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
|
#include <assert.h>
#include <stdio.h>
#include "generic/matrix.h"
#include "cuda.h"
#define CUDA_THREADS_N 16
#define CUDA_THREADS_NN (16 * 16)
#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
__global__ void sigmoid(const float *a, float *b,
int nrow, int ncol, int stride) {
int j = blockIdx.x * blockDim.x + threadIdx.x;
int i = blockIdx.y * blockDim.y + threadIdx.y;
long idx;
if (i >= nrow || j >= ncol) return;
idx = j + i * stride;
b[idx] = 1.0 / (1.0 + exp(-a[idx]));
}
__global__ void softmax_final(const float *a, float *b,
const float *max, const float *deno,
int nrow, int ncol, int stride, int mstride) {
int j = blockIdx.x * blockDim.x + threadIdx.x;
int i = blockIdx.y * blockDim.y + threadIdx.y;
long idx;
if (i >= nrow || j >= ncol) return;
idx = j + i * stride;
b[idx] = exp(a[idx] - max[0 + i * mstride]) / deno[0 + i * mstride];
}
__global__ void block_reduce_sum(const float *input, float *output,
const int istride, const int ostride,
const int n) {
extern __shared__ float arr[];
int j = blockIdx.x * blockDim.x + threadIdx.x;
arr[threadIdx.x] = j < n ? input[j + istride * blockIdx.y] : 0;
__syncthreads();
for (int offset = blockDim.x >> 1; offset; offset >>= 1)
{
if (threadIdx.x < offset)
arr[threadIdx.x] += arr[threadIdx.x + offset];
__syncthreads();
}
if (threadIdx.x == 0)
output[blockIdx.x + ostride * blockIdx.y] = arr[0];
}
__global__ void block_reduce_softmax_sum(const float *input, float *output,
const float *max,
const int istride, const int ostride,
const int mstride, const int n) {
extern __shared__ float arr[];
int j = blockIdx.x * blockDim.x + threadIdx.x;
arr[threadIdx.x] = j < n ? exp(input[j + istride * blockIdx.y] - \
max[0 + mstride * blockIdx.y]) : 0;
__syncthreads();
for (int offset = blockDim.x >> 1; offset; offset >>= 1)
{
if (threadIdx.x < offset)
arr[threadIdx.x] += arr[threadIdx.x + offset];
__syncthreads();
}
if (threadIdx.x == 0)
output[blockIdx.x + ostride * blockIdx.y] = arr[0];
}
__global__ void block_reduce_max(const float *input, float *output,
const int istride, const int ostride,
const int n) {
extern __shared__ float arr[];
int j = blockIdx.x * blockDim.x + threadIdx.x;
arr[threadIdx.x] = j < n ? input[j + istride * blockIdx.y] : 0;
__syncthreads();
for (int offset = blockDim.x >> 1; offset; offset >>= 1)
{
if (threadIdx.x < offset)
{
float l = arr[threadIdx.x],
r = arr[threadIdx.x + offset];
if (r > l) arr[threadIdx.x] = r;
}
__syncthreads();
}
if (threadIdx.x == 0)
output[blockIdx.x + ostride * blockIdx.y] = arr[0];
}
extern "C" {
#include "cukernel.h"
void cuda_sigmoid(const Matrix *a, Matrix *b) {
dim3 threadsPerBlock(CUDA_THREADS_N,
CUDA_THREADS_N);
dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x),
CEIL_DIV(b->nrow, threadsPerBlock.y));
sigmoid<<<numBlocks, threadsPerBlock>>>(a->data.f, b->data.f, b->nrow, b->ncol,
b->stride / sizeof(float));
}
void cuda_colsum(const Matrix *a, Matrix *b) {
dim3 block(CUDA_THREADS_NN, 1);
int ncol = a->ncol;
int blocks_per_row = CEIL_DIV(ncol, block.x);
dim3 grid(blocks_per_row, a->nrow);
float *res;
size_t stride;
cudaMallocPitch(&res, &stride, blocks_per_row * sizeof(float), a->nrow);
block_reduce_sum<<<grid, block, block.x * sizeof(float)>>> \
(a->data.f, res,
a->stride / sizeof(float), stride / sizeof(float),
ncol);
ncol = blocks_per_row;
assert((unsigned long)ncol <= block.x);
grid.x = 1;
block_reduce_sum<<<grid, block, block.x * sizeof(float)>>> \
(res, b->data.f,
stride / sizeof(float), b->stride / sizeof(float),
ncol);
cudaFree(res);
}
void cuda_softmax_final(const Matrix *a, const Matrix *max,
const Matrix *deno, Matrix *b) {
dim3 threadsPerBlock(CUDA_THREADS_N,
CUDA_THREADS_N);
dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x),
CEIL_DIV(b->nrow, threadsPerBlock.y));
softmax_final<<<numBlocks, threadsPerBlock>>>(a->data.f, b->data.f,
max->data.f, deno->data.f,
b->nrow, b->ncol,
b->stride / sizeof(float),
max->stride / sizeof(float));
}
void cuda_softmax_denominator(const Matrix *a, const Matrix *max, Matrix *b) {
dim3 block(CUDA_THREADS_NN, 1);
int ncol = a->ncol;
int blocks_per_row = CEIL_DIV(ncol, block.x);
dim3 grid(blocks_per_row, a->nrow);
float *res;
size_t stride;
assert(max->ncol == 1);
cudaMallocPitch(&res, &stride, blocks_per_row * sizeof(float), a->nrow);
block_reduce_softmax_sum<<<grid, block, block.x * sizeof(float)>>> \
(a->data.f, res, max->data.f,
a->stride / sizeof(float), stride / sizeof(float),
max->stride / sizeof(float),
ncol);
ncol = blocks_per_row;
assert((unsigned long)ncol <= block.x);
grid.x = 1;
block_reduce_sum<<<grid, block, block.x * sizeof(float)>>> \
(res, b->data.f,
stride / sizeof(float), b->stride / sizeof(float),
ncol);
cudaFree(res);
}
void cuda_colmax(const Matrix *a, Matrix *b) {
dim3 block(CUDA_THREADS_NN, 1);
int ncol = a->ncol;
int blocks_per_row = CEIL_DIV(ncol, block.x);
dim3 grid(blocks_per_row, a->nrow);
float *res;
size_t stride;
cudaMallocPitch(&res, &stride, blocks_per_row * sizeof(float), a->nrow);
block_reduce_max<<<grid, block, block.x * sizeof(float)>>> \
(a->data.f, res,
a->stride / sizeof(float), stride / sizeof(float),
ncol);
ncol = blocks_per_row;
assert((unsigned long)ncol <= block.x);
grid.x
|