source: trunk/lib/idisa_cpp/idisa_sse4_1.cpp @ 1584

Last change on this file since 1584 was 1584, checked in by huah, 8 years ago

fixed the bitblock::load/store for AVX; added testing modules for bitblock::srl/sll/srli/slli

File size: 150.4 KB
Line 
1#ifndef IDISA_SSE4_1_CPP
2#define IDISA_SSE4_1_CPP
3#include <stdint.h>
4#include "../config.hpp"
5
6#include "smmintrin.h"
7
8typedef __m128i bitblock128_t;
9template <uint32_t fw>
10class simd128
11{
12public:
13        static IDISA_ALWAYS_INLINE bitblock128_t max(bitblock128_t arg1, bitblock128_t arg2);
14        static IDISA_ALWAYS_INLINE bitblock128_t mult(bitblock128_t arg1, bitblock128_t arg2);
15        static IDISA_ALWAYS_INLINE bitblock128_t gt(bitblock128_t arg1, bitblock128_t arg2);
16        static IDISA_ALWAYS_INLINE bitblock128_t umult(bitblock128_t arg1, bitblock128_t arg2);
17        static IDISA_ALWAYS_INLINE bitblock128_t ult(bitblock128_t arg1, bitblock128_t arg2);
18        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
19        static IDISA_ALWAYS_INLINE bitblock128_t ctz(bitblock128_t arg1);
20        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t shift_mask);
21        static IDISA_ALWAYS_INLINE bitblock128_t eq(bitblock128_t arg1, bitblock128_t arg2);
22        static IDISA_ALWAYS_INLINE bitblock128_t popcount(bitblock128_t arg1);
23        static IDISA_ALWAYS_INLINE bitblock128_t neg(bitblock128_t arg1);
24        static IDISA_ALWAYS_INLINE bitblock128_t himask();
25        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
26        static IDISA_ALWAYS_INLINE bitblock128_t ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
27        static IDISA_ALWAYS_INLINE bitblock128_t sub(bitblock128_t arg1, bitblock128_t arg2);
28        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1);
29        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t shift_mask);
30        static IDISA_ALWAYS_INLINE bitblock128_t lomask();
31        static IDISA_ALWAYS_INLINE bitblock128_t umin(bitblock128_t arg1, bitblock128_t arg2);
32        template <uint64_t val> static IDISA_ALWAYS_INLINE bitblock128_t constant();
33        static IDISA_ALWAYS_INLINE bitblock128_t min(bitblock128_t arg1, bitblock128_t arg2);
34        static IDISA_ALWAYS_INLINE bitblock128_t umax(bitblock128_t arg1, bitblock128_t arg2);
35        static IDISA_ALWAYS_INLINE bitblock128_t abs(bitblock128_t arg1);
36        static IDISA_ALWAYS_INLINE bitblock128_t xor_hl(bitblock128_t arg1);
37        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srai(bitblock128_t arg1);
38        static IDISA_ALWAYS_INLINE bitblock128_t lt(bitblock128_t arg1, bitblock128_t arg2);
39        static IDISA_ALWAYS_INLINE bitblock128_t add(bitblock128_t arg1, bitblock128_t arg2);
40        static IDISA_ALWAYS_INLINE bitblock128_t ugt(bitblock128_t arg1, bitblock128_t arg2);
41};
42
43template <uint32_t fw>
44class hsimd128
45{
46public:
47        static IDISA_ALWAYS_INLINE bitblock128_t umin_hl(bitblock128_t arg1, bitblock128_t arg2);
48        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1, bitblock128_t arg2);
49        static IDISA_ALWAYS_INLINE bitblock128_t packss(bitblock128_t arg1, bitblock128_t arg2);
50        static IDISA_ALWAYS_INLINE bitblock128_t packh(bitblock128_t arg1, bitblock128_t arg2);
51        static IDISA_ALWAYS_INLINE uint64_t signmask(bitblock128_t arg1);
52        static IDISA_ALWAYS_INLINE bitblock128_t packl(bitblock128_t arg1, bitblock128_t arg2);
53        static IDISA_ALWAYS_INLINE bitblock128_t min_hl(bitblock128_t arg1, bitblock128_t arg2);
54        static IDISA_ALWAYS_INLINE bitblock128_t packus(bitblock128_t arg1, bitblock128_t arg2);
55};
56
57template <uint32_t fw>
58class esimd128
59{
60public:
61        static IDISA_ALWAYS_INLINE bitblock128_t mergel(bitblock128_t arg1, bitblock128_t arg2);
62        static IDISA_ALWAYS_INLINE bitblock128_t signextendh(bitblock128_t arg1);
63        static IDISA_ALWAYS_INLINE bitblock128_t mergeh(bitblock128_t arg1, bitblock128_t arg2);
64        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendh(bitblock128_t arg1);
65        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendl(bitblock128_t arg1);
66        static IDISA_ALWAYS_INLINE bitblock128_t signextendl(bitblock128_t arg1);
67};
68
69template <uint32_t fw>
70class mvmd128
71{
72public:
73        template <uint64_t msk> static IDISA_ALWAYS_INLINE bitblock128_t shufflei(bitblock128_t arg1);
74        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dsrli(bitblock128_t arg1, bitblock128_t arg2);
75        static IDISA_ALWAYS_INLINE bitblock128_t fill(uint64_t val1);
76        static IDISA_ALWAYS_INLINE bitblock128_t shuffle(bitblock128_t arg1, bitblock128_t arg2);
77        template <uint64_t pos> static IDISA_ALWAYS_INLINE uint64_t extract(bitblock128_t arg1);
78        template <uint64_t pos> static IDISA_ALWAYS_INLINE bitblock128_t splat(bitblock128_t arg1);
79        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
80        static IDISA_ALWAYS_INLINE bitblock128_t fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
81        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
82        static IDISA_ALWAYS_INLINE bitblock128_t fill2(uint64_t val1, uint64_t val2);
83        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dslli(bitblock128_t arg1, bitblock128_t arg2);
84        static IDISA_ALWAYS_INLINE bitblock128_t fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
85        static IDISA_ALWAYS_INLINE bitblock128_t fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
86};
87
88class bitblock128
89{
90public:
91        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
92        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(float const* arg1);
93        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
94        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
95        static IDISA_ALWAYS_INLINE void store_aligned(float* arg1, bitblock128_t arg2);
96        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
97        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
98        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
99        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
100        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(float const* arg1);
101        static IDISA_ALWAYS_INLINE void store_unaligned(float* arg1, bitblock128_t arg2);
102};
103
104//Declaration Part
105IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2);
106IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1);
107IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2);
108IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2);
109IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2);
110IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2);
111template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2);
112template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2);
113template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2);
114template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2);
115template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2);
116template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2);
117template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2);
118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2);
119template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2);
120template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2);
121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2);
122template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2);
123template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2);
124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2);
125template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2);
126template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2);
127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2);
128template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2);
129template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2);
130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2);
131template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2);
132template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2);
133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2);
134template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2);
135template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2);
136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2);
137template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2);
138template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2);
139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2);
140template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2);
141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2);
142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2);
143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2);
144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2);
145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2);
146template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2);
147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2);
148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2);
149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2);
150template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2);
151template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2);
152template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2);
153template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2);
154template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2);
155template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2);
156template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2);
157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2);
158template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1);
159template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1);
160template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1);
161template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1);
162template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1);
163template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1);
164template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1);
165template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1);
166template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1);
167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1);
168template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1);
169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1);
170template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1);
171template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1);
172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1);
173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2);
176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2);
177template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2);
178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2);
179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2);
180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2);
181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2);
182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2);
183template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1);
184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1);
185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1);
186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1);
187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1);
188template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1);
189template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1);
190template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1);
191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1);
192template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1);
193template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1);
194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1);
195template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1);
196template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1);
197template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1);
198template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1);
199template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1);
200template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1);
201template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1);
202template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1);
203template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1);
204template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1);
205template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1);
206template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1);
207template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1);
208template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1);
209template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1);
210template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1);
211template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1);
212template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
213template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
214template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
215template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
216template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
217template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
219template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
220template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2);
221template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2);
222template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2);
223template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2);
224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2);
225template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2);
226template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2);
227template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2);
228template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1);
229template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1);
230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1);
231template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1);
232template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1);
233template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1);
234template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1);
235template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
236template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
237template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
238template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
239template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
240template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
241template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
242template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
243template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
244template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
245template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2);
246template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2);
247template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2);
248template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2);
249template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2);
250template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2);
251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2);
252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2);
253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask();
254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask();
