source: trunk/lib/idisa_cpp/idisa_sse4_2.cpp @ 1548

Last change on this file since 1548 was 1548, checked in by cameron, 8 years ago

IDISA fixes

File size: 149.1 KB
Line 
1#ifndef IDISA_SSE4_2_CPP
2#define IDISA_SSE4_2_CPP
3#include <stdint.h>
4#include "../config.hpp"
5
6#include "smmintrin.h"
7
8typedef __m128i bitblock128_t;
9template <uint32_t fw>
10class simd128
11{
12public:
13        static IDISA_ALWAYS_INLINE bitblock128_t max(bitblock128_t arg1, bitblock128_t arg2);
14        static IDISA_ALWAYS_INLINE bitblock128_t mult(bitblock128_t arg1, bitblock128_t arg2);
15        static IDISA_ALWAYS_INLINE bitblock128_t gt(bitblock128_t arg1, bitblock128_t arg2);
16        static IDISA_ALWAYS_INLINE bitblock128_t umult(bitblock128_t arg1, bitblock128_t arg2);
17        static IDISA_ALWAYS_INLINE bitblock128_t ult(bitblock128_t arg1, bitblock128_t arg2);
18        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
19        static IDISA_ALWAYS_INLINE bitblock128_t ctz(bitblock128_t arg1);
20        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t shift_mask);
21        static IDISA_ALWAYS_INLINE bitblock128_t eq(bitblock128_t arg1, bitblock128_t arg2);
22        static IDISA_ALWAYS_INLINE bitblock128_t popcount(bitblock128_t arg1);
23        static IDISA_ALWAYS_INLINE bitblock128_t neg(bitblock128_t arg1);
24        static IDISA_ALWAYS_INLINE bitblock128_t himask();
25        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
26        static IDISA_ALWAYS_INLINE bitblock128_t ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
27        static IDISA_ALWAYS_INLINE bitblock128_t sub(bitblock128_t arg1, bitblock128_t arg2);
28        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1);
29        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t shift_mask);
30        static IDISA_ALWAYS_INLINE bitblock128_t 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 uint64_t signmask(bitblock128_t arg1);
51        static IDISA_ALWAYS_INLINE bitblock128_t packh(bitblock128_t arg1, bitblock128_t arg2);
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};
94
95//Declaration Part
96IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2);
97IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1);
98IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2);
99IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2);
100IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2);
101IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2);
102template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2);
103template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2);
104template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2);
105template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2);
106template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2);
107template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2);
108template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2);
109template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2);
110template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2);
111template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2);
112template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2);
113template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2);
114template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2);
115template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2);
116template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2);
117template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2);
118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2);
119template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2);
120template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2);
121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2);
122template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2);
123template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2);
124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2);
125template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2);
126template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2);
127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2);
128template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2);
129template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2);
130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2);
131template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2);
132template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2);
133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2);
134template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2);
135template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2);
136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2);
137template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2);
138template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2);
139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2);
140template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2);
141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2);
142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2);
143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2);
144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2);
145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2);
146template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2);
147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2);
148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2);
149template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1);
150template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1);
151template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1);
152template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1);
153template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1);
154template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1);
155template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1);
156template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1);
157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1);
158template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1);
159template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1);
160template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1);
161template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1);
162template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1);
163template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1);
164template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
165template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
166template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2);
167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2);
168template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2);
169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2);
170template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2);
171template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2);
172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2);
173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2);
174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1);
175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1);
176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1);
177template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1);
178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1);
179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1);
180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1);
181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1);
182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1);
183template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1);
184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1);
185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1);
186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1);
187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1);
188template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1);
189template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::neg(bitblock128_t arg1);
190template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1);
191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1);
192template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1);
193template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1);
194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1);
195template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1);
196template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1);
197template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1);
198template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1);
199template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1);
200template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1);
201template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1);
202template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1);
203template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1);
204template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
205template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
206template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
207template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
208template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
209template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
210template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
211template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
212template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2);
213template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2);
214template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2);
215template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2);
216template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2);
217template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2);
218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2);
219template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2);
220template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1);
221template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1);
222template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1);
223template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1);
224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1);
225template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1);
226template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1);
227template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
228template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
229template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask();
230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask();
231template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask();
232template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask();
233template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask();
234template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask();
235template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask();
236template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
237template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
238template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
239template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
240template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
241template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
242template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
243template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
244template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2);
245template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2);
246template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2);
247template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2);
248template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2);
249template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2);
250template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2);
251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2);
252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2);
253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2);
254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2);
255template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2);
256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2);
257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2);
258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2);
259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2);
260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2);
261template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2);
262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2);
263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2);
264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2);
265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2);
266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2);
267template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2);
268template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2);
269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2);
270template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2);
271template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2);
272template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2);
273template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2);
274template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2);
275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2);
276template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1);
277template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1);
278template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1);
279template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1);
280template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1);
281template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1);
282template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1);
283template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask();
284template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask();
285template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask();
286template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask();
287template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask();
288template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask();
289template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask();
290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2);
291template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2);
292template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2);
293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2);
294template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2);
295template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2);
