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
|
#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 MATRIX_BASE_TNAME nerv_matrix_cuda_tname
#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 *c = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 3, nerv_matrix_(tname));
MATRIX_ELEM alpha = luaL_checknumber(L, 4); /* alpha */
MATRIX_ELEM beta = luaL_checknumber(L, 5); /* alpha */
if (!(a->nrow == b->nrow && a->ncol == b->ncol))
nerv_error(L, "Matrices should be of the same dimension");
nerv_matrix_(add_)(a, b, c, alpha, beta);
luaT_pushudata(L, c, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(get_cublas_op)(char ch) {
return (ch == 'T' || ch == 't') ? CUBLAS_OP_T : CUBLAS_OP_N;
}
static int nerv_matrix_(mul)(lua_State *L) {
Matrix *c = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 3, nerv_matrix_(tname));
int nargs = lua_gettop(L);
int ta = nargs > 3 ? nerv_matrix_(get_cublas_op)(*luaL_checkstring(L, 4)) \
: CUBLAS_OP_N;
int tb = nargs > 4 ? nerv_matrix_(get_cublas_op)(*luaL_checkstring(L, 5)) \
: CUBLAS_OP_N;
printf("%d %d\n", ta, tb);
if (a->ncol != b->nrow)
nerv_error(L, "Wrong dimension of multipliers");
MATRIX_ELEM alpha = 1.0f, beta = 0.0f;
NERV_CUBLAS_(gemm)(cublas_handle, tb, ta,
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_(create)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = nerv_matrix_(new_)(a->nrow, a->ncol);
luaT_pushudata(L, b, 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_rowmax)(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_(rowsum)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = nerv_matrix_(new_)(a->nrow, 1);
cudak_(cuda_rowsum)(a, 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_)(1, a->ncol);
cudak_(cuda_colsum)(a, b);
luaT_pushudata(L, b, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(rowmax)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = nerv_matrix_(new_)(a->nrow, 1);
cudak_(cuda_rowmax)(a, b);
luaT_pushudata(L, b, nerv_matrix_(tname));
return 1;
}
static int nerv_matrix_(add_row)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 1, nerv_matrix_(tname));
double beta = luaL_checknumber(L, 3);
cudak_(cuda_add_row)(a, b, beta);
return 0;
}
extern const char *MATRIX_CUMATRIX_HOST_TNAME;
static int nerv_matrix_(copy_from)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 2, MATRIX_CUMATRIX_HOST_TNAME);
if (!(a->nrow == b->nrow && a->ncol == b->ncol))
nerv_error(L, "Matrices should be of the same dimension");
cudaMemcpy2D(MATRIX_ELEM_PTR(a), a->stride,
MATRIX_ELEM_PTR(b), b->stride,
sizeof(MATRIX_ELEM) * b->ncol, b->nrow,
cudaMemcpyHostToDevice);
return 0;
}
static int nerv_matrix_(copy_to)(lua_State *L) {
Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
Matrix *b = luaT_checkudata(L, 2, MATRIX_CUMATRIX_HOST_TNAME);
if (!(a->nrow == b->nrow && a->ncol == b->ncol))
nerv_error(L, "Matrices should be of the same dimension");
cudaMemcpy2D(MATRIX_ELEM_PTR(b), b->stride,
MATRIX_ELEM_PTR(a), a->stride,
sizeof(MATRIX_ELEM) * a->ncol, a->nrow,
cudaMemcpyDeviceToHost);
return 0;
}
static const luaL_Reg nerv_matrix_(extra_methods)[] = {
{"add", nerv_matrix_(add)},
{"mul", nerv_matrix_(mul)},
{"create", nerv_matrix_(create)},
{"sigmoid", nerv_matrix_(sigmoid)},
{"softmax", nerv_matrix_(softmax)},
{"colsum", nerv_matrix_(colsum)},
{"rowsum", nerv_matrix_(rowsum)},
{"rowmax", nerv_matrix_(rowmax)},
{"copy_from", nerv_matrix_(copy_from)},
{"copy_to", nerv_matrix_(copy_to)},
/* in-place calc */
{"add_row", nerv_matrix_(add_row)},
{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),
|