255template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask();
256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask();
257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask();
258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask();
259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask();
260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2);
261template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2);
262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2);
263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2);
264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2);
265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2);
266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2);
267template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2);
268template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1);
269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1);
270template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1);
271template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1);
272template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1);
273template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1);
274template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1);
275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2);
276template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2);
277template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2);
278template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2);
279template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2);
280template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2);
281template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2);
282template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2);
283template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1);
284template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1);
285template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1);
286template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1);
287template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1);
288template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1);
289template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1);
290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask();
291template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask();
292template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask();
293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask();
294template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask();
295template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask();
296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask();
297template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2);
298template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2);
299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2);
300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2);
301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2);
302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2);
303template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2);
304template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2);
305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2);
306template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2);
307template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2);
308template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2);
309template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2);
310template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2);
311template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2);
312template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2);
313template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
314template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
315template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
316template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
317template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
318template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
319template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
320template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
321template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
322template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
323template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
324template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
325template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
326template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
327template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2);
328template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2);
329template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2);
330template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2);
331template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2);
332template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2);
333template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2);
334template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1);
335template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1);
336template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1);
337template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1);
338template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1);
339template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1);
340template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2);
341template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2);
342template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2);
343template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2);
344template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2);
345template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2);
346template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2);
347template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2);
348template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2);
349template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2);
350template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2);
351template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2);
352template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2);
353template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2);
354template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
355template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
356template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
357template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
358template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
359template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
360template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
361template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2);
362template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2);
363template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2);
364template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2);
365template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2);
366template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2);
367template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2);
368template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2);
369template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2);
370template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2);
371template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2);
372template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2);
373template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2);
374template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2);
375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
376template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
377template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
379template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
380template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
382template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1);
383template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1);
384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1);
385template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1);
386template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1);
387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1);
388template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1);
389template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1);
390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1);
391template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1);
392template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1);
393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1);
394template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1);
395template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1);
396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1);
397template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1);
398template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1);
399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1);
400template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1);
401template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1);
402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1);
403template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1);
404template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1);
405template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1);
406template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1);
407template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1);
408template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1);
409template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1);
410template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1);
411template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1);
412template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1);
413template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
414template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
415template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
416template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
417template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
418template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
419template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
420template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1);
421template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1);
422template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1);
423template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1);
424template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1);
425template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1);
426template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1);
427template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1);
428template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
429template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
430template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
431template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2);
432template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1);
433template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1);
434template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1);
435template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1);
436template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1);
437template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1);
438template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1);
439template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1);
440template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1);
441template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1);
442template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1);
443template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1);
444template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1);
445template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1);
446template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1);
447template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
448template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
449template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
450template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
451template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
452template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
453template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
454template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
455template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
456template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
457template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1);
458template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1);
459template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1);
460template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1);
461template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1);
462template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1);
463template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1);
464template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2);
465template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2);
466template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2);
467template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2);
468template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2);
469template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2);
470template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2);
471template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2);
472template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2);
473template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2);
474template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2);
475template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2);
476template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2);
477template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2);
478template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1);
479template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1);
480template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1);
481template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1);
482template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1);
483template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1);
484template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1);
485template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
486template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
487template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
488template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
489template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
490
491//Implementation Part
492
493#define shufflemask4(s1, s2, s3, s4) \
494        ((s1<<6) | (s2<<4) | (s3<<2) | s4)
495
496#define shufflemask4_from_shufflemask2(msk) \
497        (msk==3 ? 238 : (msk==2 ? 228 : (msk==1 ? 78 : 68)))
498
499#define shufflemask8_to_shufflemask4(msk) \
500        ((msk&3) | (((msk>>3)&3)<<2) | (((msk>>6)&3)<<4) | (((msk>>9)&3)<<6) | (((msk>>12)&3)<<8) | (((msk>>15)&3)<<10) | (((msk>>18)&3)<<12) | (((msk>>21)&3)<<14))
501//The total number of operations is 2
502IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2)
503{
504        return simd_not(simd_or(arg1, arg2));
505}
506
507//The total number of operations is 1
508IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1)
509{
510        return simd_xor(arg1, simd128<32>::constant<-1>());
511}
512
513//The total number of operations is 1
514IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2)
515{
516        return _mm_andnot_si128(arg2, arg1);
517}
518
519//The total number of operations is 1
520IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2)
521{
522        return _mm_or_si128(arg1, arg2);
523}
524
525//The total number of operations is 1
526IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2)
527{
528        return _mm_and_si128(arg1, arg2);
529}
530
531//The total number of operations is 1
532IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2)
533{
534        return _mm_xor_si128(arg1, arg2);
535}
536
537//The total number of operations is 1
538template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2)
539{
540        return simd_and(arg1, arg2);
541}
542
543//The total number of operations is 18
544template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2)
545{
546        return simd128<1>::ifh(simd128<2>::lt(arg1, arg2), arg2, arg1);
547}
548
549//The total number of operations is 9
550template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2)
551{
552        bitblock128_t high_bit = simd128<4>::constant<(8)>();
553        return simd_xor(simd128<4>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
554}
555
556//The total number of operations is 1
557template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2)
558{
559        return _mm_max_epi8(arg1, arg2);
560}
561
562//The total number of operations is 1
563template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2)
564{
565        return _mm_max_epi16(arg1, arg2);
566}
567
568//The total number of operations is 1
569template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2)
570{
571        return _mm_max_epi32(arg1, arg2);
572}
573
574//The total number of operations is 15
575template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2)
576{
577        bitblock128_t hiAns = simd128<(32)>::max(arg1, arg2);
578        bitblock128_t loAns = simd128<(32)>::umax(arg1, arg2);
579        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(hiAns, arg1));
580        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(hiAns, arg2));
581        return simd128<1>::ifh(simd128<64>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
582}
583
584//The total number of operations is 48
585template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2)
586{
587        bitblock128_t hiAns = simd128<(64)>::max(arg1, arg2);
588        bitblock128_t loAns = simd128<(64)>::umax(arg1, arg2);
589        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
590        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
591        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
592}
593
594//The total number of operations is 1
595template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2)
596{
597        return simd_and(arg1, arg2);
598}
599
600//The total number of operations is 23
601template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2)
602{
603        bitblock128_t tmp1 = simd128<128>::slli<1>(arg1);
604        bitblock128_t tmp2 = simd128<128>::slli<1>(arg2);
605        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));
606}
607
608//The total number of operations is 31
609template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2)
610{
611        bitblock128_t loMask = simd128<(8)>::lomask();
612        bitblock128_t tmpAns1 = simd128<(8)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
613        bitblock128_t tmpAns2 = simd128<(8)>::mult(simd128<(8)>::srli<4>(arg1), simd128<(8)>::srli<4>(arg2));
614        return simd128<1>::ifh(loMask, tmpAns1, simd128<(8)>::slli<4>(tmpAns2));
615}
616
617//The total number of operations is 10
618template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2)
619{
620        bitblock128_t loMask = simd128<(16)>::lomask();
621        bitblock128_t tmpAns1 = simd128<(16)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
622        bitblock128_t tmpAns2 = simd128<(16)>::mult(simd128<(16)>::srli<8>(arg1), simd128<(16)>::srli<8>(arg2));
623        return simd128<1>::ifh(loMask, tmpAns1, simd128<(16)>::slli<8>(tmpAns2));
624}
625
626//The total number of operations is 1
627template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2)
628{
629        return _mm_mullo_epi16(arg1, arg2);
630}
631
632//The total number of operations is 1
633template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2)
634{
635        return _mm_mullo_epi32(arg1, arg2);
636}
637
638//The total number of operations is 11
639template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2)
640{
641        bitblock128_t loMask = simd128<64>::lomask();
642        bitblock128_t arg1_low = simd_and(arg1, loMask);
643        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg1);
644        bitblock128_t arg2_low = simd_and(arg2, loMask);
645        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg2);
646        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
647        bitblock128_t tmpAns2 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
648        bitblock128_t tmpAns3 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
649        return simd128<64>::add(tmpAns1, simd128<64>::add(tmpAns2, tmpAns3));
650}
651
652//The total number of operations is 205
653template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2)
654{
655        bitblock128_t loMask = simd128<128>::lomask();
656        bitblock128_t arg1_low = simd_and(arg1, loMask);
657        bitblock128_t arg1_high = simd128<128>::srli<(64)>(arg1);
658        bitblock128_t arg2_low = simd_and(arg2, loMask);
659        bitblock128_t arg2_high = simd128<128>::srli<(64)>(arg2);
660        bitblock128_t tmpAns1 = simd128<(64)>::umult(arg1_low, arg2_low);
661        bitblock128_t tmpAns2 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_low, arg2_high));
662        bitblock128_t tmpAns3 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_high, arg2_low));
663        return simd128<128>::add(tmpAns1, simd128<128>::add(tmpAns2, tmpAns3));
664}
665
666//The total number of operations is 1
667template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2)
668{
669        return simd_andc(arg2, arg1);
670}
671
672//The total number of operations is 15
673template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2)
674{
675        bitblock128_t hiAns = simd128<(1)>::gt(arg1, arg2);
676        bitblock128_t loAns = simd128<(1)>::ugt(arg1, arg2);
677        bitblock128_t mask = simd_and(loAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
678        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
679        return simd_or(simd128<2>::srai<(1)>(hiAns), mask);
680}
681
682//The total number of operations is 10
683template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2)
684{
685        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::gt(simd_and(simd128<(8)>::himask(), arg1), arg2), simd128<(8)>::gt(simd128<(8)>::slli<4>(arg1), simd128<(8)>::slli<4>(arg2)));
686}
687
688//The total number of operations is 1
689template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2)
690{
691        return _mm_cmpgt_epi8(arg1, arg2);
692}
693
694//The total number of operations is 1
695template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2)
696{
697        return _mm_cmpgt_epi16(arg1, arg2);
698}
699
700//The total number of operations is 1
701template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2)
702{
703        return _mm_cmpgt_epi32(arg1, arg2);
704}
705
706//The total number of operations is 15
707template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2)
708{
709        bitblock128_t hiAns = simd128<(32)>::gt(arg1, arg2);
710        bitblock128_t loAns = simd128<(32)>::ugt(arg1, arg2);
711        bitblock128_t mask = simd_and(loAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
712        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
713        return simd_or(simd128<64>::srai<(32)>(hiAns), mask);
714}
715
716//The total number of operations is 62
717template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2)
718{
719        bitblock128_t hiAns = simd128<(64)>::gt(arg1, arg2);
720        bitblock128_t loAns = simd128<(64)>::ugt(arg1, arg2);
721        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
722        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
723        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
724}
725
726//The total number of operations is 289
727template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2)
728{
729        bitblock128_t loMask = simd128<(2)>::lomask();
730        bitblock128_t tmpAns1 = simd128<(2)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
731        bitblock128_t tmpAns2 = simd128<(2)>::umult(simd_and(loMask, simd128<(4)>::srli<(2)>(arg1)), simd_and(loMask, simd128<(4)>::srli<(2)>(arg2)));
732        return simd_or(tmpAns1, simd128<(4)>::slli<(2)>(tmpAns2));
733}
734
735//The total number of operations is 139
736template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2)
737{
738        bitblock128_t loMask = simd128<(4)>::lomask();
739        bitblock128_t tmpAns1 = simd128<(4)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
740        bitblock128_t tmpAns2 = simd128<(4)>::umult(simd_and(loMask, simd128<(8)>::srli<(4)>(arg1)), simd_and(loMask, simd128<(8)>::srli<(4)>(arg2)));
741        return simd_or(tmpAns1, simd128<(8)>::slli<(4)>(tmpAns2));
742}
743
744//The total number of operations is 64
745template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2)
746{
747        bitblock128_t loMask = simd128<(8)>::lomask();
748        bitblock128_t tmpAns1 = simd128<(8)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
749        bitblock128_t tmpAns2 = simd128<(8)>::umult(simd_and(loMask, simd128<(16)>::srli<(8)>(arg1)), simd_and(loMask, simd128<(16)>::srli<(8)>(arg2)));
750        return simd_or(tmpAns1, simd128<(16)>::slli<(8)>(tmpAns2));
751}
752
753//The total number of operations is 28
754template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2)
755{
756        bitblock128_t loMask = simd128<(16)>::lomask();
757        bitblock128_t tmpAns1 = simd128<(16)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
758        bitblock128_t tmpAns2 = simd128<(16)>::umult(simd_and(loMask, simd128<(32)>::srli<(16)>(arg1)), simd_and(loMask, simd128<(32)>::srli<(16)>(arg2)));
759        return simd_or(tmpAns1, simd128<(32)>::slli<(16)>(tmpAns2));
760}
761
762//The total number of operations is 10
763template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2)
764{
765        bitblock128_t loMask = simd128<(32)>::lomask();
766        bitblock128_t tmpAns1 = simd128<(32)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
767        bitblock128_t tmpAns2 = simd128<(32)>::umult(simd_and(loMask, simd128<(64)>::srli<(32)>(arg1)), simd_and(loMask, simd128<(64)>::srli<(32)>(arg2)));
768        return simd_or(tmpAns1, simd128<(64)>::slli<(32)>(tmpAns2));
769}
770
771//The total number of operations is 1
772template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2)
773{
774        return _mm_mul_epu32(arg1, arg2);
775}
776
777//The total number of operations is 55
778template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2)
779{
780        bitblock128_t loMask1 = simd128<(128)>::lomask();
781        bitblock128_t arg11 = simd_and(arg1, loMask1);
782        bitblock128_t arg22 = simd_and(arg2, loMask1);
783        bitblock128_t loMask2 = simd128<64>::lomask();
784        bitblock128_t arg1_low = simd_and(arg11, loMask2);
785        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg11);
786        bitblock128_t arg2_low = simd_and(arg22, loMask2);
787        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg22);
788        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
789        bitblock128_t tmpAns2 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
790        bitblock128_t tmpAns3 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
791        bitblock128_t tmpAns4 = simd128<(128)>::slli<64>(simd128<(32)>::umult(arg1_high, arg2_high));
792        return simd128<(128)>::add(tmpAns1, simd128<(128)>::add(tmpAns2, simd128<(128)>::add(tmpAns3, tmpAns4)));
793}
794
795//The total number of operations is 1
796template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2)
797{
798        return simd_andc(arg2, arg1);
799}
800
801//The total number of operations is 14
802template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2)
803{
804        bitblock128_t tmpAns = simd128<(1)>::ult(arg1, arg2);
805        bitblock128_t mask = simd_and(tmpAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
806        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
807        return simd_or(simd128<2>::srai<(1)>(tmpAns), mask);
808}
809
810//The total number of operations is 20
811template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2)
812{
813        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::ult(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::ult(simd_andc(arg1, simd128<(8)>::himask()), simd_andc(arg2, simd128<(8)>::himask())));
814}
815
816//The total number of operations is 7
817template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2)
818{
819        bitblock128_t high_bit = simd128<8>::constant<(128)>();
820        return simd128<8>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
821}
822
823//The total number of operations is 7
824template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2)
825{
826        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
827        return simd128<16>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
828}
829
830//The total number of operations is 7
831template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2)
832{
833        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
834        return simd128<32>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
835}
836
837//The total number of operations is 15
838template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2)
839{
840        return simd_and(simd128<64>::srai<(63)>(simd_or(simd_and(simd_not(arg1), arg2), simd_and(simd_not(simd_xor(arg1, arg2)), simd128<64>::sub(arg1, arg2)))), simd_not(simd128<64>::eq(arg1, arg2)));
841}
842
843//The total number of operations is 48
844template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2)
845{
846        bitblock128_t tmpAns = simd128<(64)>::ult(arg1, arg2);
847        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
848        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
849        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
850}
851
852//The total number of operations is 1
853template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2)
854{
855        return simd_andc(arg1, arg2);
856}
857
858//The total number of operations is 15
859template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2)
860{
861        bitblock128_t hiAns = simd128<(1)>::lt(arg1, arg2);
862        bitblock128_t loAns = simd128<(1)>::ult(arg1, arg2);
863        bitblock128_t mask = simd_and(loAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
864        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
865        return simd_or(simd128<2>::srai<(1)>(hiAns), mask);
866}
867
868//The total number of operations is 18
869template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2)
870{
871        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)));
872}
873
874//The total number of operations is 5
875template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2)
876{
877        return simd_and(simd_not(simd128<8>::gt(arg1, arg2)), simd_not(simd128<8>::eq(arg1, arg2)));