296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2);
297template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2);
298template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::abs(bitblock128_t arg1);
299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1);
300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1);
301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1);
302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1);
303template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1);
304template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1);
305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1);
306template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
307template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
308template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
309template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
310template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
311template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
312template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
313template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
314template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
315template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
316template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
317template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
318template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
319template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
320template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2);
321template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2);
322template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2);
323template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2);
324template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2);
325template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2);
326template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2);
327template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1);
328template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1);
329template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1);
330template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1);
331template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1);
332template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1);
333template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2);
334template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2);
335template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2);
336template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2);
337template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2);
338template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2);
339template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2);
340template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2);
341template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2);
342template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2);
343template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2);
344template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2);
345template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2);
346template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2);
347template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
348template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
349template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
350template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
351template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
352template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
353template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
354template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2);
355template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2);
356template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2);
357template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2);
358template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2);
359template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2);
360template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2);
361template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2);
362template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2);
363template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2);
364template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2);
365template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2);
366template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2);
367template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2);
368template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
369template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
370template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
371template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
372template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
373template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
374template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1);
376template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1);
377template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1);
378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1);
379template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1);
380template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1);
381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1);
382template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1);
383template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1);
384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1);
385template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1);
386template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1);
387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1);
388template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1);
389template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1);
390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1);
391template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1);
392template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1);
393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1);
394template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1);
395template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1);
396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1);
397template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1);
398template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1);
399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1);
400template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1);
401template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1);
402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1);
403template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1);
404template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1);
405template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1);
406template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
407template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
408template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
409template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
410template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
411template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
412template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
413template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1);
414template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1);
415template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1);
416template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1);
417template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1);
418template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1);
419template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1);
420template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1);
421template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
422template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
423template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
424template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
425template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1);
426template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1);
427template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1);
428template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1);
429template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1);
430template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1);
431template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1);
432template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1);
433template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1);
434template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1);
435template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1);
436template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1);
437template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1);
438template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1);
439template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1);
440template <> 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);
441template <> 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);
442template <> 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);
443template <> 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);
444template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1);
445template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1);
446template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1);
447template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1);
448template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1);
449template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1);
450template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1);
451template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
452template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
453template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
454template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
455template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
456template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
457template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1);
458template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1);
459template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1);
460template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1);
461template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1);
462template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1);
463template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1);
464template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2);
465template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2);
466template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2);
467template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2);
468template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2);
469template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2);
470template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2);
471template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2);
472template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2);
473template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2);
474template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2);
475template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2);
476template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2);
477template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2);
478template <> 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);
479template <> 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);
480template <> 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);
481template <> 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);
482template <> 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);
483
484//Implementation Part
485
486#define shufflemask4(s1, s2, s3, s4) \
487        ((s1<<6) | (s2<<4) | (s3<<2) | s4)
488
489#define shufflemask4_from_shufflemask2(msk) \
490        (msk==3 ? 238 : (msk==2 ? 228 : (msk==1 ? 78 : 68)))
491
492#define shufflemask8_to_shufflemask4(msk) \
493        ((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))
494//The total number of operations is 2
495IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2)
496{
497        return simd_not(simd_or(arg1, arg2));
498}
499
500//The total number of operations is 1
501IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1)
502{
503        return simd_xor(arg1, simd128<32>::constant<-1>());
504}
505
506//The total number of operations is 1
507IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2)
508{
509        return _mm_andnot_si128(arg2, arg1);
510}
511
512//The total number of operations is 1
513IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2)
514{
515        return _mm_or_si128(arg1, arg2);
516}
517
518//The total number of operations is 1
519IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2)
520{
521        return _mm_and_si128(arg1, arg2);
522}
523
524//The total number of operations is 1
525IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2)
526{
527        return _mm_xor_si128(arg1, arg2);
528}
529
530//The total number of operations is 39
531template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2)
532{
533        bitblock128_t high_bit = simd128<1>::constant<(1)>();
534        return simd_xor(simd128<1>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
535}
536
537//The total number of operations is 19
538template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2)
539{
540        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)))));
541}
542
543//The total number of operations is 9
544template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2)
545{
546        bitblock128_t high_bit = simd128<4>::constant<(8)>();
547        return simd_xor(simd128<4>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
548}
549
550//The total number of operations is 1
551template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2)
552{
553        return _mm_max_epi8(arg1, arg2);
554}
555
556//The total number of operations is 1
557template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2)
558{
559        return _mm_max_epi16(arg1, arg2);
560}
561
562//The total number of operations is 1
563template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2)
564{
565        return _mm_max_epi32(arg1, arg2);
566}
567
568//The total number of operations is 4
569template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2)
570{
571        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg1, arg2);
572}
573
574//The total number of operations is 30
