source: trunk/lib/idisa_cpp/idisa_sse2.cpp @ 2546

Last change on this file since 2546 was 2330, checked in by cameron, 7 years ago

Use typename only within template declarations

File size: 153.4 KB
Line 
1
2/* Copyright (c) 2011, Hua Huang and Robert D. Cameron.
3   Licensed under the Academic Free License 3.0.
4   This file is generated by the IDISA+ generator;
5   modifications should be made only by changing the
6   generator configuration and data files. */
7
8#ifndef IDISA_SSE2_CPP
9#define IDISA_SSE2_CPP
10#include <stdint.h>
11#include "../config.hpp"
12
13#include "emmintrin.h"
14
15typedef __m128i bitblock128_t;
16
17template <uint32_t fw> struct FieldType {
18   typedef int T;  //default for FieldType::T is int
19};
20
21template <> struct FieldType<1> {typedef uint8_t T;};
22template <> struct FieldType<2> {typedef uint8_t T;};
23template <> struct FieldType<4> {typedef uint8_t T;};
24template <> struct FieldType<8> {typedef uint8_t T;};
25template <> struct FieldType<16> {typedef uint16_t T;};
26template <> struct FieldType<32> {typedef uint32_t T;};
27template <> struct FieldType<64> {typedef uint64_t T;};
28template <> struct FieldType<128> {typedef uint64_t T;};
29
30typedef FieldType<1>::T fw1_t;
31typedef FieldType<2>::T fw2_t;
32typedef FieldType<4>::T fw4_t;
33typedef FieldType<8>::T fw8_t;
34typedef FieldType<16>::T fw16_t;
35typedef FieldType<32>::T fw32_t;
36typedef FieldType<64>::T fw64_t;
37typedef FieldType<128>::T fw128_t;
38
39template <uint32_t fw>
40class simd128
41{
42public:
43        static IDISA_ALWAYS_INLINE bitblock128_t max(bitblock128_t arg1, bitblock128_t arg2);
44        static IDISA_ALWAYS_INLINE bitblock128_t mult(bitblock128_t arg1, bitblock128_t arg2);
45        static IDISA_ALWAYS_INLINE bitblock128_t gt(bitblock128_t arg1, bitblock128_t arg2);
46        static IDISA_ALWAYS_INLINE bitblock128_t umult(bitblock128_t arg1, bitblock128_t arg2);
47        static IDISA_ALWAYS_INLINE bitblock128_t ult(bitblock128_t arg1, bitblock128_t arg2);
48        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
49        static IDISA_ALWAYS_INLINE bitblock128_t ctz(bitblock128_t arg1);
50        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t shift_mask);
51        static IDISA_ALWAYS_INLINE bitblock128_t eq(bitblock128_t arg1, bitblock128_t arg2);
52        static IDISA_ALWAYS_INLINE bitblock128_t popcount(bitblock128_t arg1);
53        static IDISA_ALWAYS_INLINE bitblock128_t neg(bitblock128_t arg1);
54        static IDISA_ALWAYS_INLINE bitblock128_t himask();
55        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
56        static IDISA_ALWAYS_INLINE bitblock128_t ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
57        static IDISA_ALWAYS_INLINE bitblock128_t sub(bitblock128_t arg1, bitblock128_t arg2);
58        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1);
59        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t shift_mask);
60        static IDISA_ALWAYS_INLINE bitblock128_t lomask();
61        static IDISA_ALWAYS_INLINE bitblock128_t umin(bitblock128_t arg1, bitblock128_t arg2);
62//      template <typename FieldType<fw>::T> static IDISA_ALWAYS_INLINE bitblock128_t constant();
63        template <uint64_t> static IDISA_ALWAYS_INLINE bitblock128_t constant();
64        static IDISA_ALWAYS_INLINE bitblock128_t min(bitblock128_t arg1, bitblock128_t arg2);
65        static IDISA_ALWAYS_INLINE bitblock128_t umax(bitblock128_t arg1, bitblock128_t arg2);
66        static IDISA_ALWAYS_INLINE bitblock128_t abs(bitblock128_t arg1);
67        static IDISA_ALWAYS_INLINE bitblock128_t xor_hl(bitblock128_t arg1);
68        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srai(bitblock128_t arg1);
69        static IDISA_ALWAYS_INLINE bitblock128_t lt(bitblock128_t arg1, bitblock128_t arg2);
70        static IDISA_ALWAYS_INLINE bitblock128_t add(bitblock128_t arg1, bitblock128_t arg2);
71        static IDISA_ALWAYS_INLINE bitblock128_t ugt(bitblock128_t arg1, bitblock128_t arg2);
72};
73
74template <uint32_t fw>
75class hsimd128
76{
77public:
78        static IDISA_ALWAYS_INLINE bitblock128_t umin_hl(bitblock128_t arg1, bitblock128_t arg2);
79        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1, bitblock128_t arg2);
80        static IDISA_ALWAYS_INLINE bitblock128_t packss(bitblock128_t arg1, bitblock128_t arg2);
81        static IDISA_ALWAYS_INLINE bitblock128_t packh(bitblock128_t arg1, bitblock128_t arg2);
82        static IDISA_ALWAYS_INLINE typename FieldType<128/fw>::T signmask(bitblock128_t arg1);
83        static IDISA_ALWAYS_INLINE bitblock128_t packl(bitblock128_t arg1, bitblock128_t arg2);
84        static IDISA_ALWAYS_INLINE bitblock128_t min_hl(bitblock128_t arg1, bitblock128_t arg2);
85        static IDISA_ALWAYS_INLINE bitblock128_t packus(bitblock128_t arg1, bitblock128_t arg2);
86};
87
88template <uint32_t fw>
89class esimd128
90{
91public:
92        static IDISA_ALWAYS_INLINE bitblock128_t mergel(bitblock128_t arg1, bitblock128_t arg2);
93        static IDISA_ALWAYS_INLINE bitblock128_t signextendh(bitblock128_t arg1);
94        static IDISA_ALWAYS_INLINE bitblock128_t mergeh(bitblock128_t arg1, bitblock128_t arg2);
95        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendh(bitblock128_t arg1);
96        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendl(bitblock128_t arg1);
97        static IDISA_ALWAYS_INLINE bitblock128_t signextendl(bitblock128_t arg1);
98};
99
100template <uint32_t fw>
101class mvmd128
102{
103public:
104        template <uint8_t msk> static IDISA_ALWAYS_INLINE bitblock128_t shufflei(bitblock128_t arg1);
105        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dsrli(bitblock128_t arg1, bitblock128_t arg2);
106        static IDISA_ALWAYS_INLINE bitblock128_t fill(typename FieldType<fw>::T val1);
107        template <uint8_t pos> static IDISA_ALWAYS_INLINE typename FieldType<fw>::T extract(bitblock128_t arg1);
108        template <uint8_t pos> static IDISA_ALWAYS_INLINE bitblock128_t splat(bitblock128_t arg1);
109        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
110        static IDISA_ALWAYS_INLINE bitblock128_t fill4(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2, typename FieldType<fw>::T val3, typename FieldType<fw>::T val4);
111        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
112        static IDISA_ALWAYS_INLINE bitblock128_t fill2(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2);
113        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dslli(bitblock128_t arg1, bitblock128_t arg2);
114        static IDISA_ALWAYS_INLINE bitblock128_t fill8(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2, typename FieldType<fw>::T val3, typename FieldType<fw>::T val4, typename FieldType<fw>::T val5, typename FieldType<fw>::T val6, typename FieldType<fw>::T val7, typename FieldType<fw>::T val8);
115        static IDISA_ALWAYS_INLINE bitblock128_t fill16(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2, typename FieldType<fw>::T val3, typename FieldType<fw>::T val4, typename FieldType<fw>::T val5, typename FieldType<fw>::T val6, typename FieldType<fw>::T val7, typename FieldType<fw>::T val8, typename FieldType<fw>::T val9, typename FieldType<fw>::T val10, typename FieldType<fw>::T val11, typename FieldType<fw>::T val12, typename FieldType<fw>::T val13, typename FieldType<fw>::T val14, typename FieldType<fw>::T val15, typename FieldType<fw>::T val16);
116};
117
118class bitblock128
119{
120public:
121        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
122        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(const bitblock128_t* arg1);
123        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
124        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
125        static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t arg1, bitblock128_t* arg2);
126        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
127        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
128        static IDISA_ALWAYS_INLINE uint8_t popcount(bitblock128_t arg1);
129        template <uint8_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
130        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(const bitblock128_t* arg1);
131        static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t arg1, bitblock128_t* arg2);
132};
133
134//Declaration Part
135IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2);
136IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1);
137IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2);
138IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2);
139IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2);
140IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2);
141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2);
142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2);
143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2);
144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2);
145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2);
146template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2);
147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2);
148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2);
149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2);
150template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2);
151template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2);
152template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2);
153template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2);
154template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2);
155template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2);
156template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2);
157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2);
158template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2);
159template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2);
160template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2);
161template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2);
162template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2);
163template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2);
164template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2);
165template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2);
166template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2);
167template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2);
168template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2);
169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2);
170template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2);
171template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2);
172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2);
173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2);
174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2);
175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2);
176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2);
177template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2);
178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2);
179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2);
180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2);
181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2);
182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2);
183template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2);
184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2);
185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2);
186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2);
187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2);
188template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1);
189template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1);
190template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1);
191template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1);
192template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1);
193template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1);
194template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1);
195template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1);
196template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1);
197template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1);
198template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1);
199template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1);
200template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1);
201template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1);
202template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1);
203template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
204template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
205template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2);
206template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2);
207template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2);
208template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2);
209template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2);
210template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2);
211template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2);
212template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2);
213template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1);
214template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1);
215template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1);
216template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1);
217template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1);
218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1);
219template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1);
220template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1);
221template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1);
222template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1);
223template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1);
224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1);
225template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1);
226template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1);
227template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1);
228template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1);
229template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1);
230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1);
231template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1);
232template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1);
233template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1);
234template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1);
235template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1);
236template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1);
237template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1);
238template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1);
239template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1);
240template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1);
241template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1);
242template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
243template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
244template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
245template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
246template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
247template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
248template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
249template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
250template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2);
251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2);
252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2);
253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2);
254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2);
255template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2);
256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2);
257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2);
258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1);
259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1);
260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1);
261template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1);
262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1);
263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1);
264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1);
265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
267
268template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
269template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
270template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
271template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
272template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
273template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