878}
879
880//The total number of operations is 5
881template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2)
882{
883        return simd_and(simd_not(simd128<16>::gt(arg1, arg2)), simd_not(simd128<16>::eq(arg1, arg2)));
884}
885
886//The total number of operations is 5
887template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2)
888{
889        return simd_and(simd_not(simd128<32>::gt(arg1, arg2)), simd_not(simd128<32>::eq(arg1, arg2)));
890}
891
892//The total number of operations is 17
893template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2)
894{
895        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808UL)>();
896        return simd128<64>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
897}
898
899//The total number of operations is 65
900template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2)
901{
902        bitblock128_t hiAns = simd128<(64)>::lt(arg1, arg2);
903        bitblock128_t loAns = simd128<(64)>::ult(arg1, arg2);
904        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
905        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
906        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
907}
908
909//The total number of operations is 2
910template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1)
911{
912        return simd_and(simd128<32>::srli<sh>(arg1), simd128<2>::constant<((3)>>sh)>());
913}
914
915//The total number of operations is 2
916template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1)
917{
918        return simd_and(simd128<32>::srli<sh>(arg1), simd128<4>::constant<((15)>>sh)>());
919}
920
921//The total number of operations is 2
922template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1)
923{
924        return simd_and(simd128<32>::srli<sh>(arg1), simd128<8>::constant<((255)>>sh)>());
925}
926
927//The total number of operations is 1
928template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1)
929{
930        return _mm_srli_epi16(arg1, (int32_t)(sh));
931}
932
933//The total number of operations is 1
934template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1)
935{
936        return _mm_srli_epi32(arg1, (int32_t)(sh));
937}
938
939//The total number of operations is 1
940template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1)
941{
942        return _mm_srli_epi64(arg1, (int32_t)(sh));
943}
944
945//The total number of operations is 4
946template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1)
947{
948        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)))));
949}
950
951//The total number of operations is 1
952template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1)
953{
954        return simd_not(arg1);
955}
956
957//The total number of operations is 14
958template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1)
959{
960        bitblock128_t tmp = simd_not(arg1);
961        return simd128<1>::ifh(simd128<2>::himask(), simd_and(tmp, simd128<128>::slli<1>(tmp)), simd_and(simd128<128>::srli<1>(arg1), tmp));
962}
963
964//The total number of operations is 14
965template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1)
966{
967        return simd128<4>::popcount(simd_andc(simd128<4>::sub(arg1, simd128<4>::constant<1>()), arg1));
968}
969
970//The total number of operations is 13
971template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1)
972{
973        return simd128<8>::popcount(simd_andc(simd128<8>::sub(arg1, simd128<8>::constant<1>()), arg1));
974}
975
976//The total number of operations is 16
977template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1)
978{
979        return simd128<16>::popcount(simd_andc(simd128<16>::sub(arg1, simd128<16>::constant<1>()), arg1));
980}
981
982//The total number of operations is 19
983template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1)
984{
985        return simd128<32>::popcount(simd_andc(simd128<32>::sub(arg1, simd128<32>::constant<1>()), arg1));
986}
987
988//The total number of operations is 14
989template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1)
990{
991        return simd128<64>::popcount(simd_andc(simd128<64>::sub(arg1, simd128<64>::constant<1>()), arg1));
992}
993
994//The total number of operations is 30
995template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1)
996{
997        return simd128<128>::popcount(simd_andc(simd128<128>::sub(arg1, simd128<128>::constant<1>()), arg1));
998}
999
1000//The total number of operations is 7
1001template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1002{
1003        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())));
1004}
1005
1006//The total number of operations is 11
1007template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1008{
1009        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))));
1010}
1011
1012//The total number of operations is 1
1013template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1014{
1015        return simd_andc(arg1, arg2);
1016}
1017
1018//The total number of operations is 14
1019template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1020{
1021        bitblock128_t tmpAns = simd128<(1)>::ugt(arg1, arg2);
1022        bitblock128_t mask = simd_and(tmpAns, simd128<2>::srli<(1)>(simd128<(1)>::eq(arg1, arg2)));
1023        mask = simd_or(mask, simd128<2>::slli<(1)>(mask));
1024        return simd_or(simd128<2>::srai<(1)>(tmpAns), mask);
1025}
1026
1027//The total number of operations is 12
1028template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1029{
1030        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::ugt(simd_and(simd128<(8)>::himask(), arg1), arg2), simd128<(8)>::ugt(simd_andc(arg1, simd128<(8)>::himask()), simd_andc(arg2, simd128<(8)>::himask())));
1031}
1032
1033//The total number of operations is 3
1034template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1035{
1036        bitblock128_t high_bit = simd128<8>::constant<(128)>();
1037        return simd128<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1038}
1039
1040//The total number of operations is 3
1041template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1042{
1043        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1044        return simd128<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1045}
1046
1047//The total number of operations is 3
1048template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1049{
1050        bitblock128_t high_bit = simd128<32>::constant<(2147483648UL)>();
1051        return simd128<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1052}
1053
1054//The total number of operations is 14
1055template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1056{
1057        bitblock128_t tmpAns = simd128<(32)>::ugt(arg1, arg2);
1058        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
1059        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
1060        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
1061}
1062
1063//The total number of operations is 47
1064template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1065{
1066        bitblock128_t tmpAns = simd128<(64)>::ugt(arg1, arg2);
1067        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
1068        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
1069        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
1070}
1071
1072//The total number of operations is 4
1073template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1)
1074{
1075        return simd_xor(simd128<2>::srli<(1)>(arg1), simd_and(arg1, simd128<2>::lomask()));
1076}
1077
1078//The total number of operations is 4
1079template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1)
1080{
1081        return simd_xor(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1082}
1083
1084//The total number of operations is 4
1085template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1)
1086{
1087        return simd_xor(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1088}
1089
1090//The total number of operations is 3
1091template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1)
1092{
1093        return simd_xor(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1094}
1095
1096//The total number of operations is 3
1097template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1)
1098{
1099        return simd_xor(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1100}
1101
1102//The total number of operations is 3
1103template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1)
1104{
1105        return simd_xor(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1106}
1107
1108//The total number of operations is 6
1109template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1)
1110{
1111        return simd_xor(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1112}
1113
1114//The total number of operations is 0
1115template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1)
1116{
1117        return arg1;
1118}
1119
1120//The total number of operations is 3
1121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1)
1122{
1123        return simd128<2>::add_hl(simd128<(1)>::popcount(arg1));
1124}
1125
1126//The total number of operations is 7
1127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1)
1128{
1129        return simd128<4>::add_hl(simd128<(2)>::popcount(arg1));
1130}
1131
1132//The total number of operations is 11
1133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1)
1134{
1135        return simd128<8>::add_hl(simd128<(4)>::popcount(arg1));
1136}
1137
1138//The total number of operations is 14
1139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1)
1140{
1141        return simd128<16>::add_hl(simd128<(8)>::popcount(arg1));
1142}
1143
1144//The total number of operations is 17
1145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1)
1146{
1147        return simd128<32>::add_hl(simd128<(16)>::popcount(arg1));
1148}
1149
1150//The total number of operations is 12
1151template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1)
1152{
1153        return _mm_sad_epu8(simd128<8>::popcount(arg1), simd128<8>::constant<0>());
1154}
1155
1156//The total number of operations is 18
1157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1)
1158{
1159        bitblock128_t tmpAns = simd128<(64)>::popcount(arg1);
1160        return simd128<(64)>::add(simd_and(tmpAns, simd128<128>::lomask()), simd128<128>::srli<(64)>(tmpAns));
1161}
1162
1163//The total number of operations is 8
1164template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1)
1165{
1166        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(arg1, simd128<128>::slli<1>(arg1)), arg1);
1167}
1168
1169//The total number of operations is 6
1170template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1)
1171{
1172        return simd128<4>::sub(simd128<4>::constant<0>(), arg1);
1173}
1174
1175//The total number of operations is 1
1176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1)
1177{
1178        return simd128<8>::sub(simd128<8>::constant<0>(), arg1);
1179}
1180
1181//The total number of operations is 1
1182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1)
1183{
1184        return simd128<16>::sub(simd128<16>::constant<0>(), arg1);
1185}
1186
1187//The total number of operations is 1
1188template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1)
1189{
1190        return _mm_sign_epi32(arg1, simd128<32>::constant<-1>());
1191}
1192
1193//The total number of operations is 1
1194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1)
1195{
1196        return simd128<64>::sub(simd128<64>::constant<0>(), arg1);
1197}
1198
1199//The total number of operations is 11
1200template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1)
1201{
1202        return simd128<128>::sub(simd128<128>::constant<0>(), arg1);
1203}
1204
1205//The total number of operations is 2
1206template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1)
1207{
1208        return simd_and(simd128<32>::slli<sh>(arg1), simd128<2>::constant<(((3)<<sh)&(3))>());
1209}
1210
1211//The total number of operations is 2
1212template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1)
1213{
1214        return simd_and(simd128<32>::slli<sh>(arg1), simd128<4>::constant<(((15)<<sh)&(15))>());
1215}
1216
1217//The total number of operations is 2
1218template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1)
1219{
1220        return simd_and(simd128<32>::slli<sh>(arg1), simd128<8>::constant<(((255)<<sh)&(255))>());
1221}
1222
1223//The total number of operations is 1
1224template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1)
1225{
1226        return _mm_slli_epi16(arg1, (int32_t)(sh));
1227}
1228
1229//The total number of operations is 1
1230template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1)
1231{
1232        return _mm_slli_epi32(arg1, (int32_t)(sh));
1233}
1234
1235//The total number of operations is 1
1236template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1)
1237{
1238        return _mm_slli_epi64(arg1, (int32_t)(sh));
1239}
1240
1241//The total number of operations is 4
1242template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1)
1243{
1244        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)))));
1245}
1246
1247//The total number of operations is 3
1248template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1249{
1250        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
1251}
1252
1253//The total number of operations is 8
1254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1255{
1256        return simd128<(1)>::ifh(simd128<1>::ifh(simd128<2>::himask(), arg1, simd128<2>::srli<(1)>(arg1)), arg2, arg3);
1257}
1258
1259//The total number of operations is 13
1260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1261{
1262        return simd128<1>::ifh(simd128<4>::gt(simd128<4>::constant<0>(), arg1), arg2, arg3);
1263}
1264
1265//The total number of operations is 1
1266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1267{
1268        return _mm_blendv_epi8(arg3, arg2, arg1);
1269}
1270
1271//The total number of operations is 4
1272template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1273{
1274        return simd128<1>::ifh(simd128<16>::gt(simd128<16>::constant<0>(), arg1), arg2, arg3);
1275}
1276
1277//The total number of operations is 4
1278template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1279{
1280        return simd128<1>::ifh(simd128<32>::gt(simd128<32>::constant<0>(), arg1), arg2, arg3);
1281}
1282
1283//The total number of operations is 8
1284template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1285{
1286        return simd128<(32)>::ifh(simd128<1>::ifh(simd128<64>::himask(), arg1, simd128<64>::srli<(32)>(arg1)), arg2, arg3);
1287}
1288
1289//The total number of operations is 15
1290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1291{
1292        return simd128<(64)>::ifh(simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg1)), arg2, arg3);
1293}
1294
1295//The total number of operations is 1
1296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2)
1297{
1298        return simd_xor(arg1, arg2);
1299}
1300
1301//The total number of operations is 11
1302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2)
1303{
1304        bitblock128_t tmp = simd_xor(arg1, arg2);
1305        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2))), tmp);
1306}
1307
1308//The total number of operations is 6
1309template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2)
1310{
1311        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::sub(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::sub(arg1, arg2));
1312}
1313
1314//The total number of operations is 1
1315template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2)
1316{
1317        return _mm_sub_epi8(arg1, arg2);
1318}
1319
1320//The total number of operations is 1
1321template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2)