575template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2)
576{
577        bitblock128_t hiAns = simd128<(64)>::max(arg1, arg2);
578        bitblock128_t loAns = simd128<(64)>::umax(arg1, arg2);
579        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
580        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
581        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
582}
583
584//The total number of operations is 57
585template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2)
586{
587        bitblock128_t loMask = simd128<(2)>::lomask();
588        bitblock128_t tmpAns1 = simd128<(2)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
589        bitblock128_t tmpAns2 = simd128<(2)>::mult(simd128<(2)>::srli<1>(arg1), simd128<(2)>::srli<1>(arg2));
590        return simd128<1>::ifh(loMask, tmpAns1, simd128<(2)>::slli<1>(tmpAns2));
591}
592
593//The total number of operations is 23
594template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2)
595{
596        bitblock128_t tmp1 = simd128<128>::slli<1>(arg1);
597        bitblock128_t tmp2 = simd128<128>::slli<1>(arg2);
598        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));
599}
600
601//The total number of operations is 31
602template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2)
603{
604        bitblock128_t loMask = simd128<(8)>::lomask();
605        bitblock128_t tmpAns1 = simd128<(8)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
606        bitblock128_t tmpAns2 = simd128<(8)>::mult(simd128<(8)>::srli<4>(arg1), simd128<(8)>::srli<4>(arg2));
607        return simd128<1>::ifh(loMask, tmpAns1, simd128<(8)>::slli<4>(tmpAns2));
608}
609
610//The total number of operations is 10
611template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2)
612{
613        bitblock128_t loMask = simd128<(16)>::lomask();
614        bitblock128_t tmpAns1 = simd128<(16)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
615        bitblock128_t tmpAns2 = simd128<(16)>::mult(simd128<(16)>::srli<8>(arg1), simd128<(16)>::srli<8>(arg2));
616        return simd128<1>::ifh(loMask, tmpAns1, simd128<(16)>::slli<8>(tmpAns2));
617}
618
619//The total number of operations is 1
620template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2)
621{
622        return _mm_mullo_epi16(arg1, arg2);
623}
624
625//The total number of operations is 1
626template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2)
627{
628        return _mm_mullo_epi32(arg1, arg2);
629}
630
631//The total number of operations is 11
632template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2)
633{
634        bitblock128_t loMask = simd128<64>::lomask();
635        bitblock128_t arg1_low = simd_and(arg1, loMask);
636        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg1);
637        bitblock128_t arg2_low = simd_and(arg2, loMask);
638        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg2);
639        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
640        bitblock128_t tmpAns2 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
641        bitblock128_t tmpAns3 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
642        return simd128<64>::add(tmpAns1, simd128<64>::add(tmpAns2, tmpAns3));
643}
644
645//The total number of operations is 205
646template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2)
647{
648        bitblock128_t loMask = simd128<128>::lomask();
649        bitblock128_t arg1_low = simd_and(arg1, loMask);
650        bitblock128_t arg1_high = simd128<128>::srli<(64)>(arg1);
651        bitblock128_t arg2_low = simd_and(arg2, loMask);
652        bitblock128_t arg2_high = simd128<128>::srli<(64)>(arg2);
653        bitblock128_t tmpAns1 = simd128<(64)>::umult(arg1_low, arg2_low);
654        bitblock128_t tmpAns2 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_low, arg2_high));
655        bitblock128_t tmpAns3 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_high, arg2_low));
656        return simd128<128>::add(tmpAns1, simd128<128>::add(tmpAns2, tmpAns3));
657}
658
659//The total number of operations is 43
660template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2)
661{
662        bitblock128_t high_bit = simd128<1>::constant<(1)>();
663        return simd128<1>::ugt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
664}
665
666//The total number of operations is 18
667template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2)
668{
669        bitblock128_t tmp = simd_not(arg1);
670        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)));
671        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
672}
673
674//The total number of operations is 10
675template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2)
676{
677        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)));
678}
679
680//The total number of operations is 1
681template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2)
682{
683        return _mm_cmpgt_epi8(arg1, arg2);
684}
685
686//The total number of operations is 1
687template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2)
688{
689        return _mm_cmpgt_epi16(arg1, arg2);
690}
691
692//The total number of operations is 1
693template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2)
694{
695        return _mm_cmpgt_epi32(arg1, arg2);
696}
697
698//The total number of operations is 1
699template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2)
700{
701        return _mm_cmpgt_epi64(arg1, arg2);
702}
703
704//The total number of operations is 37
705template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2)
706{
707        bitblock128_t hiAns = simd128<(64)>::gt(arg1, arg2);
708        bitblock128_t loAns = simd128<(64)>::ugt(arg1, arg2);
709        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
710        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
711        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
712}
713
714//The total number of operations is 289
715template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2)
716{
717        bitblock128_t loMask = simd128<(2)>::lomask();
718        bitblock128_t tmpAns1 = simd128<(2)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
719        bitblock128_t tmpAns2 = simd128<(2)>::umult(simd_and(loMask, simd128<(4)>::srli<(2)>(arg1)), simd_and(loMask, simd128<(4)>::srli<(2)>(arg2)));
720        return simd_or(tmpAns1, simd128<(4)>::slli<(2)>(tmpAns2));
721}
722
723//The total number of operations is 139
724template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2)
725{
726        bitblock128_t loMask = simd128<(4)>::lomask();
727        bitblock128_t tmpAns1 = simd128<(4)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
728        bitblock128_t tmpAns2 = simd128<(4)>::umult(simd_and(loMask, simd128<(8)>::srli<(4)>(arg1)), simd_and(loMask, simd128<(8)>::srli<(4)>(arg2)));
729        return simd_or(tmpAns1, simd128<(8)>::slli<(4)>(tmpAns2));
730}
731
732//The total number of operations is 64
733template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2)
734{
735        bitblock128_t loMask = simd128<(8)>::lomask();
736        bitblock128_t tmpAns1 = simd128<(8)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
737        bitblock128_t tmpAns2 = simd128<(8)>::umult(simd_and(loMask, simd128<(16)>::srli<(8)>(arg1)), simd_and(loMask, simd128<(16)>::srli<(8)>(arg2)));
738        return simd_or(tmpAns1, simd128<(16)>::slli<(8)>(tmpAns2));
739}
740
741//The total number of operations is 28
742template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2)
743{
744        bitblock128_t loMask = simd128<(16)>::lomask();
745        bitblock128_t tmpAns1 = simd128<(16)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
746        bitblock128_t tmpAns2 = simd128<(16)>::umult(simd_and(loMask, simd128<(32)>::srli<(16)>(arg1)), simd_and(loMask, simd128<(32)>::srli<(16)>(arg2)));
747        return simd_or(tmpAns1, simd128<(32)>::slli<(16)>(tmpAns2));
748}
749
750//The total number of operations is 10
751template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2)
752{
753        bitblock128_t loMask = simd128<(32)>::lomask();
754        bitblock128_t tmpAns1 = simd128<(32)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
755        bitblock128_t tmpAns2 = simd128<(32)>::umult(simd_and(loMask, simd128<(64)>::srli<(32)>(arg1)), simd_and(loMask, simd128<(64)>::srli<(32)>(arg2)));
756        return simd_or(tmpAns1, simd128<(64)>::slli<(32)>(tmpAns2));
757}
758
759//The total number of operations is 1
760template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2)
761{
762        return _mm_mul_epu32(arg1, arg2);
763}
764
765//The total number of operations is 55
766template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2)
767{
768        bitblock128_t loMask1 = simd128<(128)>::lomask();
769        bitblock128_t arg11 = simd_and(arg1, loMask1);
770        bitblock128_t arg22 = simd_and(arg2, loMask1);
771        bitblock128_t loMask2 = simd128<64>::lomask();
772        bitblock128_t arg1_low = simd_and(arg11, loMask2);
773        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg11);
774        bitblock128_t arg2_low = simd_and(arg22, loMask2);
775        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg22);
776        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
777        bitblock128_t tmpAns2 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
778        bitblock128_t tmpAns3 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
779        bitblock128_t tmpAns4 = simd128<(128)>::slli<64>(simd128<(32)>::umult(arg1_high, arg2_high));
780        return simd128<(128)>::add(tmpAns1, simd128<(128)>::add(tmpAns2, simd128<(128)>::add(tmpAns3, tmpAns4)));
781}
782
783//The total number of operations is 41
784template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2)
785{
786        return simd_or(simd_and(simd128<(2)>::himask(), simd128<(2)>::ult(simd_and(simd128<(2)>::himask(), arg1), simd_and(simd128<(2)>::himask(), arg2))), simd_and(simd128<(2)>::lomask(), simd128<(2)>::ult(simd_and(simd128<(2)>::lomask(), arg1), simd_and(simd128<(2)>::lomask(), arg2))));
787}
788
789//The total number of operations is 17
790template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2)
791{
792        bitblock128_t tmp = simd_not(arg1);
793        bitblock128_t tmpAns = simd_or(simd_and(tmp, arg2), simd_and(simd128<128>::slli<1>(simd_and(tmp, arg2)), simd_or(tmp, arg2)));
794        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
795}
796
797//The total number of operations is 20
798template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2)
799{
800        bitblock128_t high_bit = simd128<4>::constant<(8)>();
801        return simd128<4>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
802}
803
804//The total number of operations is 7
805template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2)
806{
807        bitblock128_t high_bit = simd128<8>::constant<(128)>();
808        return simd128<8>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
809}
810
811//The total number of operations is 7
812template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2)
813{
814        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
815        return simd128<16>::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<32>::ult(bitblock128_t arg1, bitblock128_t arg2)
820{
821        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
822        return simd128<32>::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<64>::ult(bitblock128_t arg1, bitblock128_t arg2)
827{
828        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808UL)>();
829        return simd128<64>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
830}
831
832//The total number of operations is 40
833template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2)
834{
835        bitblock128_t tmpAns = simd128<(64)>::ult(arg1, arg2);
836        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
837        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
838        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
839}
840
841//The total number of operations is 43
842template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2)
843{
844        bitblock128_t high_bit = simd128<1>::constant<(1)>();
845        return simd128<1>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
846}
847
848//The total number of operations is 18
849template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2)
850{
851        bitblock128_t tmp = simd_not(arg2);
852        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)));
853        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
854}
855
856//The total number of operations is 18
857template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2)
858{
859        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)));
860}
861
862//The total number of operations is 5
863template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2)
864{
865        return simd_and(simd_not(simd128<8>::gt(arg1, arg2)), simd_not(simd128<8>::eq(arg1, arg2)));
866}
867
868//The total number of operations is 5
869template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2)
870{
871        return simd_and(simd_not(simd128<16>::gt(arg1, arg2)), simd_not(simd128<16>::eq(arg1, arg2)));
872}
873
874//The total number of operations is 5
875template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2)
876{
877        return simd_and(simd_not(simd128<32>::gt(arg1, arg2)), simd_not(simd128<32>::eq(arg1, arg2)));
878}
879
880//The total number of operations is 5
881template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2)
882{
883        return simd_and(simd_not(simd128<64>::gt(arg1, arg2)), simd_not(simd128<64>::eq(arg1, arg2)));
884}
885
886//The total number of operations is 45
887template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2)
888{
889        bitblock128_t hiAns = simd128<(64)>::lt(arg1, arg2);
890        bitblock128_t loAns = simd128<(64)>::ult(arg1, arg2);
891        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
892        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
893        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
894}
895
896//The total number of operations is 2
897template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1)
898{
899        return simd_and(simd128<32>::srli<sh>(arg1), simd128<2>::constant<((3)>>sh)>());
900}
901
902//The total number of operations is 2
903template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1)
904{
905        return simd_and(simd128<32>::srli<sh>(arg1), simd128<4>::constant<((15)>>sh)>());
906}
907
908//The total number of operations is 2
909template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1)
910{
911        return simd_and(simd128<32>::srli<sh>(arg1), simd128<8>::constant<((255)>>sh)>());
912}
913
914//The total number of operations is 1
915template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1)
916{
917        return _mm_srli_epi16(arg1, (int32_t)(sh));
918}
919
920//The total number of operations is 1
921template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1)
922{
923        return _mm_srli_epi32(arg1, (int32_t)(sh));