274template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
275template <> template <uint64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
276#if 0
277template <> template <fw1_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
278template <> template <fw2_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
279template <> template <fw4_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
280template <> template <fw8_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
281template <> template <fw16_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
282template <> template <fw32_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
283template <> template <fw64_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
284template <> template <fw128_t> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
285#endif
286template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2);
287template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2);
288template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2);
289template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2);
290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2);
291template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2);
292template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2);
293template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2);
294template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask();
295template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask();
296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask();
297template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask();
298template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask();
299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask();
300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask();
301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2);
302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2);
303template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2);
304template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2);
305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2);
306template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2);
307template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2);
308template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2);
309template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1);
310template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1);
311template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1);
312template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1);
313template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1);
314template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1);
315template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1);
316template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2);
317template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2);
318template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2);
319template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2);
320template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2);
321template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2);
322template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2);
323template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2);
324template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1);
325template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1);
326template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1);
327template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1);
328template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1);
329template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1);
330template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1);
331template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask();
332template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask();
333template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask();
334template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask();
335template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask();
336template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask();
337template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask();
338template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2);
339template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2);
340template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2);
341template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2);
342template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2);
343template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2);
344template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2);
345template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2);
346template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2);
347template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2);
348template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2);
349template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2);
350template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2);
351template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2);
352template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2);
353template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2);
354template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
355template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
356template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
357template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
358template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
359template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
360template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
361template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
362template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
363template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
364template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
365template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
366template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
367template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
368template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2);
369template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2);
370template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2);
371template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2);
372template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2);
373template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2);
374template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2);
375template <> IDISA_ALWAYS_INLINE fw32_t hsimd128<4>::signmask(bitblock128_t arg1);
376template <> IDISA_ALWAYS_INLINE FieldType<128/8>::T hsimd128<8>::signmask(bitblock128_t arg1);
377template <> IDISA_ALWAYS_INLINE FieldType<128/16>::T hsimd128<16>::signmask(bitblock128_t arg1);
378template <> IDISA_ALWAYS_INLINE FieldType<128/32>::T hsimd128<32>::signmask(bitblock128_t arg1);
379template <> IDISA_ALWAYS_INLINE FieldType<128/64>::T hsimd128<64>::signmask(bitblock128_t arg1);
380template <> IDISA_ALWAYS_INLINE FieldType<128/128>::T hsimd128<128>::signmask(bitblock128_t arg1);
381template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2);
382template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2);
383template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2);
384template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2);
385template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2);
386template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2);
387template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2);
388template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2);
389template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2);
390template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2);
391template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2);
392template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2);
393template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2);
394template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2);
395template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
396template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
397template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
398template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
399template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
400template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
401template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
402template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2);
403template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2);
404template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2);
405template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2);
406template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2);
407template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2);
408template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2);
409template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2);
410template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2);
411template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2);
412template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2);
413template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2);
414template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2);
415template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2);
416template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
417template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
418template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
419template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
420template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
421template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
422template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
423template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1);
424template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1);
425template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1);
426template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1);
427template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1);
428template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1);
429template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1);
430template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1);
431template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1);
432template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1);
433template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1);
434template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1);
435template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1);
436template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1);
437template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1);
438template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1);
439template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1);
440template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1);
441template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1);
442template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1);
443template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1);
444template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1);
445template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1);
446template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1);
447template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1);
448template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1);
449template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1);
450template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1);
451template <> template <uint8_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1);
452template <> template <uint8_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1);
453template <> template <uint8_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1);
454template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
455template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
456template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
457template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
458template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
459template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
460template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
461template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(fw1_t val1);
462template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(fw2_t val1);
463template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(fw4_t val1);
464template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(fw8_t val1);
465template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(fw16_t val1);
466template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(fw32_t val1);
467template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(fw64_t val1);
468template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(fw128_t val1);
469template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw1_t mvmd128<1>::extract(bitblock128_t arg1);
470template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw2_t mvmd128<2>::extract(bitblock128_t arg1);
471template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw4_t mvmd128<4>::extract(bitblock128_t arg1);
472template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw8_t mvmd128<8>::extract(bitblock128_t arg1);
473template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw16_t mvmd128<16>::extract(bitblock128_t arg1);
474template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw32_t mvmd128<32>::extract(bitblock128_t arg1);
475template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw64_t mvmd128<64>::extract(bitblock128_t arg1);
476template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1);
477template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1);
478template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1);
479template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1);
480template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1);
481template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1);
482template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1);
483template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1);
484template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill16(fw1_t val1, fw1_t val2, fw1_t val3, fw1_t val4, fw1_t val5, fw1_t val6, fw1_t val7, fw1_t val8, fw1_t val9, fw1_t val10, fw1_t val11, fw1_t val12, fw1_t val13, fw1_t val14, fw1_t val15, fw1_t val16);
485template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill16(fw2_t val1, fw2_t val2, fw2_t val3, fw2_t val4, fw2_t val5, fw2_t val6, fw2_t val7, fw2_t val8, fw2_t val9, fw2_t val10, fw2_t val11, fw2_t val12, fw2_t val13, fw2_t val14, fw2_t val15, fw2_t val16);
486template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill16(fw4_t val1, fw4_t val2, fw4_t val3, fw4_t val4, fw4_t val5, fw4_t val6, fw4_t val7, fw4_t val8, fw4_t val9, fw4_t val10, fw4_t val11, fw4_t val12, fw4_t val13, fw4_t val14, fw4_t val15, fw4_t val16);
487template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill16(fw8_t val1, fw8_t val2, fw8_t val3, fw8_t val4, fw8_t val5, fw8_t val6, fw8_t val7, fw8_t val8, fw8_t val9, fw8_t val10, fw8_t val11, fw8_t val12, fw8_t val13, fw8_t val14, fw8_t val15, fw8_t val16);
488template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(fw1_t val1, fw1_t val2, fw1_t val3, fw1_t val4);
489template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(fw2_t val1, fw2_t val2, fw2_t val3, fw2_t val4);
490template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(fw4_t val1, fw4_t val2, fw4_t val3, fw4_t val4);
491template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(fw8_t val1, fw8_t val2, fw8_t val3, fw8_t val4);
492template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(fw16_t val1, fw16_t val2, fw16_t val3, fw16_t val4);
493template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(fw32_t val1, fw32_t val2, fw32_t val3, fw32_t val4);
494template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1);
495template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1);
496template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1);
497template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1);
498template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1);
499template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1);
500template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1);
501template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(fw1_t val1, fw1_t val2);
502template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(fw2_t val1, fw2_t val2);
503template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(fw4_t val1, fw4_t val2);
