sp1-gpu-sys 6.2.2

FFI bindings and CUDA build system for SP1-GPU
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200

// Copyright 2022-2025 Dag Arne Osvik
// Licensed under the Apache License, Version 2.0, see LICENSE for details.
// SPDX-License-Identifier: Apache-2.0

#pragma once

#include <cstdint>

// Pack

__device__ __forceinline__ void pack(uint64_t& d, uint32_t a0, uint32_t a1) {
    asm("\n\tmov.b64 %0, {%1, %2};" : "=l"(d) : "r"(a0), "r"(a1));
}

// Unpack

__device__ __forceinline__ void unpack(uint32_t& d0, uint32_t& d1, uint64_t a) {
    asm("\n\tmov.b64 {%0, %1}, %2;" : "=r"(d0), "=r"(d1) : "l"(a));
}

// Add

__device__ __forceinline__ void add(uint32_t& d, uint32_t a, uint32_t b) {
    asm("\n\tadd.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void add(uint64_t& d, uint64_t a, uint64_t b) {
    asm("\n\tadd.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void add_cc(uint32_t& d, uint32_t a, uint32_t b) {
    asm volatile("\n\tadd.cc.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void add_cc(uint64_t& d, uint64_t a, uint64_t b) {
    asm volatile("\n\tadd.cc.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void addc_cc(uint32_t& d, uint32_t a, uint32_t b) {
    asm volatile("\n\taddc.cc.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void addc_cc(uint64_t& d, uint64_t a, uint64_t b) {
    asm volatile("\n\taddc.cc.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void addc(uint32_t& d, uint32_t a, uint32_t b) {
    asm volatile("\n\taddc.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void addc(uint64_t& d, uint64_t a, uint64_t b) {
    asm volatile("\n\taddc.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

// Sub

__device__ __forceinline__ void sub(uint32_t& d, uint32_t a, uint32_t b) {
    asm("\n\tsub.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void sub(uint64_t& d, uint64_t a, uint64_t b) {
    asm("\n\tsub.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void sub_cc(uint32_t& d, uint32_t a, uint32_t b) {
    asm volatile("\n\tsub.cc.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void sub_cc(uint64_t& d, uint64_t a, uint64_t b) {
    asm volatile("\n\tsub.cc.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void subc_cc(uint32_t& d, uint32_t a, uint32_t b) {
    asm volatile("\n\tsubc.cc.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void subc_cc(uint64_t& d, uint64_t a, uint64_t b) {
    asm volatile("\n\tsubc.cc.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void subc(uint32_t& d, uint32_t a, uint32_t b) {
    asm volatile("\n\tsubc.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void subc(uint64_t& d, uint64_t a, uint64_t b) {
    asm volatile("\n\tsubc.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

// Mul

__device__ __forceinline__ void mul_lo(uint32_t& d, uint32_t a, uint32_t b) {
    asm("\n\tmul.lo.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void mul_lo(uint64_t& d, uint64_t a, uint64_t b) {
    asm("\n\tmul.lo.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void mul_hi(uint32_t& d, uint32_t a, uint32_t b) {
    asm("\n\tmul.hi.u32 %0, %1, %2;" : "=r"(d) : "r"(a), "r"(b));
}
__device__ __forceinline__ void mul_hi(uint64_t& d, uint64_t a, uint64_t b) {
    asm("\n\tmul.hi.u64 %0, %1, %2;" : "=l"(d) : "l"(a), "l"(b));
}

__device__ __forceinline__ void mul_wide(uint64_t& d, uint32_t a, uint32_t b) {
    asm("\n\tmul.wide.u32 %0, %1, %2;" : "=l"(d) : "r"(a), "r"(b));
}

// Mad

__device__ __forceinline__ void mad_lo(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm("\n\tmad.lo.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void mad_lo(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm("\n\tmad.lo.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

__device__ __forceinline__ void mad_hi(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm("\n\tmad.hi.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void mad_hi(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm("\n\tmad.hi.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

__device__ __forceinline__ void mad_wide(uint64_t& d, uint32_t a, uint32_t b, uint64_t c) {
    asm("\n\tmad.wide.u32 %0, %1, %2, %3;" : "=l"(d) : "r"(a), "r"(b), "l"(c));
}

// Mad.cc

__device__ __forceinline__ void mad_lo_cc(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm volatile("\n\tmad.lo.cc.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void mad_lo_cc(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm volatile("\n\tmad.lo.cc.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

__device__ __forceinline__ void mad_hi_cc(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm volatile("\n\tmad.hi.cc.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void mad_hi_cc(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm volatile("\n\tmad.hi.cc.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

// Madc

__device__ __forceinline__ void madc_lo(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm volatile("\n\tmadc.lo.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void madc_lo(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm volatile("\n\tmadc.lo.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

__device__ __forceinline__ void madc_hi(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm volatile("\n\tmadc.hi.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void madc_hi(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm volatile("\n\tmadc.hi.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

// Madc.cc

__device__ __forceinline__ void madc_lo_cc(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm volatile("\n\tmadc.lo.cc.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void madc_lo_cc(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm volatile("\n\tmadc.lo.cc.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

__device__ __forceinline__ void madc_hi_cc(uint32_t& d, uint32_t a, uint32_t b, uint32_t c) {
    asm volatile("\n\tmadc.hi.cc.u32 %0, %1, %2, %3;" : "=r"(d) : "r"(a), "r"(b), "r"(c));
}
__device__ __forceinline__ void madc_hi_cc(uint64_t& d, uint64_t a, uint64_t b, uint64_t c) {
    asm volatile("\n\tmadc.hi.cc.u64 %0, %1, %2, %3;" : "=l"(d) : "l"(a), "l"(b), "l"(c));
}

// Wide mad.cc, madc.cc, madc
// Two PTX instructions each, assembled to one SASS instruction (tested on Ampere)

__device__ __forceinline__ void mad_wide_cc(uint64_t& d, uint32_t a, uint32_t b, uint64_t c) {
    asm volatile("\n\t{"
                 "\n\t.reg.u64 tmp;"
                 "\n\tmul.wide.u32 tmp, %1, %2;"
                 "\n\tadd.cc.u64   %0, tmp, %3;"
                 "\n\t}"
                 : "=l"(d)
                 : "r"(a), "r"(b), "l"(c));
}

__device__ __forceinline__ void madc_wide_cc(uint64_t& d, uint32_t a, uint32_t b, uint64_t c) {
    asm volatile("\n\t{"
                 "\n\t.reg.u64 tmp;"
                 "\n\tmul.wide.u32 tmp, %1, %2;"
                 "\n\taddc.cc.u64  %0, tmp, %3;"
                 "\n\t}"
                 : "=l"(d)
                 : "r"(a), "r"(b), "l"(c));
}

__device__ __forceinline__ void madc_wide(uint64_t& d, uint32_t a, uint32_t b, uint64_t c) {
    asm volatile("\n\t{"
                 "\n\t.reg.u64 tmp;"
                 "\n\tmul.wide.u32 tmp, %1, %2;"
                 "\n\taddc.u64     %0, tmp, %3;"
                 "\n\t}"
                 : "=l"(d)
                 : "r"(a), "r"(b), "l"(c));
}