source: trunk/lib/idisa_cpp/idisa_sse2.cpp @ 1937

Last change on this file since 1937 was 1924, checked in by cameron, 7 years ago

Update simd<128>::srl/sll

File size: 151.4 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;
16template <uint32_t fw>
17class simd128
18{
19public:
20        static IDISA_ALWAYS_INLINE bitblock128_t max(bitblock128_t arg1, bitblock128_t arg2);
21        static IDISA_ALWAYS_INLINE bitblock128_t mult(bitblock128_t arg1, bitblock128_t arg2);
22        static IDISA_ALWAYS_INLINE bitblock128_t gt(bitblock128_t arg1, bitblock128_t arg2);
23        static IDISA_ALWAYS_INLINE bitblock128_t umult(bitblock128_t arg1, bitblock128_t arg2);
24        static IDISA_ALWAYS_INLINE bitblock128_t ult(bitblock128_t arg1, bitblock128_t arg2);
25        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
26        static IDISA_ALWAYS_INLINE bitblock128_t ctz(bitblock128_t arg1);
27        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t shift_mask);
28        static IDISA_ALWAYS_INLINE bitblock128_t eq(bitblock128_t arg1, bitblock128_t arg2);
29        static IDISA_ALWAYS_INLINE bitblock128_t popcount(bitblock128_t arg1);
30        static IDISA_ALWAYS_INLINE bitblock128_t neg(bitblock128_t arg1);
31        static IDISA_ALWAYS_INLINE bitblock128_t himask();
32        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
33        static IDISA_ALWAYS_INLINE bitblock128_t ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
34        static IDISA_ALWAYS_INLINE bitblock128_t sub(bitblock128_t arg1, bitblock128_t arg2);
35        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1);
36        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t shift_mask);
37        static IDISA_ALWAYS_INLINE bitblock128_t lomask();
38        static IDISA_ALWAYS_INLINE bitblock128_t umin(bitblock128_t arg1, bitblock128_t arg2);
39        template <uint64_t val> static IDISA_ALWAYS_INLINE bitblock128_t constant();
40        static IDISA_ALWAYS_INLINE bitblock128_t min(bitblock128_t arg1, bitblock128_t arg2);
41        static IDISA_ALWAYS_INLINE bitblock128_t umax(bitblock128_t arg1, bitblock128_t arg2);
42        static IDISA_ALWAYS_INLINE bitblock128_t abs(bitblock128_t arg1);
43        static IDISA_ALWAYS_INLINE bitblock128_t xor_hl(bitblock128_t arg1);
44        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srai(bitblock128_t arg1);
45        static IDISA_ALWAYS_INLINE bitblock128_t lt(bitblock128_t arg1, bitblock128_t arg2);
46        static IDISA_ALWAYS_INLINE bitblock128_t add(bitblock128_t arg1, bitblock128_t arg2);
47        static IDISA_ALWAYS_INLINE bitblock128_t ugt(bitblock128_t arg1, bitblock128_t arg2);
48};
49
50template <uint32_t fw>
51class hsimd128
52{
53public:
54        static IDISA_ALWAYS_INLINE bitblock128_t umin_hl(bitblock128_t arg1, bitblock128_t arg2);
55        static IDISA_ALWAYS_INLINE bitblock128_t add_hl(bitblock128_t arg1, bitblock128_t arg2);
56        static IDISA_ALWAYS_INLINE bitblock128_t packss(bitblock128_t arg1, bitblock128_t arg2);
57        static IDISA_ALWAYS_INLINE bitblock128_t packh(bitblock128_t arg1, bitblock128_t arg2);
58        static IDISA_ALWAYS_INLINE uint64_t signmask(bitblock128_t arg1);
59        static IDISA_ALWAYS_INLINE bitblock128_t packl(bitblock128_t arg1, bitblock128_t arg2);
60        static IDISA_ALWAYS_INLINE bitblock128_t min_hl(bitblock128_t arg1, bitblock128_t arg2);
61        static IDISA_ALWAYS_INLINE bitblock128_t packus(bitblock128_t arg1, bitblock128_t arg2);
62};
63
64template <uint32_t fw>
65class esimd128
66{
67public:
68        static IDISA_ALWAYS_INLINE bitblock128_t mergel(bitblock128_t arg1, bitblock128_t arg2);
69        static IDISA_ALWAYS_INLINE bitblock128_t signextendh(bitblock128_t arg1);
70        static IDISA_ALWAYS_INLINE bitblock128_t mergeh(bitblock128_t arg1, bitblock128_t arg2);
71        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendh(bitblock128_t arg1);
72        static IDISA_ALWAYS_INLINE bitblock128_t zeroextendl(bitblock128_t arg1);
73        static IDISA_ALWAYS_INLINE bitblock128_t signextendl(bitblock128_t arg1);
74};
75
76template <uint32_t fw>
77class mvmd128
78{
79public:
80        template <uint64_t msk> static IDISA_ALWAYS_INLINE bitblock128_t shufflei(bitblock128_t arg1);
81        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dsrli(bitblock128_t arg1, bitblock128_t arg2);
82        static IDISA_ALWAYS_INLINE bitblock128_t fill(uint64_t val1);
83        template <uint64_t pos> static IDISA_ALWAYS_INLINE uint64_t extract(bitblock128_t arg1);
84        template <uint64_t pos> static IDISA_ALWAYS_INLINE bitblock128_t splat(bitblock128_t arg1);
85        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
86        static IDISA_ALWAYS_INLINE bitblock128_t fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
87        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
88        static IDISA_ALWAYS_INLINE bitblock128_t fill2(uint64_t val1, uint64_t val2);
89        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t dslli(bitblock128_t arg1, bitblock128_t arg2);
90        static IDISA_ALWAYS_INLINE bitblock128_t fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
91        static IDISA_ALWAYS_INLINE bitblock128_t fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
92};
93
94class bitblock128
95{
96public:
97        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
98        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(const bitblock128_t* arg1);
99        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
100        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
101        static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t arg1, bitblock128_t* arg2);
102        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
103        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
104        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
105        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
106        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(const bitblock128_t* arg1);
107        static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t arg1, bitblock128_t* arg2);
108};
109
110//Declaration Part
111IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2);
112IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1);
113IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2);
114IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2);
115IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2);
116IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2);
117template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2);
118template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2);
119template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2);
120template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2);
121template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2);
122template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2);
123template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2);
124template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2);
125template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2);
126template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2);
127template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2);
128template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2);
129template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2);
130template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2);
131template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2);
132template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2);
133template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2);
134template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2);
135template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2);
136template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2);
137template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2);
138template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2);
139template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2);
140template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2);
141template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2);
142template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2);
143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2);
144template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2);
145template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2);
146template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2);
147template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2);
148template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2);
149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2);
150template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2);
151template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2);
152template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2);
153template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2);
154template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2);
155template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2);
156template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2);
157template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2);
158template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2);
159template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2);
160template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2);
161template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2);
162template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2);
163template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2);
164template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1);
165template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1);
166template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1);
167template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1);
168template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1);
169template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1);
170template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1);
171template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1);
172template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1);
173template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1);
174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1);
175template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1);
176template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1);
177template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1);
178template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1);
179template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask);
181template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2);
182template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2);
183template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2);
184template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2);
185template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2);
186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2);
187template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2);
188template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2);
189template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1);
190template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1);
191template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1);
192template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1);
193template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1);
194template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1);
195template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1);
196template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1);
197template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1);
198template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1);
199template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1);
200template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1);
201template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1);
202template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1);
203template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1);
204template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1);
205template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1);
206template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1);
207template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1);
208template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1);
209template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1);
210template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1);
211template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1);
212template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1);
213template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1);
214template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1);
215template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1);
216template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1);
217template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1);
218template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
219template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
220template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
221template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
222template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
223template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
224template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
225template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3);
226template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2);
227template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2);
228template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2);
229template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2);
230template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2);
231template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2);
232template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2);
233template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2);
234template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1);
235template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1);
236template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1);
237template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1);
238template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1);
239template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1);
240template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1);
241template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
242template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask);
243template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant();
244template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant();
245template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant();
246template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant();
247template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant();
248template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant();
249template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant();
250template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant();
251template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2);
252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2);
253template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2);
254template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2);
255template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2);
256template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2);
257template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2);
258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2);
259template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask();
260template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask();
261template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask();
262template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask();
263template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask();
264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask();
265template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask();
266template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2);
267template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2);
268template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2);
269template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2);
270template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2);
271template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2);
272template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2);
273template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2);
274template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1);
275template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1);
276template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1);
277template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1);
278template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1);