504template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(fw8_t val1, fw8_t val2);
505template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(fw16_t val1, fw16_t val2);
506template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(fw32_t val1, fw32_t val2);
507template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(fw64_t val1, fw64_t val2);
508template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2);
509template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2);
510template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2);
511template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2);
512template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2);
513template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2);
514template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2);
515template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1);
516template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1);
517template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1);
518template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1);
519template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1);
520template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1);
521template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1);
522template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill8(fw1_t val1, fw1_t val2, fw1_t val3, fw1_t val4, fw1_t val5, fw1_t val6, fw1_t val7, fw1_t val8);
523template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill8(fw2_t val1, fw2_t val2, fw2_t val3, fw2_t val4, fw2_t val5, fw2_t val6, fw2_t val7, fw2_t val8);
524template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill8(fw4_t val1, fw4_t val2, fw4_t val3, fw4_t val4, fw4_t val5, fw4_t val6, fw4_t val7, fw4_t val8);
525template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill8(fw8_t val1, fw8_t val2, fw8_t val3, fw8_t val4, fw8_t val5, fw8_t val6, fw8_t val7, fw8_t val8);
526template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill8(fw16_t val1, fw16_t val2, fw16_t val3, fw16_t val4, fw16_t val5, fw16_t val6, fw16_t val7, fw16_t val8);
527
528//Implementation Part
529const uint64_t highbit_64 = ((uint64_t) 1) << 63;
530const uint32_t highbit_32 = 0x80000000;
531const uint16_t highbit_16 = 0x8000;
532
533#define shufflemask4(s1, s2, s3, s4) \
534        ((s1<<6) | (s2<<4) | (s3<<2) | s4)
535
536#define shufflemask4_from_shufflemask2(msk) \
537        (msk==3 ? 238 : (msk==2 ? 228 : (msk==1 ? 78 : 68)))
538
539#define shufflemask8_to_shufflemask4(msk) \
540        ((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))
541//The total number of operations is 2.0
542IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2)
543{
544        return simd_not(simd_or(arg1, arg2));
545}
546
547//The total number of operations is 1.0
548IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1)
549{
550        return simd_xor(arg1, simd128<32>::constant<-1>());
551}
552
553//The total number of operations is 1.0
554IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2)
555{
556        return _mm_andnot_si128(arg2, arg1);
557}
558
559//The total number of operations is 1.0
560IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2)
561{
562        return _mm_or_si128(arg1, arg2);
563}
564
565//The total number of operations is 1.0
566IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2)
567{
568        return _mm_xor_si128(arg1, arg2);
569}
570
571//The total number of operations is 1.0
572IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2)
573{
574        return _mm_and_si128(arg1, arg2);
575}
576
577//The total number of operations is 1.0
578template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2)
579{
580        return simd_and(arg1, arg2);
581}
582
583//The total number of operations is 15.6666666667
584template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2)
585{
586        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, arg2), simd_or(simd_and(arg2, simd128<128>::srli<1>(simd_or(arg1, simd_not(arg2)))), simd_and(arg1, simd128<128>::srli<1>(simd_or(simd_not(arg1), arg2)))));
587}
588
589//The total number of operations is 9.0
590template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2)
591{
592        bitblock128_t high_bit = simd128<4>::constant<(8)>();
593        return simd_xor(simd128<4>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
594}
595
596//The total number of operations is 4.0
597template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2)
598{
599        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg1, arg2);
600}
601
602//The total number of operations is 1.0
603template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2)
604{
605        return _mm_max_epi16(arg1, arg2);
606}
607
608//The total number of operations is 4.0
609template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2)
610{
611        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg1, arg2);
612}
613
614//The total number of operations is 17.5
615template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2)
616{
617        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg1, arg2);
618}
619
620//The total number of operations is 54.75
621template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2)
622{
623        return simd128<1>::ifh(simd128<128>::gt(arg1, arg2), arg1, arg2);
624}
625
626//The total number of operations is 1.0
627template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2)
628{
629        return simd_and(arg1, arg2);
630}
631
632//The total number of operations is 19.6666666667
633template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2)
634{
635        bitblock128_t tmp1 = simd128<128>::slli<1>(arg1);
636        bitblock128_t tmp2 = simd128<128>::slli<1>(arg2);
637        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));
638}
639
640//The total number of operations is 31.0
641template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2)
642{
643        bitblock128_t loMask = simd128<(8)>::lomask();
644        bitblock128_t tmpAns1 = simd128<(8)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
645        bitblock128_t tmpAns2 = simd128<(8)>::mult(simd128<(8)>::srli<4>(arg1), simd128<(8)>::srli<4>(arg2));
646        return simd128<1>::ifh(loMask, tmpAns1, simd128<(8)>::slli<4>(tmpAns2));
647}
648
649//The total number of operations is 10.0
650template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2)
651{
652        bitblock128_t loMask = simd128<(16)>::lomask();
653        bitblock128_t tmpAns1 = simd128<(16)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
654        bitblock128_t tmpAns2 = simd128<(16)>::mult(simd128<(16)>::srli<8>(arg1), simd128<(16)>::srli<8>(arg2));
655        return simd128<1>::ifh(loMask, tmpAns1, simd128<(16)>::slli<8>(tmpAns2));
656}
657
658//The total number of operations is 1.0
659template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2)
660{
661        return _mm_mullo_epi16(arg1, arg2);
662}
663
664//The total number of operations is 30.0
665template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2)
666{
667        bitblock128_t loMask = simd128<(64)>::lomask();
668        bitblock128_t tmpAns1 = simd128<(64)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
669        bitblock128_t tmpAns2 = simd128<(64)>::mult(simd128<(64)>::srli<32>(arg1), simd128<(64)>::srli<32>(arg2));
670        return simd128<1>::ifh(loMask, tmpAns1, simd128<(64)>::slli<32>(tmpAns2));
671}
672
673//The total number of operations is 11.0
674template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2)
675{
676        bitblock128_t loMask = simd128<64>::lomask();
677        bitblock128_t arg1_low = simd_and(arg1, loMask);
678        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg1);
679        bitblock128_t arg2_low = simd_and(arg2, loMask);
680        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg2);
681        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
682        bitblock128_t tmpAns2 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
683        bitblock128_t tmpAns3 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
684        return simd128<64>::add(tmpAns1, simd128<64>::add(tmpAns2, tmpAns3));
685}
686
687//The total number of operations is 165.0
688template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2)
689{
690        bitblock128_t loMask = simd128<128>::lomask();
691        bitblock128_t arg1_low = simd_and(arg1, loMask);
692        bitblock128_t arg1_high = simd128<128>::srli<(64)>(arg1);
693        bitblock128_t arg2_low = simd_and(arg2, loMask);
694        bitblock128_t arg2_high = simd128<128>::srli<(64)>(arg2);
695        bitblock128_t tmpAns1 = simd128<(64)>::umult(arg1_low, arg2_low);
696        bitblock128_t tmpAns2 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_low, arg2_high));
697        bitblock128_t tmpAns3 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_high, arg2_low));
698        return simd128<128>::add(tmpAns1, simd128<128>::add(tmpAns2, tmpAns3));
699}
700
701//The total number of operations is 1.0
702template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2)
703{
704        return simd_andc(arg2, arg1);
705}
706
707//The total number of operations is 14.6666666667
708template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2)
709{
710        bitblock128_t tmp = simd_not(arg1);
711        bitblock128_t tmpAns = simd_or(simd_and(tmp, arg2), simd_and(simd128<128>::slli<1>(simd_and(arg1, simd_not(arg2))), simd_or(tmp, arg2)));
712        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
713}
714
715//The total number of operations is 10.0
716template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2)
717{
718        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)));
719}
720
721//The total number of operations is 1.0
722template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2)
723{
724        return _mm_cmpgt_epi8(arg1, arg2);
725}
726
727//The total number of operations is 1.0
728template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2)
729{
730        return _mm_cmpgt_epi16(arg1, arg2);
731}
732
733//The total number of operations is 1.0
734template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2)
735{
736        return _mm_cmpgt_epi32(arg1, arg2);
737}
738
739//The total number of operations is 14.5
740template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2)
741{
742        bitblock128_t hiAns = simd128<(32)>::gt(arg1, arg2);
743        bitblock128_t loAns = simd128<(32)>::ugt(arg1, arg2);
744        bitblock128_t mask = simd_and(loAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
745        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
746        return simd_or(simd128<64>::srai<(32)>(hiAns), mask);
747}
748
749//The total number of operations is 51.75
750template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2)
751{
752        bitblock128_t hiAns = simd128<(64)>::gt(arg1, arg2);
753        bitblock128_t loAns = simd128<(64)>::ugt(arg1, arg2);
754        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
755        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
756        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
757}
758
759//The total number of operations is 289.0
760template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2)
761{
762        bitblock128_t loMask = simd128<(2)>::lomask();
763        bitblock128_t tmpAns1 = simd128<(2)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
764        bitblock128_t tmpAns2 = simd128<(2)>::umult(simd_and(loMask, simd128<(4)>::srli<(2)>(arg1)), simd_and(loMask, simd128<(4)>::srli<(2)>(arg2)));
765        return simd_or(tmpAns1, simd128<(4)>::slli<(2)>(tmpAns2));
766}
767
768//The total number of operations is 139.0
769template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2)
770{
771        bitblock128_t loMask = simd128<(4)>::lomask();
772        bitblock128_t tmpAns1 = simd128<(4)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
773        bitblock128_t tmpAns2 = simd128<(4)>::umult(simd_and(loMask, simd128<(8)>::srli<(4)>(arg1)), simd_and(loMask, simd128<(8)>::srli<(4)>(arg2)));
774        return simd_or(tmpAns1, simd128<(8)>::slli<(4)>(tmpAns2));
775}
776
777//The total number of operations is 64.0
778template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2)
779{
780        bitblock128_t loMask = simd128<(8)>::lomask();
781        bitblock128_t tmpAns1 = simd128<(8)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
782        bitblock128_t tmpAns2 = simd128<(8)>::umult(simd_and(loMask, simd128<(16)>::srli<(8)>(arg1)), simd_and(loMask, simd128<(16)>::srli<(8)>(arg2)));
783        return simd_or(tmpAns1, simd128<(16)>::slli<(8)>(tmpAns2));
784}
785
786//The total number of operations is 28.0
787template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2)
788{
789        bitblock128_t loMask = simd128<(16)>::lomask();
790        bitblock128_t tmpAns1 = simd128<(16)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
791        bitblock128_t tmpAns2 = simd128<(16)>::umult(simd_and(loMask, simd128<(32)>::srli<(16)>(arg1)), simd_and(loMask, simd128<(32)>::srli<(16)>(arg2)));
792        return simd_or(tmpAns1, simd128<(32)>::slli<(16)>(tmpAns2));
793}
794
795//The total number of operations is 10.0
796template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2)
797{
798        bitblock128_t loMask = simd128<(32)>::lomask();
799        bitblock128_t tmpAns1 = simd128<(32)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
800        bitblock128_t tmpAns2 = simd128<(32)>::umult(simd_and(loMask, simd128<(64)>::srli<(32)>(arg1)), simd_and(loMask, simd128<(64)>::srli<(32)>(arg2)));
801        return simd_or(tmpAns1, simd128<(64)>::slli<(32)>(tmpAns2));
802}
803
804//The total number of operations is 1.0
805template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2)
806{
807        return _mm_mul_epu32(arg1, arg2);
808}
809
810//The total number of operations is 45.0
811template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2)
812{
813        bitblock128_t loMask1 = simd128<(128)>::lomask();
814        bitblock128_t arg11 = simd_and(arg1, loMask1);
815        bitblock128_t arg22 = simd_and(arg2, loMask1);
816        bitblock128_t loMask2 = simd128<64>::lomask();
817        bitblock128_t arg1_low = simd_and(arg11, loMask2);
818        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg11);
819        bitblock128_t arg2_low = simd_and(arg22, loMask2);
820        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg22);
821        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
822        bitblock128_t tmpAns2 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
823        bitblock128_t tmpAns3 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
824        bitblock128_t tmpAns4 = simd128<(128)>::slli<64>(simd128<(32)>::umult(arg1_high, arg2_high));
825        return simd128<(128)>::add(tmpAns1, simd128<(128)>::add(tmpAns2, simd128<(128)>::add(tmpAns3, tmpAns4)));
826}
827
828//The total number of operations is 1.0
829template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2)
830{
831        return simd_andc(arg2, arg1);
832}
833
834//The total number of operations is 13.6666666667
835template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2)
836{
837        bitblock128_t tmp = simd_not(arg1);
838        bitblock128_t tmpAns = simd_or(simd_and(tmp, arg2), simd_and(simd128<128>::slli<1>(simd_and(tmp, arg2)), simd_or(tmp, arg2)));
839        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
840}
841
842//The total number of operations is 20.0
843template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2)
844{
845        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())));
846}
847
848//The total number of operations is 7.0
849template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2)
850{
851        bitblock128_t high_bit = simd128<8>::constant<(128)>();
852        return simd128<8>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
853}
854
855//The total number of operations is 7.0
856template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2)
857{
858        bitblock128_t high_bit = simd128<16>::constant<(highbit_16)>();
859        return simd128<16>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
860}
861
862//The total number of operations is 7.0
863template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2)
864{
865        bitblock128_t high_bit = simd128<32>::constant<(highbit_32)>();
866        return simd128<32>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
867}
868
869//The total number of operations is 17.5
870template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2)
871{
872        bitblock128_t tmpAns = simd128<(32)>::ult(arg1, arg2);
873        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
874        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
875        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
876}
877
878//The total number of operations is 40.0833333333
879template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2)
880{
