-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest.cu
199 lines (181 loc) · 8.51 KB
/
test.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
#include <stdio.h>
#include <chrono>
#include <cuda.h>
#include <cuda_runtime.h>
#include <iostream>
#include <cudnn.h>
#include <cuda_fp16.h>
using namespace std::chrono;
static const char *_cudaGetErrorEnum(cudaError_t error) {
return cudaGetErrorName(error);
}
template <typename T>
void check(T result, char const *const func, const char *const file,
int const line) {
if (result) {
fprintf(stderr, "CUDA error at %s:%d code=%d(%s) \"%s\" \n", file, line,
static_cast<unsigned int>(result), _cudaGetErrorEnum(result), func);
exit(EXIT_FAILURE);
}
}
#define checkCudaErrors(val) check((val), #val, __FILE__, __LINE__)
#define checkCUDNN(expression) \
{ \
cudnnStatus_t status = (expression); \
if (status != CUDNN_STATUS_SUCCESS) { \
std::cerr << "Error on line " << __LINE__ << ": " \
<< cudnnGetErrorString(status) << std::endl; \
std::exit(EXIT_FAILURE); \
} \
}
#define checkCUDA(expression) \
{ \
cudaError_t status = (expression); \
if (status != cudaSuccess) { \
std::cerr << "Error on line " << __LINE__ << ": " \
<< cudaGetErrorString(status) << std::endl; \
std::exit(EXIT_FAILURE); \
} \
}
void print_array(float *array, int size, const char *name) {
std::cout << name;
for (int i = 0; i < size; i++) {
std::cout << array[i] << " ";
}
std::cout << std::endl;
}
void init_array(float *array, int size, float val) {
for (int i = 0; i < size; i++) {
array[i] = val;
}
}
void test(){
cudaStream_t stream;
cudaStreamCreate(&stream);
cudaStream_t stream_1;
cudaStreamCreate(&stream_1);
CUmemoryPool pool_;
cuDeviceGetDefaultMemPool(&pool_, 0);
uint64_t threshold = UINT64_MAX;
cuMemPoolSetAttribute(pool_, CU_MEMPOOL_ATTR_RELEASE_THRESHOLD, &threshold);
int _n = 128, _c= 2048, _h = 7, _w = 7;
int x_size = _n * _c * _h * _w;
int x_size_bytes = x_size * sizeof(float);
int iter = 8;
int mean_size = _c;
int mean_size_bytes = mean_size * sizeof(float);
cudaGraph_t graph;
cudaGraphExec_t instance;
bool graphCreated=false;
float* h_x = (float*)malloc(x_size_bytes);
float* h_y = (float*)malloc(x_size_bytes);
init_array(h_x, x_size, 2.5);
init_array(h_y, x_size, 0.0);
float *x, *y;
checkCUDA(cudaMalloc(&x, x_size_bytes));
checkCUDA(cudaMalloc(&y, x_size_bytes));
cudaMemcpy(x, reinterpret_cast<const float *>(h_x), x_size_bytes, cudaMemcpyHostToDevice);
float *scale, *offset;
float *saved_mean, *saved_inv_var;
float* h_scale = (float*)malloc(mean_size_bytes);
float* h_offset = (float*)malloc(mean_size_bytes);
init_array(h_scale, mean_size, 1.5);
init_array(h_offset, mean_size, 2.0);
checkCUDA(cudaMallocManaged(&scale, mean_size_bytes));
checkCUDA(cudaMallocManaged(&offset, mean_size_bytes));
cudaMemcpy(scale, reinterpret_cast<const float *>(h_scale), mean_size_bytes, cudaMemcpyHostToDevice);
cudaMemcpy(offset, reinterpret_cast<const float *>(h_offset), mean_size_bytes, cudaMemcpyHostToDevice);
checkCUDA(cudaMallocManaged(&scale, mean_size_bytes));
checkCUDA(cudaMallocManaged(&offset, mean_size_bytes));
checkCUDA(cudaMallocManaged(&saved_mean, mean_size_bytes));
checkCUDA(cudaMallocManaged(&saved_inv_var, mean_size_bytes));
float *a_x, *a_y, *a_scale, *a_offset, *a_saved_mean, *a_saved_inv_var;
cudnnHandle_t cudnn;
checkCUDNN(cudnnCreate(&cudnn));
for (int i =0; i < iter; i++){
if (!graphCreated){
