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
|
#pragma once
// This header provides C++ wrappers around commonly used CUDA API functions.
// The benefit of using C++ here is that we can raise an exception in the
// event of an error, rather than explicitly pass around error codes. This
// leads to more natural APIs.
//
// The naming convention used here matches the naming convention of torch.cuda
#include <c10/core/Device.h>
#include <c10/core/impl/GPUTrace.h>
#include <c10/cuda/CUDAException.h>
#include <c10/cuda/CUDAMacros.h>
#include <cuda_runtime_api.h>
namespace c10 {
namespace cuda {
// NB: In the past, we were inconsistent about whether or not this reported
// an error if there were driver problems are not. Based on experience
// interacting with users, it seems that people basically ~never want this
// function to fail; it should just return zero if things are not working.
// Oblige them.
// It still might log a warning for user first time it's invoked
C10_CUDA_API DeviceIndex device_count() noexcept;
// Version of device_count that throws is no devices are detected
C10_CUDA_API DeviceIndex device_count_ensure_non_zero();
C10_CUDA_API DeviceIndex current_device();
C10_CUDA_API void set_device(DeviceIndex device);
C10_CUDA_API void device_synchronize();
C10_CUDA_API void warn_or_error_on_sync();
enum class SyncDebugMode { L_DISABLED = 0, L_WARN, L_ERROR };
// this is a holder for c10 global state (similar to at GlobalContext)
// currently it's used to store cuda synchronization warning state,
// but can be expanded to hold other related global state, e.g. to
// record stream usage
class WarningState {
public:
void set_sync_debug_mode(SyncDebugMode l) {
sync_debug_mode = l;
}
SyncDebugMode get_sync_debug_mode() {
return sync_debug_mode;
}
private:
SyncDebugMode sync_debug_mode = SyncDebugMode::L_DISABLED;
};
C10_CUDA_API __inline__ WarningState& warning_state() {
static WarningState warning_state_;
return warning_state_;
}
// the subsequent functions are defined in the header because for performance
// reasons we want them to be inline
C10_CUDA_API void __inline__ memcpy_and_sync(
void* dst,
void* src,
int64_t nbytes,
cudaMemcpyKind kind,
cudaStream_t stream) {
if (C10_UNLIKELY(
warning_state().get_sync_debug_mode() != SyncDebugMode::L_DISABLED)) {
warn_or_error_on_sync();
}
const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
if (C10_UNLIKELY(interp)) {
(*interp)->trace_gpu_stream_synchronization(
reinterpret_cast<uintptr_t>(stream));
}
#if defined(TORCH_HIP_VERSION) && (TORCH_HIP_VERSION >= 301)
C10_CUDA_CHECK(hipMemcpyWithStream(dst, src, nbytes, kind, stream));
#else
C10_CUDA_CHECK(cudaMemcpyAsync(dst, src, nbytes, kind, stream));
C10_CUDA_CHECK(cudaStreamSynchronize(stream));
#endif
}
C10_CUDA_API void __inline__ stream_synchronize(cudaStream_t stream) {
if (C10_UNLIKELY(
warning_state().get_sync_debug_mode() != SyncDebugMode::L_DISABLED)) {
warn_or_error_on_sync();
}
const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
if (C10_UNLIKELY(interp)) {
(*interp)->trace_gpu_stream_synchronization(
reinterpret_cast<uintptr_t>(stream));
}
C10_CUDA_CHECK(cudaStreamSynchronize(stream));
}
} // namespace cuda
} // namespace c10
|
