}
2969
2970//The total number of operations is 9
2971template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2972{
2973        return simd_or(mvmd128<16>::slli<sh>(arg1), mvmd128<16>::srli<((8)-sh)>(arg2));
2974}
2975
2976//The total number of operations is 9
2977template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2978{
2979        return simd_or(mvmd128<32>::slli<sh>(arg1), mvmd128<32>::srli<((4)-sh)>(arg2));
2980}
2981
2982//The total number of operations is 9
2983template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2984{
2985        return simd_or(mvmd128<64>::slli<sh>(arg1), mvmd128<64>::srli<((2)-sh)>(arg2));
2986}
2987
2988//The total number of operations is 9
2989template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2990{
2991        return simd_or(mvmd128<128>::slli<sh>(arg1), mvmd128<128>::srli<((1)-sh)>(arg2));
2992}
2993
2994//The total number of operations is 13
2995template <> 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)
2996{
2997        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<1>::fill4(val1, val2, val3, val4), mvmd128<1>::fill4(val5, val6, val7, val8));
2998}
2999
3000//The total number of operations is 13
3001template <> 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)
3002{
3003        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<2>::fill4(val1, val2, val3, val4), mvmd128<2>::fill4(val5, val6, val7, val8));
3004}
3005
3006//The total number of operations is 7
3007template <> 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)
3008{
3009        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))));
3010}
3011
3012//The total number of operations is 3
3013template <> 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)
3014{
3015        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))));
3016}
3017
3018//The total number of operations is 1
3019template <> 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)
3020{
3021        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));
3022}
3023
3024//The total number of operations is 22
3025IDISA_ALWAYS_INLINE uint64_t bitblock128::popcount(bitblock128_t arg1)
3026{
3027        return mvmd128<64>::extract<0>(simd128<128>::popcount(arg1));
3028}
3029
3030//The total number of operations is 2
3031IDISA_ALWAYS_INLINE bool bitblock128::all(bitblock128_t arg1)
3032{
3033        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<-1>())) == 65535;
3034}
3035
3036//The total number of operations is 2
3037IDISA_ALWAYS_INLINE bool bitblock128::any(bitblock128_t arg1)
3038{
3039        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<0>())) != 65535;
3040}
3041
3042#endif
Note: See TracBrowser for help on using the repository browser.