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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
|
use std::borrow::Cow;
use std::collections::HashMap;
use std::ffi::{CStr, CString};
use std::fs::File;
use std::io::{self, Read, Write};
use std::ops::Add;
use std::os::raw::c_char;
use std::path::{Path, PathBuf};
use std::process::Command;
use std::{env, fs, iter, mem, ptr, slice};
use hip_runtime_sys::{
hipCtxGetCurrent, hipCtxGetDevice, hipDeviceGetAttribute, hipDeviceGetName, hipDeviceProp_t,
hipError_t, hipGetDeviceProperties, hipGetStreamDeviceId, hipModuleLoadData,
};
use tempfile::NamedTempFile;
use crate::cuda::CUmodule;
pub struct SpirvModule {
pub binaries: Vec<u32>,
pub kernel_info: HashMap<String, ptx::KernelInfo>,
pub should_link_ptx_impl: Option<(&'static [u8], &'static [u8])>,
pub build_options: CString,
}
impl SpirvModule {
pub fn new_raw<'a>(text: *const c_char) -> Result<Self, hipError_t> {
let u8_text = unsafe { CStr::from_ptr(text) };
let ptx_text = u8_text
.to_str()
.map_err(|_| hipError_t::hipErrorInvalidImage)?;
Self::new(ptx_text)
}
pub fn new<'a>(ptx_text: &str) -> Result<Self, hipError_t> {
let mut errors = Vec::new();
let ast = ptx::ModuleParser::new()
.parse(&mut errors, ptx_text)
.map_err(|_| hipError_t::hipErrorInvalidImage)?;
if errors.len() > 0 {
return Err(hipError_t::hipErrorInvalidImage);
}
let spirv_module =
ptx::to_spirv_module(ast).map_err(|_| hipError_t::hipErrorInvalidImage)?;
Ok(SpirvModule {
binaries: spirv_module.assemble(),
kernel_info: spirv_module.kernel_info,
should_link_ptx_impl: spirv_module.should_link_ptx_impl,
build_options: spirv_module.build_options,
})
}
}
pub(crate) fn load(module: *mut CUmodule, fname: *const i8) -> Result<(), hipError_t> {
let length = (0..)
.position(|i| unsafe { *fname.add(i) == 0 })
.ok_or(hipError_t::hipErrorInvalidValue)?;
let file_name = CStr::from_bytes_with_nul(unsafe { slice::from_raw_parts(fname as _, length) })
.map_err(|_| hipError_t::hipErrorInvalidValue)?;
let valid_file_name = file_name
.to_str()
.map_err(|_| hipError_t::hipErrorInvalidValue)?;
let mut file = File::open(valid_file_name).map_err(|_| hipError_t::hipErrorFileNotFound)?;
let mut file_buffer = Vec::new();
file.read_to_end(&mut file_buffer)
.map_err(|_| hipError_t::hipErrorUnknown)?;
drop(file);
load_data(module, file_buffer.as_ptr() as _)
}
pub(crate) fn load_data(
module: *mut CUmodule,
image: *const std::ffi::c_void,
) -> Result<(), hipError_t> {
let spirv_data = SpirvModule::new_raw(image as *const _)?;
load_data_impl(module, spirv_data)
}
pub fn load_data_impl(pmod: *mut CUmodule, spirv_data: SpirvModule) -> Result<(), hipError_t> {
let mut dev = 0;
let err = unsafe { hipCtxGetDevice(&mut dev) };
if err != hipError_t::hipSuccess {
return Err(err);
}
let mut props = unsafe { mem::zeroed() };
let err = unsafe { hipGetDeviceProperties(&mut props, dev) };
let arch_binary = compile_amd(
&props,
iter::once(&spirv_data.binaries[..]),
spirv_data.should_link_ptx_impl,
)
.map_err(|_| hipError_t::hipErrorUnknown)?;
let err = unsafe { hipModuleLoadData(pmod as _, arch_binary.as_ptr() as _) };
if err != hipError_t::hipSuccess {
return Err(err);
}
Ok(())
}
const LLVM_SPIRV: &'static str = "/home/vosen/amd/llvm-project/build/bin/llvm-spirv";
const AMDGPU: &'static str = "/opt/rocm/";
const AMDGPU_TARGET: &'static str = "amdgcn-amd-amdhsa";
const AMDGPU_BITCODE: [&'static str; 8] = [
"opencl.bc",
"ocml.bc",
"ockl.bc",
"oclc_correctly_rounded_sqrt_off.bc",
"oclc_daz_opt_on.bc",
"oclc_finite_only_off.bc",
"oclc_unsafe_math_off.bc",
"oclc_wavefrontsize64_off.bc",
];
const AMDGPU_BITCODE_DEVICE_PREFIX: &'static str = "oclc_isa_version_";
pub(crate) fn compile_amd<'a>(