1322{
1323        return _mm_sub_epi16(arg1, arg2);
1324}
1325
1326//The total number of operations is 1
1327template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2)
1328{
1329        return _mm_sub_epi32(arg1, arg2);
1330}
1331
1332//The total number of operations is 1
1333template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2)
1334{
1335        return _mm_sub_epi64(arg1, arg2);
1336}
1337
1338//The total number of operations is 11
1339template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2)
1340{
1341        bitblock128_t partial = simd128<(64)>::sub(arg1, arg2);
1342        bitblock128_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
1343        bitblock128_t borrow = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(borrowMask));
1344        return simd128<(64)>::sub(partial, borrow);
1345}
1346
1347//The total number of operations is 3
1348template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1)
1349{
1350        return simd128<16>::sub(arg1, simd_and(simd128<2>::lomask(), simd128<16>::srli<1>(arg1)));
1351}
1352
1353//The total number of operations is 4
1354template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1)
1355{
1356        return simd128<(8)>::add(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1357}
1358
1359//The total number of operations is 4
1360template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1)
1361{
1362        return simd128<(16)>::add(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1363}
1364
1365//The total number of operations is 3
1366template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1)
1367{
1368        return simd128<(32)>::add(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1369}
1370
1371//The total number of operations is 3
1372template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1)
1373{
1374        return simd128<(64)>::add(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1375}
1376
1377//The total number of operations is 3
1378template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1)
1379{
1380        return simd128<64>::add(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1381}
1382
1383//The total number of operations is 16
1384template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1)
1385{
1386        return simd128<128>::add(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1387}
1388
1389//The total number of operations is 7
1390template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1391{
1392        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())));
1393}
1394
1395//The total number of operations is 11
1396template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1397{
1398        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))));
1399}
1400
1401//The total number of operations is 0
1402template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1403{
1404        return simd128<32>::constant<(-1*val)>();
1405}
1406
1407//The total number of operations is 0
1408template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1409{
1410        return ((val < 0) ? simd128<(4)>::constant<((val<<2)|(val^(-4)))>() : simd128<(4)>::constant<((val<<2)|val)>());
1411}
1412
1413//The total number of operations is 0
1414template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1415{
1416        return ((val < 0) ? simd128<(8)>::constant<((val<<4)|(val^(-16)))>() : simd128<(8)>::constant<((val<<4)|val)>());
1417}
1418
1419//The total number of operations is 0
1420template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1421{
1422        return _mm_set1_epi8((int32_t)(val));
1423}
1424
1425//The total number of operations is 0
1426template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1427{
1428        return _mm_set1_epi16((int32_t)(val));
1429}
1430
1431//The total number of operations is 0
1432template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1433{
1434        return _mm_set1_epi32((int32_t)(val));
1435}
1436
1437//The total number of operations is 0
1438template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1439{
1440        return _mm_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val));
1441}
1442
1443//The total number of operations is 0
1444template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1445{
1446        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val));
1447}
1448
1449//The total number of operations is 1
1450template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2)
1451{
1452        return simd_or(arg1, arg2);
1453}
1454
1455//The total number of operations is 18
1456template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2)
1457{
1458        return simd128<1>::ifh(simd128<2>::lt(arg1, arg2), arg1, arg2);
1459}
1460
1461//The total number of operations is 9
1462template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2)
1463{
1464        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1465        return simd_xor(simd128<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1466}
1467
1468//The total number of operations is 1
1469template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2)
1470{
1471        return _mm_min_epi8(arg1, arg2);
1472}
1473
1474//The total number of operations is 1
1475template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2)
1476{
1477        return _mm_min_epi16(arg1, arg2);
1478}
1479
1480//The total number of operations is 1
1481template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2)
1482{
1483        return _mm_min_epi32(arg1, arg2);
1484}
1485
1486//The total number of operations is 15
1487template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2)
1488{
1489        bitblock128_t hiAns = simd128<(32)>::min(arg1, arg2);
1490        bitblock128_t loAns = simd128<(32)>::umin(arg1, arg2);
1491        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(hiAns, arg1));
1492        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(hiAns, arg2));
1493        return simd128<1>::ifh(simd128<64>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
1494}
1495
1496//The total number of operations is 48
1497template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2)
1498{
1499        bitblock128_t hiAns = simd128<(64)>::min(arg1, arg2);
1500        bitblock128_t loAns = simd128<(64)>::umin(arg1, arg2);
1501        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg1));
1502        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(hiAns, arg2));
1503        return simd128<1>::ifh(simd128<128>::himask(), hiAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, loAns, arg1), arg2));
1504}
1505
1506//The total number of operations is 0
1507template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask()
1508{
1509        return simd128<2>::constant<(1)>();
1510}
1511
1512//The total number of operations is 0
1513template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask()
1514{
1515        return simd128<4>::constant<(3)>();
1516}
1517
1518//The total number of operations is 0
1519template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask()
1520{
1521        return simd128<8>::constant<(15)>();
1522}
1523
1524//The total number of operations is 0
1525template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask()
1526{
1527        return simd128<16>::constant<(255)>();
1528}
1529
1530//The total number of operations is 0
1531template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask()
1532{
1533        return simd128<32>::constant<(65535)>();
1534}
1535
1536//The total number of operations is 0
1537template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask()
1538{
1539        return _mm_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1));
1540}
1541
1542//The total number of operations is 0
1543template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask()
1544{
1545        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1));
1546}
1547
1548//The total number of operations is 1
1549template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2)
1550{
1551        return simd_and(arg1, arg2);
1552}
1553
1554//The total number of operations is 16
1555template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2)
1556{
1557        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)));
1558}
1559
1560//The total number of operations is 6
1561template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2)
1562{
1563        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)));
1564}
1565
1566//The total number of operations is 1
1567template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2)
1568{
1569        return _mm_min_epu8(arg1, arg2);
1570}
1571
1572//The total number of operations is 1
1573template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2)
1574{
1575        return _mm_min_epu16(arg1, arg2);
1576}
1577
1578//The total number of operations is 1
1579template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2)
1580{
1581        return _mm_min_epu32(arg1, arg2);
1582}
1583
1584//The total number of operations is 14
1585template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2)
1586{
1587        bitblock128_t tmpAns = simd128<(32)>::umin(arg1, arg2);
1588        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1589        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1590        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1591}
1592
1593//The total number of operations is 33
1594template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2)
1595{
1596        bitblock128_t tmpAns = simd128<(64)>::umin(arg1, arg2);
1597        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1598        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1599        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1600}
1601
1602//The total number of operations is 9
1603template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1)
1604{
1605        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, simd128<128>::slli<1>(simd_not(arg1))), arg1);
1606}
1607
1608//The total number of operations is 19
1609template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1)
1610{
1611        bitblock128_t gtMask = simd128<4>::gt(arg1, simd128<4>::constant<0>());
1612        return simd128<1>::ifh(gtMask, arg1, simd128<4>::sub(gtMask, arg1));
1613}
1614
1615//The total number of operations is 1
1616template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1)
1617{
1618        return _mm_abs_epi8(arg1);
1619}
1620
1621//The total number of operations is 1
1622template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1)
1623{
1624        return _mm_abs_epi16(arg1);
1625}
1626
1627//The total number of operations is 1
1628template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1)
1629{
1630        return _mm_abs_epi32(arg1);
1631}
1632
1633//The total number of operations is 9
1634template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1)
1635{
1636        bitblock128_t eqMask = simd128<64>::eq(simd128<1>::ifh(simd128<64>::himask(), simd128<(32)>::abs(arg1), arg1), arg1);
1637        return simd128<1>::ifh(eqMask, arg1, simd128<64>::sub(eqMask, arg1));
1638}
1639
1640//The total number of operations is 37
1641template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1)
1642{
1643        bitblock128_t eqMask = simd128<128>::eq(simd128<1>::ifh(simd128<128>::himask(), simd128<(64)>::abs(arg1), arg1), arg1);
1644        return simd128<1>::ifh(eqMask, arg1, simd128<128>::sub(eqMask, arg1));
1645}
1646
1647//The total number of operations is 2
1648template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2)
1649{
1650        return simd_not(simd_xor(arg1, arg2));
1651}
1652
1653//The total number of operations is 8
1654template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2)
1655{
1656        bitblock128_t tmpAns = simd128<(1)>::eq(arg1, arg2);
1657        bitblock128_t loMask = simd_and(tmpAns, simd128<2>::srli<(1)>(tmpAns));
1658        bitblock128_t hiMask = simd128<2>::slli<(1)>(loMask);
1659        return simd_or(loMask, hiMask);
1660}
1661
1662//The total number of operations is 9
1663template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2)
1664{
1665        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))));
1666}
1667
1668//The total number of operations is 1
1669template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2)
1670{
1671        return _mm_cmpeq_epi8(arg1, arg2);
1672}
1673
1674//The total number of operations is 1
1675template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2)
1676{
1677        return _mm_cmpeq_epi16(arg1, arg2);
1678}
1679
1680//The total number of operations is 1
1681template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2)
1682{
1683        return _mm_cmpeq_epi32(arg1, arg2);
1684}
1685
1686//The total number of operations is 1
1687template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2)
1688{
1689        return _mm_cmpeq_epi64(arg1, arg2);
1690}
1691
1692//The total number of operations is 11
1693template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2)
1694{
1695        bitblock128_t tmpAns = simd128<(64)>::eq(arg1, arg2);
1696        bitblock128_t loMask = simd_and(tmpAns, simd128<128>::srli<(64)>(tmpAns));
1697        bitblock128_t hiMask = simd128<128>::slli<(64)>(loMask);
1698        return simd_or(loMask, hiMask);
1699}
1700
1701//The total number of operations is 4
1702template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1)
1703{
1704        return ((sh == 0) ? arg1 : simd_or(simd_and(simd128<2>::himask(), arg1), simd128<2>::srli<1>(arg1)));
1705}
1706
1707//The total number of operations is 10
1708template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1)
1709{
1710        bitblock128_t tmp = simd128<4>::srli<((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh))>(arg1);
1711        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)));
1712}
1713
1714//The total number of operations is 5
1715template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1)
1716{
1717        bitblock128_t tmp = simd128<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
1718        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)));
1719}
1720
1721//The total number of operations is 1
1722template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1)
1723{
1724        return _mm_srai_epi16(arg1, (int32_t)(sh));
1725}
1726
1727//The total number of operations is 1
1728template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1)
1729{
1730        return _mm_srai_epi32(arg1, (int32_t)(sh));
1731}
1732
1733//The total number of operations is 5
1734template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1)
1735{
1736        bitblock128_t tmp = simd128<64>::srli<((sh >= 64) ? (63) : ((sh < 0) ? 0 : sh))>(arg1);
1737        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)));
1738}
1739
1740//The total number of operations is 21
1741template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1)
1742{
1743        bitblock128_t tmp = simd128<128>::srli<((sh >= 128) ? (127) : ((sh < 0) ? 0 : sh))>(arg1);
1744        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)));
1745}
1746
1747//The total number of operations is 0
1748template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask()
1749{
1750        return simd128<2>::constant<(2)>();
1751}
1752
1753//The total number of operations is 0
1754template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask()
1755{
1756        return simd128<4>::constant<(12)>();
1757}
1758
1759//The total number of operations is 0
1760template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask()
1761{
1762        return simd128<8>::constant<(240)>();
1763}
1764
1765//The total number of operations is 0
1766template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask()
1767{
1768        return simd128<16>::constant<(65280)>();
1769}
1770
1771//The total number of operations is 0
1772template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask()
1773{