279template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1);
280template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1);
281template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2);
282template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2);
283template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2);
284template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2);
285template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2);
286template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2);
287template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2);
288template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2);
289template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1);
290template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1);
291template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1);
292template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1);
293template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1);
294template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1);
295template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1);
296template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask();
297template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask();
298template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask();
299template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask();
300template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask();
301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask();
302template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask();
303template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2);
304template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2);
305template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2);
306template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2);
307template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2);
308template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2);
309template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2);
310template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2);
311template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2);
312template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2);
313template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2);
314template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2);
315template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2);
316template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2);
317template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2);
318template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2);
319template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
320template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
321template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
322template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
323template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
324template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
325template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2);
326template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
327template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
328template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
329template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
330template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
331template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
332template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2);
333template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2);
334template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2);
335template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2);
336template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2);
337template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2);
338template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2);
339template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2);
340template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1);
341template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1);
342template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1);
343template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1);
344template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1);
345template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1);
346template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2);
347template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2);
348template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2);
349template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2);
350template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2);
351template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2);
352template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2);
353template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2);
354template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2);
355template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2);
356template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2);
357template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2);
358template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2);
359template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2);
360template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
361template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
362template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
363template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
364template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
365template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
366template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2);
367template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2);
368template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2);
369template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2);
370template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2);
371template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2);
372template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2);
373template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2);
374template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2);
375template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2);
376template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2);
377template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2);
378template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2);
379template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2);
380template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2);
381template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
382template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
383template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
384template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
385template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
386template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
387template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2);
388template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1);
389template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1);
390template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1);
391template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1);
392template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1);
393template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1);
394template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1);
395template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1);
396template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1);
397template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1);
398template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1);
399template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1);
400template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1);
401template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1);
402template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1);
403template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1);
404template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1);
405template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1);
406template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1);
407template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1);
408template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1);
409template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1);
410template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1);
411template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1);
412template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1);
413template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1);
414template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1);
415template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1);
416template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1);
417template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1);
418template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1);
419template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
420template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
421template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
422template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
423template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
424template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
425template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2);
426template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1);
427template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1);
428template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1);
429template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1);
430template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1);
431template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1);
432template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1);
433template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1);
434template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1);
435template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1);
436template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1);
437template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1);
438template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1);
439template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1);
440template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1);
441template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1);
442template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1);
443template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1);
444template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1);
445template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1);
446template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1);
447template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1);
448template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1);
449template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
450template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
451template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
452template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
453template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
454template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
455template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
456template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
457template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
458template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
459template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1);
460template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1);
461template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1);
462template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1);
463template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1);
464template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1);
465template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1);
466template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2);
467template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2);
468template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2);
469template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2);
470template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2);
471template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2);
472template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2);
473template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2);
474template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2);
475template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2);
476template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2);
477template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2);
478template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2);
479template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2);
480template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1);
481template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1);
482template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1);
483template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1);
484template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1);
485template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1);
486template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1);
487template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
488template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
489template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
490template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
491template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
492
493//Implementation Part
494
495#define shufflemask4(s1, s2, s3, s4) \
496        ((s1<<6) | (s2<<4) | (s3<<2) | s4)
497
498#define shufflemask4_from_shufflemask2(msk) \
499        (msk==3 ? 238 : (msk==2 ? 228 : (msk==1 ? 78 : 68)))
500
501#define shufflemask8_to_shufflemask4(msk) \
502        ((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))
503//The total number of operations is 2.0
504IDISA_ALWAYS_INLINE bitblock128_t simd_nor(bitblock128_t arg1, bitblock128_t arg2)
505{
506        return simd_not(simd_or(arg1, arg2));
507}
508
509//The total number of operations is 1.0
510IDISA_ALWAYS_INLINE bitblock128_t simd_not(bitblock128_t arg1)
511{
512        return simd_xor(arg1, simd128<32>::constant<-1>());
513}
514
515//The total number of operations is 1.0
516IDISA_ALWAYS_INLINE bitblock128_t simd_andc(bitblock128_t arg1, bitblock128_t arg2)
517{
518        return _mm_andnot_si128(arg2, arg1);
519}
520
521//The total number of operations is 1.0
522IDISA_ALWAYS_INLINE bitblock128_t simd_or(bitblock128_t arg1, bitblock128_t arg2)
523{
524        return _mm_or_si128(arg1, arg2);
525}
526
527//The total number of operations is 1.0
528IDISA_ALWAYS_INLINE bitblock128_t simd_and(bitblock128_t arg1, bitblock128_t arg2)
529{
530        return _mm_and_si128(arg1, arg2);
531}
532
533//The total number of operations is 1.0
534IDISA_ALWAYS_INLINE bitblock128_t simd_xor(bitblock128_t arg1, bitblock128_t arg2)
535{
536        return _mm_xor_si128(arg1, arg2);
537}
538
539//The total number of operations is 1.0
540template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::max(bitblock128_t arg1, bitblock128_t arg2)
541{
542        return simd_and(arg1, arg2);
543}
544
545//The total number of operations is 15.6666666667
546template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::max(bitblock128_t arg1, bitblock128_t arg2)
547{
548        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)))));
549}
550
551//The total number of operations is 9.0
552template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::max(bitblock128_t arg1, bitblock128_t arg2)
553{
554        bitblock128_t high_bit = simd128<4>::constant<(8)>();
555        return simd_xor(simd128<4>::umax(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
556}
557
558//The total number of operations is 4.0
559template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::max(bitblock128_t arg1, bitblock128_t arg2)
560{
561        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg1, arg2);
562}
563
564//The total number of operations is 1.0
565template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::max(bitblock128_t arg1, bitblock128_t arg2)
566{
567        return _mm_max_epi16(arg1, arg2);
568}
569
570//The total number of operations is 4.0
571template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::max(bitblock128_t arg1, bitblock128_t arg2)
572{
573        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg1, arg2);
574}
575
576//The total number of operations is 17.5
577template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::max(bitblock128_t arg1, bitblock128_t arg2)