device_pros: &hipDeviceProp_t,
spirv_il: impl Iterator<Item = &'a [u32]>,
ptx_lib: Option<(&'static [u8], &'static [u8])>,
) -> io::Result<Vec<u8>> {
let null_terminator = device_pros
.gcnArchName
.iter()
.position(|&x| x == 0)
.unwrap();
let gcn_arch_slice = unsafe {
slice::from_raw_parts(device_pros.gcnArchName.as_ptr() as _, null_terminator + 1)
};
let device_name =
if let Ok(Ok(name)) = CStr::from_bytes_with_nul(gcn_arch_slice).map(|x| x.to_str()) {
name
} else {
return Err(io::Error::new(io::ErrorKind::Other, ""));
};
let dir = tempfile::tempdir()?;
let spirv_files = spirv_il
.map(|spirv| {
let mut spirv_file = NamedTempFile::new_in(&dir)?;
let spirv_u8 = unsafe {
slice::from_raw_parts(
spirv.as_ptr() as *const u8,
spirv.len() * mem::size_of::<u32>(),
)
};
spirv_file.write_all(spirv_u8)?;
Ok::<_, io::Error>(spirv_file)
})
.collect::<Result<Vec<_>, _>>()?;
let llvm_spirv_path = match env::var("LLVM_SPIRV") {
Ok(path) => Cow::Owned(path),
Err(_) => Cow::Borrowed(LLVM_SPIRV),
};
let llvm = NamedTempFile::new_in(&dir)?;
let to_llvm_cmd = Command::new(&*llvm_spirv_path)
.arg("-r")
.arg("-o")
.arg(llvm.path())
.args(spirv_files.iter().map(|f| f.path()))
.status()?;
assert!(to_llvm_cmd.success());
if cfg!(debug_assertions) {
persist_file(llvm.path())?;
}
let linked_binary = NamedTempFile::new_in(&dir)?;
let mut llvm_link = PathBuf::from(AMDGPU);
llvm_link.push("llvm");
llvm_link.push("bin");
llvm_link.push("llvm-link");
let mut linker_cmd = Command::new(&llvm_link);
linker_cmd
.arg("--only-needed")
.arg("-o")
.arg(linked_binary.path())
.arg(llvm.path())
.args(get_bitcode_paths(device_name));
if cfg!(debug_assertions) {
linker_cmd.arg("-v");
}
let status = linker_cmd.status()?;
assert!(status.success());
if cfg!(debug_assertions) {
persist_file(linked_binary.path())?;
}
let mut ptx_lib_bitcode = NamedTempFile::new_in(&dir)?;
let compiled_binary = NamedTempFile::new_in(&dir)?;
let mut clang_exe = PathBuf::from(AMDGPU);
clang_exe.push("llvm");
clang_exe.push("bin");
clang_exe.push("clang");
let mut compiler_cmd = Command::new(&clang_exe);
compiler_cmd
.arg(format!("-mcpu={}", device_name))
.arg("-ffp-contract=off")
.arg("-nogpulib")
.arg("-mno-wavefrontsize64")
.arg("-O3")
.arg("-Xlinker")
.arg("--no-undefined")
.arg("-target")
.arg(AMDGPU_TARGET)
.arg("-o")
.arg(compiled_binary.path())
.arg("-x")
.arg("ir")
.arg(linked_binary.path());
if let Some((_, bitcode)) = ptx_lib {
ptx_lib_bitcode.write_all(bitcode)?;
compiler_cmd.arg(ptx_lib_bitcode.path());
};
if cfg!(debug_assertions) {
compiler_cmd.arg("-v");
}
let status = compiler_cmd.status()?;
assert!(status.success());
let mut result = Vec::new();
let compiled_bin_path = compiled_binary.path();
let mut compiled_binary = File::open(compiled_bin_path)?;
compiled_binary.read_to_end(&mut result)?;
if cfg!(debug_assertions) {
persist_file(compiled_bin_path)?;
}
Ok(result)
}
fn persist_file(path: &Path) -> io::Result<()> {
let mut persistent = PathBuf::from("/tmp/zluda");
std::fs::create_dir_all(&persistent)?;
persistent.push(path.file_name().unwrap());
std::fs::copy(path, persistent)?;
Ok(())
}
fn get_bitcode_paths(device_name: &str) -> impl Iterator<Item = PathBuf> {
let generic_paths = AMDGPU_BITCODE.iter().map(|x| {
let mut path = PathBuf::from(AMDGPU);
path.push("amdgcn");
path.push("bitcode");
path.push(x);
path
});
let suffix = if let Some(suffix_idx) = device_name.find(':') {
suffix_idx
} else {
device_name.len()
};
let mut additional_path = PathBuf::from(AMDGPU);
additional_path.push("amdgcn");
additional_path.push("bitcode");
additional_path.push(format!(
"{}{}{}",
AMDGPU_BITCODE_DEVICE_PREFIX,
&device_name[3..suffix],
".bc"
));
generic_paths.chain(std::iter::once(additional_path))
}
|