924}
925
926//The total number of operations is 1
927template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1)
928{
929        return _mm_srli_epi64(arg1, (int32_t)(sh));
930}
931
932//The total number of operations is 4
933template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1)
934{
935        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)))));
936}
937
938//The total number of operations is 1
939template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1)
940{
941        return simd_not(arg1);
942}
943
944//The total number of operations is 14
945template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1)
946{
947        bitblock128_t tmp = simd_not(arg1);
948        return simd128<1>::ifh(simd128<2>::himask(), simd_and(tmp, simd128<128>::slli<1>(tmp)), simd_and(simd128<128>::srli<1>(arg1), tmp));
949}
950
951//The total number of operations is 14
952template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1)
953{
954        return simd128<4>::popcount(simd_andc(simd128<4>::sub(arg1, simd128<4>::constant<1>()), arg1));
955}
956
957//The total number of operations is 13
958template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1)
959{
960        return simd128<8>::popcount(simd_andc(simd128<8>::sub(arg1, simd128<8>::constant<1>()), arg1));
961}
962
963//The total number of operations is 16
964template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1)
965{
966        return simd128<16>::popcount(simd_andc(simd128<16>::sub(arg1, simd128<16>::constant<1>()), arg1));
967}
968
969//The total number of operations is 19
970template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1)
971{
972        return simd128<32>::popcount(simd_andc(simd128<32>::sub(arg1, simd128<32>::constant<1>()), arg1));
973}
974
975//The total number of operations is 14
976template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1)
977{
978        return simd128<64>::popcount(simd_andc(simd128<64>::sub(arg1, simd128<64>::constant<1>()), arg1));
979}
980
981//The total number of operations is 30
982template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1)
983{
984        return simd128<128>::popcount(simd_andc(simd128<128>::sub(arg1, simd128<128>::constant<1>()), arg1));
985}
986
987//The total number of operations is 7
988template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
989{
990        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())));
991}
992
993//The total number of operations is 11
994template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
995{
996        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))));
997}
998
999//The total number of operations is 41
1000template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1001{
1002        return simd_or(simd_and(simd128<(2)>::himask(), simd128<(2)>::ugt(simd_and(simd128<(2)>::himask(), arg1), simd_and(simd128<(2)>::himask(), arg2))), simd_and(simd128<(2)>::lomask(), simd128<(2)>::ugt(simd_and(simd128<(2)>::lomask(), arg1), simd_and(simd128<(2)>::lomask(), arg2))));
1003}
1004
1005//The total number of operations is 17
1006template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1007{
1008        bitblock128_t tmp = simd_not(arg2);
1009        bitblock128_t tmpAns = simd_or(simd_and(arg1, tmp), simd_and(simd128<128>::slli<1>(simd_and(arg1, tmp)), simd_or(arg1, tmp)));
1010        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1011}
1012
1013//The total number of operations is 12
1014template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1015{
1016        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1017        return simd128<4>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1018}
1019
1020//The total number of operations is 3
1021template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1022{
1023        bitblock128_t high_bit = simd128<8>::constant<(128)>();
1024        return simd128<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1025}
1026
1027//The total number of operations is 3
1028template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1029{
1030        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1031        return simd128<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1032}
1033
1034//The total number of operations is 3
1035template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1036{
1037        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1038        return simd128<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1039}
1040
1041//The total number of operations is 3
1042template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1043{
1044        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808UL)>();
1045        return simd128<64>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1046}
1047
1048//The total number of operations is 36
1049template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1050{
1051        bitblock128_t tmpAns = simd128<(64)>::ugt(arg1, arg2);
1052        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
1053        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
1054        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
1055}
1056
1057//The total number of operations is 4
1058template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1)
1059{
1060        return simd_xor(simd128<2>::srli<(1)>(arg1), simd_and(arg1, simd128<2>::lomask()));
1061}
1062
1063//The total number of operations is 4
1064template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1)
1065{
1066        return simd_xor(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1067}
1068
1069//The total number of operations is 4
1070template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1)
1071{
1072        return simd_xor(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1073}
1074
1075//The total number of operations is 3
1076template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1)
1077{
1078        return simd_xor(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1079}
1080
1081//The total number of operations is 3
1082template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1)
1083{
1084        return simd_xor(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1085}
1086
1087//The total number of operations is 3
1088template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1)
1089{
1090        return simd_xor(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1091}
1092
1093//The total number of operations is 6
1094template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1)
1095{
1096        return simd_xor(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1097}
1098
1099//The total number of operations is 0
1100template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1)
1101{
1102        return arg1;
1103}
1104
1105//The total number of operations is 3
1106template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1)
1107{
1108        return simd128<2>::add_hl(simd128<(1)>::popcount(arg1));
1109}
1110
1111//The total number of operations is 7
1112template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1)
1113{
1114        return simd128<4>::add_hl(simd128<(2)>::popcount(arg1));
1115}
1116
1117//The total number of operations is 11
1118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1)
1119{
1120        return simd128<8>::add_hl(simd128<(4)>::popcount(arg1));
1121}
1122
1123//The total number of operations is 14
1124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1)
1125{
1126        return simd128<16>::add_hl(simd128<(8)>::popcount(arg1));
1127}
1128
1129//The total number of operations is 17
1130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1)
1131{
1132        return simd128<32>::add_hl(simd128<(16)>::popcount(arg1));
1133}
1134
1135//The total number of operations is 12
1136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1)
1137{
1138        return _mm_sad_epu8(simd128<8>::popcount(arg1), simd128<8>::constant<0>());
1139}
1140
1141//The total number of operations is 18
1142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1)
1143{
1144        bitblock128_t tmpAns = simd128<(64)>::popcount(arg1);
1145        return simd128<(64)>::add(simd_and(tmpAns, simd128<128>::lomask()), simd128<128>::srli<(64)>(tmpAns));
1146}
1147
1148//The total number of operations is 26
1149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::neg(bitblock128_t arg1)
1150{
1151        return simd128<1>::sub(simd128<1>::constant<0>(), arg1);
1152}
1153
1154//The total number of operations is 8
1155template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1)
1156{
1157        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(arg1, simd128<128>::slli<1>(arg1)), arg1);
1158}
1159
1160//The total number of operations is 6
1161template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1)
1162{
1163        return simd128<4>::sub(simd128<4>::constant<0>(), arg1);
1164}
1165
1166//The total number of operations is 1
1167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1)
1168{
1169        return simd128<8>::sub(simd128<8>::constant<0>(), arg1);
1170}
1171
1172//The total number of operations is 1
1173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1)
1174{
1175        return simd128<16>::sub(simd128<16>::constant<0>(), arg1);
1176}
1177
1178//The total number of operations is 1
1179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1)
1180{
1181        return _mm_sign_epi32(arg1, simd128<32>::constant<-1>());
1182}
1183
1184//The total number of operations is 1
1185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1)
1186{
1187        return simd128<64>::sub(simd128<64>::constant<0>(), arg1);
1188}
1189
1190//The total number of operations is 11
1191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1)
1192{
1193        return simd128<128>::sub(simd128<128>::constant<0>(), arg1);
1194}
1195
1196//The total number of operations is 2
1197template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1)
1198{
1199        return simd_and(simd128<32>::slli<sh>(arg1), simd128<2>::constant<(((3)<<sh)&(3))>());
1200}
1201
1202//The total number of operations is 2
1203template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1)
1204{
1205        return simd_and(simd128<32>::slli<sh>(arg1), simd128<4>::constant<(((15)<<sh)&(15))>());
1206}
1207
1208//The total number of operations is 2
1209template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1)
1210{
1211        return simd_and(simd128<32>::slli<sh>(arg1), simd128<8>::constant<(((255)<<sh)&(255))>());
1212}
1213
1214//The total number of operations is 1
1215template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1)
1216{
1217        return _mm_slli_epi16(arg1, (int32_t)(sh));
1218}
1219
1220//The total number of operations is 1
1221template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1)
1222{
1223        return _mm_slli_epi32(arg1, (int32_t)(sh));
1224}
1225
1226//The total number of operations is 1
1227template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1)
1228{
1229        return _mm_slli_epi64(arg1, (int32_t)(sh));
1230}
1231
1232//The total number of operations is 4
1233template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1)
1234{
1235        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)))));
1236}
1237
1238//The total number of operations is 3
1239template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1240{
1241        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
1242}
1243
1244//The total number of operations is 8
1245template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1246{
1247        return simd128<(1)>::ifh(simd128<1>::ifh(simd128<2>::himask(), arg1, simd128<2>::srli<(1)>(arg1)), arg2, arg3);
1248}
1249
1250//The total number of operations is 13
1251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1252{
1253        return simd128<1>::ifh(simd128<4>::gt(simd128<4>::constant<0>(), arg1), arg2, arg3);
1254}
1255
1256//The total number of operations is 1
1257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1258{
1259        return _mm_blendv_epi8(arg3, arg2, arg1);
1260}
1261
1262//The total number of operations is 4
1263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1264{
1265        return simd128<1>::ifh(simd128<16>::gt(simd128<16>::constant<0>(), arg1), arg2, arg3);
1266}
1267
1268//The total number of operations is 4
1269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1270{
1271        return simd128<1>::ifh(simd128<32>::gt(simd128<32>::constant<0>(), arg1), arg2, arg3);
1272}
1273
1274//The total number of operations is 4
1275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1276{
1277        return simd128<1>::ifh(simd128<64>::gt(simd128<64>::constant<0>(), arg1), arg2, arg3);
1278}
1279
1280//The total number of operations is 11
1281template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1282{
1283        return simd128<(64)>::ifh(simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg1)), arg2, arg3);
1284}
1285
1286//The total number of operations is 26
1287template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2)
1288{
1289        return simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::sub(arg1, simd_and(simd128<(2)>::himask(), arg2)), simd128<(2)>::sub(arg1, arg2));
1290}
1291
1292//The total number of operations is 11
1293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2)
1294{
1295        bitblock128_t tmp = simd_xor(arg1, arg2);
1296        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2))), tmp);
1297}
1298
1299//The total number of operations is 6
1300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2)
1301{
1302        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::sub(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::sub(arg1, arg2));
1303}
1304
1305//The total number of operations is 1
1306template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2)
1307{
1308        return _mm_sub_epi8(arg1, arg2);
1309}
1310
1311//The total number of operations is 1
1312template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2)
1313{
1314        return _mm_sub_epi16(arg1, arg2);
1315}
1316
1317//The total number of operations is 1
1318template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2)
1319{
1320        return _mm_sub_epi32(arg1, arg2);
1321}
1322
1323//The total number of operations is 1
1324template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2)
1325{
1326        return _mm_sub_epi64(arg1, arg2);
1327}
1328
1329//The total number of operations is 11
1330template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2)
1331{
1332        bitblock128_t partial = simd128<(64)>::sub(arg1, arg2);
1333        bitblock128_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
1334        bitblock128_t borrow = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(borrowMask));
1335        return simd128<(64)>::sub(partial, borrow);
1336}
1337
1338//The total number of operations is 3
1339template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1)
1340{
1341        return simd128<16>::sub(arg1, simd_and(simd128<2>::lomask(), simd128<16>::srli<1>(arg1)));
1342}
1343
1344//The total number of operations is 4
1345template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1)
