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
|
{
lib,
fetchFromGitHub,
llvmPackages_20,
python3,
cmake,
boost,
libxml2,
libffi,
numactl,
makeWrapper,
config,
cudaPackages,
rocmPackages,
ompSupport ? true,
openclSupport ? false,
rocmSupport ? config.rocmSupport,
cudaSupport ? config.cudaSupport,
autoAddDriverRunpath,
runCommand,
callPackage,
symlinkJoin,
nix-update-script,
}:
let
inherit (llvmPackages) stdenv;
llvmPackages = llvmPackages_20;
in
stdenv.mkDerivation (finalAttrs: {
pname = "adaptivecpp";
version = "25.10.0";
src = fetchFromGitHub {
owner = "AdaptiveCpp";
repo = "AdaptiveCpp";
tag = "v${finalAttrs.version}";
sha256 = "sha256-Z3YuBtR6TVCLQHZCA88oA5N10SnLATVv0/cvb8xwZWs=";
};
# do not use old FindCUDA cmake module
postPatch = ''
rm cmake/FindCUDA.cmake
'';
# we may be able to get away with just wrapping hipcc and nothing more
# this is mainly so that if acpp tries doing <PATH_TO_HIPCC>/../amdgcn/bitcode
rocmMerged = symlinkJoin {
name = "rocm-merged";
paths = with rocmPackages; [
clr
rocm-core
rocm-device-libs
rocm-runtime
];
buildInputs = [ makeWrapper ];
postBuild = ''
wrapProgram $out/bin/hipcc \
--add-flags "--rocm-device-lib-path=$out/amdgcn/bitcode"
'';
};
nativeBuildInputs = [
cmake
makeWrapper
]
++ lib.optionals cudaSupport [
autoAddDriverRunpath
cudaPackages.cuda_nvcc
];
buildInputs = [
libxml2
libffi
boost
python3
numactl
llvmPackages.openmp
llvmPackages.libclang.dev
llvmPackages.llvm
]
++ lib.optionals cudaSupport [
cudaPackages.cuda_cudart
(lib.getOutput "stubs" cudaPackages.cuda_cudart)
];
# adaptivecpp makes use of clangs internal headers. Its cmake does not successfully discover them automatically on nixos, so we supply the path manually
cmakeFlags = [
(lib.cmakeFeature "ACPP_LLD_PATH" (lib.getExe' llvmPackages.lld "ld.lld"))
(lib.cmakeFeature "CLANG_INCLUDE_PATH" "${llvmPackages.libclang.dev}/include")
(lib.cmakeBool "WITH_CPU_BACKEND" ompSupport)
(lib.cmakeBool "WITH_CUDA_BACKEND" cudaSupport)
(lib.cmakeBool "WITH_ROCM_BACKEND" rocmSupport)
(lib.cmakeBool "WITH_OPENCL_BACKEND" openclSupport)
]
++ lib.optionals rocmSupport [
(lib.cmakeFeature "HIPCC_COMPILER" "${finalAttrs.rocmMerged}/bin/hipcc")
(lib.cmakeFeature "ROCM_PATH" "${finalAttrs.rocmMerged}")
];
# this hardening option breaks rocm builds
hardeningDisable = [ "zerocallusedregs" ];
passthru = {
tests =
# Loosely based on the AdaptiveCpp GitHub CI: https://github.com/AdaptiveCpp/AdaptiveCpp/blob/develop/.github/workflows/linux.yml
# This may be overkill, especially as this won't be run on GPU on the CI
let
runner = targets: enablePstlTests: callPackage ./tests.nix { inherit enablePstlTests; };
run =
runner: cmd: mask:
runCommand cmd { } ''
# the test runner wants to write to $HOME/.acpp, so we need to have it point to a real directory
mkdir home
export HOME=`pwd`/home
ACPP_VISIBILITY_MASK="${mask}" ${runner}/bin/${cmd} && touch $out
'';
runSycl = runner: mask: run runner "sycl_tests" mask;
runPstl = runner: mask: run runner "pstl_tests" mask;
runner-omp = runner "omp" false;
runner-sscp = runner "generic" true;
in
{
inherit runner-omp runner-sscp;
sycl-omp = runSycl runner-omp "omp";
sycl-sscp = runSycl runner-sscp "omp";
pstl-sscp = runPstl runner-sscp "omp";
}
// lib.optionalAttrs rocmSupport (
let
runner-rocm = runner "hip:gfx906" false;
runner-rocm-integrated-multipass = runner "omp;hip:gfx906" false;
runner-rocm-explicit-multipass = runner "omp;hip.explicit-multipass:gfx906;cuda:sm_61" false;
in
{
inherit runner-rocm runner-rocm-integrated-multipass runner-rocm-explicit-multipass;
sycl-rocm = runSycl runner-rocm "omp;hip";
sycl-rocm-imp = runSycl runner-rocm-integrated-multipass "omp;hip";
sycl-rocm-emp = runSycl runner-rocm-explicit-multipass "omp;hip";
sycl-rocm-sscp = runSycl runner-sscp "omp;hip";
pstl-rocm-sscp = runPstl runner-sscp "omp;hip";
}
)
// lib.optionalAttrs cudaSupport (
let
runner-cuda = runner "cuda:sm_60" false;
runner-cuda-integrated-multipass = runner "omp;cuda_61" true;
runner-cuda-explicit-multipass = runner "omp;cuda.explicit-multipass:sm_61;hip:gfx906" false;
in
{
inherit runner-cuda runner-cuda-integrated-multipass runner-cuda-explicit-multipass;
sycl-cuda = runSycl runner-cuda "omp;cuda";
sycl-cuda-imp = runSycl runner-cuda-integrated-multipass "omp;cuda";
sycl-cuda-emp = runSycl runner-cuda-explicit-multipass "omp;cuda";
sycl-cuda-sscp = runSycl runner-sscp "omp;cuda";
pstl-cuda-imp = runPstl runner-cuda-integrated-multipass "omp;cuda";
pstl-cuda-sscp = runPstl runner-sscp "omp;cuda";
}
);
updateScript = nix-update-script { };
};
meta = {
homepage = "https://github.com/AdaptiveCpp/AdaptiveCpp";
description = "Multi-backend implementation of SYCL for CPUs and GPUs";
mainProgram = "acpp";
maintainers = with lib.maintainers; [ yboettcher ];
license = lib.licenses.bsd2;
};
})
|