578{
579        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg1, arg2);
580}
581
582//The total number of operations is 54.75
583template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::max(bitblock128_t arg1, bitblock128_t arg2)
584{
585        return simd128<1>::ifh(simd128<128>::gt(arg1, arg2), arg1, arg2);
586}
587
588//The total number of operations is 1.0
589template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::mult(bitblock128_t arg1, bitblock128_t arg2)
590{
591        return simd_and(arg1, arg2);
592}
593
594//The total number of operations is 19.6666666667
595template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::mult(bitblock128_t arg1, bitblock128_t arg2)
596{
597        bitblock128_t tmp1 = simd128<128>::slli<1>(arg1);
598        bitblock128_t tmp2 = simd128<128>::slli<1>(arg2);
599        return simd128<1>::ifh(simd128<2>::himask(), simd_or(simd_and(tmp1, simd_and(arg2, simd_or(simd_not(arg1), simd_not(tmp2)))), simd_and(arg1, simd_and(tmp2, simd_or(simd_not(tmp1), simd_not(arg2))))), simd_and(arg1, arg2));
600}
601
602//The total number of operations is 31.0
603template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::mult(bitblock128_t arg1, bitblock128_t arg2)
604{
605        bitblock128_t loMask = simd128<(8)>::lomask();
606        bitblock128_t tmpAns1 = simd128<(8)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
607        bitblock128_t tmpAns2 = simd128<(8)>::mult(simd128<(8)>::srli<4>(arg1), simd128<(8)>::srli<4>(arg2));
608        return simd128<1>::ifh(loMask, tmpAns1, simd128<(8)>::slli<4>(tmpAns2));
609}
610
611//The total number of operations is 10.0
612template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::mult(bitblock128_t arg1, bitblock128_t arg2)
613{
614        bitblock128_t loMask = simd128<(16)>::lomask();
615        bitblock128_t tmpAns1 = simd128<(16)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
616        bitblock128_t tmpAns2 = simd128<(16)>::mult(simd128<(16)>::srli<8>(arg1), simd128<(16)>::srli<8>(arg2));
617        return simd128<1>::ifh(loMask, tmpAns1, simd128<(16)>::slli<8>(tmpAns2));
618}
619
620//The total number of operations is 1.0
621template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::mult(bitblock128_t arg1, bitblock128_t arg2)
622{
623        return _mm_mullo_epi16(arg1, arg2);
624}
625
626//The total number of operations is 30.0
627template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::mult(bitblock128_t arg1, bitblock128_t arg2)
628{
629        bitblock128_t loMask = simd128<(64)>::lomask();
630        bitblock128_t tmpAns1 = simd128<(64)>::mult(simd_and(loMask, arg1), simd_and(loMask, arg2));
631        bitblock128_t tmpAns2 = simd128<(64)>::mult(simd128<(64)>::srli<32>(arg1), simd128<(64)>::srli<32>(arg2));
632        return simd128<1>::ifh(loMask, tmpAns1, simd128<(64)>::slli<32>(tmpAns2));
633}
634
635//The total number of operations is 11.0
636template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::mult(bitblock128_t arg1, bitblock128_t arg2)
637{
638        bitblock128_t loMask = simd128<64>::lomask();
639        bitblock128_t arg1_low = simd_and(arg1, loMask);
640        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg1);
641        bitblock128_t arg2_low = simd_and(arg2, loMask);
642        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg2);
643        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
644        bitblock128_t tmpAns2 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
645        bitblock128_t tmpAns3 = simd128<64>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
646        return simd128<64>::add(tmpAns1, simd128<64>::add(tmpAns2, tmpAns3));
647}
648
649//The total number of operations is 165.0
650template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::mult(bitblock128_t arg1, bitblock128_t arg2)
651{
652        bitblock128_t loMask = simd128<128>::lomask();
653        bitblock128_t arg1_low = simd_and(arg1, loMask);
654        bitblock128_t arg1_high = simd128<128>::srli<(64)>(arg1);
655        bitblock128_t arg2_low = simd_and(arg2, loMask);
656        bitblock128_t arg2_high = simd128<128>::srli<(64)>(arg2);
657        bitblock128_t tmpAns1 = simd128<(64)>::umult(arg1_low, arg2_low);
658        bitblock128_t tmpAns2 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_low, arg2_high));
659        bitblock128_t tmpAns3 = simd128<128>::slli<(64)>(simd128<(64)>::umult(arg1_high, arg2_low));
660        return simd128<128>::add(tmpAns1, simd128<128>::add(tmpAns2, tmpAns3));
661}
662
663//The total number of operations is 1.0
664template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::gt(bitblock128_t arg1, bitblock128_t arg2)
665{
666        return simd_andc(arg2, arg1);
667}
668
669//The total number of operations is 14.6666666667
670template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::gt(bitblock128_t arg1, bitblock128_t arg2)
671{
672        bitblock128_t tmp = simd_not(arg1);
673        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)));
674        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
675}
676
677//The total number of operations is 10.0
678template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::gt(bitblock128_t arg1, bitblock128_t arg2)
679{
680        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)));
681}
682
683//The total number of operations is 1.0
684template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::gt(bitblock128_t arg1, bitblock128_t arg2)
685{
686        return _mm_cmpgt_epi8(arg1, arg2);
687}
688
689//The total number of operations is 1.0
690template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::gt(bitblock128_t arg1, bitblock128_t arg2)
691{
692        return _mm_cmpgt_epi16(arg1, arg2);
693}
694
695//The total number of operations is 1.0
696template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::gt(bitblock128_t arg1, bitblock128_t arg2)
697{
698        return _mm_cmpgt_epi32(arg1, arg2);
699}
700
701//The total number of operations is 14.5
702template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::gt(bitblock128_t arg1, bitblock128_t arg2)
703{
704        bitblock128_t hiAns = simd128<(32)>::gt(arg1, arg2);
705        bitblock128_t loAns = simd128<(32)>::ugt(arg1, arg2);
706        bitblock128_t mask = simd_and(loAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
707        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
708        return simd_or(simd128<64>::srai<(32)>(hiAns), mask);
709}
710
711//The total number of operations is 51.75
712template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::gt(bitblock128_t arg1, bitblock128_t arg2)
713{
714        bitblock128_t hiAns = simd128<(64)>::gt(arg1, arg2);
715        bitblock128_t loAns = simd128<(64)>::ugt(arg1, arg2);
716        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
717        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
718        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
719}
720
721//The total number of operations is 289.0
722template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umult(bitblock128_t arg1, bitblock128_t arg2)
723{
724        bitblock128_t loMask = simd128<(2)>::lomask();
725        bitblock128_t tmpAns1 = simd128<(2)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
726        bitblock128_t tmpAns2 = simd128<(2)>::umult(simd_and(loMask, simd128<(4)>::srli<(2)>(arg1)), simd_and(loMask, simd128<(4)>::srli<(2)>(arg2)));
727        return simd_or(tmpAns1, simd128<(4)>::slli<(2)>(tmpAns2));
728}
729
730//The total number of operations is 139.0
731template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umult(bitblock128_t arg1, bitblock128_t arg2)
732{
733        bitblock128_t loMask = simd128<(4)>::lomask();
734        bitblock128_t tmpAns1 = simd128<(4)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
735        bitblock128_t tmpAns2 = simd128<(4)>::umult(simd_and(loMask, simd128<(8)>::srli<(4)>(arg1)), simd_and(loMask, simd128<(8)>::srli<(4)>(arg2)));
736        return simd_or(tmpAns1, simd128<(8)>::slli<(4)>(tmpAns2));
737}
738
739//The total number of operations is 64.0
740template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umult(bitblock128_t arg1, bitblock128_t arg2)
741{
742        bitblock128_t loMask = simd128<(8)>::lomask();
743        bitblock128_t tmpAns1 = simd128<(8)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
744        bitblock128_t tmpAns2 = simd128<(8)>::umult(simd_and(loMask, simd128<(16)>::srli<(8)>(arg1)), simd_and(loMask, simd128<(16)>::srli<(8)>(arg2)));
745        return simd_or(tmpAns1, simd128<(16)>::slli<(8)>(tmpAns2));
746}
747
748//The total number of operations is 28.0
749template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umult(bitblock128_t arg1, bitblock128_t arg2)
750{
751        bitblock128_t loMask = simd128<(16)>::lomask();
752        bitblock128_t tmpAns1 = simd128<(16)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
753        bitblock128_t tmpAns2 = simd128<(16)>::umult(simd_and(loMask, simd128<(32)>::srli<(16)>(arg1)), simd_and(loMask, simd128<(32)>::srli<(16)>(arg2)));
754        return simd_or(tmpAns1, simd128<(32)>::slli<(16)>(tmpAns2));
755}
756
757//The total number of operations is 10.0
758template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umult(bitblock128_t arg1, bitblock128_t arg2)
759{
760        bitblock128_t loMask = simd128<(32)>::lomask();
761        bitblock128_t tmpAns1 = simd128<(32)>::umult(simd_and(loMask, arg1), simd_and(loMask, arg2));
762        bitblock128_t tmpAns2 = simd128<(32)>::umult(simd_and(loMask, simd128<(64)>::srli<(32)>(arg1)), simd_and(loMask, simd128<(64)>::srli<(32)>(arg2)));
763        return simd_or(tmpAns1, simd128<(64)>::slli<(32)>(tmpAns2));
764}
765
766//The total number of operations is 1.0
767template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umult(bitblock128_t arg1, bitblock128_t arg2)
768{
769        return _mm_mul_epu32(arg1, arg2);
770}
771
772//The total number of operations is 45.0
773template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umult(bitblock128_t arg1, bitblock128_t arg2)
774{
775        bitblock128_t loMask1 = simd128<(128)>::lomask();
776        bitblock128_t arg11 = simd_and(arg1, loMask1);
777        bitblock128_t arg22 = simd_and(arg2, loMask1);
778        bitblock128_t loMask2 = simd128<64>::lomask();
779        bitblock128_t arg1_low = simd_and(arg11, loMask2);
780        bitblock128_t arg1_high = simd128<64>::srli<(32)>(arg11);
781        bitblock128_t arg2_low = simd_and(arg22, loMask2);
782        bitblock128_t arg2_high = simd128<64>::srli<(32)>(arg22);
783        bitblock128_t tmpAns1 = simd128<(32)>::umult(arg1_low, arg2_low);
784        bitblock128_t tmpAns2 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_low, arg2_high));
785        bitblock128_t tmpAns3 = simd128<(128)>::slli<(32)>(simd128<(32)>::umult(arg1_high, arg2_low));
786        bitblock128_t tmpAns4 = simd128<(128)>::slli<64>(simd128<(32)>::umult(arg1_high, arg2_high));
787        return simd128<(128)>::add(tmpAns1, simd128<(128)>::add(tmpAns2, simd128<(128)>::add(tmpAns3, tmpAns4)));
788}
789
790//The total number of operations is 1.0
791template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ult(bitblock128_t arg1, bitblock128_t arg2)
792{
793        return simd_andc(arg2, arg1);
794}
795
796//The total number of operations is 13.6666666667
797template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ult(bitblock128_t arg1, bitblock128_t arg2)
798{
799        bitblock128_t tmp = simd_not(arg1);
800        bitblock128_t tmpAns = simd_or(simd_and(tmp, arg2), simd_and(simd128<128>::slli<1>(simd_and(tmp, arg2)), simd_or(tmp, arg2)));
801        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
802}
803
804//The total number of operations is 20.0
805template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ult(bitblock128_t arg1, bitblock128_t arg2)
806{
807        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())));
808}
809
810//The total number of operations is 7.0
811template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ult(bitblock128_t arg1, bitblock128_t arg2)
812{
813        bitblock128_t high_bit = simd128<8>::constant<(128)>();
814        return simd128<8>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
815}
816
817//The total number of operations is 7.0
818template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ult(bitblock128_t arg1, bitblock128_t arg2)
819{
820        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
821        return simd128<16>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
822}
823
824//The total number of operations is 7.0
825template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ult(bitblock128_t arg1, bitblock128_t arg2)
826{
827        bitblock128_t high_bit = simd128<32>::constant<(2147483648ULL)>();
828        return simd128<32>::lt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
829}
830
831//The total number of operations is 17.5
832template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ult(bitblock128_t arg1, bitblock128_t arg2)
833{
834        bitblock128_t tmpAns = simd128<(32)>::ult(arg1, arg2);
835        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
836        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
837        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
838}
839
840//The total number of operations is 40.0833333333
841template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ult(bitblock128_t arg1, bitblock128_t arg2)
842{
843        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)));
844}
845
846//The total number of operations is 1.0
847template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::lt(bitblock128_t arg1, bitblock128_t arg2)
848{
849        return simd_andc(arg1, arg2);
850}
851
852//The total number of operations is 14.6666666667
853template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lt(bitblock128_t arg1, bitblock128_t arg2)
854{
855        bitblock128_t tmp = simd_not(arg2);
856        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)));
857        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
858}
859
860//The total number of operations is 18.0
861template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lt(bitblock128_t arg1, bitblock128_t arg2)
862{
863        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)));
864}
865
866//The total number of operations is 5.0
867template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lt(bitblock128_t arg1, bitblock128_t arg2)
868{
869        return simd_and(simd_not(simd128<8>::gt(arg1, arg2)), simd_not(simd128<8>::eq(arg1, arg2)));
870}
871
872//The total number of operations is 5.0
873template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lt(bitblock128_t arg1, bitblock128_t arg2)
874{
875        return simd_and(simd_not(simd128<16>::gt(arg1, arg2)), simd_not(simd128<16>::eq(arg1, arg2)));
876}
877
878//The total number of operations is 5.0
879template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lt(bitblock128_t arg1, bitblock128_t arg2)
880{
881        return simd_and(simd_not(simd128<32>::gt(arg1, arg2)), simd_not(simd128<32>::eq(arg1, arg2)));
882}
883
884//The total number of operations is 19.5
885template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lt(bitblock128_t arg1, bitblock128_t arg2)
886{
887        bitblock128_t high_bit = simd128<64>::constant<(9223372036854775808ULL)>();
888        return simd128<64>::ult(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
889}
890
891//The total number of operations is 60.75
892template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lt(bitblock128_t arg1, bitblock128_t arg2)
893{
894        bitblock128_t hiAns = simd128<(64)>::lt(arg1, arg2);
895        bitblock128_t loAns = simd128<(64)>::ult(arg1, arg2);
896        bitblock128_t mask = simd_and(loAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
897        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
898        return simd_or(simd128<128>::srai<(64)>(hiAns), mask);
899}
900
901//The total number of operations is 2.0
902template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srli(bitblock128_t arg1)
903{
904        return simd_and(simd128<32>::srli<sh>(arg1), simd128<2>::constant<((3)>>sh)>());
905}
906
907//The total number of operations is 2.0
908template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srli(bitblock128_t arg1)
909{
910        return simd_and(simd128<32>::srli<sh>(arg1), simd128<4>::constant<((15)>>sh)>());
911}
912
913//The total number of operations is 2.0
914template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srli(bitblock128_t arg1)
915{
916        return simd_and(simd128<32>::srli<sh>(arg1), simd128<8>::constant<((255)>>sh)>());
917}
918
919//The total number of operations is 1.0
920template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srli(bitblock128_t arg1)
921{
922        return _mm_srli_epi16(arg1, (int32_t)(sh));
923}
924
925//The total number of operations is 1.0
926template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srli(bitblock128_t arg1)
927{
928        return _mm_srli_epi32(arg1, (int32_t)(sh));
929}
930
931//The total number of operations is 1.0
932template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srli(bitblock128_t arg1)
933{
934        return _mm_srli_epi64(arg1, (int32_t)(sh));
935}
936