1346{
1347        return simd128<(8)>::add(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1348}
1349
1350//The total number of operations is 4
1351template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1)
1352{
1353        return simd128<(16)>::add(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1354}
1355
1356//The total number of operations is 3
1357template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1)
1358{
1359        return simd128<(32)>::add(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1360}
1361
1362//The total number of operations is 3
1363template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1)
1364{
1365        return simd128<(64)>::add(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1366}
1367
1368//The total number of operations is 3
1369template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1)
1370{
1371        return simd128<64>::add(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1372}
1373
1374//The total number of operations is 16
1375template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1)
1376{
1377        return simd128<128>::add(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1378}
1379
1380//The total number of operations is 7
1381template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1382{
1383        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())));
1384}
1385
1386//The total number of operations is 11
1387template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1388{
1389        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))));
1390}
1391
1392//The total number of operations is 0
1393template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask()
1394{
1395        return simd128<2>::constant<(1)>();
1396}
1397
1398//The total number of operations is 0
1399template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask()
1400{
1401        return simd128<4>::constant<(3)>();
1402}
1403
1404//The total number of operations is 0
1405template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask()
1406{
1407        return simd128<8>::constant<(15)>();
1408}
1409
1410//The total number of operations is 0
1411template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask()
1412{
1413        return simd128<16>::constant<(255)>();
1414}
1415
1416//The total number of operations is 0
1417template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask()
1418{
1419        return simd128<32>::constant<(65535)>();
1420}
1421
1422//The total number of operations is 0
1423template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask()
1424{
1425        return _mm_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1));
1426}
1427
1428//The total number of operations is 0
1429template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask()
1430{
1431        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1));
1432}
1433
1434//The total number of operations is 0
1435template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1436{
1437        return simd128<32>::constant<(-1*val)>();
1438}
1439
1440//The total number of operations is 0
1441template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1442{
1443        return ((val < 0) ? simd128<(4)>::constant<((val<<2)|(val^(-4)))>() : simd128<(4)>::constant<((val<<2)|val)>());
1444}
1445
1446//The total number of operations is 0
1447template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1448{
1449        return ((val < 0) ? simd128<(8)>::constant<((val<<4)|(val^(-16)))>() : simd128<(8)>::constant<((val<<4)|val)>());
1450}
1451
1452//The total number of operations is 0
1453template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1454{
1455        return _mm_set1_epi8((int32_t)(val));
1456}
1457
1458//The total number of operations is 0
1459template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1460{
1461        return _mm_set1_epi16((int32_t)(val));
1462}
1463
1464//The total number of operations is 0
1465template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1466{
1467        return _mm_set1_epi32((int32_t)(val));
1468}
1469
1470//The total number of operations is 0
1471template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1472{
1473        return _mm_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val));
1474}
1475
1476//The total number of operations is 0
1477template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1478{
1479        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val));
1480}
1481
1482//The total number of operations is 39
1483template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2)
1484{
1485        bitblock128_t high_bit = simd128<1>::constant<(1)>();
1486        return simd_xor(simd128<1>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1487}
1488
1489//The total number of operations is 19
1490template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2)
1491{
1492        bitblock128_t high_bit = simd128<2>::constant<(2)>();
1493        return simd_xor(simd128<2>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1494}
1495
1496//The total number of operations is 9
1497template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2)
1498{
1499        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1500        return simd_xor(simd128<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1501}
1502
1503//The total number of operations is 1
1504template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2)
1505{
1506        return _mm_min_epi8(arg1, arg2);
1507}
1508
1509//The total number of operations is 1
1510template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2)
1511{
1512        return _mm_min_epi16(arg1, arg2);
1513}
1514
1515//The total number of operations is 1
1516template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2)
1517{
1518        return _mm_min_epi32(arg1, arg2);
1519}
1520
1521//The total number of operations is 4
1522template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2)
1523{
1524        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg2, arg1);
1525}
1526
1527//The total number of operations is 30
1528template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2)
1529{
1530        bitblock128_t hiAns = simd128<(64)>::min(arg1, arg2);
1531        bitblock128_t loAns = simd128<(64)>::umin(arg1, arg2);
1532        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
1533        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
1534        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
1535}
1536
1537//The total number of operations is 36
1538template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2)
1539{
1540        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)));
1541}
1542
1543//The total number of operations is 16
1544template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2)
1545{
1546        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)));
1547}
1548
1549//The total number of operations is 6
1550template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2)
1551{
1552        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)));
1553}
1554
1555//The total number of operations is 1
1556template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2)
1557{
1558        return _mm_min_epu8(arg1, arg2);
1559}
1560
1561//The total number of operations is 1
1562template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2)
1563{
1564        return _mm_min_epu16(arg1, arg2);
1565}
1566
1567//The total number of operations is 1
1568template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2)
1569{
1570        return _mm_min_epu32(arg1, arg2);
1571}
1572
1573//The total number of operations is 7
1574template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2)
1575{
1576        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808UL)>();
1577        return simd_xor(simd128<64>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1578}
1579
1580//The total number of operations is 26
1581template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2)
1582{
1583        bitblock128_t tmpAns = simd128<(64)>::umin(arg1, arg2);
1584        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1585        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1586        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1587}
1588
1589//The total number of operations is 36
1590template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2)
1591{
1592        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)));
1593}
1594
1595//The total number of operations is 16
1596template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2)
1597{
1598        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)));
1599}
1600
1601//The total number of operations is 6
1602template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2)
1603{
1604        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)));
1605}
1606
1607//The total number of operations is 1
1608template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2)
1609{
1610        return _mm_max_epu8(arg1, arg2);
1611}
1612
1613//The total number of operations is 1
1614template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2)
1615{
1616        return _mm_max_epu16(arg1, arg2);
1617}
1618
1619//The total number of operations is 1
1620template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2)
1621{
1622        return _mm_max_epu32(arg1, arg2);
1623}
1624
1625//The total number of operations is 7
1626template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2)
1627{
1628        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808UL)>();
1629        return simd_xor(simd128<64>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1630}
1631
1632//The total number of operations is 26
1633template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2)
1634{
1635        bitblock128_t tmpAns = simd128<(64)>::umax(arg1, arg2);
1636        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1637        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1638        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1639}
1640
1641//The total number of operations is 37
1642template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2)
1643{
1644        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))));
1645}
1646
1647//The total number of operations is 15
1648template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2)
1649{
1650        bitblock128_t tmp = simd_xor(arg1, arg2);
1651        bitblock128_t tmpAns = simd_and(simd_not(simd128<128>::slli<1>(tmp)), simd_not(tmp));
1652        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1653}
1654
1655//The total number of operations is 9
1656template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2)
1657{
1658        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))));
1659}
1660
1661//The total number of operations is 1
1662template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2)
1663{
1664        return _mm_cmpeq_epi8(arg1, arg2);
1665}
1666
1667//The total number of operations is 1
1668template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2)
1669{
1670        return _mm_cmpeq_epi16(arg1, arg2);
1671}
1672
1673//The total number of operations is 1
1674template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2)
1675{
1676        return _mm_cmpeq_epi32(arg1, arg2);
1677}
1678
1679//The total number of operations is 1
1680template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2)
1681{
1682        return _mm_cmpeq_epi64(arg1, arg2);
1683}
1684
1685//The total number of operations is 11
1686template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2)
1687{
1688        bitblock128_t tmpAns = simd128<(64)>::eq(arg1, arg2);
1689        bitblock128_t loMask = simd_and(tmpAns, simd128<128>::srli<(64)>(tmpAns));
1690        bitblock128_t hiMask = simd128<128>::slli<(64)>(loMask);
1691        return simd_or(loMask, hiMask);
1692}
1693
1694//The total number of operations is 4
1695template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1)
1696{
1697        return ((sh == 0) ? arg1 : simd_or(simd_and(simd128<2>::himask(), arg1), simd128<2>::srli<1>(arg1)));
1698}
1699
1700//The total number of operations is 10
1701template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1)
1702{
1703        bitblock128_t tmp = simd128<4>::srli<((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh))>(arg1);
1704        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)));
1705}
1706
1707//The total number of operations is 5
1708template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1)
1709{
1710        bitblock128_t tmp = simd128<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
1711        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)));
1712}
1713
1714//The total number of operations is 1
1715template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1)
1716{
1717        return _mm_srai_epi16(arg1, (int32_t)(sh));
1718}
1719
1720//The total number of operations is 1
1721template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1)
1722{
1723        return _mm_srai_epi32(arg1, (int32_t)(sh));
1724}
1725
1726//The total number of operations is 5
1727template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1)
1728{
1729        bitblock128_t tmp = simd128<64>::srli<((sh >= 64) ? (63) : ((sh < 0) ? 0 : sh))>(arg1);
1730        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)));
1731}
1732
1733//The total number of operations is 21
1734template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1)
1735{
1736        bitblock128_t tmp = simd128<128>::srli<((sh >= 128) ? (127) : ((sh < 0) ? 0 : sh))>(arg1);
1737        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)));
1738}
1739
1740//The total number of operations is 0
1741template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask()
1742{
1743        return simd128<2>::constant<(2)>();
1744}
1745
1746//The total number of operations is 0
1747template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask()
1748{
1749        return simd128<4>::constant<(12)>();
1750}
1751
1752//The total number of operations is 0
1753template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask()
1754{
1755        return simd128<8>::constant<(240)>();
1756}
1757
1758//The total number of operations is 0
1759template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask()
1760{
1761        return simd128<16>::constant<(65280)>();
1762}
1763
1764//The total number of operations is 0
1765template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask()
1766{
1767        return simd128<32>::constant<-65536>();
1768}
1769
1770//The total number of operations is 0
1771template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask()
1772{
1773        return _mm_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0));
1774}
1775
1776//The total number of operations is 0
1777template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask()
1778{
1779        return _mm_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0));
1780}
1781
1782//The total number of operations is 24
1783template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2)
1784{
