-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconv2d_impl.cu
264 lines (237 loc) · 8.17 KB
/
conv2d_impl.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
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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
#include <iostream>
#include <sstream>
#include <vector>
#include <algorithm>
#include <cuda_runtime.h>
#include <cudnn.h>
#include "utils.cuh"
#include "conv2d_impl.cuh"
#include "pybind11/pybind11.h"
#include "pybind11/numpy.h"
namespace py = pybind11;
using algo_perf_t = cudnnConvolutionFwdAlgoPerf_t;
// if exit, algo_arr[0] will be best candidate
bool get_valid_best_algo(std::vector<algo_perf_t>& algo_arr) {
auto it = std::remove_if(algo_arr.begin(), algo_arr.end(), [](algo_perf_t algo_perf){
return algo_perf.status != CUDNN_STATUS_SUCCESS;
});
algo_arr.erase(it, algo_arr.end());
if(algo_arr.size() == 0) {
std::runtime_error("Found no valid conv algorithm!");
}
std::sort(algo_arr.begin(), algo_arr.end(), [](algo_perf_t algo1, algo_perf_t algo2){
return algo1.time < algo2.time;
});
return algo_arr.size()>0;
}
void cudnn_conv2d_out(const Tensor& x_gpu, const Tensor& w_gpu, const Conv2dParam& conv_param, Tensor& y_gpu) {
cudnnHandle_t h_handle;
CHECK_CUDNN(cudnnCreate(&h_handle));
cudnnTensorDescriptor_t x_desc, y_desc;
CHECK_CUDNN(cudnnCreateTensorDescriptor(&x_desc));
CHECK_CUDNN(cudnnCreateTensorDescriptor(&y_desc));
CHECK_CUDNN(cudnnSetTensor4dDescriptor(
x_desc,
CUDNN_TENSOR_NHWC,
CUDNN_DATA_FLOAT,
x_gpu.n,
x_gpu.c,
x_gpu.h,
x_gpu.w
));
// kernel
cudnnFilterDescriptor_t w_desc;
CHECK_CUDNN(cudnnCreateFilterDescriptor(&w_desc));
CHECK_CUDNN(cudnnSetFilter4dDescriptor(
w_desc,
CUDNN_DATA_FLOAT,
CUDNN_TENSOR_NHWC,
w_gpu.n,
w_gpu.c,
w_gpu.h,
w_gpu.w
));
// conv
cudnnConvolutionDescriptor_t conv_desc;
CHECK_CUDNN(cudnnCreateConvolutionDescriptor(&conv_desc));
CHECK_CUDNN(cudnnSetConvolution2dDescriptor(
conv_desc,
conv_param.pad_h,
conv_param.pad_w,
conv_param.u,
conv_param.v,
conv_param.dilation_h,
conv_param.dilation_w,
CUDNN_CROSS_CORRELATION,
CUDNN_DATA_FLOAT
));
// output
CHECK_CUDNN(cudnnGetConvolution2dForwardOutputDim(
conv_desc,
x_desc,
w_desc,
&y_gpu.n,
&y_gpu.c,
&y_gpu.h,
&y_gpu.w
));
y_gpu.layout = TensorLayout::NHWC;
y_gpu.alloc_gpu();
CHECK_CUDNN(cudnnSetTensor4dDescriptor(
y_desc,
CUDNN_TENSOR_NHWC,
CUDNN_DATA_FLOAT,
y_gpu.n,
y_gpu.c,
y_gpu.h,
y_gpu.w
));
// conv algorithm
std::vector<algo_perf_t> algo_perf_arr;
int request_cnt = 0;
CHECK_CUDNN(cudnnGetConvolutionForwardAlgorithmMaxCount(h_handle, &request_cnt));
algo_perf_arr.resize(request_cnt);
int algo_count = 0;
CHECK_CUDNN(cudnnSetConvolutionMathType(conv_desc, CUDNN_FMA_MATH));
// cudnnGetConvolutionForwardAlgorithm_v7
CHECK_CUDNN(cudnnGetConvolutionForwardAlgorithm_v7(
h_handle,
x_desc,
w_desc,
conv_desc,
y_desc,
request_cnt,
&algo_count,
algo_perf_arr.data()
));
if(!get_valid_best_algo(algo_perf_arr)) {
std::runtime_error("Found no valid conv algorithm!");
}
cudnnConvolutionFwdAlgo_t best_algo = algo_perf_arr[0].algo;
size_t ws = 0;
CHECK_CUDNN(cudnnGetConvolutionForwardWorkspaceSize(
h_handle,
x_desc,
w_desc,
conv_desc,
y_desc,
best_algo,
&ws));
void* workspace = nullptr;
if(ws > 0) {
CHECK_CUDA(cudaMalloc(&workspace, ws));
}
// Forward
float alpha = 1.0f;
float beta = 1.0f;
float* out_ptr = y_gpu.get_ptr();
CHECK_CUDNN(cudnnConvolutionForward(
h_handle,
&alpha,
x_desc,