881        return simd_and(simd128<128>::srai<(127)>(simd_or(simd_and(simd_not(arg1), arg2), simd_and(simd_not(simd_xor(arg1, arg2)), simd128<128>::sub(arg1, arg2)))), simd_not(simd128<128>::eq(arg1, arg2)));
882}
883
884//The total number of operations is 1.0
885template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2)
886{
887        return simd_andc(arg1, arg2);
888}
889
890//The total number of operations is 14.6666666667
891template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2)
892{
893        bitblock128_t tmp = simd_not(arg2);
894        bitblock128_t tmpAns = simd_or(simd_and(arg1, tmp), simd_and(simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2)), simd_or(arg1, tmp)));
895        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
896}
897
898//The total number of operations is 18.0
899template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2)
900{
901        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)));
902}
903
904//The total number of operations is 5.0
905template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2)
906{
907        return simd_and(simd_not(simd128<8>::gt(arg1, arg2)), simd_not(simd128<8>::eq(arg1, arg2)));
908}
909
910//The total number of operations is 5.0
911template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2)
912{
913        return simd_and(simd_not(simd128<16>::gt(arg1, arg2)), simd_not(simd128<16>::eq(arg1, arg2)));
914}
915
916//The total number of operations is 5.0
917template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2)
918{
919        return simd_and(simd_not(simd128<32>::gt(arg1, arg2)), simd_not(simd128<32>::eq(arg1, arg2)));
920}
921
922//The total number of operations is 19.5
923template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2)
924{
925        bitblock128_t high_bit = simd128<64>::constant<(highbit_64)>();
926        return simd128<64>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
927}
928
929//The total number of operations is 60.75
930template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2)
931{
932        bitblock128_t hiAns = simd128<(64)>::lt(arg1, arg2);
933        bitblock128_t loAns = simd128<(64)>::ult(arg1, arg2);
934        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
935        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
936        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
937}
938
939//The total number of operations is 2.0
940template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1)
941{
942        return simd_and(simd128<32>::srli<sh>(arg1), simd128<2>::constant<((3)>>sh)>());
943}
944
945//The total number of operations is 2.0
946template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1)
947{
948        return simd_and(simd128<32>::srli<sh>(arg1), simd128<4>::constant<((15)>>sh)>());
949}
950
951//The total number of operations is 2.0
952template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1)
953{
954        return simd_and(simd128<32>::srli<sh>(arg1), simd128<8>::constant<((255)>>sh)>());
955}
956
957//The total number of operations is 1.0
958template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1)
959{
960        return _mm_srli_epi16(arg1, (int32_t)(sh));
961}
962
963//The total number of operations is 1.0
964template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1)
965{
966        return _mm_srli_epi32(arg1, (int32_t)(sh));
967}
968
969//The total number of operations is 1.0
970template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1)
971{
972        return _mm_srli_epi64(arg1, (int32_t)(sh));
973}
974
975//The total number of operations is 2.33333333333
976template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1)
977{
978        return (((sh%8) == 0) ? _mm_srli_si128(arg1, (int32_t)((sh/8))) : ((sh >= 64) ? simd128<64>::srli<(sh&63)>(_mm_srli_si128(arg1, (int32_t)(8))) : simd_or(simd128<64>::srli<sh>(arg1), _mm_srli_si128(simd128<64>::slli<((128-sh)&63)>(arg1), (int32_t)(8)))));
979}
980
981//The total number of operations is 1.0
982template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1)
983{
984        return simd_not(arg1);
985}
986
987//The total number of operations is 10.6666666667
988template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1)
989{
990        bitblock128_t tmp = simd_not(arg1);
991        return simd128<1>::ifh(simd128<2>::himask(), simd_and(tmp, simd128<128>::slli<1>(tmp)), simd_and(simd128<128>::srli<1>(arg1), tmp));
992}
993
994//The total number of operations is 14.0
995template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1)
996{
997        return simd128<4>::popcount(simd_andc(simd128<4>::sub(arg1, simd128<4>::constant<1>()), arg1));
998}
999
1000//The total number of operations is 13.0
1001template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1)
1002{
1003        return simd128<8>::popcount(simd_andc(simd128<8>::sub(arg1, simd128<8>::constant<1>()), arg1));
1004}
1005
1006//The total number of operations is 16.0
1007template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1)
1008{
1009        return simd128<16>::popcount(simd_andc(simd128<16>::sub(arg1, simd128<16>::constant<1>()), arg1));
1010}
1011
1012//The total number of operations is 19.0
1013template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1)
1014{
1015        return simd128<32>::popcount(simd_andc(simd128<32>::sub(arg1, simd128<32>::constant<1>()), arg1));
1016}
1017
1018//The total number of operations is 14.0
1019template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1)
1020{
1021        return simd128<64>::popcount(simd_andc(simd128<64>::sub(arg1, simd128<64>::constant<1>()), arg1));
1022}
1023
1024//The total number of operations is 26.6666666667
1025template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1)
1026{
1027        return simd128<128>::popcount(simd_andc(simd128<128>::sub(arg1, simd128<128>::constant<1>()), arg1));
1028}
1029
1030//The total number of operations is 10.0
1031template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1032{
1033        return simd128<1>::ifh(simd128<128>::himask(), _mm_sll_epi64(arg1, simd_and(_mm_srli_si128(shift_mask, (int32_t)(8)), _mm_cvtsi32_si128((int32_t)(63)))), _mm_sll_epi64(arg1, simd_and(shift_mask, _mm_cvtsi32_si128((int32_t)(63)))));
1034}
1035
1036//The total number of operations is 13.0
1037template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1038{
1039        bitblock128_t shift = simd_and(shift_mask, _mm_cvtsi32_si128((int32_t)(127)));
1040        return simd_or(_mm_sll_epi64(arg1, shift), simd_or(_mm_slli_si128(_mm_sll_epi64(arg1, simd128<32>::sub(shift, _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)), (int32_t)(8))));
1041}
1042
1043//The total number of operations is 1.0
1044template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1045{
1046        return simd_andc(arg1, arg2);
1047}
1048
1049//The total number of operations is 13.6666666667
1050template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1051{
1052        bitblock128_t tmp = simd_not(arg2);
1053        bitblock128_t tmpAns = simd_or(simd_and(arg1, tmp), simd_and(simd128<128>::slli<1>(simd_and(arg1, tmp)), simd_or(arg1, tmp)));
1054        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1055}
1056
1057//The total number of operations is 12.0
1058template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1059{
1060        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())));
1061}
1062
1063//The total number of operations is 3.0
1064template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1065{
1066        bitblock128_t high_bit = simd128<8>::constant<(128)>();
1067        return simd128<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1068}
1069
1070//The total number of operations is 3.0
1071template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1072{
1073        bitblock128_t high_bit = simd128<16>::constant<(highbit_16)>();
1074        return simd128<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1075}
1076
1077//The total number of operations is 3.0
1078template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1079{
1080        bitblock128_t high_bit = simd128<32>::constant<(highbit_32)>();
1081        return simd128<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1082}
1083
1084//The total number of operations is 13.5
1085template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1086{
1087        bitblock128_t tmpAns = simd128<(32)>::ugt(arg1, arg2);
1088        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
1089        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
1090        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
1091}
1092
1093//The total number of operations is 37.25
1094template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1095{
1096        bitblock128_t tmpAns = simd128<(64)>::ugt(arg1, arg2);
1097        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
1098        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
1099        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
1100}
1101
1102//The total number of operations is 4.0
1103template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1)
1104{
1105        return simd_xor(simd128<2>::srli<(1)>(arg1), simd_and(arg1, simd128<2>::lomask()));
1106}
1107
1108//The total number of operations is 4.0
1109template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1)
1110{
1111        return simd_xor(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1112}
1113
1114//The total number of operations is 4.0
1115template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1)
1116{
1117        return simd_xor(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1118}
1119
1120//The total number of operations is 3.0
1121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1)
1122{
1123        return simd_xor(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1124}
1125
1126//The total number of operations is 3.0
1127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1)
1128{
1129        return simd_xor(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1130}
1131
1132//The total number of operations is 3.0
1133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1)
1134{
1135        return simd_xor(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1136}
1137
1138//The total number of operations is 4.33333333333
1139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1)
1140{
1141        return simd_xor(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1142}
1143
1144//The total number of operations is 0
1145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1)
1146{
1147        return arg1;
1148}
1149
1150//The total number of operations is 3.0
1151template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1)
1152{
1153        return simd128<2>::add_hl(simd128<(1)>::popcount(arg1));
1154}
1155
1156//The total number of operations is 7.0
1157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1)
1158{
1159        return simd128<4>::add_hl(simd128<(2)>::popcount(arg1));
1160}
1161
1162//The total number of operations is 11.0
1163template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1)
1164{
1165        return simd128<8>::add_hl(simd128<(4)>::popcount(arg1));
1166}
1167
1168//The total number of operations is 14.0
1169template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1)
1170{
1171        return simd128<16>::add_hl(simd128<(8)>::popcount(arg1));
1172}
1173
1174//The total number of operations is 17.0
1175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1)
1176{
1177        return simd128<32>::add_hl(simd128<(16)>::popcount(arg1));
1178}
1179
1180//The total number of operations is 12.0
1181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1)
1182{
1183        return _mm_sad_epu8(simd128<8>::popcount(arg1), simd128<8>::constant<0>());
1184}
1185
1186//The total number of operations is 16.3333333333
1187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1)
1188{
1189        bitblock128_t tmpAns = simd128<(64)>::popcount(arg1);
1190        return simd128<(64)>::add(simd_and(tmpAns, simd128<128>::lomask()), simd128<128>::srli<(64)>(tmpAns));
1191}
1192
1193//The total number of operations is 6.33333333333
1194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1)
1195{
1196        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(arg1, simd128<128>::slli<1>(arg1)), arg1);
1197}
1198
1199//The total number of operations is 6.0
1200template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1)
1201{
1202        return simd128<4>::sub(simd128<4>::constant<0>(), arg1);
1203}
1204
1205//The total number of operations is 1.0
1206template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1)
1207{
1208        return simd128<8>::sub(simd128<8>::constant<0>(), arg1);
1209}
1210
1211//The total number of operations is 1.0
1212template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1)
1213{
1214        return simd128<16>::sub(simd128<16>::constant<0>(), arg1);
1215}
1216
1217//The total number of operations is 1.0
1218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1)
1219{
1220        return simd128<32>::sub(simd128<32>::constant<0>(), arg1);
1221}
1222
1223//The total number of operations is 1.0
1224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1)
1225{
1226        return simd128<64>::sub(simd128<64>::constant<0>(), arg1);
1227}
1228
1229//The total number of operations is 9.33333333333
1230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1)
1231{
1232        return simd128<128>::sub(simd128<128>::constant<0>(), arg1);
1233}
1234
1235//The total number of operations is 2.0
1236template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1)
1237{
1238        return simd_and(simd128<32>::slli<sh>(arg1), simd128<2>::constant<(((3)<<sh)&(3))>());
1239}
1240
1241//The total number of operations is 2.0
1242template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1)
1243{
1244        return simd_and(simd128<32>::slli<sh>(arg1), simd128<4>::constant<(((15)<<sh)&(15))>());
1245}
1246
1247//The total number of operations is 2.0
1248template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1)
1249{
1250        return simd_and(simd128<32>::slli<sh>(arg1), simd128<8>::constant<(((255)<<sh)&(255))>());
1251}
1252
1253//The total number of operations is 1.0
1254template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1)
1255{
1256        return _mm_slli_epi16(arg1, (int32_t)(sh));
1257}
1258
1259//The total number of operations is 1.0
1260template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1)
1261{
1262        return _mm_slli_epi32(arg1, (int32_t)(sh));
1263}
1264
1265//The total number of operations is 1.0
1266template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1)
1267{
1268        return _mm_slli_epi64(arg1, (int32_t)(sh));
1269}
1270
1271//The total number of operations is 2.33333333333
1272template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1)
1273{
1274        return (((sh%8) == 0) ? _mm_slli_si128(arg1, (int32_t)((sh/8))) : ((sh >= 64) ? simd128<64>::slli<(sh&63)>(_mm_slli_si128(arg1, (int32_t)(8))) : simd_or(simd128<64>::slli<sh>(arg1), _mm_slli_si128(simd128<64>::srli<((128-sh)&63)>(arg1), (int32_t)(8)))));
1275}
1276
1277//The total number of operations is 3.0
1278template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1279{
1280        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
1281}
1282
1283//The total number of operations is 8.0
1284template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1285{
1286        return simd128<(1)>::ifh(simd128<1>::ifh(simd128<2>::himask(), arg1, simd128<2>::srli<(1)>(arg1)), arg2, arg3);
1287}
1288
1289//The total number of operations is 13.0
1290template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1291{
1292        return simd128<1>::ifh(simd128<4>::gt(simd128<4>::constant<0>(), arg1), arg2, arg3);
1293}
1294
1295//The total number of operations is 4.0
1296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1297{
1298        return simd128<1>::ifh(simd128<8>::gt(simd128<8>::constant<0>(), arg1), arg2, arg3);
1299}
1300
1301//The total number of operations is 4.0
1302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1303{
1304        return simd128<1>::ifh(simd128<16>::gt(simd128<16>::constant<0>(), arg1), arg2, arg3);
1305}
1306
1307//The total number of operations is 4.0
1308template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1309{
1310        return simd128<1>::ifh(simd128<32>::gt(simd128<32>::constant<0>(), arg1), arg2, arg3);
1311}
1312
1313//The total number of operations is 8.0
1314template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1315{
1316        return simd128<(32)>::ifh(simd128<1>::ifh(simd128<64>::himask(), arg1, simd128<64>::srli<(32)>(arg1)), arg2, arg3);
1317}
1318
1319//The total number of operations is 13.3333333333
1320template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1321{
1322        return simd128<(64)>::ifh(simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg1)), arg2, arg3);
1323}
1324
1325//The total number of operations is 1.0
1326template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2)
1327{
1328        return simd_xor(arg1, arg2);
1329}
1330
1331//The total number of operations is 9.33333333333
1332template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2)
1333{
