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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
#[cfg(not(feature = "cuda"))]
fn main() {
println!("cargo:rerun-if-changed=build.rs");
}
#[cfg(feature = "cuda")]
fn main() {
build::run();
}
#[cfg(feature = "cuda")]
mod build {
use cudaforge::{CudaToolkit, KernelBuilder};
use std::env;
use std::path::{Path, PathBuf};
/// One SASS target, or a PTX-only fallback target (embedded IR, no
/// machine code, JIT-compiled by the driver on unlisted GPUs).
#[derive(Clone, Copy)]
enum ArchEntry {
Sass(u32),
Ptx(u32),
}
// Shared decision (docs/plan/README.md): one fat binary per CUDA major,
// not one build per SM. No 'a'/'f' suffix on any compiled arch - see
// resolve_archs's comment on why we build -gencode strings ourselves
// instead of going through cudaforge's GpuArch::auto_suffix.
const CUDA12_DEFAULT: &[ArchEntry] = &[
ArchEntry::Sass(75),
ArchEntry::Sass(80),
ArchEntry::Sass(86),
ArchEntry::Sass(89),
ArchEntry::Sass(90),
ArchEntry::Ptx(90),
];
const CUDA13_DEFAULT: &[ArchEntry] = &[
ArchEntry::Sass(75),
ArchEntry::Sass(80),
ArchEntry::Sass(86),
ArchEntry::Sass(89),
ArchEntry::Sass(90),
ArchEntry::Sass(100),
ArchEntry::Sass(120),
ArchEntry::Ptx(120),
];
pub fn run() {
println!("cargo:rerun-if-changed=cuda");
println!("cargo:rerun-if-env-changed=CUJSON_CUDA_ARCHS");
println!("cargo:rerun-if-env-changed=NVCC");
println!("cargo:rerun-if-env-changed=CUDA_HOME");
let out_dir = PathBuf::from(env::var("OUT_DIR").expect("OUT_DIR not set by cargo"));
let toolkit = CudaToolkit::detect().unwrap_or_else(|e| {
panic!(
"cujson-sys's `cuda` feature needs a CUDA toolkit with nvcc. \
Set the NVCC environment variable to the nvcc binary, or \
CUDA_HOME to the toolkit root (so $CUDA_HOME/bin/nvcc exists). \
Detection error: {e}"
)
});
let archs = resolve_archs(&toolkit);
let archs_display = archs_display_string(&archs);
println!("cargo:archs={archs_display}");
let mut gencode_args: Vec<String> = archs
.iter()
.map(|a| match a {
// No 'a'/'f' suffix (task 04 brief): cudaforge's GpuArch
// auto-suffixes any numeric cap >= 90 (e.g. 90 -> sm_90a),
// and sm_90a code doesn't run on later GPUs. Building the
// -gencode string ourselves keeps every compiled arch
// plain (sm_90, sm_100, sm_120, not *_a).
ArchEntry::Sass(n) => format!("-gencode=arch=compute_{n},code=sm_{n}"),
ArchEntry::Ptx(n) => format!("-gencode=arch=compute_{n},code=compute_{n}"),
})
.collect();
gencode_args.push(format!("-DCUJSON_COMPILED_ARCHS=\"{archs_display}\""));
for extra in ["-std=c++17", "-O3", "-w", "-Xcompiler", "-fPIC"] {
gencode_args.push(extra.to_string());
}
let lib_path = out_dir.join("libcujson.a");
KernelBuilder::new()
.source_files([
"cuda/capi_standard.cu",
"cuda/capi_lines.cu",
"cuda/capi_common.cu",
])
// Every file, not the directory: cudaforge hashes only .h/.cuh/.hpp
// under a watched directory, but the unity TUs also #include
// upstream .cu and .cpp files, whose edits would otherwise leave
// stale objects in place.
.watch(files_under(Path::new("cuda")))
// Forced low base purely so cudaforge's own per-file default
// -gencode (added automatically, one per file, see builder.rs)
// is always valid and never runs nvidia-smi (no GPU in this
// container - and CI/most build machines building this crate
// won't have one attached either). The full arch list above is
// what actually ends up in the fat binary; this one is
// redundant with (or a superset-neutral addition to) it.