1785        return simd128<1>::ifh(simd128<(2)>::himask(), simd128<(2)>::add(arg1, simd_and(simd128<(2)>::himask(), arg2)), simd128<(2)>::add(arg1, arg2));
1786}
1787
1788//The total number of operations is 10
1789template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2)
1790{
1791        bitblock128_t tmp = simd_xor(arg1, arg2);
1792        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(arg1, arg2))), tmp);
1793}
1794
1795//The total number of operations is 6
1796template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2)
1797{
1798        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::add(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::add(arg1, arg2));
1799}
1800
1801//The total number of operations is 1
1802template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2)
1803{
1804        return _mm_add_epi8(arg1, arg2);
1805}
1806
1807//The total number of operations is 1
1808template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2)
1809{
1810        return _mm_add_epi16(arg1, arg2);
1811}
1812
1813//The total number of operations is 1
1814template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2)
1815{
1816        return _mm_add_epi32(arg1, arg2);
1817}
1818
1819//The total number of operations is 1
1820template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2)
1821{
1822        return _mm_add_epi64(arg1, arg2);
1823}
1824
1825//The total number of operations is 11
1826template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2)
1827{
1828        bitblock128_t partial = simd128<(64)>::add(arg1, arg2);
1829        bitblock128_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
1830        bitblock128_t carry = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(carryMask));
1831        return simd128<(64)>::add(partial, carry);
1832}
1833
1834//The total number of operations is 72
1835template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::abs(bitblock128_t arg1)
1836{
1837        bitblock128_t gtMask = simd128<1>::gt(arg1, simd128<1>::constant<0>());
1838        return simd128<1>::ifh(gtMask, arg1, simd128<1>::sub(gtMask, arg1));
1839}
1840
1841//The total number of operations is 9
1842template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1)
1843{
1844        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, simd128<128>::slli<1>(simd_not(arg1))), arg1);
1845}
1846
1847//The total number of operations is 19
1848template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1)
1849{
1850        bitblock128_t gtMask = simd128<4>::gt(arg1, simd128<4>::constant<0>());
1851        return simd128<1>::ifh(gtMask, arg1, simd128<4>::sub(gtMask, arg1));
1852}
1853
1854//The total number of operations is 1
1855template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1)
1856{
1857        return _mm_abs_epi8(arg1);
1858}
1859
1860//The total number of operations is 1
1861template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1)
1862{
1863        return _mm_abs_epi16(arg1);
1864}
1865
1866//The total number of operations is 1
1867template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1)
1868{
1869        return _mm_abs_epi32(arg1);
1870}
1871
1872//The total number of operations is 5
1873template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1)
1874{
1875        bitblock128_t gtMask = simd128<64>::gt(arg1, simd128<64>::constant<0>());
1876        return simd128<1>::ifh(gtMask, arg1, simd128<64>::sub(gtMask, arg1));
1877}
1878
1879//The total number of operations is 33
1880template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1)
1881{
1882        bitblock128_t eqMask = simd128<128>::eq(simd128<1>::ifh(simd128<128>::himask(), simd128<(64)>::abs(arg1), arg1), arg1);
1883        return simd128<1>::ifh(eqMask, arg1, simd128<128>::sub(eqMask, arg1));
1884}
1885
1886//The total number of operations is 128
1887template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1888{
1889        return simd128<(1)>::umin(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1890}
1891
1892//The total number of operations is 80
1893template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1894{
1895        return simd128<(2)>::umin(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1896}
1897
1898//The total number of operations is 42
1899template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1900{
1901        return simd128<(4)>::umin(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1902}
1903
1904//The total number of operations is 7
1905template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1906{
1907        return simd128<(8)>::umin(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1908}
1909
1910//The total number of operations is 7
1911template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1912{
1913        return simd128<(16)>::umin(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1914}
1915
1916//The total number of operations is 7
1917template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1918{
1919        return simd128<(32)>::umin(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1920}
1921
1922//The total number of operations is 21
1923template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1924{
1925        return simd128<(64)>::umin(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1926}
1927
1928//The total number of operations is 116
1929template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1930{
1931        return simd128<(1)>::add(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1932}
1933
1934//The total number of operations is 74
1935template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1936{
1937        return simd128<(2)>::add(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1938}
1939
1940//The total number of operations is 42
1941template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1942{
1943        return simd128<(4)>::add(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1944}
1945
1946//The total number of operations is 7
1947template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1948{
1949        return simd128<(8)>::add(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1950}
1951
1952//The total number of operations is 1
1953template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1954{
1955        return _mm_hadd_epi16(arg2, arg1);
1956}
1957
1958//The total number of operations is 1
1959template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1960{
1961        return _mm_hadd_epi32(arg2, arg1);
1962}
1963
1964//The total number of operations is 15
1965template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1966{
1967        return simd128<(64)>::add(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1968}
1969
1970//The total number of operations is 132
1971template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2)
1972{
1973        bitblock128_t hiBound = simd128<2>::srli<1>(simd128<2>::lomask());
1974        bitblock128_t loBound = simd_not(hiBound);
1975        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)));
1976}
1977
1978//The total number of operations is 86
1979template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2)
1980{
1981        bitblock128_t hiBound = simd128<4>::srli<1>(simd128<4>::lomask());
1982        bitblock128_t loBound = simd_not(hiBound);
1983        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)));
1984}
1985
1986//The total number of operations is 36
1987template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2)
1988{
1989        bitblock128_t hiBound = simd128<8>::srli<1>(simd128<8>::lomask());
1990        bitblock128_t loBound = simd_not(hiBound);
1991        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)));
1992}
1993
1994//The total number of operations is 1
1995template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2)
1996{
1997        return _mm_packs_epi16(arg2, arg1);
1998}
1999
2000//The total number of operations is 1
2001template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2)
2002{
2003        return _mm_packs_epi32(arg2, arg1);
2004}
2005
2006//The total number of operations is 21
2007template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2)
2008{
2009        bitblock128_t hiBound = simd128<64>::srli<1>(simd128<64>::lomask());
2010        bitblock128_t loBound = simd_not(hiBound);
2011        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)));
2012}
2013
2014//The total number of operations is 172
2015template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2)
2016{
2017        bitblock128_t hiBound = simd128<128>::srli<1>(simd128<128>::lomask());
2018        bitblock128_t loBound = simd_not(hiBound);
2019        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)));
2020}
2021
2022//The total number of operations is 24
2023template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1)
2024{
2025        uint64_t tmpAns1 = hsimd128<(8)>::signmask(esimd128<4>::mergeh(arg1, simd128<4>::constant<0>()));
2026        uint64_t tmpAns2 = hsimd128<(8)>::signmask(esimd128<4>::mergel(arg1, simd128<4>::constant<0>()));
2027        return ((tmpAns1<<(16))+tmpAns2);
2028}
2029
2030//The total number of operations is 1
2031template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1)
2032{
2033        return _mm_movemask_epi8(arg1);
2034}
2035
2036//The total number of operations is 4
2037template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1)
2038{
2039        return hsimd128<(8)>::signmask(hsimd128<16>::packh(simd128<16>::constant<0>(), arg1));
2040}
2041
2042//The total number of operations is 7
2043template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1)
2044{
2045        return hsimd128<(16)>::signmask(hsimd128<32>::packh(simd128<32>::constant<0>(), arg1));
2046}
2047
2048//The total number of operations is 10
2049template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1)
2050{
2051        return hsimd128<(32)>::signmask(hsimd128<64>::packh(simd128<64>::constant<0>(), arg1));
2052}
2053
2054//The total number of operations is 17
2055template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1)
2056{
2057        return hsimd128<(64)>::signmask(hsimd128<128>::packh(simd128<128>::constant<0>(), arg1));
2058}
2059
2060//The total number of operations is 45
2061template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2)
2062{
2063        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));
2064}
2065
2066//The total number of operations is 31
2067template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2)
2068{
2069        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));
2070}
2071
2072//The total number of operations is 17
2073template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2)
2074{
2075        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));
2076}
2077
2078//The total number of operations is 3
2079template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2)
2080{
2081        return hsimd128<16>::packus(simd_and(arg1, simd128<16>::lomask()), simd_and(arg2, simd128<16>::lomask()));
2082}
2083
2084//The total number of operations is 3
2085template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2)
2086{
2087        return hsimd128<32>::packus(simd_and(arg1, simd128<32>::lomask()), simd_and(arg2, simd128<32>::lomask()));
2088}
2089
2090//The total number of operations is 3
2091template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2)
2092{
2093        return _mm_hsub_epi32(simd_and(arg2, simd128<64>::lomask()), simd_and(arg1, simd128<64>::lomask()));
2094}
2095
2096//The total number of operations is 7
2097template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2)
2098{
2099        return simd128<1>::ifh(simd128<128>::himask(), simd128<128>::slli<(64)>(arg1), arg2);
2100}
2101
2102//The total number of operations is 47
2103template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2)
2104{
2105        return hsimd128<2>::packl(simd128<64>::srli<(1)>(arg1), simd128<64>::srli<(1)>(arg2));
2106}
2107
2108//The total number of operations is 33
2109template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2)
2110{
2111        return hsimd128<4>::packl(simd128<64>::srli<(2)>(arg1), simd128<64>::srli<(2)>(arg2));
2112}
2113
2114//The total number of operations is 19
2115template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2)
2116{
2117        return hsimd128<8>::packl(simd128<64>::srli<(4)>(arg1), simd128<64>::srli<(4)>(arg2));
2118}
2119
2120//The total number of operations is 3
2121template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2)
2122{
2123        return hsimd128<16>::packus(simd128<16>::srli<(8)>(arg1), simd128<16>::srli<(8)>(arg2));
2124}
2125
2126//The total number of operations is 3
2127template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2)
2128{
2129        return _mm_hsub_epi16(simd128<32>::srli<(16)>(arg2), simd128<32>::srli<(16)>(arg1));
2130}
2131
2132//The total number of operations is 3
2133template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2)
2134{
2135        return _mm_hsub_epi32(simd128<64>::srli<(32)>(arg2), simd128<64>::srli<(32)>(arg1));
2136}
2137
2138//The total number of operations is 7
2139template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2)
2140{
2141        return simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg2));
2142}
2143
2144//The total number of operations is 131
2145template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2146{
2147        return simd128<(1)>::min(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
2148}
2149
2150//The total number of operations is 83
2151template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2152{
2153        return simd128<(2)>::min(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
2154}
2155
2156//The total number of operations is 45
2157template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2158{
2159        return simd128<(4)>::min(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
2160}
2161
2162//The total number of operations is 7
2163template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2164{
2165        return simd128<(8)>::min(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2166}
2167
2168//The total number of operations is 7
2169template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2170{
2171        return simd128<(16)>::min(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2172}
2173
2174//The total number of operations is 7
2175template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2176{
2177        return simd128<(32)>::min(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2178}
2179
2180//The total number of operations is 18
2181template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2182{
2183        return simd128<(64)>::min(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2184}
2185
2186//The total number of operations is 99
2187template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2)
2188{
2189        bitblock128_t arg11 = simd128<2>::ifh(arg1, simd128<2>::constant<0>(), arg1);
2190        bitblock128_t arg12 = simd_and(simd128<2>::lomask(), arg11);
2191        bitblock128_t arg21 = simd128<2>::ifh(arg2, simd128<2>::constant<0>(), arg2);
2192        bitblock128_t arg22 = simd_and(simd128<2>::lomask(), arg21);