1774        return simd128<32>::constant<-65536>();
1775}
1776
1777//The total number of operations is 0
1778template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask()
1779{
1780        return _mm_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0));
1781}
1782
1783//The total number of operations is 0
1784template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask()
1785{
1786        return _mm_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0));
1787}
1788
1789//The total number of operations is 1
1790template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2)
1791{
1792        return simd_xor(arg1, arg2);
1793}
1794
1795//The total number of operations is 10
1796template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2)
1797{
1798        bitblock128_t tmp = simd_xor(arg1, arg2);
1799        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(arg1, arg2))), tmp);
1800}
1801
1802//The total number of operations is 6
1803template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2)
1804{
1805        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::add(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::add(arg1, arg2));
1806}
1807
1808//The total number of operations is 1
1809template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2)
1810{
1811        return _mm_add_epi8(arg1, arg2);
1812}
1813
1814//The total number of operations is 1
1815template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2)
1816{
1817        return _mm_add_epi16(arg1, arg2);
1818}
1819
1820//The total number of operations is 1
1821template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2)
1822{
1823        return _mm_add_epi32(arg1, arg2);
1824}
1825
1826//The total number of operations is 1
1827template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2)
1828{
1829        return _mm_add_epi64(arg1, arg2);
1830}
1831
1832//The total number of operations is 11
1833template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2)
1834{
1835        bitblock128_t partial = simd128<(64)>::add(arg1, arg2);
1836        bitblock128_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
1837        bitblock128_t carry = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(carryMask));
1838        return simd128<(64)>::add(partial, carry);
1839}
1840
1841//The total number of operations is 1
1842template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2)
1843{
1844        return simd_or(arg1, arg2);
1845}
1846
1847//The total number of operations is 16
1848template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2)
1849{
1850        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)));
1851}
1852
1853//The total number of operations is 6
1854template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2)
1855{
1856        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)));
1857}
1858
1859//The total number of operations is 1
1860template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2)
1861{
1862        return _mm_max_epu8(arg1, arg2);
1863}
1864
1865//The total number of operations is 1
1866template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2)
1867{
1868        return _mm_max_epu16(arg1, arg2);
1869}
1870
1871//The total number of operations is 1
1872template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2)
1873{
1874        return _mm_max_epu32(arg1, arg2);
1875}
1876
1877//The total number of operations is 14
1878template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2)
1879{
1880        bitblock128_t tmpAns = simd128<(32)>::umax(arg1, arg2);
1881        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1882        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1883        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1884}
1885
1886//The total number of operations is 33
1887template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2)
1888{
1889        bitblock128_t tmpAns = simd128<(64)>::umax(arg1, arg2);
1890        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1891        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1892        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1893}
1894
1895//The total number of operations is 93
1896template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1897{
1898        return simd128<(1)>::umin(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1899}
1900
1901//The total number of operations is 80
1902template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1903{
1904        return simd128<(2)>::umin(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1905}
1906
1907//The total number of operations is 42
1908template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1909{
1910        return simd128<(4)>::umin(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1911}
1912
1913//The total number of operations is 7
1914template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1915{
1916        return simd128<(8)>::umin(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1917}
1918
1919//The total number of operations is 7
1920template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1921{
1922        return simd128<(16)>::umin(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1923}
1924
1925//The total number of operations is 7
1926template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1927{
1928        return simd128<(32)>::umin(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1929}
1930
1931//The total number of operations is 28
1932template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1933{
1934        return simd128<(64)>::umin(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1935}
1936
1937//The total number of operations is 93
1938template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1939{
1940        return simd128<(1)>::add(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1941}
1942
1943//The total number of operations is 74
1944template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1945{
1946        return simd128<(2)>::add(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1947}
1948
1949//The total number of operations is 42
1950template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1951{
1952        return simd128<(4)>::add(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1953}
1954
1955//The total number of operations is 7
1956template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1957{
1958        return simd128<(8)>::add(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1959}
1960
1961//The total number of operations is 1
1962template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1963{
1964        return _mm_hadd_epi16(arg2, arg1);
1965}
1966
1967//The total number of operations is 1
1968template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1969{
1970        return _mm_hadd_epi32(arg2, arg1);
1971}
1972
1973//The total number of operations is 15
1974template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1975{
1976        return simd128<(64)>::add(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1977}
1978
1979//The total number of operations is 120
1980template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2)
1981{
1982        bitblock128_t hiBound = simd128<2>::srli<1>(simd128<2>::lomask());
1983        bitblock128_t loBound = simd_not(hiBound);
1984        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)));
1985}
1986
1987//The total number of operations is 86
1988template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2)
1989{
1990        bitblock128_t hiBound = simd128<4>::srli<1>(simd128<4>::lomask());
1991        bitblock128_t loBound = simd_not(hiBound);
1992        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)));
1993}
1994
1995//The total number of operations is 36
1996template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2)
1997{
1998        bitblock128_t hiBound = simd128<8>::srli<1>(simd128<8>::lomask());
1999        bitblock128_t loBound = simd_not(hiBound);
2000        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)));
2001}
2002
2003//The total number of operations is 1
2004template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2)
2005{
2006        return _mm_packs_epi16(arg2, arg1);
2007}
2008
2009//The total number of operations is 1
2010template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2)
2011{
2012        return _mm_packs_epi32(arg2, arg1);
2013}
2014
2015//The total number of operations is 77
2016template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2)
2017{
2018        bitblock128_t hiBound = simd128<64>::srli<1>(simd128<64>::lomask());
2019        bitblock128_t loBound = simd_not(hiBound);
2020        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)));
2021}
2022
2023//The total number of operations is 272
2024template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2)
2025{
2026        bitblock128_t hiBound = simd128<128>::srli<1>(simd128<128>::lomask());
2027        bitblock128_t loBound = simd_not(hiBound);
2028        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)));
2029}
2030
2031//The total number of operations is 24
2032template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1)
2033{
2034        uint64_t tmpAns1 = hsimd128<(8)>::signmask(esimd128<4>::mergeh(arg1, simd128<4>::constant<0>()));
2035        uint64_t tmpAns2 = hsimd128<(8)>::signmask(esimd128<4>::mergel(arg1, simd128<4>::constant<0>()));
2036        return ((tmpAns1<<(16))+tmpAns2);
2037}
2038
2039//The total number of operations is 1
2040template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1)
2041{
2042        return _mm_movemask_epi8(arg1);
2043}
2044
2045//The total number of operations is 4
2046template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1)
2047{
2048        return hsimd128<(8)>::signmask(hsimd128<16>::packh(simd128<16>::constant<0>(), arg1));
2049}
2050
2051//The total number of operations is 7
2052template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1)
2053{
2054        return hsimd128<(16)>::signmask(hsimd128<32>::packh(simd128<32>::constant<0>(), arg1));
2055}
2056
2057//The total number of operations is 10
2058template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1)
2059{
2060        return hsimd128<(32)>::signmask(hsimd128<64>::packh(simd128<64>::constant<0>(), arg1));
2061}
2062
2063//The total number of operations is 17
2064template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1)
2065{
2066        return hsimd128<(64)>::signmask(hsimd128<128>::packh(simd128<128>::constant<0>(), arg1));
2067}
2068
2069//The total number of operations is 45
2070template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2)
2071{
2072        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));
2073}
2074
2075//The total number of operations is 31
2076template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2)
2077{
2078        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));
2079}
2080
2081//The total number of operations is 17
2082template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2)
2083{
2084        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));
2085}
2086
2087//The total number of operations is 3
2088template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2)
2089{
2090        return hsimd128<16>::packus(simd_and(arg1, simd128<16>::lomask()), simd_and(arg2, simd128<16>::lomask()));
2091}
2092
2093//The total number of operations is 3
2094template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2)
2095{
2096        return hsimd128<32>::packus(simd_and(arg1, simd128<32>::lomask()), simd_and(arg2, simd128<32>::lomask()));
2097}
2098
2099//The total number of operations is 3
2100template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2)
2101{
2102        return _mm_hsub_epi32(simd_and(arg2, simd128<64>::lomask()), simd_and(arg1, simd128<64>::lomask()));
2103}
2104
2105//The total number of operations is 7
2106template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2)
2107{
2108        return simd128<1>::ifh(simd128<128>::himask(), simd128<128>::slli<(64)>(arg1), arg2);
2109}
2110
2111//The total number of operations is 47
2112template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2)
2113{
2114        return hsimd128<2>::packl(simd128<64>::srli<(1)>(arg1), simd128<64>::srli<(1)>(arg2));
2115}
2116
2117//The total number of operations is 33
2118template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2)
2119{
2120        return hsimd128<4>::packl(simd128<64>::srli<(2)>(arg1), simd128<64>::srli<(2)>(arg2));
2121}
2122
2123//The total number of operations is 19
2124template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2)
2125{
2126        return hsimd128<8>::packl(simd128<64>::srli<(4)>(arg1), simd128<64>::srli<(4)>(arg2));
2127}
2128
2129//The total number of operations is 3
2130template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2)
2131{
2132        return hsimd128<16>::packus(simd128<16>::srli<(8)>(arg1), simd128<16>::srli<(8)>(arg2));
2133}
2134
2135//The total number of operations is 3
2136template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2)
2137{
2138        return _mm_hsub_epi16(simd128<32>::srli<(16)>(arg2), simd128<32>::srli<(16)>(arg1));
2139}
2140
2141//The total number of operations is 3
2142template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2)
2143{
2144        return _mm_hsub_epi32(simd128<64>::srli<(32)>(arg2), simd128<64>::srli<(32)>(arg1));
2145}
2146
2147//The total number of operations is 7
2148template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2)
2149{
2150        return simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg2));
2151}
2152
2153//The total number of operations is 93
2154template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2155{
2156        return simd128<(1)>::min(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
2157}
2158
2159//The total number of operations is 82
2160template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2161{
2162        return simd128<(2)>::min(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
2163}
2164
2165//The total number of operations is 45
2166template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2167{
2168        return simd128<(4)>::min(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
2169}
2170
2171//The total number of operations is 7
2172template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2173{
2174        return simd128<(8)>::min(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2175}
2176
2177//The total number of operations is 7
2178template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2179{
2180        return simd128<(16)>::min(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2181}
2182
2183//The total number of operations is 7
2184template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2185{
2186        return simd128<(32)>::min(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2187}
2188
2189//The total number of operations is 29
2190template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2191{
2192        return simd128<(64)>::min(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2193}
2194
2195//The total number of operations is 85
2196template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2)
2197{
2198        bitblock128_t arg11 = simd128<2>::ifh(arg1, simd128<2>::constant<0>(), arg1);
2199        bitblock128_t arg12 = simd_and(simd128<2>::lomask(), arg11);
2200        bitblock128_t arg21 = simd128<2>::ifh(arg2, simd128<2>::constant<0>(), arg2);
2201        bitblock128_t arg22 = simd_and(simd128<2>::lomask(), arg21);
2202        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()));
2203}
2204
2205//The total number of operations is 83
2206template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2)
2207{
2208        bitblock128_t arg11 = simd128<4>::ifh(arg1, simd128<4>::constant<0>(), arg1);
2209        bitblock128_t arg12 = simd_and(simd128<4>::lomask(), arg11);