x_gpu.get_ptr(),
w_desc,
w_gpu.get_ptr(),
conv_desc,
best_algo,
workspace,
ws,
&beta,
y_desc,
out_ptr
));
cudaDeviceSynchronize();
cudnnDestroyTensorDescriptor(x_desc);
cudnnDestroyTensorDescriptor(y_desc);
cudnnDestroyFilterDescriptor(w_desc);
cudnnDestroyConvolutionDescriptor(conv_desc);
cudnnDestroy(h_handle);
}
#if 0
Tensor cudnn_conv2d(const Tensor& x_gpu, const Tensor& w_gpu, const Conv2dParam& conv_param) {
// BUG, GPU memory released
Tensor y_gpu;
cudnn_conv2d_out(x_gpu,w_gpu, conv_param, y_gpu);
y_gpu.save("tensor_out2.dat");
return y_gpu;
}
#endif //0
PYBIND11_MODULE(libconv2d, m) {
m.doc() = "cudnn lib test";
py::enum_<TensorLayout>(m, "TensorLayout")
.value("NCHW", TensorLayout::NCHW)
.value("NHWC", TensorLayout::NHWC)
.export_values();
py::class_<Tensor>(m, "Tensor")
.def(py::init<>())
// .def("alloc_gpu", py::overload_cast<>(&Tensor::alloc_gpu))
// .def("alloc_gpu", py::overload_cast<int, int, int, int>(&Tensor::alloc_gpu))
.def_readwrite("n", &Tensor::n)
.def_readwrite("c", &Tensor::c)
.def_readwrite("h", &Tensor::h)
.def_readwrite("w", &Tensor::w)
.def("__repr__",
[](const Tensor& tensor){
std::stringstream ss;
ss << "shape:" << "[" << tensor.n << "," << tensor.c <<","
<< tensor.h <<"," << tensor.w <<"]" << std::endl
<< "dtype: " << "float32" << std::endl
<< "size_byte: " << tensor.size_byte << std::endl
<< "is_gpu: " << tensor.is_gpu;
return ss.str();
}
)
.def("from_numpy",
[](Tensor& tensor, py::array_t<float>& ndarray, TensorLayout layout = TensorLayout::NHWC){
assert(ndarray.ndim()==4 && "Dim must be 4 (n,c,h,w) or (n, h, w, c)");
assert(ndarray.dtype().char_() == 'f' && "dtype must be float32");
if (layout == TensorLayout::NCHW) {
tensor.n = ndarray.shape(0);
tensor.c = ndarray.shape(1);
tensor.h = ndarray.shape(2);
tensor.w = ndarray.shape(3);
}else {
tensor.n = ndarray.shape(0);
tensor.h = ndarray.shape(1);
tensor.w = ndarray.shape(2);
tensor.c = ndarray.shape(3);
}
tensor.alloc_gpu();
py::buffer_info buf = ndarray.request();
float* ptr = tensor.get_ptr();
CHECK_CUDA(cudaMemcpy(ptr, buf.ptr, ndarray.nbytes(), cudaMemcpyHostToDevice));
return true;
},
py::arg("ndarray"), py::arg("layout") = TensorLayout::NHWC
)
.def("numpy",
[](const Tensor& tensor) {
assert(tensor.size_byte > 0);
py::array::ShapeContainer sc;
if (tensor.layout == TensorLayout::NCHW)
sc->assign({tensor.n, tensor.c, tensor.h, tensor.w});
else
sc->assign({tensor.n, tensor.h, tensor.w, tensor.c});
auto ndarray = py::array_t<float>(sc);
assert(tensor.size_byte == ndarray.nbytes());
float* ptr = static_cast<float*>(ndarray.request(true).ptr);
if(tensor.is_gpu) {
CHECK_CUDA(cudaMemcpy(ptr, tensor.get_ptr(), ndarray.nbytes(), cudaMemcpyDeviceToHost));
}else {
memcpy(ptr, tensor.get_ptr(), ndarray.nbytes());
}
return ndarray;
}
);
py::class_<Conv2dParam>(m, "Conv2dParam")
.def(py::init<>())
.def(py::init<int, int, int, int, int, int>())
.def_readwrite("pad_h", &Conv2dParam::pad_h)
.def_readwrite("pad_w", &Conv2dParam::pad_w)
.def_readwrite("dilation_h", &Conv2dParam::dilation_h)
.def_readwrite("dilation_w", &Conv2dParam::dilation_w)
.def_readwrite("u", &Conv2dParam::u)
.def_readwrite("v", &Conv2dParam::v);
// m.def("cudnn_conv2d", cudnn_conv2d, py::arg("input_gpu"), py::arg("weight_gpu"), py::arg("params"));
m.def("cudnn_conv2d_out", cudnn_conv2d_out, py::arg("input_gpu"), py::arg("weight_gpu"), py::arg("params"), py::arg("output_gpu"));
}