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
|
#ifdef NERV_GENERIC_CUMATRIX
#include "matrix.h"
#include "elem_type.h"
#define MATRIX_DATA_FREE(ptr) cuda_matrix_(free)(ptr)
#define MATRIX_DATA_ALLOC(dptr, stride, width, height) \
cuda_matrix_(alloc)(dptr, stride, width, height)
#define MATRIX_DATA_WRITE(data, idx, val) cuda_matrix_(write)(data, idx, val)
#define MATRIX_DATA_READ(data, idx) cuda_matrix_(read)(data, idx)
#define MATRIX_INIT(L) cuda_matrix_(init)(L)
#define NERV_GENERIC_MATRIX
#define NERV_GENERIC_CUKERNEL
#include "../../common.h"
#include "../cukernel.h"
#include "cuda.h"
#include "cuda_runtime.h"
#include "driver_types.h"
#include "cublas_v2.h"
static cublasHandle_t cublas_handle;
Matrix *nerv_matrix_(new_)(long nrow, long ncol);
static void nerv_matrix_(add_)(const Matrix *a, const Matrix *b,
const Matrix *c,
MATRIX_ELEM alpha, MATRIX_ELEM beta) {
NERV_CUBLAS_(geam)(cublas_handle, CUBLAS_OP_N, CUBLAS_OP_N,
a->ncol, a->nrow,
&alpha,
MATRIX_ELEM_PTR(a), a->stride / sizeof(MATRIX_ELEM),
&beta,
MATRIX_ELEM_PTR(b), b->stride / sizeof(MATRIX_ELEM),
MATRIX_ELEM_PTR(c), c->stride / sizeof(MATRIX_ELEM));
}
static int nerv_matrix_(add)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 2, nerv_matrix_(tname));
Matrix *c;
if (!(a->nrow == b->nrow && a->ncol == b->ncol))
nerv_error(L, "Matrices should be of the same dimension");
c = nerv_matrix_(new_)(a->nrow, a->ncol);
nerv_matrix_(add_)(a, b, c, 1.0f, 1.0f);
luaT_pushudata(L, c, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(sub)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 2, nerv_matrix_(tname));
Matrix *c;
if (!(a->nrow == b->nrow && a->ncol == b->ncol))
nerv_error(L, "Matrices should be of the same dimension");
c = nerv_matrix_(new_)(a->nrow, a->ncol);
nerv_matrix_(add_)(a, b, c, 1.0f, -1.0f);
luaT_pushudata(L, c, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(mul)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 2, nerv_matrix_(tname));
Matrix *c;
if (a->ncol != b->nrow)
nerv_error(L, "Wrong dimension of multipliers");
c = nerv_matrix_(new_)(a->nrow, b->ncol);
MATRIX_ELEM alpha = 1.0f, beta = 0.0f;
NERV_CUBLAS_(gemm)(cublas_handle, CUBLAS_OP_N, CUBLAS_OP_N,
b->ncol, a->nrow, b->nrow,
&alpha,
MATRIX_ELEM_PTR(b), b->stride / sizeof(MATRIX_ELEM),
MATRIX_ELEM_PTR(a), a->stride / sizeof(MATRIX_ELEM),
&beta,
MATRIX_ELEM_PTR(c), c->stride / sizeof(MATRIX_ELEM));
luaT_pushudata(L, c, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(sigmoid)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = nerv_matrix_(new_)(a->nrow, a->ncol);
cudak_(cuda_sigmoid)(a, b);
luaT_pushudata(L, b, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(softmax)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *max = nerv_matrix_(new_)(a->nrow, 1);
Matrix *dno = nerv_matrix_(new_)(a->nrow, 1);
Matrix *b = nerv_matrix_(new_)(a->nrow, a->ncol);
cudak_(cuda_colmax)(a, max);
cudak_(cuda_softmax_denominator)(a, max, dno);
cudak_(cuda_softmax_final)(a, max, dno, b);
luaT_pushudata(L, b, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(colsum)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = nerv_matrix_(new_)(a->nrow, 1);
cudak_(cuda_colsum)(a, b);
luaT_pushudata(L, b, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(colmax)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = nerv_matrix_(new_)(a->nrow, 1);
cudak_(cuda_colmax)(a, b);
luaT_pushudata(L, b, nerv_matrix_(tname));
return 1;
}
static const luaL_Reg nerv_matrix_(extra_methods)[] = {
{"__add__", nerv_matrix_(add)},
{"__sub__", nerv_matrix_(sub)},
{"__mul__", nerv_matrix_(mul)},
{"sigmoid", nerv_matrix_(sigmoid)},
{"softmax", nerv_matrix_(softmax)},
{"colsum", nerv_matrix_(colsum)},
{"colmax", nerv_matrix_(colmax)},
{NULL, NULL}
};
static void cuda_matrix_(init)(lua_State *L) {
luaN_append_methods(L, nerv_matrix_(extra_methods));
cublasCreate(&cublas_handle);
}
static void cuda_matrix_(free)(MATRIX_ELEM *ptr) {
cudaFree(ptr);
}
static void cuda_matrix_(alloc)(MATRIX_ELEM **dptr, size_t *stride,
long width, long height) {
cudaMallocPitch((void **)dptr, stride, width, height);
}
static MATRIX_ELEM cuda_matrix_(read)(MATRIX_ELEM *data, int idx) {
MATRIX_ELEM res;
cudaMemcpy(&res, data + idx, sizeof(MATRIX_ELEM), cudaMemcpyDeviceToHost);
return res;
}
static void cuda_matrix_(write)(MATRIX_ELEM *data, int idx, MATRIX_ELEM val) {
cudaMemcpy(data + idx, &val, sizeof(MATRIX_ELEM), cudaMemcpyHostToDevice);
}
int nerv_matrix_(get_elem)(lua_State *L) {
return nerv_error_method_not_implemented(L);
}
int nerv_matrix_(set_elem)(lua_State *L) {
return nerv_error_method_not_implemented(L);
}
#include "matrix.c"
#endif
|