-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconvKernel.cu
161 lines (122 loc) · 7.03 KB
/
convKernel.cu
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
#include <stdint.h>
#include <stdio.h>
#include "convKernel.h"
#define FILTER_HEIGHT 3
#define FILTER_WIDTH 3
#define INPUT_CHANNELS 3
#define OUTPUT_CHANNELS 64
#define BLOCK_SIZE 16
__global__ void ConvForward00(const struct Filter filter, const struct Data_Tensor input_tensor, struct Data_Tensor output_tensor) {
int32_t thread_x = threadIdx.x;
int32_t thread_y = threadIdx.y;
int32_t thread_z = threadIdx.z;
int32_t x_stride = blockDim.x;
int32_t y_stride = blockDim.y;
int32_t z_stride = blockDim.z;
int32_t grid_size_x = output_tensor.channels / gridDim.x;
int32_t grid_offset_x = blockIdx.x * grid_size_x;
int32_t grid_size_y = output_tensor.height / gridDim.y;
int32_t grid_offset_y = blockIdx.y * grid_size_y;
int32_t grid_size_z = output_tensor.width / gridDim.z;
int32_t grid_offset_z = blockIdx.z * grid_size_z;
for (int32_t o_channel = thread_x + grid_offset_x; o_channel < grid_offset_x + grid_size_x; o_channel += x_stride) {
for (int32_t o_row = thread_y + grid_offset_y; o_row < grid_offset_y + grid_size_y; o_row += y_stride) {
for (int32_t o_col = thread_z + grid_offset_z; o_col < grid_offset_z + grid_size_z; o_col += z_stride) {
double output_element_value = 0.0;
for (int32_t i_channel = 0; i_channel < filter.input_channels; i_channel += 1) {
for (int32_t row = 0; row < filter.height; row += 1) {
for (int32_t col = 0; col < filter.width; col += 1) {
output_element_value += filter.elements[
o_channel * (filter.input_channels * filter.height * filter.width) +
i_channel * (filter.height * filter.width) +
(filter.height - 1 - row) * filter.width +
(filter.width - 1 - col)
] * input_tensor.elements[
i_channel * (input_tensor.true_height * input_tensor.true_width) +
(o_row + row) * (input_tensor.true_width) +
(o_col + col)
];
}
}
}
output_tensor.elements[
o_channel * (output_tensor.true_height * output_tensor.true_width) +
o_row * (output_tensor.true_width) +
o_col
] = output_element_value;
}
}
}
}
__global__ void ConvForward01(const struct Filter filter, const struct Data_Tensor input_tensor, struct Data_Tensor output_tensor) {
int32_t thread_x = threadIdx.x;
int32_t thread_y = threadIdx.y;
int32_t thread_z = threadIdx.z;
int32_t x_stride = blockDim.x;
int32_t y_stride = blockDim.y;
int32_t z_stride = blockDim.z;
__shared__ double FilterTile[OUTPUT_CHANNELS][INPUT_CHANNELS][FILTER_HEIGHT][FILTER_WIDTH];
__shared__ double InputTile[INPUT_CHANNELS][BLOCK_SIZE + FILTER_HEIGHT - 1][BLOCK_SIZE + FILTER_WIDTH - 1];
int32_t grid_size_x = output_tensor.channels / gridDim.x;
int32_t grid_offset_x = blockIdx.x * grid_size_x;
int32_t grid_size_y = output_tensor.height / gridDim.y;
int32_t grid_offset_y = blockIdx.y * grid_size_y;
int32_t grid_size_z = output_tensor.width / gridDim.z;
int32_t grid_offset_z = blockIdx.z * grid_size_z;
for (int32_t o_channel = thread_x + grid_offset_x; o_channel < grid_offset_x + grid_size_x; o_channel += x_stride) {
for (int32_t o_row = thread_y + grid_offset_y; o_row < grid_offset_y + grid_size_y; o_row += y_stride) {
for (int32_t o_col = thread_z + grid_offset_z; o_col < grid_offset_z + grid_size_z; o_col += z_stride) {
if (thread_y < 3 && thread_z < 3) {
#pragma unroll
for (int32_t i_channel = 0; i_channel < INPUT_CHANNELS; i_channel += 1) {
FilterTile[o_channel][i_channel][thread_y][thread_z] = filter.elements[
o_channel * (filter.input_channels * filter.height * filter.width) +
i_channel * (filter.height * filter.width) +
(filter.height - 1 - thread_y) * filter.width +
(filter.width - 1 - thread_z)
];
}
}
double* input_val;
if (thread_x < INPUT_CHANNELS) {
input_val = &input_tensor.elements[
thread_x * (input_tensor.true_height * input_tensor.true_width) +
(o_row) * (input_tensor.true_width) +
(o_col)
];
InputTile[thread_x][thread_y][thread_z] = input_val[0];
if (thread_y == BLOCK_SIZE - 1) {
InputTile[thread_x][thread_y + 1][thread_z] = input_val[input_tensor.true_width];
InputTile[thread_x][thread_y + 2][thread_z] = input_val[2 * input_tensor.true_width];
}
if (thread_z == BLOCK_SIZE - 1) {
InputTile[thread_x][thread_y][thread_z + 1] = input_val[1];
InputTile[thread_x][thread_y][thread_z + 2] = input_val[2];
}
if (thread_y == (BLOCK_SIZE - 1) && thread_z == (BLOCK_SIZE - 1)) {
InputTile[thread_x][thread_y + 1][thread_z + 1] = input_val[input_tensor.true_width + 1];
InputTile[thread_x][thread_y + 1][thread_z + 2] = input_val[input_tensor.true_width + 2];
InputTile[thread_x][thread_y + 2][thread_z + 1] = input_val[2 * input_tensor.true_width + 1];
InputTile[thread_x][thread_y + 2][thread_z + 2] = input_val[2 * input_tensor.true_width + 2];
}
}
__syncthreads();
double output_element_value = 0.0;
for (int32_t i_channel = 0; i_channel < INPUT_CHANNELS; i_channel += 1) {
for (int32_t row = 0; row < FILTER_HEIGHT; row += 1) {
for (int32_t col = 0; col < FILTER_WIDTH; col += 1) {
output_element_value += FilterTile[o_channel][i_channel][row][col] *
InputTile[i_channel][thread_y + row][thread_z + col];
}
}
}
output_tensor.elements[
o_channel * (output_tensor.true_height * output_tensor.true_width) +
o_row * (output_tensor.true_width) +
o_col
] = output_element_value;
__syncthreads();
}
}
}
}