2210        bitblock128_t arg21 = simd128<4>::ifh(arg2, simd128<4>::constant<0>(), arg2);
2211        bitblock128_t arg22 = simd_and(simd128<4>::lomask(), arg21);
2212        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()));
2213}
2214
2215//The total number of operations is 29
2216template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2)
2217{
2218        bitblock128_t arg11 = simd128<8>::ifh(arg1, simd128<8>::constant<0>(), arg1);
2219        bitblock128_t arg12 = simd_and(simd128<8>::lomask(), arg11);
2220        bitblock128_t arg21 = simd128<8>::ifh(arg2, simd128<8>::constant<0>(), arg2);
2221        bitblock128_t arg22 = simd_and(simd128<8>::lomask(), arg21);
2222        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()));
2223}
2224
2225//The total number of operations is 1
2226template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2)
2227{
2228        return _mm_packus_epi16(arg2, arg1);
2229}
2230
2231//The total number of operations is 1
2232template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2)
2233{
2234        return _mm_packus_epi32(arg2, arg1);
2235}
2236
2237//The total number of operations is 12
2238template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2)
2239{
2240        bitblock128_t hiPart = hsimd128<64>::packh(arg1, arg2);
2241        return simd128<(32)>::ifh(hiPart, simd128<(32)>::constant<0>(), simd_or(simd128<(32)>::gt(hiPart, simd128<(32)>::constant<0>()), hsimd128<64>::packl(arg1, arg2)));
2242}
2243
2244//The total number of operations is 38
2245template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2)
2246{
2247        bitblock128_t hiPart = hsimd128<128>::packh(arg1, arg2);
2248        return simd128<(64)>::ifh(hiPart, simd128<(64)>::constant<0>(), simd_or(simd128<(64)>::gt(hiPart, simd128<(64)>::constant<0>()), hsimd128<128>::packl(arg1, arg2)));
2249}
2250
2251//The total number of operations is 31
2252template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2253{
2254        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));
2255}
2256
2257//The total number of operations is 21
2258template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2259{
2260        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));
2261}
2262
2263//The total number of operations is 11
2264template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2265{
2266        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));
2267}
2268
2269//The total number of operations is 1
2270template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2271{
2272        return _mm_unpacklo_epi8(arg2, arg1);
2273}
2274
2275//The total number of operations is 1
2276template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2277{
2278        return _mm_unpacklo_epi16(arg2, arg1);
2279}
2280
2281//The total number of operations is 1
2282template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2283{
2284        return _mm_unpacklo_epi32(arg2, arg1);
2285}
2286
2287//The total number of operations is 1
2288template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2289{
2290        return _mm_unpacklo_epi64(arg2, arg1);
2291}
2292
2293//The total number of operations is 31
2294template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2295{
2296        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));
2297}
2298
2299//The total number of operations is 21
2300template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2301{
2302        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));
2303}
2304
2305//The total number of operations is 11
2306template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2307{
2308        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));
2309}
2310
2311//The total number of operations is 1
2312template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2313{
2314        return _mm_unpackhi_epi8(arg2, arg1);
2315}
2316
2317//The total number of operations is 1
2318template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2319{
2320        return _mm_unpackhi_epi16(arg2, arg1);
2321}
2322
2323//The total number of operations is 1
2324template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2325{
2326        return _mm_unpackhi_epi32(arg2, arg1);
2327}
2328
2329//The total number of operations is 1
2330template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2331{
2332        return _mm_unpackhi_epi64(arg2, arg1);
2333}
2334
2335//The total number of operations is 24
2336template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1)
2337{
2338        return esimd128<(2)>::mergeh(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2339}
2340
2341//The total number of operations is 14
2342template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1)
2343{
2344        return esimd128<(4)>::mergeh(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2345}
2346
2347//The total number of operations is 4
2348template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1)
2349{
2350        return esimd128<(8)>::mergeh(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2351}
2352
2353//The total number of operations is 3
2354template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1)
2355{
2356        return esimd128<(16)>::mergeh(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2357}
2358
2359//The total number of operations is 3
2360template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1)
2361{
2362        return esimd128<(32)>::mergeh(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2363}
2364
2365//The total number of operations is 3
2366template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1)
2367{
2368        return esimd128<(64)>::mergeh(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2369}
2370
2371//The total number of operations is 4
2372template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1)
2373{
2374        return simd128<(128)>::srli<64>(arg1);
2375}
2376
2377//The total number of operations is 24
2378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1)
2379{
2380        return esimd128<(2)>::mergel(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2381}
2382
2383//The total number of operations is 14
2384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1)
2385{
2386        return esimd128<(4)>::mergel(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2387}
2388
2389//The total number of operations is 4
2390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1)
2391{
2392        return esimd128<(8)>::mergel(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2393}
2394
2395//The total number of operations is 1
2396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1)
2397{
2398        return _mm_cvtepu8_epi16(arg1);
2399}
2400
2401//The total number of operations is 1
2402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1)
2403{
2404        return _mm_cvtepu16_epi32(arg1);
2405}
2406
2407//The total number of operations is 1
2408template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1)
2409{
2410        return _mm_cvtepu32_epi64(arg1);
2411}
2412
2413//The total number of operations is 1
2414template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1)
2415{
2416        return simd_and(simd128<(128)>::lomask(), arg1);
2417}
2418
2419//The total number of operations is 31
2420template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1)
2421{
2422        return esimd128<(2)>::mergeh(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2423}
2424
2425//The total number of operations is 33
2426template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1)
2427{
2428        return esimd128<(4)>::mergeh(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2429}
2430
2431//The total number of operations is 13
2432template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1)
2433{
2434        return esimd128<(8)>::mergeh(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2435}
2436
2437//The total number of operations is 4
2438template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1)
2439{
2440        return esimd128<(16)>::mergeh(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2441}
2442
2443//The total number of operations is 4
2444template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1)
2445{
2446        return esimd128<(32)>::mergeh(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2447}
2448
2449//The total number of operations is 5
2450template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1)
2451{
2452        return esimd128<32>::signextendl(simd128<128>::srli<(64)>(arg1));
2453}
2454
2455//The total number of operations is 21
2456template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1)
2457{
2458        return simd128<(128)>::srai<64>(arg1);
2459}
2460
2461//The total number of operations is 31
2462template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1)
2463{
2464        return esimd128<(2)>::mergel(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2465}
2466
2467//The total number of operations is 33
2468template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1)
2469{
2470        return esimd128<(4)>::mergel(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2471}
2472
2473//The total number of operations is 13
2474template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1)
2475{
2476        return esimd128<(8)>::mergel(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2477}
2478
2479//The total number of operations is 1
2480template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1)
2481{
2482        return _mm_cvtepi8_epi16(arg1);
2483}
2484
2485//The total number of operations is 1
2486template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1)
2487{
2488        return _mm_cvtepi16_epi32(arg1);
2489}
2490
2491//The total number of operations is 1
2492template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1)
2493{
2494        return _mm_cvtepi32_epi64(arg1);
2495}
2496
2497//The total number of operations is 25
2498template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1)
2499{
2500        return simd128<(128)>::srai<64>(simd128<(128)>::slli<64>(arg1));
2501}
2502
2503//The total number of operations is 17
2504template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1)
2505{
2506        bitblock128_t tmphi = _mm_shufflehi_epi16(arg1, (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2507        bitblock128_t tmpAns = _mm_shufflelo_epi16(tmphi, (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2508        bitblock128_t tmplh = _mm_shufflehi_epi16(simd128<128>::slli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2509        bitblock128_t tmphl = _mm_shufflelo_epi16(simd128<128>::srli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2510        uint32_t a1 = ((((msk>>21)&4) == 0) ? 0 : (131071));
2511        uint32_t a2 = ((((msk>>18)&4) == 0) ? 0 : (131071));
2512        uint32_t a3 = ((((msk>>15)&4) == 0) ? 0 : (131071));
2513        uint32_t a4 = ((((msk>>12)&4) == 0) ? 0 : (131071));
2514        uint32_t a5 = ((((msk>>9)&4) == 0) ? (131071) : 0);
2515        uint32_t a6 = ((((msk>>6)&4) == 0) ? (131071) : 0);
2516        uint32_t a7 = ((((msk>>3)&4) == 0) ? (131071) : 0);
2517        uint32_t a8 = (((msk&4) == 0) ? (131071) : 0);
2518        return simd128<1>::ifh(mvmd128<16>::fill8(a1, a2, a3, a4, a5, a6, a7, a8), tmpAns, simd_or(tmplh, tmphl));
2519}
2520
2521//The total number of operations is 1
2522template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1)
2523{
2524        return _mm_shuffle_epi32(arg1, (int32_t)(msk));
2525}
2526
2527//The total number of operations is 1
2528template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1)
2529{
2530        return mvmd128<32>::shufflei<shufflemask4_from_shufflemask2(msk)>(arg1);
2531}
2532
2533//The total number of operations is 9
2534template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2535{
2536        return simd_or(mvmd128<2>::srli<sh>(arg1), mvmd128<2>::slli<((64)-sh)>(arg2));
2537}
2538
2539//The total number of operations is 9
2540template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2541{
2542        return simd_or(mvmd128<4>::srli<sh>(arg1), mvmd128<4>::slli<((32)-sh)>(arg2));
2543}
2544
2545//The total number of operations is 9
2546template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2547{
2548        return simd_or(mvmd128<8>::srli<sh>(arg1), mvmd128<8>::slli<((16)-sh)>(arg2));
2549}
2550
2551//The total number of operations is 9
2552template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2553{
2554        return simd_or(mvmd128<16>::srli<sh>(arg1), mvmd128<16>::slli<((8)-sh)>(arg2));
2555}
2556
2557//The total number of operations is 9
2558template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2559{
2560        return simd_or(mvmd128<32>::srli<sh>(arg1), mvmd128<32>::slli<((4)-sh)>(arg2));
2561}
2562
2563//The total number of operations is 9
2564template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2565{
2566        return simd_or(mvmd128<64>::srli<sh>(arg1), mvmd128<64>::slli<((2)-sh)>(arg2));
2567}
2568
2569//The total number of operations is 9
2570template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2571{
2572        return simd_or(mvmd128<128>::srli<sh>(arg1), mvmd128<128>::slli<((1)-sh)>(arg2));
2573}
2574
2575//The total number of operations is 1
2576template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1)
2577{
2578        return mvmd128<32>::fill((-1*val1));
2579}
2580
2581//The total number of operations is 1
2582template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1)
2583{
2584        return mvmd128<(4)>::fill(((val1<<2)|val1));
2585}
2586
2587//The total number of operations is 1
2588template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1)
2589{
2590        return mvmd128<(8)>::fill(((val1<<4)|val1));
2591}
2592
2593//The total number of operations is 1
2594template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1)
2595{
2596        return _mm_set1_epi8((int32_t)(val1));
2597}
2598
2599//The total number of operations is 1
2600template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1)
2601{
2602        return _mm_set1_epi16((int32_t)(val1));
2603}
2604
2605//The total number of operations is 1
2606template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1)
2607{
2608        return _mm_set1_epi32((int32_t)(val1));
2609}
2610
2611//The total number of operations is 1
2612template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1)
2613{
2614        return _mm_set_epi32((int32_t)((val1>>32)), (int32_t)(val1), (int32_t)((val1>>32)), (int32_t)(val1));
2615}
2616
2617//The total number of operations is 1
2618template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1)
2619{
2620        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val1>>32)), (int32_t)(val1));
2621}
2622
2623//The total number of operations is 1
2624template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2625{
2626        return _mm_shuffle_epi8(arg1, arg2);
2627}
2628
2629//The total number of operations is 10
2630template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2631{
2632        bitblock128_t tmp1 = simd_and(simd128<16>::constant<(7)>(), arg2);
2633        bitblock128_t msk1 = simd128<16>::add(tmp1, tmp1);
2634        bitblock128_t msk2 = simd128<16>::add(msk1, simd128<16>::constant<1>());
2635        bitblock128_t msk = simd_or(msk1, simd128<16>::slli<(8)>(msk2));
2636        return simd128<16>::ifh(arg2, simd128<16>::constant<0>(), mvmd128<(8)>::shuffle(arg1, msk));
2637}
2638
2639//The total number of operations is 19
2640template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2641{
2642        bitblock128_t tmp1 = simd_and(simd128<32>::constant<(3)>(), arg2);
2643        bitblock128_t msk1 = simd128<32>::add(tmp1, tmp1);
2644        bitblock128_t msk2 = simd128<32>::add(msk1, simd128<32>::constant<1>());
2645        bitblock128_t msk = simd_or(msk1, simd128<32>::slli<(16)>(msk2));
2646        return simd128<32>::ifh(arg2, simd128<32>::constant<0>(), mvmd128<(16)>::shuffle(arg1, msk));