2193        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()));
2194}
2195
2196//The total number of operations is 83
2197template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2)
2198{
2199        bitblock128_t arg11 = simd128<4>::ifh(arg1, simd128<4>::constant<0>(), arg1);
2200        bitblock128_t arg12 = simd_and(simd128<4>::lomask(), arg11);
2201        bitblock128_t arg21 = simd128<4>::ifh(arg2, simd128<4>::constant<0>(), arg2);
2202        bitblock128_t arg22 = simd_and(simd128<4>::lomask(), arg21);
2203        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()));
2204}
2205
2206//The total number of operations is 29
2207template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2)
2208{
2209        bitblock128_t arg11 = simd128<8>::ifh(arg1, simd128<8>::constant<0>(), arg1);
2210        bitblock128_t arg12 = simd_and(simd128<8>::lomask(), arg11);
2211        bitblock128_t arg21 = simd128<8>::ifh(arg2, simd128<8>::constant<0>(), arg2);
2212        bitblock128_t arg22 = simd_and(simd128<8>::lomask(), arg21);
2213        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()));
2214}
2215
2216//The total number of operations is 1
2217template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2)
2218{
2219        return _mm_packus_epi16(arg2, arg1);
2220}
2221
2222//The total number of operations is 1
2223template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2)
2224{
2225        return _mm_packus_epi32(arg2, arg1);
2226}
2227
2228//The total number of operations is 12
2229template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2)
2230{
2231        bitblock128_t hiPart = hsimd128<64>::packh(arg1, arg2);
2232        return simd128<(32)>::ifh(hiPart, simd128<(32)>::constant<0>(), simd_or(simd128<(32)>::gt(hiPart, simd128<(32)>::constant<0>()), hsimd128<64>::packl(arg1, arg2)));
2233}
2234
2235//The total number of operations is 20
2236template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2)
2237{
2238        bitblock128_t hiPart = hsimd128<128>::packh(arg1, arg2);
2239        return simd128<(64)>::ifh(hiPart, simd128<(64)>::constant<0>(), simd_or(simd128<(64)>::gt(hiPart, simd128<(64)>::constant<0>()), hsimd128<128>::packl(arg1, arg2)));
2240}
2241
2242//The total number of operations is 31
2243template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2244{
2245        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));
2246}
2247
2248//The total number of operations is 21
2249template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2250{
2251        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));
2252}
2253
2254//The total number of operations is 11
2255template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2256{
2257        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));
2258}
2259
2260//The total number of operations is 1
2261template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2262{
2263        return _mm_unpacklo_epi8(arg2, arg1);
2264}
2265
2266//The total number of operations is 1
2267template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2268{
2269        return _mm_unpacklo_epi16(arg2, arg1);
2270}
2271
2272//The total number of operations is 1
2273template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2274{
2275        return _mm_unpacklo_epi32(arg2, arg1);
2276}
2277
2278//The total number of operations is 1
2279template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2280{
2281        return _mm_unpacklo_epi64(arg2, arg1);
2282}
2283
2284//The total number of operations is 31
2285template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2286{
2287        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));
2288}
2289
2290//The total number of operations is 21
2291template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2292{
2293        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));
2294}
2295
2296//The total number of operations is 11
2297template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2298{
2299        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));
2300}
2301
2302//The total number of operations is 1
2303template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2304{
2305        return _mm_unpackhi_epi8(arg2, arg1);
2306}
2307
2308//The total number of operations is 1
2309template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2310{
2311        return _mm_unpackhi_epi16(arg2, arg1);
2312}
2313
2314//The total number of operations is 1
2315template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2316{
2317        return _mm_unpackhi_epi32(arg2, arg1);
2318}
2319
2320//The total number of operations is 1
2321template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2322{
2323        return _mm_unpackhi_epi64(arg2, arg1);
2324}
2325
2326//The total number of operations is 24
2327template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1)
2328{
2329        return esimd128<(2)>::mergeh(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2330}
2331
2332//The total number of operations is 14
2333template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1)
2334{
2335        return esimd128<(4)>::mergeh(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2336}
2337
2338//The total number of operations is 4
2339template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1)
2340{
2341        return esimd128<(8)>::mergeh(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2342}
2343
2344//The total number of operations is 3
2345template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1)
2346{
2347        return esimd128<(16)>::mergeh(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2348}
2349
2350//The total number of operations is 3
2351template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1)
2352{
2353        return esimd128<(32)>::mergeh(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2354}
2355
2356//The total number of operations is 3
2357template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1)
2358{
2359        return esimd128<(64)>::mergeh(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2360}
2361
2362//The total number of operations is 4
2363template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1)
2364{
2365        return simd128<(128)>::srli<64>(arg1);
2366}
2367
2368//The total number of operations is 24
2369template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1)
2370{
2371        return esimd128<(2)>::mergel(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2372}
2373
2374//The total number of operations is 14
2375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1)
2376{
2377        return esimd128<(4)>::mergel(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2378}
2379
2380//The total number of operations is 4
2381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1)
2382{
2383        return esimd128<(8)>::mergel(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2384}
2385
2386//The total number of operations is 1
2387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1)
2388{
2389        return _mm_cvtepu8_epi16(arg1);
2390}
2391
2392//The total number of operations is 1
2393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1)
2394{
2395        return _mm_cvtepu16_epi32(arg1);
2396}
2397
2398//The total number of operations is 1
2399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1)
2400{
2401        return _mm_cvtepu32_epi64(arg1);
2402}
2403
2404//The total number of operations is 1
2405template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1)
2406{
2407        return simd_and(simd128<(128)>::lomask(), arg1);
2408}
2409
2410//The total number of operations is 31
2411template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1)
2412{
2413        return esimd128<(2)>::mergeh(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2414}
2415
2416//The total number of operations is 33
2417template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1)
2418{
2419        return esimd128<(4)>::mergeh(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2420}
2421
2422//The total number of operations is 13
2423template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1)
2424{
2425        return esimd128<(8)>::mergeh(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2426}
2427
2428//The total number of operations is 4
2429template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1)
2430{
2431        return esimd128<(16)>::mergeh(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2432}
2433
2434//The total number of operations is 4
2435template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1)
2436{
2437        return esimd128<(32)>::mergeh(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2438}
2439
2440//The total number of operations is 5
2441template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1)
2442{
2443        return esimd128<32>::signextendl(simd128<128>::srli<(64)>(arg1));
2444}
2445
2446//The total number of operations is 21
2447template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1)
2448{
2449        return simd128<(128)>::srai<64>(arg1);
2450}
2451
2452//The total number of operations is 31
2453template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1)
2454{
2455        return esimd128<(2)>::mergel(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2456}
2457
2458//The total number of operations is 33
2459template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1)
2460{
2461        return esimd128<(4)>::mergel(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2462}
2463
2464//The total number of operations is 13
2465template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1)
2466{
2467        return esimd128<(8)>::mergel(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2468}
2469
2470//The total number of operations is 1
2471template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1)
2472{
2473        return _mm_cvtepi8_epi16(arg1);
2474}
2475
2476//The total number of operations is 1
2477template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1)
2478{
2479        return _mm_cvtepi16_epi32(arg1);
2480}
2481
2482//The total number of operations is 1
2483template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1)
2484{
2485        return _mm_cvtepi32_epi64(arg1);
2486}
2487
2488//The total number of operations is 25
2489template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1)
2490{
2491        return simd128<(128)>::srai<64>(simd128<(128)>::slli<64>(arg1));
2492}
2493
2494//The total number of operations is 17
2495template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1)
2496{
2497        bitblock128_t tmphi = _mm_shufflehi_epi16(arg1, (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2498        bitblock128_t tmpAns = _mm_shufflelo_epi16(tmphi, (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2499        bitblock128_t tmplh = _mm_shufflehi_epi16(simd128<128>::slli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2500        bitblock128_t tmphl = _mm_shufflelo_epi16(simd128<128>::srli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2501        uint32_t a1 = ((((msk>>21)&4) == 0) ? 0 : (131071));
2502        uint32_t a2 = ((((msk>>18)&4) == 0) ? 0 : (131071));
2503        uint32_t a3 = ((((msk>>15)&4) == 0) ? 0 : (131071));
2504        uint32_t a4 = ((((msk>>12)&4) == 0) ? 0 : (131071));
2505        uint32_t a5 = ((((msk>>9)&4) == 0) ? (131071) : 0);
2506        uint32_t a6 = ((((msk>>6)&4) == 0) ? (131071) : 0);
2507        uint32_t a7 = ((((msk>>3)&4) == 0) ? (131071) : 0);
2508        uint32_t a8 = (((msk&4) == 0) ? (131071) : 0);
2509        return simd128<1>::ifh(mvmd128<16>::fill8(a1, a2, a3, a4, a5, a6, a7, a8), tmpAns, simd_or(tmplh, tmphl));
2510}
2511
2512//The total number of operations is 1
2513template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1)
2514{
2515        return _mm_shuffle_epi32(arg1, (int32_t)(msk));
2516}
2517
2518//The total number of operations is 1
2519template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1)
2520{
2521        return mvmd128<32>::shufflei<shufflemask4_from_shufflemask2(msk)>(arg1);
2522}
2523
2524//The total number of operations is 9
2525template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2526{
2527        return simd_or(mvmd128<2>::srli<sh>(arg1), mvmd128<2>::slli<((64)-sh)>(arg2));
2528}
2529
2530//The total number of operations is 9
2531template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2532{
2533        return simd_or(mvmd128<4>::srli<sh>(arg1), mvmd128<4>::slli<((32)-sh)>(arg2));
2534}
2535
2536//The total number of operations is 9
2537template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2538{
2539        return simd_or(mvmd128<8>::srli<sh>(arg1), mvmd128<8>::slli<((16)-sh)>(arg2));
2540}
2541
2542//The total number of operations is 9
2543template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2544{
2545        return simd_or(mvmd128<16>::srli<sh>(arg1), mvmd128<16>::slli<((8)-sh)>(arg2));
2546}
2547
2548//The total number of operations is 9
2549template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2550{
2551        return simd_or(mvmd128<32>::srli<sh>(arg1), mvmd128<32>::slli<((4)-sh)>(arg2));
2552}
2553
2554//The total number of operations is 9
2555template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2556{
2557        return simd_or(mvmd128<64>::srli<sh>(arg1), mvmd128<64>::slli<((2)-sh)>(arg2));
2558}
2559
2560//The total number of operations is 9
2561template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2562{
2563        return simd_or(mvmd128<128>::srli<sh>(arg1), mvmd128<128>::slli<((1)-sh)>(arg2));
2564}
2565
2566//The total number of operations is 1
2567template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1)
2568{
2569        return mvmd128<32>::fill((-1*val1));
2570}
2571
2572//The total number of operations is 1
2573template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1)
2574{
2575        return mvmd128<(4)>::fill(((val1<<2)|val1));
2576}
2577
2578//The total number of operations is 1
2579template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1)
2580{
2581        return mvmd128<(8)>::fill(((val1<<4)|val1));
2582}
2583
2584//The total number of operations is 1
2585template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1)
2586{
2587        return _mm_set1_epi8((int32_t)(val1));
2588}
2589
2590//The total number of operations is 1
2591template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1)
2592{
2593        return _mm_set1_epi16((int32_t)(val1));
2594}
2595
2596//The total number of operations is 1
2597template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1)
2598{
2599        return _mm_set1_epi32((int32_t)(val1));
2600}
2601
2602//The total number of operations is 1
2603template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1)
2604{
2605        return _mm_set_epi32((int32_t)((val1>>32)), (int32_t)(val1), (int32_t)((val1>>32)), (int32_t)(val1));
2606}
2607
2608//The total number of operations is 1
2609template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1)
2610{
2611        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val1>>32)), (int32_t)(val1));
2612}
2613
2614//The total number of operations is 1