937//The total number of operations is 2.33333333333
938template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srli(bitblock128_t arg1)
939{
940        return (((sh%8) == 0) ? _mm_srli_si128(arg1, (int32_t)((sh/8))) : ((sh >= 64) ? simd128<64>::srli<(sh-64)>(_mm_srli_si128(arg1, (int32_t)(8))) : simd_or(simd128<64>::srli<sh>(arg1), _mm_srli_si128(simd128<64>::slli<(64-sh)>(arg1), (int32_t)(8)))));
941}
942
943//The total number of operations is 1.0
944template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ctz(bitblock128_t arg1)
945{
946        return simd_not(arg1);
947}
948
949//The total number of operations is 10.6666666667
950template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ctz(bitblock128_t arg1)
951{
952        bitblock128_t tmp = simd_not(arg1);
953        return simd128<1>::ifh(simd128<2>::himask(), simd_and(tmp, simd128<128>::slli<1>(tmp)), simd_and(simd128<128>::srli<1>(arg1), tmp));
954}
955
956//The total number of operations is 14.0
957template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ctz(bitblock128_t arg1)
958{
959        return simd128<4>::popcount(simd_andc(simd128<4>::sub(arg1, simd128<4>::constant<1>()), arg1));
960}
961
962//The total number of operations is 13.0
963template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ctz(bitblock128_t arg1)
964{
965        return simd128<8>::popcount(simd_andc(simd128<8>::sub(arg1, simd128<8>::constant<1>()), arg1));
966}
967
968//The total number of operations is 16.0
969template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ctz(bitblock128_t arg1)
970{
971        return simd128<16>::popcount(simd_andc(simd128<16>::sub(arg1, simd128<16>::constant<1>()), arg1));
972}
973
974//The total number of operations is 19.0
975template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ctz(bitblock128_t arg1)
976{
977        return simd128<32>::popcount(simd_andc(simd128<32>::sub(arg1, simd128<32>::constant<1>()), arg1));
978}
979
980//The total number of operations is 14.0
981template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ctz(bitblock128_t arg1)
982{
983        return simd128<64>::popcount(simd_andc(simd128<64>::sub(arg1, simd128<64>::constant<1>()), arg1));
984}
985
986//The total number of operations is 26.6666666667
987template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ctz(bitblock128_t arg1)
988{
989        return simd128<128>::popcount(simd_andc(simd128<128>::sub(arg1, simd128<128>::constant<1>()), arg1));
990}
991
992//The total number of operations is 10.0
993template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
994{
995        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)))));
996}
997
998//The total number of operations is 13.0
999template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sll(bitblock128_t arg1, bitblock128_t shift_mask)
1000{
1001        bitblock128_t shift = simd_and(shift_mask, _mm_cvtsi32_si128((int32_t)(127)));
1002        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))));
1003}
1004
1005//The total number of operations is 1.0
1006template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1007{
1008        return simd_andc(arg1, arg2);
1009}
1010
1011//The total number of operations is 13.6666666667
1012template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1013{
1014        bitblock128_t tmp = simd_not(arg2);
1015        bitblock128_t tmpAns = simd_or(simd_and(arg1, tmp), simd_and(simd128<128>::slli<1>(simd_and(arg1, tmp)), simd_or(arg1, tmp)));
1016        return simd128<1>::ifh(simd128<2>::himask(), tmpAns, simd128<128>::srli<1>(tmpAns));
1017}
1018
1019//The total number of operations is 12.0
1020template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1021{
1022        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())));
1023}
1024
1025//The total number of operations is 3.0
1026template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1027{
1028        bitblock128_t high_bit = simd128<8>::constant<(128)>();
1029        return simd128<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1030}
1031
1032//The total number of operations is 3.0
1033template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1034{
1035        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1036        return simd128<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1037}
1038
1039//The total number of operations is 3.0
1040template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1041{
1042        bitblock128_t high_bit = simd128<32>::constant<(2147483648ULL)>();
1043        return simd128<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
1044}
1045
1046//The total number of operations is 13.5
1047template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1048{
1049        bitblock128_t tmpAns = simd128<(32)>::ugt(arg1, arg2);
1050        bitblock128_t mask = simd_and(tmpAns, simd128<64>::srli<(32)>(simd128<(32)>::eq(arg1, arg2)));
1051        mask = simd_or(mask, simd128<64>::slli<(32)>(mask));
1052        return simd_or(simd128<64>::srai<(32)>(tmpAns), mask);
1053}
1054
1055//The total number of operations is 37.25
1056template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ugt(bitblock128_t arg1, bitblock128_t arg2)
1057{
1058        bitblock128_t tmpAns = simd128<(64)>::ugt(arg1, arg2);
1059        bitblock128_t mask = simd_and(tmpAns, simd128<128>::srli<(64)>(simd128<(64)>::eq(arg1, arg2)));
1060        mask = simd_or(mask, simd128<128>::slli<(64)>(mask));
1061        return simd_or(simd128<128>::srai<(64)>(tmpAns), mask);
1062}
1063
1064//The total number of operations is 4.0
1065template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::xor_hl(bitblock128_t arg1)
1066{
1067        return simd_xor(simd128<2>::srli<(1)>(arg1), simd_and(arg1, simd128<2>::lomask()));
1068}
1069
1070//The total number of operations is 4.0
1071template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::xor_hl(bitblock128_t arg1)
1072{
1073        return simd_xor(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1074}
1075
1076//The total number of operations is 4.0
1077template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::xor_hl(bitblock128_t arg1)
1078{
1079        return simd_xor(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1080}
1081
1082//The total number of operations is 3.0
1083template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::xor_hl(bitblock128_t arg1)
1084{
1085        return simd_xor(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1086}
1087
1088//The total number of operations is 3.0
1089template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::xor_hl(bitblock128_t arg1)
1090{
1091        return simd_xor(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1092}
1093
1094//The total number of operations is 3.0
1095template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::xor_hl(bitblock128_t arg1)
1096{
1097        return simd_xor(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1098}
1099
1100//The total number of operations is 4.33333333333
1101template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::xor_hl(bitblock128_t arg1)
1102{
1103        return simd_xor(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1104}
1105
1106//The total number of operations is 0
1107template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::popcount(bitblock128_t arg1)
1108{
1109        return arg1;
1110}
1111
1112//The total number of operations is 3.0
1113template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::popcount(bitblock128_t arg1)
1114{
1115        return simd128<2>::add_hl(simd128<(1)>::popcount(arg1));
1116}
1117
1118//The total number of operations is 7.0
1119template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::popcount(bitblock128_t arg1)
1120{
1121        return simd128<4>::add_hl(simd128<(2)>::popcount(arg1));
1122}
1123
1124//The total number of operations is 11.0
1125template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::popcount(bitblock128_t arg1)
1126{
1127        return simd128<8>::add_hl(simd128<(4)>::popcount(arg1));
1128}
1129
1130//The total number of operations is 14.0
1131template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::popcount(bitblock128_t arg1)
1132{
1133        return simd128<16>::add_hl(simd128<(8)>::popcount(arg1));
1134}
1135
1136//The total number of operations is 17.0
1137template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::popcount(bitblock128_t arg1)
1138{
1139        return simd128<32>::add_hl(simd128<(16)>::popcount(arg1));
1140}
1141
1142//The total number of operations is 12.0
1143template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::popcount(bitblock128_t arg1)
1144{
1145        return _mm_sad_epu8(simd128<8>::popcount(arg1), simd128<8>::constant<0>());
1146}
1147
1148//The total number of operations is 16.3333333333
1149template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::popcount(bitblock128_t arg1)
1150{
1151        bitblock128_t tmpAns = simd128<(64)>::popcount(arg1);
1152        return simd128<(64)>::add(simd_and(tmpAns, simd128<128>::lomask()), simd128<128>::srli<(64)>(tmpAns));
1153}
1154
1155//The total number of operations is 6.33333333333
1156template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::neg(bitblock128_t arg1)
1157{
1158        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(arg1, simd128<128>::slli<1>(arg1)), arg1);
1159}
1160
1161//The total number of operations is 6.0
1162template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::neg(bitblock128_t arg1)
1163{
1164        return simd128<4>::sub(simd128<4>::constant<0>(), arg1);
1165}
1166
1167//The total number of operations is 1.0
1168template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::neg(bitblock128_t arg1)
1169{
1170        return simd128<8>::sub(simd128<8>::constant<0>(), arg1);
1171}
1172
1173//The total number of operations is 1.0
1174template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::neg(bitblock128_t arg1)
1175{
1176        return simd128<16>::sub(simd128<16>::constant<0>(), arg1);
1177}
1178
1179//The total number of operations is 1.0
1180template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::neg(bitblock128_t arg1)
1181{
1182        return simd128<32>::sub(simd128<32>::constant<0>(), arg1);
1183}
1184
1185//The total number of operations is 1.0
1186template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::neg(bitblock128_t arg1)
1187{
1188        return simd128<64>::sub(simd128<64>::constant<0>(), arg1);
1189}
1190
1191//The total number of operations is 9.33333333333
1192template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::neg(bitblock128_t arg1)
1193{
1194        return simd128<128>::sub(simd128<128>::constant<0>(), arg1);
1195}
1196
1197//The total number of operations is 2.0
1198template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::slli(bitblock128_t arg1)
1199{
1200        return simd_and(simd128<32>::slli<sh>(arg1), simd128<2>::constant<(((3)<<sh)&(3))>());
1201}
1202
1203//The total number of operations is 2.0
1204template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::slli(bitblock128_t arg1)
1205{
1206        return simd_and(simd128<32>::slli<sh>(arg1), simd128<4>::constant<(((15)<<sh)&(15))>());
1207}
1208
1209//The total number of operations is 2.0
1210template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::slli(bitblock128_t arg1)
1211{
1212        return simd_and(simd128<32>::slli<sh>(arg1), simd128<8>::constant<(((255)<<sh)&(255))>());
1213}
1214
1215//The total number of operations is 1.0
1216template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::slli(bitblock128_t arg1)
1217{
1218        return _mm_slli_epi16(arg1, (int32_t)(sh));
1219}
1220
1221//The total number of operations is 1.0
1222template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::slli(bitblock128_t arg1)
1223{
1224        return _mm_slli_epi32(arg1, (int32_t)(sh));
1225}
1226
1227//The total number of operations is 1.0
1228template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::slli(bitblock128_t arg1)
1229{
1230        return _mm_slli_epi64(arg1, (int32_t)(sh));
1231}
1232
1233//The total number of operations is 2.33333333333
1234template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::slli(bitblock128_t arg1)
1235{
1236        return (((sh%8) == 0) ? _mm_slli_si128(arg1, (int32_t)((sh/8))) : ((sh >= 64) ? simd128<64>::slli<(sh-64)>(_mm_slli_si128(arg1, (int32_t)(8))) : simd_or(simd128<64>::slli<sh>(arg1), _mm_slli_si128(simd128<64>::srli<(64-sh)>(arg1), (int32_t)(8)))));
1237}
1238
1239//The total number of operations is 3.0
1240template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1241{
1242        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
1243}
1244
1245//The total number of operations is 8.0
1246template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1247{
1248        return simd128<(1)>::ifh(simd128<1>::ifh(simd128<2>::himask(), arg1, simd128<2>::srli<(1)>(arg1)), arg2, arg3);
1249}
1250
1251//The total number of operations is 13.0
1252template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1253{
1254        return simd128<1>::ifh(simd128<4>::gt(simd128<4>::constant<0>(), arg1), arg2, arg3);
1255}
1256
1257//The total number of operations is 4.0
1258template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1259{
1260        return simd128<1>::ifh(simd128<8>::gt(simd128<8>::constant<0>(), arg1), arg2, arg3);
1261}
1262
1263//The total number of operations is 4.0
1264template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1265{
1266        return simd128<1>::ifh(simd128<16>::gt(simd128<16>::constant<0>(), arg1), arg2, arg3);
1267}
1268
1269//The total number of operations is 4.0
1270template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1271{
1272        return simd128<1>::ifh(simd128<32>::gt(simd128<32>::constant<0>(), arg1), arg2, arg3);
1273}
1274
1275//The total number of operations is 8.0
1276template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1277{
1278        return simd128<(32)>::ifh(simd128<1>::ifh(simd128<64>::himask(), arg1, simd128<64>::srli<(32)>(arg1)), arg2, arg3);
1279}
1280
1281//The total number of operations is 13.3333333333
1282template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::ifh(bitblock128_t arg1, bitblock128_t arg2, bitblock128_t arg3)
1283{
1284        return simd128<(64)>::ifh(simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg1)), arg2, arg3);
1285}
1286
1287//The total number of operations is 1.0
1288template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::sub(bitblock128_t arg1, bitblock128_t arg2)
1289{
1290        return simd_xor(arg1, arg2);
1291}
1292
1293//The total number of operations is 9.33333333333
1294template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::sub(bitblock128_t arg1, bitblock128_t arg2)
1295{
1296        bitblock128_t tmp = simd_xor(arg1, arg2);
1297        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(simd_not(arg1), arg2))), tmp);
1298}
1299
1300//The total number of operations is 6.0
1301template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::sub(bitblock128_t arg1, bitblock128_t arg2)
1302{
1303        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::sub(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::sub(arg1, arg2));
1304}
1305
1306//The total number of operations is 1.0
1307template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::sub(bitblock128_t arg1, bitblock128_t arg2)
1308{
1309        return _mm_sub_epi8(arg1, arg2);
1310}
1311
1312//The total number of operations is 1.0
1313template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::sub(bitblock128_t arg1, bitblock128_t arg2)
1314{
1315        return _mm_sub_epi16(arg1, arg2);
1316}
1317
1318//The total number of operations is 1.0
1319template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::sub(bitblock128_t arg1, bitblock128_t arg2)
1320{
1321        return _mm_sub_epi32(arg1, arg2);
1322}
1323
1324//The total number of operations is 1.0
1325template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::sub(bitblock128_t arg1, bitblock128_t arg2)
1326{
1327        return _mm_sub_epi64(arg1, arg2);
1328}
1329
1330//The total number of operations is 9.33333333333
1331template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::sub(bitblock128_t arg1, bitblock128_t arg2)
1332{
1333        bitblock128_t partial = simd128<(64)>::sub(arg1, arg2);
1334        bitblock128_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
1335        bitblock128_t borrow = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(borrowMask));
1336        return simd128<(64)>::sub(partial, borrow);
1337}
1338
1339//The total number of operations is 3.0
1340template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add_hl(bitblock128_t arg1)
1341{
1342        return simd128<16>::sub(arg1, simd_and(simd128<2>::lomask(), simd128<16>::srli<1>(arg1)));
1343}
1344
1345//The total number of operations is 4.0
1346template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add_hl(bitblock128_t arg1)
1347{
1348        return simd128<(8)>::add(simd128<4>::srli<(2)>(arg1), simd_and(arg1, simd128<4>::lomask()));
1349}
1350
1351//The total number of operations is 4.0
1352template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add_hl(bitblock128_t arg1)
1353{
1354        return simd128<(16)>::add(simd128<8>::srli<(4)>(arg1), simd_and(arg1, simd128<8>::lomask()));
1355}
1356
1357//The total number of operations is 3.0