2647}
2648
2649//The total number of operations is 32
2650template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shuffle(bitblock128_t arg1, bitblock128_t arg2)
2651{
2652        bitblock128_t tmp1 = simd_and(simd128<64>::constant<(1)>(), arg2);
2653        bitblock128_t msk1 = simd128<64>::add(tmp1, tmp1);
2654        bitblock128_t msk2 = simd128<64>::add(msk1, simd128<64>::constant<1>());
2655        bitblock128_t msk = simd_or(msk1, simd128<64>::slli<(32)>(msk2));
2656        return simd128<64>::ifh(arg2, simd128<64>::constant<0>(), mvmd128<(32)>::shuffle(arg1, msk));
2657}
2658
2659//The total number of operations is 1
2660template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1)
2661{
2662        return (((pos%2) == 0) ? (mvmd128<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd128<(2)>::extract<(pos/2)>(arg1)>>1));
2663}
2664
2665//The total number of operations is 1
2666template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1)
2667{
2668        return (((pos%2) == 0) ? (mvmd128<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd128<(4)>::extract<(pos/2)>(arg1)>>2));
2669}
2670
2671//The total number of operations is 1
2672template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1)
2673{
2674        return (((pos%2) == 0) ? (mvmd128<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd128<(8)>::extract<(pos/2)>(arg1)>>4));
2675}
2676
2677//The total number of operations is 1
2678template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1)
2679{
2680        return (((pos%2) == 0) ? (mvmd128<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd128<(16)>::extract<(pos/2)>(arg1)>>8));
2681}
2682
2683//The total number of operations is 1
2684template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1)
2685{
2686        return (65535&_mm_extract_epi16(arg1, (int32_t)(pos)));
2687}
2688
2689//The total number of operations is 1
2690template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1)
2691{
2692        return (((uint64_t)((4294967296UL)-1))&_mm_extract_epi32(arg1, (int32_t)(pos)));
2693}
2694
2695//The total number of operations is 2
2696template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1)
2697{
2698        return ((((uint64_t)mvmd128<(32)>::extract<((2*pos)+1)>(arg1))<<(32))|mvmd128<(32)>::extract<(2*pos)>(arg1));
2699}
2700
2701//The total number of operations is 14
2702template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1)
2703{
2704        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(2)>::slli<1>(arg1) : simd128<(2)>::srli<1>(arg1));
2705        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(2)>::lomask(), arg1) : simd_and(simd128<(2)>::himask(), arg1));
2706        return mvmd128<(2)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2707}
2708
2709//The total number of operations is 10
2710template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1)
2711{
2712        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(4)>::slli<2>(arg1) : simd128<(4)>::srli<2>(arg1));
2713        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(4)>::lomask(), arg1) : simd_and(simd128<(4)>::himask(), arg1));
2714        return mvmd128<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2715}
2716
2717//The total number of operations is 6
2718template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1)
2719{
2720        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(8)>::slli<4>(arg1) : simd128<(8)>::srli<4>(arg1));
2721        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(8)>::lomask(), arg1) : simd_and(simd128<(8)>::himask(), arg1));
2722        return mvmd128<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2723}
2724
2725//The total number of operations is 2
2726template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1)
2727{
2728        return mvmd128<8>::fill(_mm_extract_epi8(arg1, (int32_t)(pos)));
2729}
2730
2731//The total number of operations is 2
2732template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1)
2733{
2734        return mvmd128<16>::fill(_mm_extract_epi16(arg1, (int32_t)(pos)));
2735}
2736
2737//The total number of operations is 1
2738template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1)
2739{
2740        return mvmd128<32>::shufflei<shufflemask4(pos, pos, pos, pos)>(arg1);
2741}
2742
2743//The total number of operations is 5
2744template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1)
2745{
2746        return simd128<1>::ifh(simd128<64>::himask(), mvmd128<(32)>::splat<((2*pos)+1)>(arg1), mvmd128<(32)>::splat<(2*pos)>(arg1));
2747}
2748
2749//The total number of operations is 13
2750template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1)
2751{
2752        return simd128<1>::ifh(simd128<128>::himask(), mvmd128<(64)>::splat<((2*pos)+1)>(arg1), mvmd128<(64)>::splat<(2*pos)>(arg1));
2753}
2754
2755//The total number of operations is 15
2756template <> 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)
2757{
2758        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))));
2759}
2760
2761//The total number of operations is 7
2762template <> 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)
2763{
2764        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))));
2765}
2766
2767//The total number of operations is 3
2768template <> 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)
2769{
2770        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))));
2771}
2772
2773//The total number of operations is 1
2774template <> 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)
2775{
2776        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));
2777}
2778
2779//The total number of operations is 5
2780template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2781{
2782        return simd128<1>::ifh(simd128<(4)>::himask(), mvmd128<1>::fill2(val1, val2), mvmd128<1>::fill2(val3, val4));
2783}
2784
2785//The total number of operations is 5
2786template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2787{
2788        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<2>::fill2(val1, val2), mvmd128<2>::fill2(val3, val4));
2789}
2790
2791//The total number of operations is 5
2792template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2793{
2794        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<4>::fill2(val1, val2), mvmd128<4>::fill2(val3, val4));
2795}
2796
2797//The total number of operations is 5
2798template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2799{
2800        return simd128<1>::ifh(simd128<(32)>::himask(), mvmd128<8>::fill2(val1, val2), mvmd128<8>::fill2(val3, val4));
2801}
2802
2803//The total number of operations is 3
2804template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2805{
2806        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))));
2807}
2808
2809//The total number of operations is 1
2810template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2811{
2812        return _mm_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
2813}
2814
2815//The total number of operations is 4
2816template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1)
2817{
2818        return simd128<128>::srli<(sh*2)>(arg1);
2819}
2820
2821//The total number of operations is 4
2822template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1)
2823{
2824        return simd128<128>::srli<(sh*4)>(arg1);
2825}
2826
2827//The total number of operations is 4
2828template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1)
2829{
2830        return simd128<128>::srli<(sh*8)>(arg1);
2831}
2832
2833//The total number of operations is 4
2834template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1)
2835{
2836        return simd128<128>::srli<(sh*16)>(arg1);
2837}
2838
2839//The total number of operations is 4
2840template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1)
2841{
2842        return simd128<128>::srli<(sh*32)>(arg1);
2843}
2844
2845//The total number of operations is 4
2846template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1)
2847{
2848        return simd128<128>::srli<(sh*64)>(arg1);
2849}
2850
2851//The total number of operations is 4
2852template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1)
2853{
2854        return simd128<128>::srli<(sh*128)>(arg1);
2855}
2856
2857//The total number of operations is 1
2858template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2)
2859{
2860        return mvmd128<(2)>::fill(((val1<<1)|(val2&(1))));
2861}
2862
2863//The total number of operations is 1
2864template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2)
2865{
2866        return mvmd128<(4)>::fill(((val1<<2)|(val2&(3))));
2867}
2868
2869//The total number of operations is 1
2870template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2)
2871{
2872        return mvmd128<(8)>::fill(((val1<<4)|(val2&(15))));
2873}
2874
2875//The total number of operations is 1
2876template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2)
2877{
2878        return mvmd128<(16)>::fill(((val1<<8)|(val2&(255))));
2879}
2880
2881//The total number of operations is 1
2882template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2)
2883{
2884        return mvmd128<(32)>::fill(((val1<<16)|(val2&(65535))));
2885}
2886
2887//The total number of operations is 5
2888template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2)
2889{
2890        return simd128<1>::ifh(simd128<(64)>::himask(), mvmd128<32>::fill(val1), mvmd128<32>::fill(val2));
2891}
2892
2893//The total number of operations is 5
2894template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2)
2895{
2896        return simd128<1>::ifh(simd128<(128)>::himask(), mvmd128<64>::fill(val1), mvmd128<64>::fill(val2));
2897}
2898
2899//The total number of operations is 9
2900template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2901{
2902        return simd_or(mvmd128<2>::slli<sh>(arg1), mvmd128<2>::srli<((64)-sh)>(arg2));
2903}
2904
2905//The total number of operations is 9
2906template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2907{
2908        return simd_or(mvmd128<4>::slli<sh>(arg1), mvmd128<4>::srli<((32)-sh)>(arg2));
2909}
2910
2911//The total number of operations is 9
2912template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2913{
2914        return simd_or(mvmd128<8>::slli<sh>(arg1), mvmd128<8>::srli<((16)-sh)>(arg2));
2915}
2916
2917//The total number of operations is 9
2918template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2919{
2920        return simd_or(mvmd128<16>::slli<sh>(arg1), mvmd128<16>::srli<((8)-sh)>(arg2));
2921}
2922
2923//The total number of operations is 9
2924template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2925{
2926        return simd_or(mvmd128<32>::slli<sh>(arg1), mvmd128<32>::srli<((4)-sh)>(arg2));
2927}
2928
2929//The total number of operations is 9
2930template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2931{
2932        return simd_or(mvmd128<64>::slli<sh>(arg1), mvmd128<64>::srli<((2)-sh)>(arg2));
2933}
2934
2935//The total number of operations is 9
2936template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2937{
2938        return simd_or(mvmd128<128>::slli<sh>(arg1), mvmd128<128>::srli<((1)-sh)>(arg2));
2939}
2940
2941//The total number of operations is 4
2942template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1)
2943{
2944        return simd128<128>::slli<(sh*2)>(arg1);
2945}
2946
2947//The total number of operations is 4
2948template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1)
2949{
2950        return simd128<128>::slli<(sh*4)>(arg1);
2951}
2952
2953//The total number of operations is 4
2954template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1)
2955{
2956        return simd128<128>::slli<(sh*8)>(arg1);
2957}
2958
2959//The total number of operations is 4
2960template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1)
2961{
2962        return simd128<128>::slli<(sh*16)>(arg1);
2963}
2964
2965//The total number of operations is 4
2966template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1)
2967{
2968        return simd128<128>::slli<(sh*32)>(arg1);
2969}
2970
2971//The total number of operations is 4
2972template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1)
2973{
2974        return simd128<128>::slli<(sh*64)>(arg1);
2975}
2976
2977//The total number of operations is 4
2978template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1)
2979{
2980        return simd128<128>::slli<(sh*128)>(arg1);
2981}
2982
2983//The total number of operations is 13
2984template <> 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)
2985{
2986        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<1>::fill4(val1, val2, val3, val4), mvmd128<1>::fill4(val5, val6, val7, val8));
2987}
2988
2989//The total number of operations is 13
2990template <> 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)
2991{
2992        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<2>::fill4(val1, val2, val3, val4), mvmd128<2>::fill4(val5, val6, val7, val8));
2993}
2994
2995//The total number of operations is 7
2996template <> 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)
2997{
2998        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))));
2999}
3000
3001//The total number of operations is 3
3002template <> 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)
3003{
3004        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))));
3005}
3006
3007//The total number of operations is 1
3008template <> 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)
3009{
3010        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));
3011}
3012
3013//The total number of operations is 11
3014IDISA_ALWAYS_INLINE bitblock128_t bitblock128::sll(bitblock128_t arg1, bitblock128_t arg2)
3015{
3016        return simd128<128>::sll(arg1, arg2);
3017}
3018
3019//The total number of operations is 1
3020IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(float const* arg1)
3021{
3022        return _mm_loadu_si128((bitblock128_t*)(arg1));
3023}
3024
3025//The total number of operations is 4
3026template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srli(bitblock128_t arg1)
3027{
3028        return simd128<128>::srli<sh>(arg1);
3029}
3030
3031//The total number of operations is 1
3032IDISA_ALWAYS_INLINE void bitblock128::store_aligned(float* arg1, bitblock128_t arg2)
3033{
3034        _mm_store_si128((bitblock128_t*)(arg1), arg2);
3035}
3036
3037//The total number of operations is 20
3038IDISA_ALWAYS_INLINE uint64_t bitblock128::popcount(bitblock128_t arg1)
3039{
3040        return mvmd128<64>::extract<0>(simd128<128>::popcount(arg1));
3041}
3042
3043//The total number of operations is 2
3044IDISA_ALWAYS_INLINE bool bitblock128::all(bitblock128_t arg1)
3045{
3046        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<-1>())) == 65535;
3047}
3048
3049//The total number of operations is 11
3050IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srl(bitblock128_t arg1, bitblock128_t arg2)
3051{
3052        return simd128<128>::srl(arg1, arg2);
3053}
3054
3055//The total number of operations is 4
3056template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::slli(bitblock128_t arg1)
3057{
3058        return simd128<128>::slli<sh>(arg1);
3059}
3060
3061//The total number of operations is 2
3062IDISA_ALWAYS_INLINE bool bitblock128::any(bitblock128_t arg1)
3063{
3064        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<0>())) != 65535;
3065}
3066
3067//The total number of operations is 1
3068IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(float const* arg1)
3069{
3070        return _mm_load_si128((bitblock128_t*)(arg1));
3071}
3072
3073//The total number of operations is 1
3074IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(float* arg1, bitblock128_t arg2)
3075{
3076        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
3077}
3078
3079#endif
Note: See TracBrowser for help on using the repository browser.