2615template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2616{
2617        return _mm_shuffle_epi8(arg1, arg2);
2618}
2619
2620//The total number of operations is 10
2621template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2622{
2623        bitblock128_t tmp1 = simd_and(simd128<16>::constant<(7)>(), arg2);
2624        bitblock128_t msk1 = simd128<16>::add(tmp1, tmp1);
2625        bitblock128_t msk2 = simd128<16>::add(msk1, simd128<16>::constant<1>());
2626        bitblock128_t msk = simd_or(msk1, simd128<16>::slli<(8)>(msk2));
2627        return simd128<16>::ifh(arg2, simd128<16>::constant<0>(), mvmd128<(8)>::shuffle(arg1, msk));
2628}
2629
2630//The total number of operations is 19
2631template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2632{
2633        bitblock128_t tmp1 = simd_and(simd128<32>::constant<(3)>(), arg2);
2634        bitblock128_t msk1 = simd128<32>::add(tmp1, tmp1);
2635        bitblock128_t msk2 = simd128<32>::add(msk1, simd128<32>::constant<1>());
2636        bitblock128_t msk = simd_or(msk1, simd128<32>::slli<(16)>(msk2));
2637        return simd128<32>::ifh(arg2, simd128<32>::constant<0>(), mvmd128<(16)>::shuffle(arg1, msk));
2638}
2639
2640//The total number of operations is 28
2641template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2642{
2643        bitblock128_t tmp1 = simd_and(simd128<64>::constant<(1)>(), arg2);
2644        bitblock128_t msk1 = simd128<64>::add(tmp1, tmp1);
2645        bitblock128_t msk2 = simd128<64>::add(msk1, simd128<64>::constant<1>());
2646        bitblock128_t msk = simd_or(msk1, simd128<64>::slli<(32)>(msk2));
2647        return simd128<64>::ifh(arg2, simd128<64>::constant<0>(), mvmd128<(32)>::shuffle(arg1, msk));
2648}
2649
2650//The total number of operations is 1
2651template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1)
2652{
2653        return (((pos%2) == 0) ? (mvmd128<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd128<(2)>::extract<(pos/2)>(arg1)>>1));
2654}
2655
2656//The total number of operations is 1
2657template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1)
2658{
2659        return (((pos%2) == 0) ? (mvmd128<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd128<(4)>::extract<(pos/2)>(arg1)>>2));
2660}
2661
2662//The total number of operations is 1
2663template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1)
2664{
2665        return (((pos%2) == 0) ? (mvmd128<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd128<(8)>::extract<(pos/2)>(arg1)>>4));
2666}
2667
2668//The total number of operations is 1
2669template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1)
2670{
2671        return (((pos%2) == 0) ? (mvmd128<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd128<(16)>::extract<(pos/2)>(arg1)>>8));
2672}
2673
2674//The total number of operations is 1
2675template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1)
2676{
2677        return (65535&_mm_extract_epi16(arg1, (int32_t)(pos)));
2678}
2679
2680//The total number of operations is 1
2681template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1)
2682{
2683        return (((uint64_t)((4294967296UL)-1))&_mm_extract_epi32(arg1, (int32_t)(pos)));
2684}
2685
2686//The total number of operations is 2
2687template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1)
2688{
2689        return ((((uint64_t)mvmd128<(32)>::extract<((2*pos)+1)>(arg1))<<(32))|mvmd128<(32)>::extract<(2*pos)>(arg1));
2690}
2691
2692//The total number of operations is 14
2693template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1)
2694{
2695        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(2)>::slli<1>(arg1) : simd128<(2)>::srli<1>(arg1));
2696        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(2)>::lomask(), arg1) : simd_and(simd128<(2)>::himask(), arg1));
2697        return mvmd128<(2)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2698}
2699
2700//The total number of operations is 10
2701template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1)
2702{
2703        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(4)>::slli<2>(arg1) : simd128<(4)>::srli<2>(arg1));
2704        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(4)>::lomask(), arg1) : simd_and(simd128<(4)>::himask(), arg1));
2705        return mvmd128<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2706}
2707
2708//The total number of operations is 6
2709template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1)
2710{
2711        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(8)>::slli<4>(arg1) : simd128<(8)>::srli<4>(arg1));
2712        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(8)>::lomask(), arg1) : simd_and(simd128<(8)>::himask(), arg1));
2713        return mvmd128<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2714}
2715
2716//The total number of operations is 2
2717template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1)
2718{
2719        return mvmd128<8>::fill(_mm_extract_epi8(arg1, (int32_t)(pos)));
2720}
2721
2722//The total number of operations is 2
2723template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1)
2724{
2725        return mvmd128<16>::fill(_mm_extract_epi16(arg1, (int32_t)(pos)));
2726}
2727
2728//The total number of operations is 1
2729template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1)
2730{
2731        return mvmd128<32>::shufflei<shufflemask4(pos, pos, pos, pos)>(arg1);
2732}
2733
2734//The total number of operations is 5
2735template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1)
2736{
2737        return simd128<1>::ifh(simd128<64>::himask(), mvmd128<(32)>::splat<((2*pos)+1)>(arg1), mvmd128<(32)>::splat<(2*pos)>(arg1));
2738}
2739
2740//The total number of operations is 13
2741template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1)
2742{
2743        return simd128<1>::ifh(simd128<128>::himask(), mvmd128<(64)>::splat<((2*pos)+1)>(arg1), mvmd128<(64)>::splat<(2*pos)>(arg1));
2744}
2745
2746//The total number of operations is 15
2747template <> 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)
2748{
2749        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))));
2750}
2751
2752//The total number of operations is 7
2753template <> 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)
2754{
2755        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))));
2756}
2757
2758//The total number of operations is 3
2759template <> 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)
2760{
2761        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))));
2762}
2763
2764//The total number of operations is 1
2765template <> 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)
2766{
2767        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));
2768}
2769
2770//The total number of operations is 4
2771template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1)
2772{
2773        return simd128<128>::slli<(sh*2)>(arg1);
2774}
2775
2776//The total number of operations is 4
2777template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1)
2778{
2779        return simd128<128>::slli<(sh*4)>(arg1);
2780}
2781
2782//The total number of operations is 4
2783template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1)
2784{
2785        return simd128<128>::slli<(sh*8)>(arg1);
2786}
2787
2788//The total number of operations is 4
2789template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1)
2790{
2791        return simd128<128>::slli<(sh*16)>(arg1);
2792}
2793
2794//The total number of operations is 4
2795template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1)
2796{
2797        return simd128<128>::slli<(sh*32)>(arg1);
2798}
2799
2800//The total number of operations is 4
2801template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1)
2802{
2803        return simd128<128>::slli<(sh*64)>(arg1);
2804}
2805
2806//The total number of operations is 4
2807template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1)
2808{
2809        return simd128<128>::slli<(sh*128)>(arg1);
2810}
2811
2812//The total number of operations is 5
2813template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2814{
2815        return simd128<1>::ifh(simd128<(4)>::himask(), mvmd128<1>::fill2(val1, val2), mvmd128<1>::fill2(val3, val4));
2816}
2817
2818//The total number of operations is 5
2819template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2820{
2821        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<2>::fill2(val1, val2), mvmd128<2>::fill2(val3, val4));
2822}
2823
2824//The total number of operations is 5
2825template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2826{
2827        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<4>::fill2(val1, val2), mvmd128<4>::fill2(val3, val4));
2828}
2829
2830//The total number of operations is 5
2831template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2832{
2833        return simd128<1>::ifh(simd128<(32)>::himask(), mvmd128<8>::fill2(val1, val2), mvmd128<8>::fill2(val3, val4));
2834}
2835
2836//The total number of operations is 3
2837template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2838{
2839        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))));
2840}
2841
2842//The total number of operations is 1
2843template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2844{
2845        return _mm_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
2846}
2847
2848//The total number of operations is 4
2849template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1)
2850{
2851        return simd128<128>::srli<(sh*2)>(arg1);
2852}
2853
2854//The total number of operations is 4
2855template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1)
2856{
2857        return simd128<128>::srli<(sh*4)>(arg1);
2858}
2859
2860//The total number of operations is 4
2861template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1)
2862{
2863        return simd128<128>::srli<(sh*8)>(arg1);
2864}
2865
2866//The total number of operations is 4
2867template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1)
2868{
2869        return simd128<128>::srli<(sh*16)>(arg1);
2870}
2871
2872//The total number of operations is 4
2873template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1)
2874{
2875        return simd128<128>::srli<(sh*32)>(arg1);
2876}
2877
2878//The total number of operations is 4
2879template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1)
2880{
2881        return simd128<128>::srli<(sh*64)>(arg1);
2882}
2883
2884//The total number of operations is 4
2885template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1)
2886{
2887        return simd128<128>::srli<(sh*128)>(arg1);
2888}
2889
2890//The total number of operations is 1
2891template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2)
2892{
2893        return mvmd128<(2)>::fill(((val1<<1)|(val2&(1))));
2894}
2895
2896//The total number of operations is 1
2897template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2)
2898{
2899        return mvmd128<(4)>::fill(((val1<<2)|(val2&(3))));
2900}
2901
2902//The total number of operations is 1
2903template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2)
2904{
2905        return mvmd128<(8)>::fill(((val1<<4)|(val2&(15))));
2906}
2907
2908//The total number of operations is 1
2909template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2)
2910{
2911        return mvmd128<(16)>::fill(((val1<<8)|(val2&(255))));
2912}
2913
2914//The total number of operations is 1
2915template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2)
2916{
2917        return mvmd128<(32)>::fill(((val1<<16)|(val2&(65535))));
2918}
2919
2920//The total number of operations is 5
2921template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2)
2922{
2923        return simd128<1>::ifh(simd128<(64)>::himask(), mvmd128<32>::fill(val1), mvmd128<32>::fill(val2));
2924}
2925
2926//The total number of operations is 5
2927template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2)
2928{
2929        return simd128<1>::ifh(simd128<(128)>::himask(), mvmd128<64>::fill(val1), mvmd128<64>::fill(val2));
2930}
2931
2932//The total number of operations is 9
2933template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2934{
2935        return simd_or(mvmd128<2>::slli<sh>(arg1), mvmd128<2>::srli<((64)-sh)>(arg2));
2936}
2937
2938//The total number of operations is 9
2939template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2940{
2941        return simd_or(mvmd128<4>::slli<sh>(arg1), mvmd128<4>::srli<((32)-sh)>(arg2));
2942}
2943
2944//The total number of operations is 9
2945template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2946{
2947        return simd_or(mvmd128<8>::slli<sh>(arg1), mvmd128<8>::srli<((16)-sh)>(arg2));
2948}
2949
2950//The total number of operations is 9
2951template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2952{
2953        return simd_or(mvmd128<16>::slli<sh>(arg1), mvmd128<16>::srli<((8)-sh)>(arg2));
2954}
2955
2956//The total number of operations is 9
2957template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2958{
2959        return simd_or(mvmd128<32>::slli<sh>(arg1), mvmd128<32>::srli<((4)-sh)>(arg2));
2960}
2961
2962//The total number of operations is 9
2963template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2964{
2965        return simd_or(mvmd128<64>::slli<sh>(arg1), mvmd128<64>::srli<((2)-sh)>(arg2));
2966}
2967
2968//The total number of operations is 9
2969template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2970{
2971        return simd_or(mvmd128<128>::slli<sh>(arg1), mvmd128<128>::srli<((1)-sh)>(arg2));
2972}
2973
2974//The total number of operations is 13
2975template <> 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)
2976{
2977        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<1>::fill4(val1, val2, val3, val4), mvmd128<1>::fill4(val5, val6, val7, val8));
2978}
2979
2980//The total number of operations is 13
2981template <> 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)
2982{
2983        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<2>::fill4(val1, val2, val3, val4), mvmd128<2>::fill4(val5, val6, val7, val8));
2984}
2985
2986//The total number of operations is 7
2987template <> 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)
2988{
2989        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))));
2990}
2991
2992//The total number of operations is 3
2993template <> 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)
2994{
2995        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))));
2996}
2997
2998//The total number of operations is 1
2999template <> 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)
3000{
3001        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));
3002}
3003
3004//The total number of operations is 2
3005IDISA_ALWAYS_INLINE bool bitblock128::all(bitblock128_t arg1)
3006{
3007        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<-1>())) == 65535;
3008}
3009
3010//The total number of operations is 2
3011IDISA_ALWAYS_INLINE bool bitblock128::any(bitblock128_t arg1)
3012{
3013        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<0>())) != 65535;
3014}
3015
3016#endif
Note: See TracBrowser for help on using the repository browser.