cudaStreamBeginCapture(stream, cudaStreamCaptureModeGlobal);
cuMemAllocFromPoolAsync(reinterpret_cast<CUdeviceptr*>(&a_x), x_size_bytes, pool_, stream);
cuMemAllocFromPoolAsync(reinterpret_cast<CUdeviceptr*>(&a_y), x_size_bytes, pool_, stream);
cuMemAllocFromPoolAsync(reinterpret_cast<CUdeviceptr*>(&a_scale), mean_size_bytes, pool_, stream);
cuMemAllocFromPoolAsync(reinterpret_cast<CUdeviceptr*>(&a_offset), mean_size_bytes, pool_, stream);
cuMemAllocFromPoolAsync(reinterpret_cast<CUdeviceptr*>(&a_saved_mean), mean_size_bytes, pool_, stream);
cuMemAllocFromPoolAsync(reinterpret_cast<CUdeviceptr*>(&a_saved_inv_var), mean_size_bytes, pool_, stream);
cudaMemcpyAsync(a_x, x, x_size_bytes, cudaMemcpyDeviceToDevice, stream);
cudaMemcpyAsync(a_y, y, x_size_bytes, cudaMemcpyDeviceToDevice, stream);
cudaMemcpyAsync(a_scale, scale, mean_size_bytes, cudaMemcpyDeviceToDevice, stream);
cudaMemcpyAsync(a_offset, offset, mean_size_bytes, cudaMemcpyDeviceToDevice, stream);
cudaMemcpyAsync(a_saved_mean, saved_mean, mean_size_bytes, cudaMemcpyDeviceToDevice, stream);
cudaMemcpyAsync(a_saved_inv_var, saved_inv_var, mean_size_bytes, cudaMemcpyDeviceToDevice, stream);
auto mode = CUDNN_BATCHNORM_SPATIAL_PERSISTENT;
float one = 1.0;
float zero = 0.0;
//int N = 128, C = 2048, H = 7, W = 7;
cudnnTensorDescriptor_t x_descriptor;
checkCUDNN(cudnnCreateTensorDescriptor(&x_descriptor));
checkCUDNN(cudnnSetTensor4dDescriptor(x_descriptor,
/*format=*/CUDNN_TENSOR_NHWC,
/*dataType=*/CUDNN_DATA_FLOAT,
/*batch_size=*/128,
/*channels=*/2048,
/*image_height=*/7,
/*image_width=*/7));
cudnnTensorDescriptor_t mean_descriptor;
checkCUDNN(cudnnCreateTensorDescriptor(&mean_descriptor));
checkCUDNN(cudnnSetTensor4dDescriptor(mean_descriptor,
/*format=*/CUDNN_TENSOR_NHWC,
/*dataType=*/CUDNN_DATA_FLOAT,
/*batch_size=*/1,
/*channels=*/2048,
/*image_height=*/1,
/*image_width=*/1));
checkCUDNN(cudnnBatchNormalizationForwardInference(
/*handle=*/cudnn,
/*mode=*/mode,
/*alphaDataDiff=*/&one,
/*betaDataDiff=*/&zero,
/*xDesc=*/x_descriptor,
a_x,
/*xDesc=*/x_descriptor,
a_y,
/*bnScaleBiasMeanVarDesc=*/mean_descriptor,
/*bnScale=*/a_scale,
/*bnBias=*/a_offset,
/*resultSaveMean=*/a_saved_mean,
/*resultSaveInvVariance=*/a_saved_inv_var,
/*epsilon=*/0.001)
)
cuMemFreeAsync(reinterpret_cast<const CUdeviceptr&>(a_x), stream);
cuMemFreeAsync(reinterpret_cast<const CUdeviceptr&>(a_scale), stream);
cuMemFreeAsync(reinterpret_cast<const CUdeviceptr&>(a_offset), stream);
cuMemFreeAsync(reinterpret_cast<const CUdeviceptr&>(a_saved_mean), stream);
cuMemFreeAsync(reinterpret_cast<const CUdeviceptr&>(a_saved_inv_var), stream);
checkCudaErrors(cudaGraphInstantiate(&instance, graph, NULL, NULL, 0));
checkCudaErrors(cudaGraphUpload(instance, stream));
graphCreated = true;
}
checkCudaErrors(cudaGraphLaunch(instance, stream));
checkCUDA(cudaDeviceSynchronize());
float* out = (float*)malloc(x_size_bytes);
cudaMemcpy(out, reinterpret_cast<const float *>(a_y), x_size_bytes, cudaMemcpyDeviceToHost);
cuMemFreeAsync(reinterpret_cast<const CUdeviceptr&>(a_y), stream);
print_array(out, x_size, "dx NCHW format: ");
}
cudaStreamDestroy(stream);
cudaGraphDestroy(graph);
cudaGraphExecDestroy(instance);
checkCUDA(cudaFree(x));
checkCUDA(cudaFree(y));
checkCUDA(cudaFree(scale));
checkCUDA(cudaFree(offset));
checkCUDA(cudaFree(saved_mean));
checkCUDA(cudaFree(saved_inv_var));
free(h_x);
free(h_y);
free(h_scale);
free(h_offset);
}
int main() {
test();
return 0;
}