1334        bitblock128_t tmp = simd_xor(arg1, arg2);
1335        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2))), tmp);
1336}
1337
1338//The total number of operations is 6.0
1339template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2)
1340{
1341        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::sub(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::sub(arg1, arg2));
1342}
1343
1344//The total number of operations is 1.0
1345template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2)
1346{
1347        return _mm_sub_epi8(arg1, arg2);
1348}
1349
1350//The total number of operations is 1.0
1351template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2)
1352{
1353        return _mm_sub_epi16(arg1, arg2);
1354}
1355
1356//The total number of operations is 1.0
1357template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2)
1358{
1359        return _mm_sub_epi32(arg1, arg2);
1360}
1361
1362//The total number of operations is 1.0
1363template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2)
1364{
1365        return _mm_sub_epi64(arg1, arg2);
1366}
1367
1368//The total number of operations is 9.33333333333
1369template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2)
1370{
1371        bitblock128_t partial = simd128<(64)>::sub(arg1, arg2);
1372        bitblock128_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
1373        bitblock128_t borrow = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(borrowMask));
1374        return simd128<(64)>::sub(partial, borrow);
1375}
1376
1377//The total number of operations is 3.0
1378template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1)
1379{
1380        return simd128<16>::sub(arg1, simd_and(simd128<2>::lomask(), simd128<16>::srli<1>(arg1)));
1381}
1382
1383//The total number of operations is 4.0
1384template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1)
1385{
1386        return simd128<(8)>::add(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1387}
1388
1389//The total number of operations is 4.0
1390template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1)
1391{
1392        return simd128<(16)>::add(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1393}
1394
1395//The total number of operations is 3.0
1396template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1)
1397{
1398        return simd128<(32)>::add(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1399}
1400
1401//The total number of operations is 3.0
1402template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1)
1403{
1404        return simd128<(64)>::add(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1405}
1406
1407//The total number of operations is 3.0
1408template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1)
1409{
1410        return simd128<64>::add(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1411}
1412
1413//The total number of operations is 12.6666666667
1414template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1)
1415{
1416        return simd128<128>::add(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1417}
1418
1419//The total number of operations is 10.0
1420template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1421{
1422        return simd128<1>::ifh(simd128<128>::himask(), _mm_srl_epi64(arg1, simd_and(_mm_srli_si128(shift_mask, (int32_t)(8)), _mm_cvtsi32_si128((int32_t)(63)))), _mm_srl_epi64(arg1, simd_and(shift_mask, _mm_cvtsi32_si128((int32_t)(63)))));
1423}
1424
1425//The total number of operations is 13.0
1426template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1427{
1428        bitblock128_t shift = simd_and(shift_mask, _mm_cvtsi32_si128((int32_t)(127)));
1429        return simd_or(_mm_srl_epi64(arg1, shift), simd_or(_mm_srli_si128(_mm_srl_epi64(arg1, simd128<32>::sub(shift, _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)), (int32_t)(8))));
1430}
1431
1432//The total number of operations is 0
1433//template <> template <fw1_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1434template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1435{
1436        return simd128<32>::constant<(-1*val)>();
1437}
1438
1439//The total number of operations is 0
1440//template <> template <fw2_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1441template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1442{
1443        return simd128<(4)>::constant<((val<<2)|(val&(3)))>();
1444}
1445
1446//The total number of operations is 0
1447//template <> template <fw4_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1448template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1449{
1450        return simd128<(8)>::constant<((val<<4)|(val&(15)))>();
1451}
1452
1453//The total number of operations is 0
1454//template <> template <fw8_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1455template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1456{
1457        return _mm_set1_epi8((uint8_t)(val));
1458}
1459
1460//The total number of operations is 0
1461//template <> template <fw16_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1462template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1463{
1464        return _mm_set1_epi16((uint16_t)(val));
1465}
1466
1467//The total number of operations is 0
1468//template <> template <fw32_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1469template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1470{
1471        return _mm_set1_epi32((uint32_t)(val));
1472}
1473
1474//The total number of operations is 0
1475//template <> template <fw64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1476template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1477{
1478        return _mm_set_epi32((uint32_t)((val>>32)), (uint32_t)(val), (uint32_t)((val>>32)), (uint32_t)(val));
1479}
1480
1481//The total number of operations is 0
1482//template <> template <fw128_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1483template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1484{
1485        return _mm_set_epi32(0, 0, (uint32_t)((val>>32)), (uint32_t)(val));
1486}
1487
1488//The total number of operations is 1.0
1489template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2)
1490{
1491        return simd_or(arg1, arg2);
1492}
1493
1494//The total number of operations is 16.6666666667
1495template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2)
1496{
1497        bitblock128_t tmp1 = simd128<128>::srli<1>(arg1);
1498        bitblock128_t tmp2 = simd128<128>::srli<1>(arg2);
1499        return simd128<1>::ifh(simd128<2>::himask(), simd_or(arg1, arg2), simd_or(simd_and(arg1, simd_and(tmp1, simd_not(tmp2))), simd_and(arg2, simd_or(simd_and(simd_not(tmp1), tmp2), arg1))));
1500}
1501
1502//The total number of operations is 9.0
1503template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2)
1504{
1505        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1506        return simd_xor(simd128<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1507}
1508
1509//The total number of operations is 4.0
1510template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2)
1511{
1512        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg2, arg1);
1513}
1514
1515//The total number of operations is 1.0
1516template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2)
1517{
1518        return _mm_min_epi16(arg1, arg2);
1519}
1520
1521//The total number of operations is 4.0
1522template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2)
1523{
1524        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg2, arg1);
1525}
1526
1527//The total number of operations is 17.5
1528template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2)
1529{
1530        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg2, arg1);
1531}
1532
1533//The total number of operations is 54.75
1534template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2)
1535{
1536        return simd128<1>::ifh(simd128<128>::gt(arg1, arg2), arg2, arg1);
1537}
1538
1539//The total number of operations is 0
1540template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask()
1541{
1542        return simd128<2>::constant<(1)>();
1543}
1544
1545//The total number of operations is 0
1546template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask()
1547{
1548        return simd128<4>::constant<(3)>();
1549}
1550
1551//The total number of operations is 0
1552template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask()
1553{
1554        return simd128<8>::constant<(15)>();
1555}
1556
1557//The total number of operations is 0
1558template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask()
1559{
1560        return simd128<16>::constant<(255)>();
1561}
1562
1563//The total number of operations is 0
1564template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask()
1565{
1566        return simd128<32>::constant<(65535)>();
1567}
1568
1569//The total number of operations is 0
1570template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask()
1571{
1572        return _mm_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1));
1573}
1574
1575//The total number of operations is 0
1576template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask()
1577{
1578        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1));
1579}
1580
1581//The total number of operations is 1.0
1582template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2)
1583{
1584        return simd_and(arg1, arg2);
1585}
1586
1587//The total number of operations is 16.0
1588template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2)
1589{
1590        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)));
1591}
1592
1593//The total number of operations is 6.0
1594template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2)
1595{
1596        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)));
1597}
1598
1599//The total number of operations is 1.0
1600template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2)
1601{
1602        return _mm_min_epu8(arg1, arg2);
1603}
1604
1605//The total number of operations is 4.0
1606template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2)
1607{
1608        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1609        return simd_xor(simd128<16>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1610}
1611
1612//The total number of operations is 7.0
1613template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2)
1614{
1615        bitblock128_t high_bit = simd128<32>::constant<(highbit_32)>();
1616        return simd_xor(simd128<32>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1617}
1618
1619//The total number of operations is 20.0
1620template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2)
1621{
1622        bitblock128_t tmpAns = simd128<(32)>::umin(arg1, arg2);
1623        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1624        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1625        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1626}
1627
1628//The total number of operations is 43.6666666667
1629template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2)
1630{
1631        bitblock128_t tmpAns = simd128<(64)>::umin(arg1, arg2);
1632        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1633        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1634        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1635}
1636
1637//The total number of operations is 7.33333333333
1638template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1)
1639{
1640        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, simd128<128>::slli<1>(simd_not(arg1))), arg1);
1641}
1642
1643//The total number of operations is 19.0
1644template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1)
1645{
1646        bitblock128_t gtMask = simd128<4>::gt(arg1, simd128<4>::constant<0>());
1647        return simd128<1>::ifh(gtMask, arg1, simd128<4>::sub(gtMask, arg1));
1648}
1649
1650//The total number of operations is 5.0
1651template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1)
1652{
1653        bitblock128_t gtMask = simd128<8>::gt(arg1, simd128<8>::constant<0>());
1654        return simd128<1>::ifh(gtMask, arg1, simd128<8>::sub(gtMask, arg1));
1655}
1656
1657//The total number of operations is 5.0
1658template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1)
1659{
1660        bitblock128_t gtMask = simd128<16>::gt(arg1, simd128<16>::constant<0>());
1661        return simd128<1>::ifh(gtMask, arg1, simd128<16>::sub(gtMask, arg1));
1662}
1663
1664//The total number of operations is 5.0
1665template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1)
1666{
1667        bitblock128_t gtMask = simd128<32>::gt(arg1, simd128<32>::constant<0>());
1668        return simd128<1>::ifh(gtMask, arg1, simd128<32>::sub(gtMask, arg1));
1669}
1670
1671//The total number of operations is 17.0
1672template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1)
1673{
1674        bitblock128_t eqMask = simd128<64>::eq(simd128<1>::ifh(simd128<64>::himask(), simd128<(32)>::abs(arg1), arg1), arg1);
1675        return simd128<1>::ifh(eqMask, arg1, simd128<64>::sub(eqMask, arg1));
1676}
1677
1678//The total number of operations is 44.0
1679template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1)
1680{
1681        bitblock128_t eqMask = simd128<128>::eq(simd128<1>::ifh(simd128<128>::himask(), simd128<(64)>::abs(arg1), arg1), arg1);
1682        return simd128<1>::ifh(eqMask, arg1, simd128<128>::sub(eqMask, arg1));
1683}
1684
1685//The total number of operations is 2.0
1686template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2)
1687{
1688        return simd_not(simd_xor(arg1, arg2));
1689}
1690
1691//The total number of operations is 8.0
1692template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2)
1693{
1694        bitblock128_t tmpAns = simd128<(1)>::eq(arg1, arg2);
1695        bitblock128_t loMask = simd_and(tmpAns, simd128<2>::srli<(1)>(tmpAns));
1696        bitblock128_t hiMask = simd128<2>::slli<(1)>(loMask);
1697        return simd_or(loMask, hiMask);
1698}
1699
1700//The total number of operations is 9.0
1701template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2)
1702{
1703        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))));
1704}
1705
1706//The total number of operations is 1.0
1707template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2)
1708{
1709        return _mm_cmpeq_epi8(arg1, arg2);
1710}
1711
1712//The total number of operations is 1.0
1713template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2)
1714{
1715        return _mm_cmpeq_epi16(arg1, arg2);
1716}
1717
1718//The total number of operations is 1.0
1719template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2)
1720{
1721        return _mm_cmpeq_epi32(arg1, arg2);
1722}
1723
1724//The total number of operations is 5.0
1725template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2)
1726{
1727        bitblock128_t tmpAns = simd128<(32)>::eq(arg1, arg2);
1728        bitblock128_t loMask = simd_and(tmpAns, simd128<64>::srli<(32)>(tmpAns));
1729        bitblock128_t hiMask = simd128<64>::slli<(32)>(loMask);
1730        return simd_or(loMask, hiMask);
1731}
1732
1733//The total number of operations is 11.6666666667
1734template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2)
1735{
1736        bitblock128_t tmpAns = simd128<(64)>::eq(arg1, arg2);
1737        bitblock128_t loMask = simd_and(tmpAns, simd128<128>::srli<(64)>(tmpAns));
1738        bitblock128_t hiMask = simd128<128>::slli<(64)>(loMask);
1739        return simd_or(loMask, hiMask);
1740}
1741
1742//The total number of operations is 4.0
1743template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1)
1744{
1745        return ((sh == 0) ? arg1 : simd_or(simd_and(simd128<2>::himask(), arg1), simd128<2>::srli<1>(arg1)));
1746}
1747
1748//The total number of operations is 10.0
1749template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1)
1750{
1751        bitblock128_t tmp = simd128<4>::srli<((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh))>(arg1);
1752        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)));
1753}
1754
1755//The total number of operations is 5.0
1756template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1)
1757{
1758        bitblock128_t tmp = simd128<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
1759        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)));
1760}
1761
1762//The total number of operations is 1.0
1763template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1)
1764{
1765        return _mm_srai_epi16(arg1, (int32_t)(sh));
1766}
1767
1768//The total number of operations is 1.0
1769template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1)
1770{
1771        return _mm_srai_epi32(arg1, (int32_t)(sh));
1772}
1773
1774//The total number of operations is 4.5
1775template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1)
1776{
1777        return simd_or(simd_and(simd128<64>::himask(), simd128<(32)>::srai<((sh < (32)) ? sh : (32))>(arg1)), ((sh <= (32)) ? simd128<64>::srli<sh>(arg1) : simd128<(32)>::srai<(sh-(32))>(simd128<64>::srli<(32)>(arg1))));
1778}
1779
1780//The total number of operations is 11.0833333333