1358template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add_hl(bitblock128_t arg1)
1359{
1360        return simd128<(32)>::add(simd128<16>::srli<(8)>(arg1), simd_and(arg1, simd128<16>::lomask()));
1361}
1362
1363//The total number of operations is 3.0
1364template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add_hl(bitblock128_t arg1)
1365{
1366        return simd128<(64)>::add(simd128<32>::srli<(16)>(arg1), simd_and(arg1, simd128<32>::lomask()));
1367}
1368
1369//The total number of operations is 3.0
1370template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add_hl(bitblock128_t arg1)
1371{
1372        return simd128<64>::add(simd128<64>::srli<(32)>(arg1), simd_and(arg1, simd128<64>::lomask()));
1373}
1374
1375//The total number of operations is 12.6666666667
1376template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add_hl(bitblock128_t arg1)
1377{
1378        return simd128<128>::add(simd128<128>::srli<(64)>(arg1), simd_and(arg1, simd128<128>::lomask()));
1379}
1380
1381//The total number of operations is 10.0
1382template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1383{
1384        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)))));
1385}
1386
1387//The total number of operations is 13.0
1388template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srl(bitblock128_t arg1, bitblock128_t shift_mask)
1389{
1390        bitblock128_t shift = simd_and(shift_mask, _mm_cvtsi32_si128((int32_t)(127)));
1391        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))));
1392}
1393
1394//The total number of operations is 0
1395template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::constant()
1396{
1397        return simd128<32>::constant<(-1*val)>();
1398}
1399
1400//The total number of operations is 0
1401template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::constant()
1402{
1403        return ((val < 0) ? simd128<(4)>::constant<((val<<2)|(val^(-4)))>() : simd128<(4)>::constant<((val<<2)|val)>());
1404}
1405
1406//The total number of operations is 0
1407template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::constant()
1408{
1409        return ((val < 0) ? simd128<(8)>::constant<((val<<4)|(val^(-16)))>() : simd128<(8)>::constant<((val<<4)|val)>());
1410}
1411
1412//The total number of operations is 0
1413template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::constant()
1414{
1415        return _mm_set1_epi8((int32_t)(val));
1416}
1417
1418//The total number of operations is 0
1419template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::constant()
1420{
1421        return _mm_set1_epi16((int32_t)(val));
1422}
1423
1424//The total number of operations is 0
1425template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::constant()
1426{
1427        return _mm_set1_epi32((int32_t)(val));
1428}
1429
1430//The total number of operations is 0
1431template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::constant()
1432{
1433        return _mm_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val));
1434}
1435
1436//The total number of operations is 0
1437template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::constant()
1438{
1439        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val));
1440}
1441
1442//The total number of operations is 1.0
1443template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::min(bitblock128_t arg1, bitblock128_t arg2)
1444{
1445        return simd_or(arg1, arg2);
1446}
1447
1448//The total number of operations is 16.6666666667
1449template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::min(bitblock128_t arg1, bitblock128_t arg2)
1450{
1451        bitblock128_t tmp1 = simd128<128>::srli<1>(arg1);
1452        bitblock128_t tmp2 = simd128<128>::srli<1>(arg2);
1453        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))));
1454}
1455
1456//The total number of operations is 9.0
1457template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::min(bitblock128_t arg1, bitblock128_t arg2)
1458{
1459        bitblock128_t high_bit = simd128<4>::constant<(8)>();
1460        return simd_xor(simd128<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1461}
1462
1463//The total number of operations is 4.0
1464template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::min(bitblock128_t arg1, bitblock128_t arg2)
1465{
1466        return simd128<1>::ifh(simd128<8>::gt(arg1, arg2), arg2, arg1);
1467}
1468
1469//The total number of operations is 1.0
1470template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::min(bitblock128_t arg1, bitblock128_t arg2)
1471{
1472        return _mm_min_epi16(arg1, arg2);
1473}
1474
1475//The total number of operations is 4.0
1476template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::min(bitblock128_t arg1, bitblock128_t arg2)
1477{
1478        return simd128<1>::ifh(simd128<32>::gt(arg1, arg2), arg2, arg1);
1479}
1480
1481//The total number of operations is 17.5
1482template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::min(bitblock128_t arg1, bitblock128_t arg2)
1483{
1484        return simd128<1>::ifh(simd128<64>::gt(arg1, arg2), arg2, arg1);
1485}
1486
1487//The total number of operations is 54.75
1488template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::min(bitblock128_t arg1, bitblock128_t arg2)
1489{
1490        return simd128<1>::ifh(simd128<128>::gt(arg1, arg2), arg2, arg1);
1491}
1492
1493//The total number of operations is 0
1494template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::lomask()
1495{
1496        return simd128<2>::constant<(1)>();
1497}
1498
1499//The total number of operations is 0
1500template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::lomask()
1501{
1502        return simd128<4>::constant<(3)>();
1503}
1504
1505//The total number of operations is 0
1506template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::lomask()
1507{
1508        return simd128<8>::constant<(15)>();
1509}
1510
1511//The total number of operations is 0
1512template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::lomask()
1513{
1514        return simd128<16>::constant<(255)>();
1515}
1516
1517//The total number of operations is 0
1518template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::lomask()
1519{
1520        return simd128<32>::constant<(65535)>();
1521}
1522
1523//The total number of operations is 0
1524template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::lomask()
1525{
1526        return _mm_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1));
1527}
1528
1529//The total number of operations is 0
1530template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::lomask()
1531{
1532        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1));
1533}
1534
1535//The total number of operations is 1.0
1536template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umin(bitblock128_t arg1, bitblock128_t arg2)
1537{
1538        return simd_and(arg1, arg2);
1539}
1540
1541//The total number of operations is 16.0
1542template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umin(bitblock128_t arg1, bitblock128_t arg2)
1543{
1544        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)));
1545}
1546
1547//The total number of operations is 6.0
1548template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umin(bitblock128_t arg1, bitblock128_t arg2)
1549{
1550        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)));
1551}
1552
1553//The total number of operations is 1.0
1554template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umin(bitblock128_t arg1, bitblock128_t arg2)
1555{
1556        return _mm_min_epu8(arg1, arg2);
1557}
1558
1559//The total number of operations is 4.0
1560template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umin(bitblock128_t arg1, bitblock128_t arg2)
1561{
1562        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1563        return simd_xor(simd128<16>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1564}
1565
1566//The total number of operations is 7.0
1567template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umin(bitblock128_t arg1, bitblock128_t arg2)
1568{
1569        bitblock128_t high_bit = simd128<32>::constant<(2147483648ULL)>();
1570        return simd_xor(simd128<32>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1571}
1572
1573//The total number of operations is 20.0
1574template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umin(bitblock128_t arg1, bitblock128_t arg2)
1575{
1576        bitblock128_t tmpAns = simd128<(32)>::umin(arg1, arg2);
1577        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1578        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1579        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1580}
1581
1582//The total number of operations is 43.6666666667
1583template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umin(bitblock128_t arg1, bitblock128_t arg2)
1584{
1585        bitblock128_t tmpAns = simd128<(64)>::umin(arg1, arg2);
1586        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1587        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1588        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1589}
1590
1591//The total number of operations is 7.33333333333
1592template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::abs(bitblock128_t arg1)
1593{
1594        return simd128<1>::ifh(simd128<2>::himask(), simd_and(arg1, simd128<128>::slli<1>(simd_not(arg1))), arg1);
1595}
1596
1597//The total number of operations is 19.0
1598template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::abs(bitblock128_t arg1)
1599{
1600        bitblock128_t gtMask = simd128<4>::gt(arg1, simd128<4>::constant<0>());
1601        return simd128<1>::ifh(gtMask, arg1, simd128<4>::sub(gtMask, arg1));
1602}
1603
1604//The total number of operations is 5.0
1605template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::abs(bitblock128_t arg1)
1606{
1607        bitblock128_t gtMask = simd128<8>::gt(arg1, simd128<8>::constant<0>());
1608        return simd128<1>::ifh(gtMask, arg1, simd128<8>::sub(gtMask, arg1));
1609}
1610
1611//The total number of operations is 5.0
1612template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::abs(bitblock128_t arg1)
1613{
1614        bitblock128_t gtMask = simd128<16>::gt(arg1, simd128<16>::constant<0>());
1615        return simd128<1>::ifh(gtMask, arg1, simd128<16>::sub(gtMask, arg1));
1616}
1617
1618//The total number of operations is 5.0
1619template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::abs(bitblock128_t arg1)
1620{
1621        bitblock128_t gtMask = simd128<32>::gt(arg1, simd128<32>::constant<0>());
1622        return simd128<1>::ifh(gtMask, arg1, simd128<32>::sub(gtMask, arg1));
1623}
1624
1625//The total number of operations is 17.0
1626template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::abs(bitblock128_t arg1)
1627{
1628        bitblock128_t eqMask = simd128<64>::eq(simd128<1>::ifh(simd128<64>::himask(), simd128<(32)>::abs(arg1), arg1), arg1);
1629        return simd128<1>::ifh(eqMask, arg1, simd128<64>::sub(eqMask, arg1));
1630}
1631
1632//The total number of operations is 44.0
1633template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::abs(bitblock128_t arg1)
1634{
1635        bitblock128_t eqMask = simd128<128>::eq(simd128<1>::ifh(simd128<128>::himask(), simd128<(64)>::abs(arg1), arg1), arg1);
1636        return simd128<1>::ifh(eqMask, arg1, simd128<128>::sub(eqMask, arg1));
1637}
1638
1639//The total number of operations is 2.0
1640template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::eq(bitblock128_t arg1, bitblock128_t arg2)
1641{
1642        return simd_not(simd_xor(arg1, arg2));
1643}
1644
1645//The total number of operations is 8.0
1646template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::eq(bitblock128_t arg1, bitblock128_t arg2)
1647{
1648        bitblock128_t tmpAns = simd128<(1)>::eq(arg1, arg2);
1649        bitblock128_t loMask = simd_and(tmpAns, simd128<2>::srli<(1)>(tmpAns));
1650        bitblock128_t hiMask = simd128<2>::slli<(1)>(loMask);
1651        return simd_or(loMask, hiMask);
1652}
1653
1654//The total number of operations is 9.0
1655template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::eq(bitblock128_t arg1, bitblock128_t arg2)
1656{
1657        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))));
1658}
1659
1660//The total number of operations is 1.0
1661template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::eq(bitblock128_t arg1, bitblock128_t arg2)
1662{
1663        return _mm_cmpeq_epi8(arg1, arg2);
1664}
1665
1666//The total number of operations is 1.0
1667template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::eq(bitblock128_t arg1, bitblock128_t arg2)
1668{
1669        return _mm_cmpeq_epi16(arg1, arg2);
1670}
1671
1672//The total number of operations is 1.0
1673template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::eq(bitblock128_t arg1, bitblock128_t arg2)
1674{
1675        return _mm_cmpeq_epi32(arg1, arg2);
1676}
1677
1678//The total number of operations is 5.0
1679template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::eq(bitblock128_t arg1, bitblock128_t arg2)
1680{
1681        bitblock128_t tmpAns = simd128<(32)>::eq(arg1, arg2);
1682        bitblock128_t loMask = simd_and(tmpAns, simd128<64>::srli<(32)>(tmpAns));
1683        bitblock128_t hiMask = simd128<64>::slli<(32)>(loMask);
1684        return simd_or(loMask, hiMask);
1685}
1686
1687//The total number of operations is 11.6666666667
1688template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::eq(bitblock128_t arg1, bitblock128_t arg2)
1689{
1690        bitblock128_t tmpAns = simd128<(64)>::eq(arg1, arg2);
1691        bitblock128_t loMask = simd_and(tmpAns, simd128<128>::srli<(64)>(tmpAns));
1692        bitblock128_t hiMask = simd128<128>::slli<(64)>(loMask);
1693        return simd_or(loMask, hiMask);
1694}
1695
1696//The total number of operations is 4.0
1697template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::srai(bitblock128_t arg1)
1698{
1699        return ((sh == 0) ? arg1 : simd_or(simd_and(simd128<2>::himask(), arg1), simd128<2>::srli<1>(arg1)));
1700}
1701
1702//The total number of operations is 10.0
1703template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::srai(bitblock128_t arg1)
1704{
1705        bitblock128_t tmp = simd128<4>::srli<((sh >= 4) ? (3) : ((sh < 0) ? 0 : sh))>(arg1);
1706        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)));
1707}
1708
1709//The total number of operations is 5.0
1710template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::srai(bitblock128_t arg1)
1711{
1712        bitblock128_t tmp = simd128<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
1713        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)));
1714}
1715
1716//The total number of operations is 1.0
1717template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::srai(bitblock128_t arg1)
1718{
1719        return _mm_srai_epi16(arg1, (int32_t)(sh));
1720}
1721
1722//The total number of operations is 1.0
1723template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::srai(bitblock128_t arg1)
1724{
1725        return _mm_srai_epi32(arg1, (int32_t)(sh));
1726}
1727
1728//The total number of operations is 4.5
1729template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::srai(bitblock128_t arg1)
1730{
1731        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))));
1732}
1733
1734//The total number of operations is 11.0833333333
1735template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::srai(bitblock128_t arg1)
1736{
1737        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))));
1738}
1739
1740//The total number of operations is 0
1741template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::himask()
1742{
1743        return simd128<2>::constant<(2)>();
1744}
1745
1746//The total number of operations is 0
1747template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::himask()
1748{
1749        return simd128<4>::constant<(12)>();
1750}
1751
1752//The total number of operations is 0
1753template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::himask()
1754{
1755        return simd128<8>::constant<(240)>();
1756}
1757
1758//The total number of operations is 0
1759template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::himask()
1760{
1761        return simd128<16>::constant<(65280)>();
1762}
1763
1764//The total number of operations is 0
1765template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::himask()
1766{
1767        return simd128<32>::constant<-65536>();
1768}
1769
1770//The total number of operations is 0
1771template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::himask()
1772{
1773        return _mm_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0));
1774}
1775
1776//The total number of operations is 0
1777template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::himask()
1778{
1779        return _mm_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0));
1780}
1781
1782//The total number of operations is 1.0
1783template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::add(bitblock128_t arg1, bitblock128_t arg2)
1784{
1785        return simd_xor(arg1, arg2);
1786}
1787
1788//The total number of operations is 8.33333333333
1789template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::add(bitblock128_t arg1, bitblock128_t arg2)
1790{
1791        bitblock128_t tmp = simd_xor(arg1, arg2);