.compute_cap_arch("75")
.args(&gencode_args)
.build_lib(&lib_path)
.expect("cuJSON CUDA kernel build failed");
let cuda_root = toolkit
.nvcc_path
.parent()
.and_then(|p| p.parent())
.expect("nvcc path has no toolkit root")
.to_path_buf();
for candidate in ["lib64", "lib", "targets/x86_64-linux/lib"] {
let dir = cuda_root.join(candidate);
if dir.is_dir() {
println!("cargo:rustc-link-search=native={}", dir.display());
}
}
println!("cargo:rustc-link-search=native={}", out_dir.display());
println!("cargo:rustc-link-lib=static=cujson");
println!("cargo:rustc-link-lib=static=cudart_static");
println!("cargo:rustc-link-lib=dylib=stdc++");
println!("cargo:rustc-link-lib=dylib=rt");
println!("cargo:rustc-link-lib=dylib=dl");
println!("cargo:rustc-link-lib=dylib=pthread");
}
/// CUJSON_CUDA_ARCHS overrides the default list; format is a comma list
/// of plain numbers (SASS) and/or "ptxNN" entries (PTX-only), e.g.
/// "89" or "80,90,ptx90". Otherwise picks CUDA12_DEFAULT/CUDA13_DEFAULT
/// from the detected toolkit's major version.
/// All files below `dir`, recursively, in a stable order.
fn files_under(dir: &Path) -> Vec<PathBuf> {
let mut files = Vec::new();
let mut entries: Vec<_> = std::fs::read_dir(dir)
.unwrap_or_else(|e| panic!("reading {}: {e}", dir.display()))
.map(|e| e.expect("reading directory entry").path())
.collect();
entries.sort();
for path in entries {
if path.is_dir() {
files.extend(files_under(&path));
} else {
files.push(path);
}
}
files
}
fn resolve_archs(toolkit: &CudaToolkit) -> Vec<ArchEntry> {
// An empty value (e.g. `CUJSON_CUDA_ARCHS: ""` in a CI matrix) means unset.
if let Some(raw) = env::var("CUJSON_CUDA_ARCHS")
.ok()
.filter(|s| !s.trim().is_empty())
{
let mut archs = Vec::new();
for tok in raw.split(',') {
let tok = tok.trim();
if tok.is_empty() {
continue;
}
if let Some(rest) = tok.strip_prefix("ptx") {
let n: u32 = rest
.parse()
.unwrap_or_else(|_| panic!("CUJSON_CUDA_ARCHS: bad ptx entry '{tok}'"));
archs.push(ArchEntry::Ptx(n));
} else {
let n: u32 = tok
.parse()
.unwrap_or_else(|_| panic!("CUJSON_CUDA_ARCHS: bad arch '{tok}'"));
archs.push(ArchEntry::Sass(n));
}
}
assert!(
!archs.is_empty(),
"CUJSON_CUDA_ARCHS is set but empty after parsing"
);
return archs;
}
let major = toolkit
.version
.as_deref()
.and_then(|v| v.split('.').next())
.and_then(|s| s.parse::<u32>().ok());
match major {
Some(13) => CUDA13_DEFAULT.to_vec(),
// Default to the CUDA 12 list for 12.x and for anything we
// couldn't parse a version out of - nvcc --version's "release
// X.Y" line is what CudaToolkit::detect() parses, and every
// toolkit this crate targets prints it.
_ => CUDA12_DEFAULT.to_vec(),
}
}
fn archs_display_string(archs: &[ArchEntry]) -> String {
let sass: Vec<String> = archs
.iter()
.filter_map(|a| match a {
ArchEntry::Sass(n) => Some(n.to_string()),
ArchEntry::Ptx(_) => None,
})
.collect();
let ptx: Vec<String> = archs
.iter()
.filter_map(|a| match a {
ArchEntry::Ptx(n) => Some(format!("ptx{n}")),
ArchEntry::Sass(_) => None,
})
.collect();
if ptx.is_empty() {
sass.join(",")
} else {
format!("{};{}", sass.join(","), ptx.join(","))
}
}
}