1781template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1)
1782{
1783        return simd_or(simd_and(simd128<128>::himask(), simd128<(64)>::srai<((sh < (64)) ? sh : (64))>(arg1)), ((sh <= (64)) ? simd128<128>::srli<sh>(arg1) : simd128<(64)>::srai<(sh-(64))>(simd128<128>::srli<(64)>(arg1))));
1784}
1785
1786//The total number of operations is 0
1787template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask()
1788{
1789        return simd128<2>::constant<(2)>();
1790}
1791
1792//The total number of operations is 0
1793template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask()
1794{
1795        return simd128<4>::constant<(12)>();
1796}
1797
1798//The total number of operations is 0
1799template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask()
1800{
1801        return simd128<8>::constant<(240)>();
1802}
1803
1804//The total number of operations is 0
1805template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask()
1806{
1807        return simd128<16>::constant<(65280)>();
1808}
1809
1810//The total number of operations is 0
1811template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask()
1812{
1813        return simd128<32>::constant<-65536>();
1814}
1815
1816//The total number of operations is 0
1817template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask()
1818{
1819        return _mm_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0));
1820}
1821
1822//The total number of operations is 0
1823template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask()
1824{
1825        return _mm_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0));
1826}
1827
1828//The total number of operations is 1.0
1829template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2)
1830{
1831        return simd_xor(arg1, arg2);
1832}
1833
1834//The total number of operations is 8.33333333333
1835template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2)
1836{
1837        bitblock128_t tmp = simd_xor(arg1, arg2);
1838        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(arg1, arg2))), tmp);
1839}
1840
1841//The total number of operations is 6.0
1842template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2)
1843{
1844        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::add(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::add(arg1, arg2));
1845}
1846
1847//The total number of operations is 1.0
1848template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2)
1849{
1850        return _mm_add_epi8(arg1, arg2);
1851}
1852
1853//The total number of operations is 1.0
1854template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2)
1855{
1856        return _mm_add_epi16(arg1, arg2);
1857}
1858
1859//The total number of operations is 1.0
1860template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2)
1861{
1862        return _mm_add_epi32(arg1, arg2);
1863}
1864
1865//The total number of operations is 1.0
1866template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2)
1867{
1868        return _mm_add_epi64(arg1, arg2);
1869}
1870
1871//The total number of operations is 9.33333333333
1872template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2)
1873{
1874        bitblock128_t partial = simd128<(64)>::add(arg1, arg2);
1875        bitblock128_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
1876        bitblock128_t carry = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(carryMask));
1877        return simd128<(64)>::add(partial, carry);
1878}
1879
1880//The total number of operations is 1.0
1881template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2)
1882{
1883        return simd_or(arg1, arg2);
1884}
1885
1886//The total number of operations is 15.6666666667
1887template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2)
1888{
1889        return simd128<1>::ifh(simd128<2>::himask(), simd_or(arg1, arg2), simd_or(simd_and(arg2, simd128<128>::srli<1>(simd_or(simd_not(arg1), arg2))), simd_and(arg1, simd128<128>::srli<1>(simd_or(arg1, simd_not(arg2))))));
1890}
1891
1892//The total number of operations is 6.0
1893template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2)
1894{
1895        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)));
1896}
1897
1898//The total number of operations is 1.0
1899template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2)
1900{
1901        return _mm_max_epu8(arg1, arg2);
1902}
1903
1904//The total number of operations is 4.0
1905template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2)
1906{
1907        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1908        return simd_xor(simd128<16>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1909}
1910
1911//The total number of operations is 7.0
1912template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2)
1913{
1914        bitblock128_t high_bit = simd128<32>::constant<(highbit_32)>();
1915        return simd_xor(simd128<32>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1916}
1917
1918//The total number of operations is 20.0
1919template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2)
1920{
1921        bitblock128_t tmpAns = simd128<(32)>::umax(arg1, arg2);
1922        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1923        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1924        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1925}
1926
1927//The total number of operations is 43.6666666667
1928template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2)
1929{
1930        bitblock128_t tmpAns = simd128<(64)>::umax(arg1, arg2);
1931        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1932        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1933        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1934}
1935
1936//The total number of operations is 73.0
1937template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1938{
1939        return simd128<(1)>::umin(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1940}
1941
1942//The total number of operations is 66.6666666667
1943template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1944{
1945        return simd128<(2)>::umin(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1946}
1947
1948//The total number of operations is 35.3333333333
1949template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1950{
1951        return simd128<(4)>::umin(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1952}
1953
1954//The total number of operations is 7.0
1955template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1956{
1957        return simd128<(8)>::umin(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1958}
1959
1960//The total number of operations is 37.3333333333
1961template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1962{
1963        return simd128<(16)>::umin(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1964}
1965
1966//The total number of operations is 19.0
1967template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1968{
1969        return simd128<(32)>::umin(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1970}
1971
1972//The total number of operations is 30.6666666667
1973template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1974{
1975        return simd128<(64)>::umin(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1976}
1977
1978//The total number of operations is 73.0
1979template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1980{
1981        return simd128<(1)>::add(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1982}
1983
1984//The total number of operations is 59.0
1985template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1986{
1987        return simd128<(2)>::add(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1988}
1989
1990//The total number of operations is 35.3333333333
1991template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1992{
1993        return simd128<(4)>::add(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1994}
1995
1996//The total number of operations is 7.0
1997template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1998{
1999        return simd128<(8)>::add(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2000}
2001
2002//The total number of operations is 34.3333333333
2003template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
2004{
2005        return simd128<(16)>::add(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2006}
2007
2008//The total number of operations is 13.0
2009template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
2010{
2011        return simd128<(32)>::add(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2012}
2013
2014//The total number of operations is 11.6666666667
2015template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
2016{
2017        return simd128<(64)>::add(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2018}
2019
2020//The total number of operations is 108.666666667
2021template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2)
2022{
2023        bitblock128_t hiBound = simd128<2>::srli<1>(simd128<2>::lomask());
2024        bitblock128_t loBound = simd_not(hiBound);
2025        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)));
2026}
2027
2028//The total number of operations is 79.3333333333
2029template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2)
2030{
2031        bitblock128_t hiBound = simd128<4>::srli<1>(simd128<4>::lomask());
2032        bitblock128_t loBound = simd_not(hiBound);
2033        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)));
2034}
2035
2036//The total number of operations is 32.6666666667
2037template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2)
2038{
2039        bitblock128_t hiBound = simd128<8>::srli<1>(simd128<8>::lomask());
2040        bitblock128_t loBound = simd_not(hiBound);
2041        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)));
2042}
2043
2044//The total number of operations is 1.0
2045template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2)
2046{
2047        return _mm_packs_epi16(arg2, arg1);
2048}
2049
2050//The total number of operations is 1.0
2051template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2)
2052{
2053        return _mm_packs_epi32(arg2, arg1);
2054}
2055
2056//The total number of operations is 77.0
2057template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2)
2058{
2059        bitblock128_t hiBound = simd128<64>::srli<1>(simd128<64>::lomask());
2060        bitblock128_t loBound = simd_not(hiBound);
2061        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)));
2062}
2063
2064//The total number of operations is 227.666666667
2065template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2)
2066{
2067        bitblock128_t hiBound = simd128<128>::srli<1>(simd128<128>::lomask());
2068        bitblock128_t loBound = simd_not(hiBound);
2069        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)));
2070}
2071
2072//The total number of operations is 24.0
2073template <> IDISA_ALWAYS_INLINE FieldType<128/4>::T hsimd128<4>::signmask(bitblock128_t arg1)
2074{
2075        uint32_t tmpAns1 = hsimd128<(8)>::signmask(esimd128<4>::mergeh(arg1, simd128<4>::constant<0>()));
2076        uint32_t tmpAns2 = hsimd128<(8)>::signmask(esimd128<4>::mergel(arg1, simd128<4>::constant<0>()));
2077        return ((tmpAns1<<(16))+tmpAns2);
2078}
2079
2080//The total number of operations is 1.0
2081template <> IDISA_ALWAYS_INLINE FieldType<128/8>::T hsimd128<8>::signmask(bitblock128_t arg1)
2082{
2083        return _mm_movemask_epi8(arg1);
2084}
2085
2086//The total number of operations is 2.0
2087template <> IDISA_ALWAYS_INLINE FieldType<128/16>::T hsimd128<16>::signmask(bitblock128_t arg1)
2088{
2089        return hsimd128<(8)>::signmask(hsimd128<16>::packss(simd128<16>::constant<0>(), arg1));
2090}
2091
2092//The total number of operations is 1.0
2093template <> IDISA_ALWAYS_INLINE FieldType<128/32>::T hsimd128<32>::signmask(bitblock128_t arg1)
2094{
2095        return _mm_movemask_ps(_mm_castsi128_ps(arg1));
2096}
2097
2098//The total number of operations is 1.0
2099template <> IDISA_ALWAYS_INLINE FieldType<128/64>::T hsimd128<64>::signmask(bitblock128_t arg1)
2100{
2101        return _mm_movemask_pd(_mm_castsi128_pd(arg1));
2102}
2103
2104//The total number of operations is 6.33333333333
2105template <> IDISA_ALWAYS_INLINE FieldType<128/128>::T hsimd128<128>::signmask(bitblock128_t arg1)
2106{
2107        return hsimd128<(64)>::signmask(hsimd128<128>::packh(simd128<128>::constant<0>(), arg1));
2108}
2109
2110//The total number of operations is 35.0
2111template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2)
2112{
2113        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));
2114}
2115
2116//The total number of operations is 24.3333333333
2117template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2)
2118{
2119        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));
2120}
2121
2122//The total number of operations is 13.6666666667
2123template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2)
2124{
2125        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));
2126}
2127
2128//The total number of operations is 3.0
2129template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2)
2130{
2131        return hsimd128<16>::packus(simd_and(arg1, simd128<16>::lomask()), simd_and(arg2, simd128<16>::lomask()));
2132}
2133
2134//The total number of operations is 15.6666666667
2135template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2)
2136{
2137        return hsimd128<(64)>::packl(simd128<1>::ifh(simd128<32>::himask(), simd128<128>::srli<(16)>(arg1), arg1), simd128<1>::ifh(simd128<32>::himask(), simd128<128>::srli<(16)>(arg2), arg2));
2138}
2139
2140//The total number of operations is 5.0
2141template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2)
2142{
2143        return simd_or(mvmd128<32>::shufflei<shufflemask4(2, 0, 3, 3)>(simd_andc(arg1, simd128<64>::himask())), mvmd128<32>::shufflei<shufflemask4(3, 3, 2, 0)>(simd_andc(arg2, simd128<64>::himask())));
2144}
2145
2146//The total number of operations is 5.33333333333
2147template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2)
2148{
2149        return simd128<1>::ifh(simd128<128>::himask(), simd128<128>::slli<(64)>(arg1), arg2);
2150}
2151
2152//The total number of operations is 37.0
2153template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2)
2154{
2155        return hsimd128<2>::packl(simd128<64>::srli<(1)>(arg1), simd128<64>::srli<(1)>(arg2));
2156}
2157
2158//The total number of operations is 26.3333333333
2159template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2)
2160{
2161        return hsimd128<4>::packl(simd128<64>::srli<(2)>(arg1), simd128<64>::srli<(2)>(arg2));
2162}
2163
2164//The total number of operations is 15.6666666667
2165template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2)
2166{
2167        return hsimd128<8>::packl(simd128<64>::srli<(4)>(arg1), simd128<64>::srli<(4)>(arg2));
2168}
2169
2170//The total number of operations is 3.0
2171template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2)
2172{
2173        return hsimd128<16>::packus(simd128<16>::srli<(8)>(arg1), simd128<16>::srli<(8)>(arg2));
2174}
2175
2176//The total number of operations is 17.6666666667
2177template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2)
2178{
2179        return hsimd128<32>::packl(simd128<64>::srli<(16)>(arg1), simd128<64>::srli<(16)>(arg2));
2180}
2181
2182//The total number of operations is 7.0
2183template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2)
2184{
2185        return hsimd128<64>::packl(simd128<64>::srli<(32)>(arg1), simd128<64>::srli<(32)>(arg2));
2186}
2187
2188//The total number of operations is 5.33333333333
2189template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2)
2190{
2191        return simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg2));
2192}
2193
2194//The total number of operations is 73.0
2195template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2196{
2197        return simd128<(1)>::min(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
2198}
2199
2200//The total number of operations is 67.3333333333
2201template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2202{
2203        return simd128<(2)>::min(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
2204}
2205
2206//The total number of operations is 38.3333333333
2207template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2208{
2209        return simd128<(4)>::min(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
2210}
2211
2212//The total number of operations is 10.0
2213template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2214{
2215        return simd128<(8)>::min(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2216}
2217
2218//The total number of operations is 34.3333333333
2219template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2220{
2221        return simd128<(16)>::min(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2222}
2223
2224//The total number of operations is 16.0
2225template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2226{
2227        return simd128<(32)>::min(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2228}
2229
2230//The total number of operations is 28.1666666667
