-
Notifications
You must be signed in to change notification settings - Fork 21.4k
/
CUDABlas.h
160 lines (134 loc) · 4.73 KB
/
CUDABlas.h
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
#pragma once
/*
Provides a subset of CUDA BLAS functions as templates:
gemm<Dtype>(transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c,
ldc)
gemv<Dtype>(transa, m, n, alpha, a, lda, x, incx, beta, y, incy)
dot<Dtype>(n, x, incx, y, incy, result)
where Dtype is double, float, at::Half or at::BFloat16 (ROCm, NOT for dot).
The functions are available in at::cuda::blas namespace.
*/
#include <ATen/cuda/CUDAContext.h>
namespace at {
namespace cuda {
namespace blas {
// RAII guard that sets the CuBLAS pointer mode and restores it to
// its previous value when the guard is destroyed
class PointerModeGuard {
public:
PointerModeGuard(cublasHandle_t handle, cublasPointerMode_t mode) :
handle(handle) {
TORCH_CUDABLAS_CHECK(cublasGetPointerMode(handle, &previous_mode));
TORCH_CUDABLAS_CHECK(cublasSetPointerMode(handle, mode));
}
~PointerModeGuard() {
cublasSetPointerMode(handle, previous_mode);
}
private:
cublasHandle_t handle;
cublasPointerMode_t previous_mode;
};
/* LEVEL 3 BLAS FUNCTIONS */
#define CUDABLAS_GEMM_ARGTYPES(Dtype) \
char transa, char transb, int64_t m, int64_t n, int64_t k, Dtype alpha, \
const Dtype *a, int64_t lda, const Dtype *b, int64_t ldb, Dtype beta, \
Dtype *c, int64_t ldc
template <typename Dtype>
inline void gemm(CUDABLAS_GEMM_ARGTYPES(Dtype)) {
AT_ERROR("at::cuda::blas::gemm: not implemented for ", typeid(Dtype).name());
}
template <>
void gemm<double>(CUDABLAS_GEMM_ARGTYPES(double));
template <>
void gemm<float>(CUDABLAS_GEMM_ARGTYPES(float));
#ifndef __HIP_PLATFORM_HCC__
template <>
void gemm<c10::complex<double>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<double>));
#endif
#ifndef __HIP_PLATFORM_HCC__
template <>
void gemm<c10::complex<float>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<float>));
#endif
template <>
void gemm<at::Half>(CUDABLAS_GEMM_ARGTYPES(at::Half));
#ifdef __HIP_PLATFORM_HCC__
template <>
void gemm<at::BFloat16>(CUDABLAS_GEMM_ARGTYPES(at::BFloat16));
#endif
#define CUDABLAS_BGEMM_ARGTYPES(Dtype) \
char transa, char transb, int64_t m, int64_t n, int64_t k, Dtype alpha, \
const Dtype *a, int64_t lda, int64_t stridea, \
const Dtype *b, int64_t ldb, int64_t strideb, \
Dtype beta, Dtype *c, int64_t ldc, int64_t stridec, int64_t num_batches
template <typename Dtype>
inline void bgemm(CUDABLAS_BGEMM_ARGTYPES(Dtype)) {
AT_ERROR("at::cuda::blas::bgemm: not implemented for ", typeid(Dtype).name());
}
template <>
void bgemm<double>(CUDABLAS_BGEMM_ARGTYPES(double));
template <>
void bgemm<float>(CUDABLAS_BGEMM_ARGTYPES(float));
#ifndef __HIP_PLATFORM_HCC__
template <>
void bgemm<c10::complex<double>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<double>));
template <>
void bgemm<c10::complex<float>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<float>));
#endif
#ifdef __HIP_PLATFORM_HCC__
template <>
void bgemm<at::BFloat16>(CUDABLAS_BGEMM_ARGTYPES(at::BFloat16));
#endif
template <>
void bgemm<at::Half>(CUDABLAS_BGEMM_ARGTYPES(at::Half));
/* LEVEL 2 BLAS FUNCTIONS */
#define CUDABLAS_GEMV_ARGTYPES(Dtype) \
char trans, int64_t m, int64_t n, Dtype alpha, const Dtype *a, int64_t lda, \
const Dtype *x, int64_t incx, Dtype beta, Dtype *y, int64_t incy
template <typename Dtype>
inline void gemv(CUDABLAS_GEMV_ARGTYPES(Dtype)) {
AT_ERROR("at::cuda::blas::gemv: not implemented for ", typeid(Dtype).name());
}
template <>
void gemv<double>(CUDABLAS_GEMV_ARGTYPES(double));
template <>
void gemv<float>(CUDABLAS_GEMV_ARGTYPES(float));
#ifndef __HIP_PLATFORM_HCC__
template <>
void gemv<c10::complex<double>>(CUDABLAS_GEMV_ARGTYPES(c10::complex<double>));
template <>
void gemv<c10::complex<float>>(CUDABLAS_GEMV_ARGTYPES(c10::complex<float>));
#endif
#ifdef __HIP_PLATFORM_HCC__
template <>
void gemv<at::BFloat16>(CUDABLAS_GEMV_ARGTYPES(at::BFloat16));
#endif
template <>
void gemv<at::Half>(CUDABLAS_GEMV_ARGTYPES(at::Half));
template <typename Dtype>
void ger(
int64_t m,
int64_t n,
Dtype alpha,
Dtype* x,
int64_t incx,
Dtype* y,
int64_t incy,
Dtype* a,
int64_t lda);
/* LEVEL 1 BLAS FUNCTIONS */
#define CUDABLAS_DOT_ARGTYPES(Dtype) \
cublasHandle_t handle, int n, const Dtype *x, int incx, const Dtype *y, \
int incy, Dtype *result
template <typename Dtype>
inline void dot(CUDABLAS_DOT_ARGTYPES(Dtype)) {
AT_ERROR("at::cuda::blas::dot: not implemented for ", typeid(Dtype).name());
}
template <>
void dot<double>(CUDABLAS_DOT_ARGTYPES(double));
template <>
void dot<float>(CUDABLAS_DOT_ARGTYPES(float));
template <>
void dot<at::Half>(CUDABLAS_DOT_ARGTYPES(at::Half));
} // namespace blas
} // namespace cuda
} // namespace at