1792        return simd128<1>::ifh(simd128<2>::himask(), simd_xor(tmp, simd128<128>::slli<1>(simd_and(arg1, arg2))), tmp);
1793}
1794
1795//The total number of operations is 6.0
1796template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::add(bitblock128_t arg1, bitblock128_t arg2)
1797{
1798        return simd128<1>::ifh(simd128<(8)>::himask(), simd128<(8)>::add(arg1, simd_and(simd128<(8)>::himask(), arg2)), simd128<(8)>::add(arg1, arg2));
1799}
1800
1801//The total number of operations is 1.0
1802template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::add(bitblock128_t arg1, bitblock128_t arg2)
1803{
1804        return _mm_add_epi8(arg1, arg2);
1805}
1806
1807//The total number of operations is 1.0
1808template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::add(bitblock128_t arg1, bitblock128_t arg2)
1809{
1810        return _mm_add_epi16(arg1, arg2);
1811}
1812
1813//The total number of operations is 1.0
1814template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::add(bitblock128_t arg1, bitblock128_t arg2)
1815{
1816        return _mm_add_epi32(arg1, arg2);
1817}
1818
1819//The total number of operations is 1.0
1820template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::add(bitblock128_t arg1, bitblock128_t arg2)
1821{
1822        return _mm_add_epi64(arg1, arg2);
1823}
1824
1825//The total number of operations is 9.33333333333
1826template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::add(bitblock128_t arg1, bitblock128_t arg2)
1827{
1828        bitblock128_t partial = simd128<(64)>::add(arg1, arg2);
1829        bitblock128_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
1830        bitblock128_t carry = simd128<128>::slli<(64)>(simd128<(64)>::srli<(63)>(carryMask));
1831        return simd128<(64)>::add(partial, carry);
1832}
1833
1834//The total number of operations is 1.0
1835template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<1>::umax(bitblock128_t arg1, bitblock128_t arg2)
1836{
1837        return simd_or(arg1, arg2);
1838}
1839
1840//The total number of operations is 15.6666666667
1841template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<2>::umax(bitblock128_t arg1, bitblock128_t arg2)
1842{
1843        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))))));
1844}
1845
1846//The total number of operations is 6.0
1847template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<4>::umax(bitblock128_t arg1, bitblock128_t arg2)
1848{
1849        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)));
1850}
1851
1852//The total number of operations is 1.0
1853template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<8>::umax(bitblock128_t arg1, bitblock128_t arg2)
1854{
1855        return _mm_max_epu8(arg1, arg2);
1856}
1857
1858//The total number of operations is 4.0
1859template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<16>::umax(bitblock128_t arg1, bitblock128_t arg2)
1860{
1861        bitblock128_t high_bit = simd128<16>::constant<(32768)>();
1862        return simd_xor(simd128<16>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1863}
1864
1865//The total number of operations is 7.0
1866template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<32>::umax(bitblock128_t arg1, bitblock128_t arg2)
1867{
1868        bitblock128_t high_bit = simd128<32>::constant<(2147483648ULL)>();
1869        return simd_xor(simd128<32>::max(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
1870}
1871
1872//The total number of operations is 20.0
1873template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<64>::umax(bitblock128_t arg1, bitblock128_t arg2)
1874{
1875        bitblock128_t tmpAns = simd128<(32)>::umax(arg1, arg2);
1876        bitblock128_t eqMask1 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg1));
1877        bitblock128_t eqMask2 = simd128<64>::srli<(32)>(simd128<(32)>::eq(tmpAns, arg2));
1878        return simd128<1>::ifh(simd128<64>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1879}
1880
1881//The total number of operations is 43.6666666667
1882template <> IDISA_ALWAYS_INLINE bitblock128_t simd128<128>::umax(bitblock128_t arg1, bitblock128_t arg2)
1883{
1884        bitblock128_t tmpAns = simd128<(64)>::umax(arg1, arg2);
1885        bitblock128_t eqMask1 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg1));
1886        bitblock128_t eqMask2 = simd128<128>::srli<(64)>(simd128<(64)>::eq(tmpAns, arg2));
1887        return simd128<1>::ifh(simd128<128>::himask(), tmpAns, simd128<1>::ifh(eqMask1, simd128<1>::ifh(eqMask2, tmpAns, arg1), arg2));
1888}
1889
1890//The total number of operations is 73.0
1891template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1892{
1893        return simd128<(1)>::umin(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1894}
1895
1896//The total number of operations is 66.6666666667
1897template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1898{
1899        return simd128<(2)>::umin(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1900}
1901
1902//The total number of operations is 35.3333333333
1903template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1904{
1905        return simd128<(4)>::umin(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1906}
1907
1908//The total number of operations is 7.0
1909template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1910{
1911        return simd128<(8)>::umin(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1912}
1913
1914//The total number of operations is 37.3333333333
1915template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1916{
1917        return simd128<(16)>::umin(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1918}
1919
1920//The total number of operations is 19.0
1921template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1922{
1923        return simd128<(32)>::umin(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1924}
1925
1926//The total number of operations is 30.6666666667
1927template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::umin_hl(bitblock128_t arg1, bitblock128_t arg2)
1928{
1929        return simd128<(64)>::umin(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1930}
1931
1932//The total number of operations is 73.0
1933template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1934{
1935        return simd128<(1)>::add(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
1936}
1937
1938//The total number of operations is 59.0
1939template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1940{
1941        return simd128<(2)>::add(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
1942}
1943
1944//The total number of operations is 35.3333333333
1945template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1946{
1947        return simd128<(4)>::add(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
1948}
1949
1950//The total number of operations is 7.0
1951template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1952{
1953        return simd128<(8)>::add(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
1954}
1955
1956//The total number of operations is 34.3333333333
1957template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1958{
1959        return simd128<(16)>::add(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
1960}
1961
1962//The total number of operations is 13.0
1963template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1964{
1965        return simd128<(32)>::add(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
1966}
1967
1968//The total number of operations is 11.6666666667
1969template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::add_hl(bitblock128_t arg1, bitblock128_t arg2)
1970{
1971        return simd128<(64)>::add(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
1972}
1973
1974//The total number of operations is 108.666666667
1975template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packss(bitblock128_t arg1, bitblock128_t arg2)
1976{
1977        bitblock128_t hiBound = simd128<2>::srli<1>(simd128<2>::lomask());
1978        bitblock128_t loBound = simd_not(hiBound);
1979        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)));
1980}
1981
1982//The total number of operations is 79.3333333333
1983template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packss(bitblock128_t arg1, bitblock128_t arg2)
1984{
1985        bitblock128_t hiBound = simd128<4>::srli<1>(simd128<4>::lomask());
1986        bitblock128_t loBound = simd_not(hiBound);
1987        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)));
1988}
1989
1990//The total number of operations is 32.6666666667
1991template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packss(bitblock128_t arg1, bitblock128_t arg2)
1992{
1993        bitblock128_t hiBound = simd128<8>::srli<1>(simd128<8>::lomask());
1994        bitblock128_t loBound = simd_not(hiBound);
1995        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)));
1996}
1997
1998//The total number of operations is 1.0
1999template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packss(bitblock128_t arg1, bitblock128_t arg2)
2000{
2001        return _mm_packs_epi16(arg2, arg1);
2002}
2003
2004//The total number of operations is 1.0
2005template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packss(bitblock128_t arg1, bitblock128_t arg2)
2006{
2007        return _mm_packs_epi32(arg2, arg1);
2008}
2009
2010//The total number of operations is 77.0
2011template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packss(bitblock128_t arg1, bitblock128_t arg2)
2012{
2013        bitblock128_t hiBound = simd128<64>::srli<1>(simd128<64>::lomask());
2014        bitblock128_t loBound = simd_not(hiBound);
2015        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)));
2016}
2017
2018//The total number of operations is 227.666666667
2019template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packss(bitblock128_t arg1, bitblock128_t arg2)
2020{
2021        bitblock128_t hiBound = simd128<128>::srli<1>(simd128<128>::lomask());
2022        bitblock128_t loBound = simd_not(hiBound);
2023        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)));
2024}
2025
2026//The total number of operations is 24.0
2027template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<4>::signmask(bitblock128_t arg1)
2028{
2029        uint64_t tmpAns1 = hsimd128<(8)>::signmask(esimd128<4>::mergeh(arg1, simd128<4>::constant<0>()));
2030        uint64_t tmpAns2 = hsimd128<(8)>::signmask(esimd128<4>::mergel(arg1, simd128<4>::constant<0>()));
2031        return ((tmpAns1<<(16))+tmpAns2);
2032}
2033
2034//The total number of operations is 1.0
2035template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<8>::signmask(bitblock128_t arg1)
2036{
2037        return _mm_movemask_epi8(arg1);
2038}
2039
2040//The total number of operations is 2.0
2041template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<16>::signmask(bitblock128_t arg1)
2042{
2043        return hsimd128<(8)>::signmask(hsimd128<16>::packss(simd128<16>::constant<0>(), arg1));
2044}
2045
2046//The total number of operations is 1.0
2047template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<32>::signmask(bitblock128_t arg1)
2048{
2049        return _mm_movemask_ps(_mm_castsi128_ps(arg1));
2050}
2051
2052//The total number of operations is 1.0
2053template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<64>::signmask(bitblock128_t arg1)
2054{
2055        return _mm_movemask_pd(_mm_castsi128_pd(arg1));
2056}
2057
2058//The total number of operations is 6.33333333333
2059template <> IDISA_ALWAYS_INLINE uint64_t hsimd128<128>::signmask(bitblock128_t arg1)
2060{
2061        return hsimd128<(64)>::signmask(hsimd128<128>::packh(simd128<128>::constant<0>(), arg1));
2062}
2063
2064//The total number of operations is 35.0
2065template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packl(bitblock128_t arg1, bitblock128_t arg2)
2066{
2067        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));
2068}
2069
2070//The total number of operations is 24.3333333333
2071template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packl(bitblock128_t arg1, bitblock128_t arg2)
2072{
2073        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));
2074}
2075
2076//The total number of operations is 13.6666666667
2077template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packl(bitblock128_t arg1, bitblock128_t arg2)
2078{
2079        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));
2080}
2081
2082//The total number of operations is 3.0
2083template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packl(bitblock128_t arg1, bitblock128_t arg2)
2084{
2085        return hsimd128<16>::packus(simd_and(arg1, simd128<16>::lomask()), simd_and(arg2, simd128<16>::lomask()));
2086}
2087
2088//The total number of operations is 15.6666666667
2089template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packl(bitblock128_t arg1, bitblock128_t arg2)
2090{
2091        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));
2092}
2093
2094//The total number of operations is 5.0
2095template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packl(bitblock128_t arg1, bitblock128_t arg2)
2096{
2097        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())));
2098}
2099
2100//The total number of operations is 5.33333333333
2101template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packl(bitblock128_t arg1, bitblock128_t arg2)
2102{
2103        return simd128<1>::ifh(simd128<128>::himask(), simd128<128>::slli<(64)>(arg1), arg2);
2104}
2105
2106//The total number of operations is 37.0
2107template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packh(bitblock128_t arg1, bitblock128_t arg2)
2108{
2109        return hsimd128<2>::packl(simd128<64>::srli<(1)>(arg1), simd128<64>::srli<(1)>(arg2));
2110}
2111
2112//The total number of operations is 26.3333333333
2113template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packh(bitblock128_t arg1, bitblock128_t arg2)
2114{
2115        return hsimd128<4>::packl(simd128<64>::srli<(2)>(arg1), simd128<64>::srli<(2)>(arg2));
2116}
2117
2118//The total number of operations is 15.6666666667
2119template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packh(bitblock128_t arg1, bitblock128_t arg2)
2120{
2121        return hsimd128<8>::packl(simd128<64>::srli<(4)>(arg1), simd128<64>::srli<(4)>(arg2));
2122}
2123
2124//The total number of operations is 3.0
2125template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packh(bitblock128_t arg1, bitblock128_t arg2)
2126{
2127        return hsimd128<16>::packus(simd128<16>::srli<(8)>(arg1), simd128<16>::srli<(8)>(arg2));
2128}
2129
2130//The total number of operations is 17.6666666667
2131template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packh(bitblock128_t arg1, bitblock128_t arg2)
2132{
2133        return hsimd128<32>::packl(simd128<64>::srli<(16)>(arg1), simd128<64>::srli<(16)>(arg2));
2134}
2135
2136//The total number of operations is 7.0
2137template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packh(bitblock128_t arg1, bitblock128_t arg2)
2138{
2139        return hsimd128<64>::packl(simd128<64>::srli<(32)>(arg1), simd128<64>::srli<(32)>(arg2));
2140}
2141
2142//The total number of operations is 5.33333333333
2143template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packh(bitblock128_t arg1, bitblock128_t arg2)
2144{
2145        return simd128<1>::ifh(simd128<128>::himask(), arg1, simd128<128>::srli<(64)>(arg2));
2146}
2147
2148//The total number of operations is 73.0
2149template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2150{
2151        return simd128<(1)>::min(hsimd128<2>::packh(arg1, arg2), hsimd128<2>::packl(arg1, arg2));
2152}
2153
2154//The total number of operations is 67.3333333333
2155template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2156{
2157        return simd128<(2)>::min(hsimd128<4>::packh(arg1, arg2), hsimd128<4>::packl(arg1, arg2));
2158}
2159
2160//The total number of operations is 38.3333333333
2161template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2162{
2163        return simd128<(4)>::min(hsimd128<8>::packh(arg1, arg2), hsimd128<8>::packl(arg1, arg2));
2164}
2165
2166//The total number of operations is 10.0
2167template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2168{
2169        return simd128<(8)>::min(hsimd128<16>::packh(arg1, arg2), hsimd128<16>::packl(arg1, arg2));
2170}
2171
2172//The total number of operations is 34.3333333333
2173template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2174{
2175        return simd128<(16)>::min(hsimd128<32>::packh(arg1, arg2), hsimd128<32>::packl(arg1, arg2));
2176}
2177
2178//The total number of operations is 16.0
2179template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2180{
2181        return simd128<(32)>::min(hsimd128<64>::packh(arg1, arg2), hsimd128<64>::packl(arg1, arg2));
2182}
2183
2184//The total number of operations is 28.1666666667
2185template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::min_hl(bitblock128_t arg1, bitblock128_t arg2)
2186{
2187        return simd128<(64)>::min(hsimd128<128>::packh(arg1, arg2), hsimd128<128>::packl(arg1, arg2));
2188}
2189
2190//The total number of operations is 75.0
2191template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<2>::packus(bitblock128_t arg1, bitblock128_t arg2)
2192{
2193        bitblock128_t arg11 = simd128<2>::ifh(arg1, simd128<2>::constant<0>(), arg1);