2231template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2232{
2233        return simd128<(64)>::min(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2234}
2235
2236//The total number of operations is 75.0
2237template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2)
2238{
2239        bitblock128_t arg11 = simd128<2>::ifh(arg1, simd128<2>::constant<0>(), arg1);
2240        bitblock128_t arg12 = simd_and(simd128<2>::lomask(), arg11);
2241        bitblock128_t arg21 = simd128<2>::ifh(arg2, simd128<2>::constant<0>(), arg2);
2242        bitblock128_t arg22 = simd_and(simd128<2>::lomask(), arg21);
2243        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()));
2244}
2245
2246//The total number of operations is 74.3333333333
2247template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2)
2248{
2249        bitblock128_t hiPart = hsimd128<4>::packh(arg1, arg2);
2250        return simd128<(2)>::ifh(hiPart, simd128<(2)>::constant<0>(), simd_or(simd128<(2)>::gt(hiPart, simd128<(2)>::constant<0>()), hsimd128<4>::packl(arg1, arg2)));
2251}
2252
2253//The total number of operations is 31.6666666667
2254template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2)
2255{
2256        bitblock128_t arg11 = simd128<8>::ifh(arg1, simd128<8>::constant<0>(), arg1);
2257        bitblock128_t arg12 = simd_and(simd128<8>::lomask(), arg11);
2258        bitblock128_t arg21 = simd128<8>::ifh(arg2, simd128<8>::constant<0>(), arg2);
2259        bitblock128_t arg22 = simd_and(simd128<8>::lomask(), arg21);
2260        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()));
2261}
2262
2263//The total number of operations is 1.0
2264template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2)
2265{
2266        return _mm_packus_epi16(arg2, arg1);
2267}
2268
2269//The total number of operations is 33.6666666667
2270template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2)
2271{
2272        bitblock128_t arg11 = simd128<32>::ifh(arg1, simd128<32>::constant<0>(), arg1);
2273        bitblock128_t arg12 = simd_and(simd128<32>::lomask(), arg11);
2274        bitblock128_t arg21 = simd128<32>::ifh(arg2, simd128<32>::constant<0>(), arg2);
2275        bitblock128_t arg22 = simd_and(simd128<32>::lomask(), arg21);
2276        return hsimd128<32>::packl(simd128<1>::ifh(simd128<32>::eq(arg12, arg11), arg12, simd128<32>::lomask()), simd128<1>::ifh(simd128<32>::eq(arg22, arg21), arg22, simd128<32>::lomask()));
2277}
2278
2279//The total number of operations is 18.0
2280template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2)
2281{
2282        bitblock128_t hiPart = hsimd128<64>::packh(arg1, arg2);
2283        return simd128<(32)>::ifh(hiPart, simd128<(32)>::constant<0>(), simd_or(simd128<(32)>::gt(hiPart, simd128<(32)>::constant<0>()), hsimd128<64>::packl(arg1, arg2)));
2284}
2285
2286//The total number of operations is 34.1666666667
2287template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2)
2288{
2289        bitblock128_t hiPart = hsimd128<128>::packh(arg1, arg2);
2290        return simd128<(64)>::ifh(hiPart, simd128<(64)>::constant<0>(), simd_or(simd128<(64)>::gt(hiPart, simd128<(64)>::constant<0>()), hsimd128<128>::packl(arg1, arg2)));
2291}
2292
2293//The total number of operations is 31.0
2294template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2295{
2296        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));
2297}
2298
2299//The total number of operations is 21.0
2300template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2301{
2302        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));
2303}
2304
2305//The total number of operations is 11.0
2306template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2307{
2308        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));
2309}
2310
2311//The total number of operations is 1.0
2312template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2313{
2314        return _mm_unpacklo_epi8(arg2, arg1);
2315}
2316
2317//The total number of operations is 1.0
2318template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2319{
2320        return _mm_unpacklo_epi16(arg2, arg1);
2321}
2322
2323//The total number of operations is 1.0
2324template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2325{
2326        return _mm_unpacklo_epi32(arg2, arg1);
2327}
2328
2329//The total number of operations is 1.0
2330template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2331{
2332        return _mm_unpacklo_epi64(arg2, arg1);
2333}
2334
2335//The total number of operations is 31.0
2336template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2337{
2338        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));
2339}
2340
2341//The total number of operations is 21.0
2342template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2343{
2344        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));
2345}
2346
2347//The total number of operations is 11.0
2348template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2349{
2350        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));
2351}
2352
2353//The total number of operations is 1.0
2354template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2355{
2356        return _mm_unpackhi_epi8(arg2, arg1);
2357}
2358
2359//The total number of operations is 1.0
2360template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2361{
2362        return _mm_unpackhi_epi16(arg2, arg1);
2363}
2364
2365//The total number of operations is 1.0
2366template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2367{
2368        return _mm_unpackhi_epi32(arg2, arg1);
2369}
2370
2371//The total number of operations is 1.0
2372template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2373{
2374        return _mm_unpackhi_epi64(arg2, arg1);
2375}
2376
2377//The total number of operations is 24.0
2378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1)
2379{
2380        return esimd128<(2)>::mergeh(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2381}
2382
2383//The total number of operations is 14.0
2384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1)
2385{
2386        return esimd128<(4)>::mergeh(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2387}
2388
2389//The total number of operations is 4.0
2390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1)
2391{
2392        return esimd128<(8)>::mergeh(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2393}
2394
2395//The total number of operations is 3.0
2396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1)
2397{
2398        return esimd128<(16)>::mergeh(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2399}
2400
2401//The total number of operations is 3.0
2402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1)
2403{
2404        return esimd128<(32)>::mergeh(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2405}
2406
2407//The total number of operations is 3.0
2408template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1)
2409{
2410        return esimd128<(64)>::mergeh(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2411}
2412
2413//The total number of operations is 2.33333333333
2414template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1)
2415{
2416        return simd128<(128)>::srli<64>(arg1);
2417}
2418
2419//The total number of operations is 24.0
2420template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1)
2421{
2422        return esimd128<(2)>::mergel(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2423}
2424
2425//The total number of operations is 14.0
2426template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1)
2427{
2428        return esimd128<(4)>::mergel(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2429}
2430
2431//The total number of operations is 4.0
2432template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1)
2433{
2434        return esimd128<(8)>::mergel(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2435}
2436
2437//The total number of operations is 3.0
2438template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1)
2439{
2440        return esimd128<(16)>::mergel(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2441}
2442
2443//The total number of operations is 3.0
2444template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1)
2445{
2446        return esimd128<(32)>::mergel(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2447}
2448
2449//The total number of operations is 3.0
2450template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1)
2451{
2452        return esimd128<(64)>::mergel(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2453}
2454
2455//The total number of operations is 1.0
2456template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1)
2457{
2458        return simd_and(simd128<(128)>::lomask(), arg1);
2459}
2460
2461//The total number of operations is 31.0
2462template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1)
2463{
2464        return esimd128<(2)>::mergeh(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2465}
2466
2467//The total number of operations is 33.0
2468template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1)
2469{
2470        return esimd128<(4)>::mergeh(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2471}
2472
2473//The total number of operations is 13.0
2474template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1)
2475{
2476        return esimd128<(8)>::mergeh(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2477}
2478
2479//The total number of operations is 4.0
2480template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1)
2481{
2482        return esimd128<(16)>::mergeh(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2483}
2484
2485//The total number of operations is 4.0
2486template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1)
2487{
2488        return esimd128<(32)>::mergeh(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2489}
2490
2491//The total number of operations is 11.0
2492template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1)
2493{
2494        return esimd128<(64)>::mergeh(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2495}
2496
2497//The total number of operations is 11.0833333333
2498template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1)
2499{
2500        return simd128<(128)>::srai<64>(arg1);
2501}
2502
2503//The total number of operations is 31.0
2504template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1)
2505{
2506        return esimd128<(2)>::mergel(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2507}
2508
2509//The total number of operations is 33.0
2510template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1)
2511{
2512        return esimd128<(4)>::mergel(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2513}
2514
2515//The total number of operations is 13.0
2516template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1)
2517{
2518        return esimd128<(8)>::mergel(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2519}
2520
2521//The total number of operations is 4.0
2522template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1)
2523{
2524        return esimd128<(16)>::mergel(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2525}
2526
2527//The total number of operations is 4.0
2528template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1)
2529{
2530        return esimd128<(32)>::mergel(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2531}
2532
2533//The total number of operations is 11.0
2534template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1)
2535{
2536        return esimd128<(64)>::mergel(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2537}
2538
2539//The total number of operations is 13.4166666667
2540template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1)
2541{
2542        return simd128<(128)>::srai<64>(simd128<(128)>::slli<64>(arg1));
2543}
2544
2545//The total number of operations is 13.6666666667
2546template <> template <uint8_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1)
2547{
2548        bitblock128_t tmphi = _mm_shufflehi_epi16(arg1, (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2549        bitblock128_t tmpAns = _mm_shufflelo_epi16(tmphi, (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2550        bitblock128_t tmplh = _mm_shufflehi_epi16(simd128<128>::slli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2551        bitblock128_t tmphl = _mm_shufflelo_epi16(simd128<128>::srli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2552        uint32_t a1 = ((((msk>>21)&4) == 0) ? 0 : (131071));
2553        uint32_t a2 = ((((msk>>18)&4) == 0) ? 0 : (131071));
2554        uint32_t a3 = ((((msk>>15)&4) == 0) ? 0 : (131071));
2555        uint32_t a4 = ((((msk>>12)&4) == 0) ? 0 : (131071));
2556        uint32_t a5 = ((((msk>>9)&4) == 0) ? (131071) : 0);
2557        uint32_t a6 = ((((msk>>6)&4) == 0) ? (131071) : 0);
2558        uint32_t a7 = ((((msk>>3)&4) == 0) ? (131071) : 0);
2559        uint32_t a8 = (((msk&4) == 0) ? (131071) : 0);
2560        return simd128<1>::ifh(mvmd128<16>::fill8(a1, a2, a3, a4, a5, a6, a7, a8), tmpAns, simd_or(tmplh, tmphl));
2561}
2562
2563//The total number of operations is 1.0
2564template <> template <uint8_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1)
2565{
2566        return _mm_shuffle_epi32(arg1, (int32_t)(msk));
2567}
2568
2569//The total number of operations is 1.0
2570template <> template <uint8_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1)
2571{
2572        return mvmd128<32>::shufflei<shufflemask4_from_shufflemask2(msk)>(arg1);
2573}
2574
2575//The total number of operations is 5.66666666667
2576template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2577{
2578        return simd_or(mvmd128<2>::srli<sh>(arg1), mvmd128<2>::slli<((64)-sh)>(arg2));
2579}
2580
2581//The total number of operations is 5.66666666667
2582template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2583{
2584        return simd_or(mvmd128<4>::srli<sh>(arg1), mvmd128<4>::slli<((32)-sh)>(arg2));
2585}
2586
2587//The total number of operations is 3.0
2588template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2589{
2590        return simd_or(mvmd128<8>::srli<sh>(arg1), mvmd128<8>::slli<((16)-sh)>(arg2));
2591}
2592
2593//The total number of operations is 3.0
2594template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2595{
2596        return simd_or(mvmd128<16>::srli<sh>(arg1), mvmd128<16>::slli<((8)-sh)>(arg2));
2597}
2598
2599//The total number of operations is 3.0
2600template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2601{
2602        return simd_or(mvmd128<32>::srli<sh>(arg1), mvmd128<32>::slli<((4)-sh)>(arg2));
2603}
2604
2605//The total number of operations is 3.0
2606template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2607{
2608        return simd_or(mvmd128<64>::srli<sh>(arg1), mvmd128<64>::slli<((2)-sh)>(arg2));
2609}
2610
2611//The total number of operations is 3.0
2612template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2613{
2614        return simd_or(mvmd128<128>::srli<sh>(arg1), mvmd128<128>::slli<((1)-sh)>(arg2));
2615}
2616
2617//The total number of operations is 1.0
2618template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(fw1_t val1)
2619{
2620        return mvmd128<32>::fill((-1*val1));
2621}
2622
2623//The total number of operations is 1.0
2624template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(fw2_t val1)
2625{
2626        return mvmd128<(4)>::fill(((val1<<2)|val1));
2627}
2628
2629//The total number of operations is 1.0
2630template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(fw4_t val1)
2631{
2632        return mvmd128<(8)>::fill(((val1<<4)|val1));
2633}
2634
2635//The total number of operations is 1.0
2636template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(fw8_t val1)
2637{
2638        return _mm_set1_epi8((int32_t)(val1));
2639}
2640
2641//The total number of operations is 1.0
2642template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(fw16_t val1)
2643{
2644        return _mm_set1_epi16((int32_t)(val1));
2645}
2646
2647//The total number of operations is 1.0
2648template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(fw32_t val1)
2649{
2650        return _mm_set1_epi32((int32_t)(val1));
2651}
2652
2653//The total number of operations is 1.0
2654template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(fw64_t val1)
2655{
2656        return _mm_set_epi32((int32_t)((val1>>32)), (int32_t)(val1), (int32_t)((val1>>32)), (int32_t)(val1));
2657}
2658
2659//The total number of operations is 1.0
2660template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(fw128_t val1)
2661{
2662        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val1>>32)), (int32_t)(val1));
2663}
2664
2665//The total number of operations is 1.0
2666template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw1_t mvmd128<1>::extract(bitblock128_t arg1)
2667{
2668        return (((pos%2) == 0) ? (mvmd128<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd128<(2)>::extract<(pos/2)>(arg1)>>1));
2669}
2670
2671//The total number of operations is 1.0
2672template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw2_t mvmd128<2>::extract(bitblock128_t arg1)
2673{
2674        return (((pos%2) == 0) ? (mvmd128<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd128<(4)>::extract<(pos/2)>(arg1)>>2));
2675}
2676
2677//The total number of operations is 1.0
2678template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw4_t mvmd128<4>::extract(bitblock128_t arg1)
2679{
2680        return (((pos%2) == 0) ? (mvmd128<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd128<(8)>::extract<(pos/2)>(arg1)>>4));
2681}
2682
2683//The total number of operations is 1.0
2684template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw8_t mvmd128<8>::extract(bitblock128_t arg1)
2685{
2686        return (((pos%2) == 0) ? (mvmd128<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd128<(16)>::extract<(pos/2)>(arg1)>>8));
2687}
2688
2689//The total number of operations is 1.0
2690template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw16_t mvmd128<16>::extract(bitblock128_t arg1)
2691{
2692        return (65535&_mm_extract_epi16(arg1, (int32_t)(pos)));
2693}
2694
2695//The total number of operations is 2.0
2696template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw32_t mvmd128<32>::extract(bitblock128_t arg1)
2697{
2698        return ((mvmd128<(16)>::extract<((2*pos)+1)>(arg1)<<(16))|mvmd128<(16)>::extract<(2*pos)>(arg1));
2699}
2700
2701//The total number of operations is 4.0
2702template <> template <uint8_t pos> IDISA_ALWAYS_INLINE fw64_t mvmd128<64>::extract(bitblock128_t arg1)
2703{
2704        return (((fw64_t)(mvmd128<(32)>::extract<((2*pos)+1)>(arg1))<<(32))|mvmd128<(32)>::extract<(2*pos)>(arg1));
2705}
2706
2707//The total number of operations is 12.6666666667
2708template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1)
2709{
2710        return simd128<128>::sub(simd128<128>::constant<0>(), simd_and(simd128<128>::constant<1>(), simd128<128>::srli<pos>(arg1)));
2711}
2712
2713//The total number of operations is 13.0
2714template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1)
2715{
2716        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(4)>::slli<2>(arg1) : simd128<(4)>::srli<2>(arg1));
2717        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(4)>::lomask(), arg1) : simd_and(simd128<(4)>::himask(), arg1));
2718        return mvmd128<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2719}
2720
2721//The total number of operations is 9.0
2722template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1)
2723{
2724        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(8)>::slli<4>(arg1) : simd128<(8)>::srli<4>(arg1));
2725        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(8)>::lomask(), arg1) : simd_and(simd128<(8)>::himask(), arg1));
2726        return mvmd128<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2727}
2728
2729//The total number of operations is 5.0
2730template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1)
2731{
2732        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(16)>::slli<8>(arg1) : simd128<(16)>::srli<8>(arg1));
2733        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(16)>::lomask(), arg1) : simd_and(simd128<(16)>::himask(), arg1));
2734        return mvmd128<(16)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2735}
2736
2737//The total number of operations is 2.0
2738template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1)
2739{
2740        return mvmd128<16>::fill(_mm_extract_epi16(arg1, (int32_t)(pos)));
2741}
2742
2743//The total number of operations is 1.0
2744template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1)
2745{
2746        return mvmd128<32>::shufflei<shufflemask4(pos, pos, pos, pos)>(arg1);
2747}
2748
2749//The total number of operations is 5.0
2750template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1)
2751{
2752        return simd128<1>::ifh(simd128<64>::himask(), mvmd128<(32)>::splat<((2*pos)+1)>(arg1), mvmd128<(32)>::splat<(2*pos)>(arg1));
2753}
2754
2755//The total number of operations is 13.0
2756template <> template <uint8_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1)
2757{
2758        return simd128<1>::ifh(simd128<128>::himask(), mvmd128<(64)>::splat<((2*pos)+1)>(arg1), mvmd128<(64)>::splat<(2*pos)>(arg1));
2759}
2760
2761//The total number of operations is 15.0
2762template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill16(fw1_t val1, fw1_t val2, fw1_t val3, fw1_t val4, fw1_t val5, fw1_t val6, fw1_t val7, fw1_t val8, fw1_t val9, fw1_t val10, fw1_t val11, fw1_t val12, fw1_t val13, fw1_t val14, fw1_t val15, fw1_t val16)
2763{
2764        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))));
2765}
2766
2767//The total number of operations is 7.0
2768template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill16(fw2_t val1, fw2_t val2, fw2_t val3, fw2_t val4, fw2_t val5, fw2_t val6, fw2_t val7, fw2_t val8, fw2_t val9, fw2_t val10, fw2_t val11, fw2_t val12, fw2_t val13, fw2_t val14, fw2_t val15, fw2_t val16)
2769{
2770        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))));
2771}
2772
2773//The total number of operations is 3.0
2774template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill16(fw4_t val1, fw4_t val2, fw4_t val3, fw4_t val4, fw4_t val5, fw4_t val6, fw4_t val7, fw4_t val8, fw4_t val9, fw4_t val10, fw4_t val11, fw4_t val12, fw4_t val13, fw4_t val14, fw4_t val15, fw4_t val16)
2775{
2776        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))));
2777}
2778
2779//The total number of operations is 1.0
2780template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill16(fw8_t val1, fw8_t val2, fw8_t val3, fw8_t val4, fw8_t val5, fw8_t val6, fw8_t val7, fw8_t val8, fw8_t val9, fw8_t val10, fw8_t val11, fw8_t val12, fw8_t val13, fw8_t val14, fw8_t val15, fw8_t val16)
2781{
2782        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));
2783}
2784
2785//The total number of operations is 5.0
2786template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(fw1_t val1, fw1_t val2, fw1_t val3, fw1_t val4)
2787{
2788        return simd128<1>::ifh(simd128<(4)>::himask(), mvmd128<1>::fill2(val1, val2), mvmd128<1>::fill2(val3, val4));
2789}
2790
2791//The total number of operations is 5.0
2792template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(fw2_t val1, fw2_t val2, fw2_t val3, fw2_t val4)
2793{
2794        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<2>::fill2(val1, val2), mvmd128<2>::fill2(val3, val4));
2795}
2796
2797//The total number of operations is 5.0
2798template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(fw4_t val1, fw4_t val2, fw4_t val3, fw4_t val4)
2799{
2800        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<4>::fill2(val1, val2), mvmd128<4>::fill2(val3, val4));
2801}
2802
2803//The total number of operations is 5.0
2804template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(fw8_t val1, fw8_t val2, fw8_t val3, fw8_t val4)
2805{
2806        return simd128<1>::ifh(simd128<(32)>::himask(), mvmd128<8>::fill2(val1, val2), mvmd128<8>::fill2(val3, val4));
2807}
2808
2809//The total number of operations is 3.0
2810template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(fw16_t val1, fw16_t val2, fw16_t val3, fw16_t val4)
2811{
2812        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))));
2813}
2814
2815//The total number of operations is 1.0
2816template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(fw32_t val1, fw32_t val2, fw32_t val3, fw32_t val4)
2817{
2818        return _mm_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
2819}
2820
2821//The total number of operations is 2.33333333333
2822template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1)
2823{
2824        return simd128<128>::srli<(sh*2)>(arg1);
2825}
2826
2827//The total number of operations is 2.33333333333
2828template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1)
2829{
2830        return simd128<128>::srli<(sh*4)>(arg1);
2831}
2832
2833//The total number of operations is 1.0
2834template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1)
2835{
2836        return _mm_srli_si128(arg1, (int32_t)(sh));
2837}
2838
2839//The total number of operations is 1.0
2840template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1)
2841{
2842        return mvmd128<(8)>::srli<(sh*2)>(arg1);
2843}
2844
2845//The total number of operations is 1.0
2846template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1)
2847{
2848        return mvmd128<(16)>::srli<(sh*2)>(arg1);
2849}
2850
2851//The total number of operations is 1.0
2852template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1)
2853{
2854        return mvmd128<(32)>::srli<(sh*2)>(arg1);
2855}
2856
2857//The total number of operations is 1.0
2858template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1)
2859{
2860        return mvmd128<(64)>::srli<(sh*2)>(arg1);
2861}
2862
2863//The total number of operations is 1.0
2864template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(fw1_t val1, fw1_t val2)
2865{
2866        return mvmd128<(2)>::fill(((val1<<1)|(val2&(1))));
2867}
2868
2869//The total number of operations is 1.0
2870template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(fw2_t val1, fw2_t val2)
2871{
2872        return mvmd128<(4)>::fill(((val1<<2)|(val2&(3))));
2873}
2874
2875//The total number of operations is 1.0
2876template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(fw4_t val1, fw4_t val2)
2877{
2878        return mvmd128<(8)>::fill(((val1<<4)|(val2&(15))));
2879}
2880
2881//The total number of operations is 1.0
2882template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(fw8_t val1, fw8_t val2)
2883{
2884        return mvmd128<(16)>::fill(((val1<<8)|(val2&(255))));
2885}
2886
2887//The total number of operations is 1.0
2888template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(fw16_t val1, fw16_t val2)
2889{
2890        return mvmd128<(32)>::fill(((val1<<16)|(val2&(65535))));
2891}
2892
2893//The total number of operations is 5.0
2894template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(fw32_t val1, fw32_t val2)
2895{
2896        return simd128<1>::ifh(simd128<(64)>::himask(), mvmd128<32>::fill(val1), mvmd128<32>::fill(val2));
2897}
2898
2899//The total number of operations is 5.0
2900template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(fw64_t val1, fw64_t val2)
2901{
2902        return simd128<1>::ifh(simd128<(128)>::himask(), mvmd128<64>::fill(val1), mvmd128<64>::fill(val2));
2903}
2904
2905//The total number of operations is 5.66666666667
2906template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2907{
2908        return simd_or(mvmd128<2>::slli<sh>(arg1), mvmd128<2>::srli<((64)-sh)>(arg2));
2909}
2910
2911//The total number of operations is 5.66666666667
2912template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2913{
2914        return simd_or(mvmd128<4>::slli<sh>(arg1), mvmd128<4>::srli<((32)-sh)>(arg2));
2915}
2916
2917//The total number of operations is 3.0
2918template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2919{
2920        return simd_or(mvmd128<8>::slli<sh>(arg1), mvmd128<8>::srli<((16)-sh)>(arg2));
2921}
2922
2923//The total number of operations is 3.0
2924template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2925{
2926        return simd_or(mvmd128<16>::slli<sh>(arg1), mvmd128<16>::srli<((8)-sh)>(arg2));
2927}
2928
2929//The total number of operations is 3.0
2930template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2931{
2932        return simd_or(mvmd128<32>::slli<sh>(arg1), mvmd128<32>::srli<((4)-sh)>(arg2));
2933}
2934
2935//The total number of operations is 3.0
2936template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2937{
2938        return simd_or(mvmd128<64>::slli<sh>(arg1), mvmd128<64>::srli<((2)-sh)>(arg2));
2939}
2940
2941//The total number of operations is 3.0
2942template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2943{
2944        return simd_or(mvmd128<128>::slli<sh>(arg1), mvmd128<128>::srli<((1)-sh)>(arg2));
2945}
2946
2947//The total number of operations is 2.33333333333
2948template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1)
2949{
2950        return simd128<128>::slli<(sh*2)>(arg1);
2951}
2952
2953//The total number of operations is 2.33333333333
2954template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1)
2955{
2956        return mvmd128<(2)>::slli<(sh*2)>(arg1);
2957}
2958
2959//The total number of operations is 1.0
2960template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1)
2961{
2962        return _mm_slli_si128(arg1, (int32_t)(sh));
2963}
2964
2965//The total number of operations is 1.0
2966template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1)
2967{
2968        return mvmd128<(8)>::slli<(sh*2)>(arg1);
2969}
2970
2971//The total number of operations is 1.0
2972template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1)
2973{
2974        return mvmd128<(16)>::slli<(sh*2)>(arg1);
2975}
2976
2977//The total number of operations is 1.0
2978template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1)
2979{
2980        return mvmd128<(32)>::slli<(sh*2)>(arg1);
2981}
2982
2983//The total number of operations is 1.0
2984template <> template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1)
2985{
2986        return mvmd128<(64)>::slli<(sh*2)>(arg1);
2987}
2988
2989//The total number of operations is 13.0
2990template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill8(fw1_t val1, fw1_t val2, fw1_t val3, fw1_t val4, fw1_t val5, fw1_t val6, fw1_t val7, fw1_t val8)
2991{
2992        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<1>::fill4(val1, val2, val3, val4), mvmd128<1>::fill4(val5, val6, val7, val8));
2993}
2994
2995//The total number of operations is 13.0
2996template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill8(fw2_t val1, fw2_t val2, fw2_t val3, fw2_t val4, fw2_t val5, fw2_t val6, fw2_t val7, fw2_t val8)
2997{
2998        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<2>::fill4(val1, val2, val3, val4), mvmd128<2>::fill4(val5, val6, val7, val8));
2999}
3000
3001//The total number of operations is 7.0
3002template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill8(fw4_t val1, fw4_t val2, fw4_t val3, fw4_t val4, fw4_t val5, fw4_t val6, fw4_t val7, fw4_t val8)
3003{
3004        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))));
3005}
3006
3007//The total number of operations is 3.0
3008template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill8(fw8_t val1, fw8_t val2, fw8_t val3, fw8_t val4, fw8_t val5, fw8_t val6, fw8_t val7, fw8_t val8)
3009{
3010        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))));
3011}
3012
3013//The total number of operations is 1.0
3014template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill8(fw16_t val1, fw16_t val2, fw16_t val3, fw16_t val4, fw16_t val5, fw16_t val6, fw16_t val7, fw16_t val8)
3015{
3016        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));
3017}
3018
3019//The total number of operations is 13.0
3020IDISA_ALWAYS_INLINE bitblock128_t bitblock128::sll(bitblock128_t arg1, bitblock128_t arg2)
3021{
3022        return simd128<128>::sll(arg1, arg2);
3023}
3024
3025//The total number of operations is 1.0
3026IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(const bitblock128_t* arg1)
3027{
3028        return _mm_loadu_si128((bitblock128_t*)(arg1));
3029}
3030
3031//The total number of operations is 2.33333333333
3032template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srli(bitblock128_t arg1)
3033{
3034        return simd128<128>::srli<sh>(arg1);
3035}
3036
3037//The total number of operations is 1.0
3038IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t arg1, bitblock128_t* arg2)
3039{
3040        _mm_store_si128((bitblock128_t*)(arg2), arg1);
3041}
3042
3043//The total number of operations is 20.3333333333
3044IDISA_ALWAYS_INLINE uint8_t bitblock128::popcount(bitblock128_t arg1)
3045{
3046        return mvmd128<8>::extract<0>(simd128<128>::popcount(arg1));
3047}
3048
3049//The total number of operations is 2.0
3050IDISA_ALWAYS_INLINE bool bitblock128::all(bitblock128_t arg1)
3051{
3052        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<-1>())) == 65535;
3053}
3054
3055//The total number of operations is 13.0
3056IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srl(bitblock128_t arg1, bitblock128_t arg2)
3057{
3058        return simd128<128>::srl(arg1, arg2);
3059}
3060
3061//The total number of operations is 2.33333333333
3062template <uint8_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::slli(bitblock128_t arg1)
3063{
3064        return simd128<128>::slli<sh>(arg1);
3065}
3066
3067//The total number of operations is 2.0
3068IDISA_ALWAYS_INLINE bool bitblock128::any(bitblock128_t arg1)
3069{
3070        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<0>())) != 65535;
3071}
3072
3073//The total number of operations is 1.0
3074IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(const bitblock128_t* arg1)
3075{
3076        return _mm_load_si128((bitblock128_t*)(arg1));
3077}
3078
3079//The total number of operations is 1.0
3080IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t arg1, bitblock128_t* arg2)
3081{
3082        _mm_storeu_si128((bitblock128_t*)(arg2), arg1);
3083}
3084
3085#endif
Note: See TracBrowser for help on using the repository browser.