2194        bitblock128_t arg12 = simd_and(simd128<2>::lomask(), arg11);
2195        bitblock128_t arg21 = simd128<2>::ifh(arg2, simd128<2>::constant<0>(), arg2);
2196        bitblock128_t arg22 = simd_and(simd128<2>::lomask(), arg21);
2197        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()));
2198}
2199
2200//The total number of operations is 74.3333333333
2201template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<4>::packus(bitblock128_t arg1, bitblock128_t arg2)
2202{
2203        bitblock128_t hiPart = hsimd128<4>::packh(arg1, arg2);
2204        return simd128<(2)>::ifh(hiPart, simd128<(2)>::constant<0>(), simd_or(simd128<(2)>::gt(hiPart, simd128<(2)>::constant<0>()), hsimd128<4>::packl(arg1, arg2)));
2205}
2206
2207//The total number of operations is 31.6666666667
2208template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<8>::packus(bitblock128_t arg1, bitblock128_t arg2)
2209{
2210        bitblock128_t arg11 = simd128<8>::ifh(arg1, simd128<8>::constant<0>(), arg1);
2211        bitblock128_t arg12 = simd_and(simd128<8>::lomask(), arg11);
2212        bitblock128_t arg21 = simd128<8>::ifh(arg2, simd128<8>::constant<0>(), arg2);
2213        bitblock128_t arg22 = simd_and(simd128<8>::lomask(), arg21);
2214        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()));
2215}
2216
2217//The total number of operations is 1.0
2218template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<16>::packus(bitblock128_t arg1, bitblock128_t arg2)
2219{
2220        return _mm_packus_epi16(arg2, arg1);
2221}
2222
2223//The total number of operations is 33.6666666667
2224template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<32>::packus(bitblock128_t arg1, bitblock128_t arg2)
2225{
2226        bitblock128_t arg11 = simd128<32>::ifh(arg1, simd128<32>::constant<0>(), arg1);
2227        bitblock128_t arg12 = simd_and(simd128<32>::lomask(), arg11);
2228        bitblock128_t arg21 = simd128<32>::ifh(arg2, simd128<32>::constant<0>(), arg2);
2229        bitblock128_t arg22 = simd_and(simd128<32>::lomask(), arg21);
2230        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()));
2231}
2232
2233//The total number of operations is 18.0
2234template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<64>::packus(bitblock128_t arg1, bitblock128_t arg2)
2235{
2236        bitblock128_t hiPart = hsimd128<64>::packh(arg1, arg2);
2237        return simd128<(32)>::ifh(hiPart, simd128<(32)>::constant<0>(), simd_or(simd128<(32)>::gt(hiPart, simd128<(32)>::constant<0>()), hsimd128<64>::packl(arg1, arg2)));
2238}
2239
2240//The total number of operations is 34.1666666667
2241template <> IDISA_ALWAYS_INLINE bitblock128_t hsimd128<128>::packus(bitblock128_t arg1, bitblock128_t arg2)
2242{
2243        bitblock128_t hiPart = hsimd128<128>::packh(arg1, arg2);
2244        return simd128<(64)>::ifh(hiPart, simd128<(64)>::constant<0>(), simd_or(simd128<(64)>::gt(hiPart, simd128<(64)>::constant<0>()), hsimd128<128>::packl(arg1, arg2)));
2245}
2246
2247//The total number of operations is 31.0
2248template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2249{
2250        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));
2251}
2252
2253//The total number of operations is 21.0
2254template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2255{
2256        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));
2257}
2258
2259//The total number of operations is 11.0
2260template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2261{
2262        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));
2263}
2264
2265//The total number of operations is 1.0
2266template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2267{
2268        return _mm_unpacklo_epi8(arg2, arg1);
2269}
2270
2271//The total number of operations is 1.0
2272template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2273{
2274        return _mm_unpacklo_epi16(arg2, arg1);
2275}
2276
2277//The total number of operations is 1.0
2278template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2279{
2280        return _mm_unpacklo_epi32(arg2, arg1);
2281}
2282
2283//The total number of operations is 1.0
2284template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergel(bitblock128_t arg1, bitblock128_t arg2)
2285{
2286        return _mm_unpacklo_epi64(arg2, arg1);
2287}
2288
2289//The total number of operations is 31.0
2290template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2291{
2292        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));
2293}
2294
2295//The total number of operations is 21.0
2296template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2297{
2298        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));
2299}
2300
2301//The total number of operations is 11.0
2302template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2303{
2304        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));
2305}
2306
2307//The total number of operations is 1.0
2308template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2309{
2310        return _mm_unpackhi_epi8(arg2, arg1);
2311}
2312
2313//The total number of operations is 1.0
2314template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2315{
2316        return _mm_unpackhi_epi16(arg2, arg1);
2317}
2318
2319//The total number of operations is 1.0
2320template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2321{
2322        return _mm_unpackhi_epi32(arg2, arg1);
2323}
2324
2325//The total number of operations is 1.0
2326template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::mergeh(bitblock128_t arg1, bitblock128_t arg2)
2327{
2328        return _mm_unpackhi_epi64(arg2, arg1);
2329}
2330
2331//The total number of operations is 24.0
2332template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendh(bitblock128_t arg1)
2333{
2334        return esimd128<(2)>::mergeh(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2335}
2336
2337//The total number of operations is 14.0
2338template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendh(bitblock128_t arg1)
2339{
2340        return esimd128<(4)>::mergeh(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2341}
2342
2343//The total number of operations is 4.0
2344template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendh(bitblock128_t arg1)
2345{
2346        return esimd128<(8)>::mergeh(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2347}
2348
2349//The total number of operations is 3.0
2350template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendh(bitblock128_t arg1)
2351{
2352        return esimd128<(16)>::mergeh(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2353}
2354
2355//The total number of operations is 3.0
2356template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendh(bitblock128_t arg1)
2357{
2358        return esimd128<(32)>::mergeh(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2359}
2360
2361//The total number of operations is 3.0
2362template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendh(bitblock128_t arg1)
2363{
2364        return esimd128<(64)>::mergeh(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2365}
2366
2367//The total number of operations is 2.33333333333
2368template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendh(bitblock128_t arg1)
2369{
2370        return simd128<(128)>::srli<64>(arg1);
2371}
2372
2373//The total number of operations is 24.0
2374template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::zeroextendl(bitblock128_t arg1)
2375{
2376        return esimd128<(2)>::mergel(simd128<(2)>::srli<1>(arg1), simd_and(simd128<(2)>::lomask(), arg1));
2377}
2378
2379//The total number of operations is 14.0
2380template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::zeroextendl(bitblock128_t arg1)
2381{
2382        return esimd128<(4)>::mergel(simd128<(4)>::srli<2>(arg1), simd_and(simd128<(4)>::lomask(), arg1));
2383}
2384
2385//The total number of operations is 4.0
2386template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::zeroextendl(bitblock128_t arg1)
2387{
2388        return esimd128<(8)>::mergel(simd128<(8)>::srli<4>(arg1), simd_and(simd128<(8)>::lomask(), arg1));
2389}
2390
2391//The total number of operations is 3.0
2392template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::zeroextendl(bitblock128_t arg1)
2393{
2394        return esimd128<(16)>::mergel(simd128<(16)>::srli<8>(arg1), simd_and(simd128<(16)>::lomask(), arg1));
2395}
2396
2397//The total number of operations is 3.0
2398template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::zeroextendl(bitblock128_t arg1)
2399{
2400        return esimd128<(32)>::mergel(simd128<(32)>::srli<16>(arg1), simd_and(simd128<(32)>::lomask(), arg1));
2401}
2402
2403//The total number of operations is 3.0
2404template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::zeroextendl(bitblock128_t arg1)
2405{
2406        return esimd128<(64)>::mergel(simd128<(64)>::srli<32>(arg1), simd_and(simd128<(64)>::lomask(), arg1));
2407}
2408
2409//The total number of operations is 1.0
2410template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::zeroextendl(bitblock128_t arg1)
2411{
2412        return simd_and(simd128<(128)>::lomask(), arg1);
2413}
2414
2415//The total number of operations is 31.0
2416template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendh(bitblock128_t arg1)
2417{
2418        return esimd128<(2)>::mergeh(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2419}
2420
2421//The total number of operations is 33.0
2422template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendh(bitblock128_t arg1)
2423{
2424        return esimd128<(4)>::mergeh(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2425}
2426
2427//The total number of operations is 13.0
2428template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendh(bitblock128_t arg1)
2429{
2430        return esimd128<(8)>::mergeh(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2431}
2432
2433//The total number of operations is 4.0
2434template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendh(bitblock128_t arg1)
2435{
2436        return esimd128<(16)>::mergeh(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2437}
2438
2439//The total number of operations is 4.0
2440template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendh(bitblock128_t arg1)
2441{
2442        return esimd128<(32)>::mergeh(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2443}
2444
2445//The total number of operations is 11.0
2446template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendh(bitblock128_t arg1)
2447{
2448        return esimd128<(64)>::mergeh(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2449}
2450
2451//The total number of operations is 11.0833333333
2452template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendh(bitblock128_t arg1)
2453{
2454        return simd128<(128)>::srai<64>(arg1);
2455}
2456
2457//The total number of operations is 31.0
2458template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<1>::signextendl(bitblock128_t arg1)
2459{
2460        return esimd128<(2)>::mergel(simd128<(2)>::srai<1>(arg1), simd128<(2)>::srai<1>(simd128<(2)>::slli<1>(arg1)));
2461}
2462
2463//The total number of operations is 33.0
2464template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<2>::signextendl(bitblock128_t arg1)
2465{
2466        return esimd128<(4)>::mergel(simd128<(4)>::srai<2>(arg1), simd128<(4)>::srai<2>(simd128<(4)>::slli<2>(arg1)));
2467}
2468
2469//The total number of operations is 13.0
2470template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<4>::signextendl(bitblock128_t arg1)
2471{
2472        return esimd128<(8)>::mergel(simd128<(8)>::srai<4>(arg1), simd128<(8)>::srai<4>(simd128<(8)>::slli<4>(arg1)));
2473}
2474
2475//The total number of operations is 4.0
2476template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<8>::signextendl(bitblock128_t arg1)
2477{
2478        return esimd128<(16)>::mergel(simd128<(16)>::srai<8>(arg1), simd128<(16)>::srai<8>(simd128<(16)>::slli<8>(arg1)));
2479}
2480
2481//The total number of operations is 4.0
2482template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<16>::signextendl(bitblock128_t arg1)
2483{
2484        return esimd128<(32)>::mergel(simd128<(32)>::srai<16>(arg1), simd128<(32)>::srai<16>(simd128<(32)>::slli<16>(arg1)));
2485}
2486
2487//The total number of operations is 11.0
2488template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<32>::signextendl(bitblock128_t arg1)
2489{
2490        return esimd128<(64)>::mergel(simd128<(64)>::srai<32>(arg1), simd128<(64)>::srai<32>(simd128<(64)>::slli<32>(arg1)));
2491}
2492
2493//The total number of operations is 13.4166666667
2494template <> IDISA_ALWAYS_INLINE bitblock128_t esimd128<64>::signextendl(bitblock128_t arg1)
2495{
2496        return simd128<(128)>::srai<64>(simd128<(128)>::slli<64>(arg1));
2497}
2498
2499//The total number of operations is 13.6666666667
2500template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::shufflei(bitblock128_t arg1)
2501{
2502        bitblock128_t tmphi = _mm_shufflehi_epi16(arg1, (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2503        bitblock128_t tmpAns = _mm_shufflelo_epi16(tmphi, (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2504        bitblock128_t tmplh = _mm_shufflehi_epi16(simd128<128>::slli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)>>8)));
2505        bitblock128_t tmphl = _mm_shufflelo_epi16(simd128<128>::srli<64>(arg1), (int32_t)((shufflemask8_to_shufflemask4(msk)&255)));
2506        uint32_t a1 = ((((msk>>21)&4) == 0) ? 0 : (131071));
2507        uint32_t a2 = ((((msk>>18)&4) == 0) ? 0 : (131071));
2508        uint32_t a3 = ((((msk>>15)&4) == 0) ? 0 : (131071));
2509        uint32_t a4 = ((((msk>>12)&4) == 0) ? 0 : (131071));
2510        uint32_t a5 = ((((msk>>9)&4) == 0) ? (131071) : 0);
2511        uint32_t a6 = ((((msk>>6)&4) == 0) ? (131071) : 0);
2512        uint32_t a7 = ((((msk>>3)&4) == 0) ? (131071) : 0);
2513        uint32_t a8 = (((msk&4) == 0) ? (131071) : 0);
2514        return simd128<1>::ifh(mvmd128<16>::fill8(a1, a2, a3, a4, a5, a6, a7, a8), tmpAns, simd_or(tmplh, tmphl));
2515}
2516
2517//The total number of operations is 1.0
2518template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::shufflei(bitblock128_t arg1)
2519{
2520        return _mm_shuffle_epi32(arg1, (int32_t)(msk));
2521}
2522
2523//The total number of operations is 1.0
2524template <> template <uint64_t msk> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::shufflei(bitblock128_t arg1)
2525{
2526        return mvmd128<32>::shufflei<shufflemask4_from_shufflemask2(msk)>(arg1);
2527}
2528
2529//The total number of operations is 5.66666666667
2530template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2531{
2532        return simd_or(mvmd128<2>::srli<sh>(arg1), mvmd128<2>::slli<((64)-sh)>(arg2));
2533}
2534
2535//The total number of operations is 5.66666666667
2536template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2537{
2538        return simd_or(mvmd128<4>::srli<sh>(arg1), mvmd128<4>::slli<((32)-sh)>(arg2));
2539}
2540
2541//The total number of operations is 3.0
2542template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2543{
2544        return simd_or(mvmd128<8>::srli<sh>(arg1), mvmd128<8>::slli<((16)-sh)>(arg2));
2545}
2546
2547//The total number of operations is 3.0
2548template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2549{
2550        return simd_or(mvmd128<16>::srli<sh>(arg1), mvmd128<16>::slli<((8)-sh)>(arg2));
2551}
2552
2553//The total number of operations is 3.0
2554template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2555{
2556        return simd_or(mvmd128<32>::srli<sh>(arg1), mvmd128<32>::slli<((4)-sh)>(arg2));
2557}
2558
2559//The total number of operations is 3.0
2560template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2561{
2562        return simd_or(mvmd128<64>::srli<sh>(arg1), mvmd128<64>::slli<((2)-sh)>(arg2));
2563}
2564
2565//The total number of operations is 3.0
2566template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dsrli(bitblock128_t arg1, bitblock128_t arg2)
2567{
2568        return simd_or(mvmd128<128>::srli<sh>(arg1), mvmd128<128>::slli<((1)-sh)>(arg2));
2569}
2570
2571//The total number of operations is 1.0
2572template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill(uint64_t val1)
2573{
2574        return mvmd128<32>::fill((-1*val1));
2575}
2576
2577//The total number of operations is 1.0
2578template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill(uint64_t val1)
2579{
2580        return mvmd128<(4)>::fill(((val1<<2)|val1));
2581}
2582
2583//The total number of operations is 1.0
2584template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill(uint64_t val1)
2585{
2586        return mvmd128<(8)>::fill(((val1<<4)|val1));
2587}
2588
2589//The total number of operations is 1.0
2590template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill(uint64_t val1)
2591{
2592        return _mm_set1_epi8((int32_t)(val1));
2593}
2594
2595//The total number of operations is 1.0
2596template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill(uint64_t val1)
2597{
2598        return _mm_set1_epi16((int32_t)(val1));
2599}
2600
2601//The total number of operations is 1.0
2602template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill(uint64_t val1)
2603{
2604        return _mm_set1_epi32((int32_t)(val1));
2605}
2606
2607//The total number of operations is 1.0
2608template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill(uint64_t val1)
2609{
2610        return _mm_set_epi32((int32_t)((val1>>32)), (int32_t)(val1), (int32_t)((val1>>32)), (int32_t)(val1));
2611}
2612
2613//The total number of operations is 1.0
2614template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::fill(uint64_t val1)
2615{
2616        return _mm_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val1>>32)), (int32_t)(val1));
2617}
2618
2619//The total number of operations is 1.0
2620template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<1>::extract(bitblock128_t arg1)
2621{
2622        return (((pos%2) == 0) ? (mvmd128<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd128<(2)>::extract<(pos/2)>(arg1)>>1));
2623}
2624
2625//The total number of operations is 1.0
2626template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<2>::extract(bitblock128_t arg1)
2627{
2628        return (((pos%2) == 0) ? (mvmd128<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd128<(4)>::extract<(pos/2)>(arg1)>>2));
2629}
2630
2631//The total number of operations is 1.0
2632template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<4>::extract(bitblock128_t arg1)
2633{
2634        return (((pos%2) == 0) ? (mvmd128<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd128<(8)>::extract<(pos/2)>(arg1)>>4));
2635}
2636
2637//The total number of operations is 1.0
2638template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<8>::extract(bitblock128_t arg1)
2639{
2640        return (((pos%2) == 0) ? (mvmd128<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd128<(16)>::extract<(pos/2)>(arg1)>>8));
2641}
2642
2643//The total number of operations is 1.0
2644template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<16>::extract(bitblock128_t arg1)
2645{
2646        return (65535&_mm_extract_epi16(arg1, (int32_t)(pos)));
2647}
2648
2649//The total number of operations is 2.0
2650template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<32>::extract(bitblock128_t arg1)
2651{
2652        return ((mvmd128<(16)>::extract<((2*pos)+1)>(arg1)<<(16))|mvmd128<(16)>::extract<(2*pos)>(arg1));
2653}
2654
2655//The total number of operations is 4.0
2656template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd128<64>::extract(bitblock128_t arg1)
2657{
2658        return ((mvmd128<(32)>::extract<((2*pos)+1)>(arg1)<<(32))|mvmd128<(32)>::extract<(2*pos)>(arg1));
2659}
2660
2661//The total number of operations is 12.6666666667
2662template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::splat(bitblock128_t arg1)
2663{
2664        return simd128<128>::sub(simd128<128>::constant<0>(), simd_and(simd128<128>::constant<1>(), simd128<128>::srli<pos>(arg1)));
2665}
2666
2667//The total number of operations is 13.0
2668template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::splat(bitblock128_t arg1)
2669{
2670        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(4)>::slli<2>(arg1) : simd128<(4)>::srli<2>(arg1));
2671        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(4)>::lomask(), arg1) : simd_and(simd128<(4)>::himask(), arg1));
2672        return mvmd128<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2673}
2674
2675//The total number of operations is 9.0
2676template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::splat(bitblock128_t arg1)
2677{
2678        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(8)>::slli<4>(arg1) : simd128<(8)>::srli<4>(arg1));
2679        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(8)>::lomask(), arg1) : simd_and(simd128<(8)>::himask(), arg1));
2680        return mvmd128<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2681}
2682
2683//The total number of operations is 5.0
2684template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::splat(bitblock128_t arg1)
2685{
2686        bitblock128_t tmpArg = (((pos%2) == 0) ? simd128<(16)>::slli<8>(arg1) : simd128<(16)>::srli<8>(arg1));
2687        bitblock128_t arg11 = (((pos%2) == 0) ? simd_and(simd128<(16)>::lomask(), arg1) : simd_and(simd128<(16)>::himask(), arg1));
2688        return mvmd128<(16)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
2689}
2690
2691//The total number of operations is 2.0
2692template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::splat(bitblock128_t arg1)
2693{
2694        return mvmd128<16>::fill(_mm_extract_epi16(arg1, (int32_t)(pos)));
2695}
2696
2697//The total number of operations is 1.0
2698template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::splat(bitblock128_t arg1)
2699{
2700        return mvmd128<32>::shufflei<shufflemask4(pos, pos, pos, pos)>(arg1);
2701}
2702
2703//The total number of operations is 5.0
2704template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::splat(bitblock128_t arg1)
2705{
2706        return simd128<1>::ifh(simd128<64>::himask(), mvmd128<(32)>::splat<((2*pos)+1)>(arg1), mvmd128<(32)>::splat<(2*pos)>(arg1));
2707}
2708
2709//The total number of operations is 13.0
2710template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::splat(bitblock128_t arg1)
2711{
2712        return simd128<1>::ifh(simd128<128>::himask(), mvmd128<(64)>::splat<((2*pos)+1)>(arg1), mvmd128<(64)>::splat<(2*pos)>(arg1));
2713}
2714
2715//The total number of operations is 15.0
2716template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2717{
2718        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))));
2719}
2720
2721//The total number of operations is 7.0
2722template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2723{
2724        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))));
2725}
2726
2727//The total number of operations is 3.0
2728template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2729{
2730        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))));
2731}
2732
2733//The total number of operations is 1.0
2734template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
2735{
2736        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));
2737}
2738
2739//The total number of operations is 5.0
2740template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2741{
2742        return simd128<1>::ifh(simd128<(4)>::himask(), mvmd128<1>::fill2(val1, val2), mvmd128<1>::fill2(val3, val4));
2743}
2744
2745//The total number of operations is 5.0
2746template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2747{
2748        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<2>::fill2(val1, val2), mvmd128<2>::fill2(val3, val4));
2749}
2750
2751//The total number of operations is 5.0
2752template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2753{
2754        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<4>::fill2(val1, val2), mvmd128<4>::fill2(val3, val4));
2755}
2756
2757//The total number of operations is 5.0
2758template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2759{
2760        return simd128<1>::ifh(simd128<(32)>::himask(), mvmd128<8>::fill2(val1, val2), mvmd128<8>::fill2(val3, val4));
2761}
2762
2763//The total number of operations is 3.0
2764template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2765{
2766        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))));
2767}
2768
2769//The total number of operations is 1.0
2770template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
2771{
2772        return _mm_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
2773}
2774
2775//The total number of operations is 2.33333333333
2776template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::srli(bitblock128_t arg1)
2777{
2778        return simd128<128>::srli<(sh*2)>(arg1);
2779}
2780
2781//The total number of operations is 2.33333333333
2782template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::srli(bitblock128_t arg1)
2783{
2784        return simd128<128>::srli<(sh*4)>(arg1);
2785}
2786
2787//The total number of operations is 1.0
2788template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::srli(bitblock128_t arg1)
2789{
2790        return _mm_srli_si128(arg1, (int32_t)(sh));
2791}
2792
2793//The total number of operations is 1.0
2794template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::srli(bitblock128_t arg1)
2795{
2796        return mvmd128<(8)>::srli<(sh*2)>(arg1);
2797}
2798
2799//The total number of operations is 1.0
2800template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::srli(bitblock128_t arg1)
2801{
2802        return mvmd128<(16)>::srli<(sh*2)>(arg1);
2803}
2804
2805//The total number of operations is 1.0
2806template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::srli(bitblock128_t arg1)
2807{
2808        return mvmd128<(32)>::srli<(sh*2)>(arg1);
2809}
2810
2811//The total number of operations is 1.0
2812template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::srli(bitblock128_t arg1)
2813{
2814        return mvmd128<(64)>::srli<(sh*2)>(arg1);
2815}
2816
2817//The total number of operations is 1.0
2818template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill2(uint64_t val1, uint64_t val2)
2819{
2820        return mvmd128<(2)>::fill(((val1<<1)|(val2&(1))));
2821}
2822
2823//The total number of operations is 1.0
2824template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill2(uint64_t val1, uint64_t val2)
2825{
2826        return mvmd128<(4)>::fill(((val1<<2)|(val2&(3))));
2827}
2828
2829//The total number of operations is 1.0
2830template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill2(uint64_t val1, uint64_t val2)
2831{
2832        return mvmd128<(8)>::fill(((val1<<4)|(val2&(15))));
2833}
2834
2835//The total number of operations is 1.0
2836template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill2(uint64_t val1, uint64_t val2)
2837{
2838        return mvmd128<(16)>::fill(((val1<<8)|(val2&(255))));
2839}
2840
2841//The total number of operations is 1.0
2842template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill2(uint64_t val1, uint64_t val2)
2843{
2844        return mvmd128<(32)>::fill(((val1<<16)|(val2&(65535))));
2845}
2846
2847//The total number of operations is 5.0
2848template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::fill2(uint64_t val1, uint64_t val2)
2849{
2850        return simd128<1>::ifh(simd128<(64)>::himask(), mvmd128<32>::fill(val1), mvmd128<32>::fill(val2));
2851}
2852
2853//The total number of operations is 5.0
2854template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::fill2(uint64_t val1, uint64_t val2)
2855{
2856        return simd128<1>::ifh(simd128<(128)>::himask(), mvmd128<64>::fill(val1), mvmd128<64>::fill(val2));
2857}
2858
2859//The total number of operations is 5.66666666667
2860template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2861{
2862        return simd_or(mvmd128<2>::slli<sh>(arg1), mvmd128<2>::srli<((64)-sh)>(arg2));
2863}
2864
2865//The total number of operations is 5.66666666667
2866template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2867{
2868        return simd_or(mvmd128<4>::slli<sh>(arg1), mvmd128<4>::srli<((32)-sh)>(arg2));
2869}
2870
2871//The total number of operations is 3.0
2872template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2873{
2874        return simd_or(mvmd128<8>::slli<sh>(arg1), mvmd128<8>::srli<((16)-sh)>(arg2));
2875}
2876
2877//The total number of operations is 3.0
2878template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2879{
2880        return simd_or(mvmd128<16>::slli<sh>(arg1), mvmd128<16>::srli<((8)-sh)>(arg2));
2881}
2882
2883//The total number of operations is 3.0
2884template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2885{
2886        return simd_or(mvmd128<32>::slli<sh>(arg1), mvmd128<32>::srli<((4)-sh)>(arg2));
2887}
2888
2889//The total number of operations is 3.0
2890template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2891{
2892        return simd_or(mvmd128<64>::slli<sh>(arg1), mvmd128<64>::srli<((2)-sh)>(arg2));
2893}
2894
2895//The total number of operations is 3.0
2896template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::dslli(bitblock128_t arg1, bitblock128_t arg2)
2897{
2898        return simd_or(mvmd128<128>::slli<sh>(arg1), mvmd128<128>::srli<((1)-sh)>(arg2));
2899}
2900
2901//The total number of operations is 2.33333333333
2902template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::slli(bitblock128_t arg1)
2903{
2904        return simd128<128>::slli<(sh*2)>(arg1);
2905}
2906
2907//The total number of operations is 2.33333333333
2908template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::slli(bitblock128_t arg1)
2909{
2910        return mvmd128<(2)>::slli<(sh*2)>(arg1);
2911}
2912
2913//The total number of operations is 1.0
2914template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::slli(bitblock128_t arg1)
2915{
2916        return _mm_slli_si128(arg1, (int32_t)(sh));
2917}
2918
2919//The total number of operations is 1.0
2920template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::slli(bitblock128_t arg1)
2921{
2922        return mvmd128<(8)>::slli<(sh*2)>(arg1);
2923}
2924
2925//The total number of operations is 1.0
2926template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<32>::slli(bitblock128_t arg1)
2927{
2928        return mvmd128<(16)>::slli<(sh*2)>(arg1);
2929}
2930
2931//The total number of operations is 1.0
2932template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<64>::slli(bitblock128_t arg1)
2933{
2934        return mvmd128<(32)>::slli<(sh*2)>(arg1);
2935}
2936
2937//The total number of operations is 1.0
2938template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<128>::slli(bitblock128_t arg1)
2939{
2940        return mvmd128<(64)>::slli<(sh*2)>(arg1);
2941}
2942
2943//The total number of operations is 13.0
2944template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<1>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2945{
2946        return simd128<1>::ifh(simd128<(8)>::himask(), mvmd128<1>::fill4(val1, val2, val3, val4), mvmd128<1>::fill4(val5, val6, val7, val8));
2947}
2948
2949//The total number of operations is 13.0
2950template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<2>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2951{
2952        return simd128<1>::ifh(simd128<(16)>::himask(), mvmd128<2>::fill4(val1, val2, val3, val4), mvmd128<2>::fill4(val5, val6, val7, val8));
2953}
2954
2955//The total number of operations is 7.0
2956template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<4>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2957{
2958        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))));
2959}
2960
2961//The total number of operations is 3.0
2962template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<8>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2963{
2964        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))));
2965}
2966
2967//The total number of operations is 1.0
2968template <> IDISA_ALWAYS_INLINE bitblock128_t mvmd128<16>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
2969{
2970        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));
2971}
2972
2973//The total number of operations is 13.0
2974IDISA_ALWAYS_INLINE bitblock128_t bitblock128::sll(bitblock128_t arg1, bitblock128_t arg2)
2975{
2976        return simd128<128>::sll(arg1, arg2);
2977}
2978
2979//The total number of operations is 1.0
2980IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(const bitblock128_t* arg1)
2981{
2982        return _mm_loadu_si128((bitblock128_t*)(arg1));
2983}
2984
2985//The total number of operations is 2.33333333333
2986template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srli(bitblock128_t arg1)
2987{
2988        return simd128<128>::srli<sh>(arg1);
2989}
2990
2991//The total number of operations is 1.0
2992IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t arg1, bitblock128_t* arg2)
2993{
2994        _mm_store_si128((bitblock128_t*)(arg2), arg1);
2995}
2996
2997//The total number of operations is 20.3333333333
2998IDISA_ALWAYS_INLINE uint64_t bitblock128::popcount(bitblock128_t arg1)
2999{
3000        return mvmd128<64>::extract<0>(simd128<128>::popcount(arg1));
3001}
3002
3003//The total number of operations is 2.0
3004IDISA_ALWAYS_INLINE bool bitblock128::all(bitblock128_t arg1)
3005{
3006        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<-1>())) == 65535;
3007}
3008
3009//The total number of operations is 13.0
3010IDISA_ALWAYS_INLINE bitblock128_t bitblock128::srl(bitblock128_t arg1, bitblock128_t arg2)
3011{
3012        return simd128<128>::srl(arg1, arg2);
3013}
3014
3015//The total number of operations is 2.33333333333
3016template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock128_t bitblock128::slli(bitblock128_t arg1)
3017{
3018        return simd128<128>::slli<sh>(arg1);
3019}
3020
3021//The total number of operations is 2.0
3022IDISA_ALWAYS_INLINE bool bitblock128::any(bitblock128_t arg1)
3023{
3024        return hsimd128<8>::signmask(simd128<8>::eq(arg1, simd128<8>::constant<0>())) != 65535;
3025}
3026
3027//The total number of operations is 1.0
3028IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(const bitblock128_t* arg1)
3029{
3030        return _mm_load_si128((bitblock128_t*)(arg1));
3031}
3032
3033//The total number of operations is 1.0
3034IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t arg1, bitblock128_t* arg2)
3035{
3036        _mm_storeu_si128((bitblock128_t*)(arg2), arg1);
3037}
3038
3039#endif
Note: See TracBrowser for help on using the repository browser.