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
|
#ifndef EIGEN_TEST_GPU_COMMON_H
#define EIGEN_TEST_GPU_COMMON_H
#ifdef EIGEN_USE_HIP
#include <hip/hip_runtime.h>
#include <hip/hip_runtime_api.h>
#else
#include <cuda.h>
#include <cuda_runtime.h>
#include <cuda_runtime_api.h>
#endif
#include <iostream>
#define EIGEN_USE_GPU
#include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
#if !defined(__CUDACC__) && !defined(__HIPCC__)
dim3 threadIdx, blockDim, blockIdx;
#endif
template<typename Kernel, typename Input, typename Output>
void run_on_cpu(const Kernel& ker, int n, const Input& in, Output& out)
{
for(int i=0; i<n; i++)
ker(i, in.data(), out.data());
}
template<typename Kernel, typename Input, typename Output>
__global__
EIGEN_HIP_LAUNCH_BOUNDS_1024
void run_on_gpu_meta_kernel(const Kernel ker, int n, const Input* in, Output* out)
{
int i = threadIdx.x + blockIdx.x*blockDim.x;
if(i<n) {
ker(i, in, out);
}
}
template<typename Kernel, typename Input, typename Output>
void run_on_gpu(const Kernel& ker, int n, const Input& in, Output& out)
{
typename Input::Scalar* d_in;
typename Output::Scalar* d_out;
std::ptrdiff_t in_bytes = in.size() * sizeof(typename Input::Scalar);
std::ptrdiff_t out_bytes = out.size() * sizeof(typename Output::Scalar);
gpuMalloc((void**)(&d_in), in_bytes);
gpuMalloc((void**)(&d_out), out_bytes);
gpuMemcpy(d_in, in.data(), in_bytes, gpuMemcpyHostToDevice);
gpuMemcpy(d_out, out.data(), out_bytes, gpuMemcpyHostToDevice);
// Simple and non-optimal 1D mapping assuming n is not too large
// That's only for unit testing!
dim3 Blocks(128);
dim3 Grids( (n+int(Blocks.x)-1)/int(Blocks.x) );
gpuDeviceSynchronize();
#ifdef EIGEN_USE_HIP
hipLaunchKernelGGL(HIP_KERNEL_NAME(run_on_gpu_meta_kernel<Kernel,
typename std::decay<decltype(*d_in)>::type,
typename std::decay<decltype(*d_out)>::type>),
dim3(Grids), dim3(Blocks), 0, 0, ker, n, d_in, d_out);
#else
run_on_gpu_meta_kernel<<<Grids,Blocks>>>(ker, n, d_in, d_out);
#endif
// Pre-launch errors.
gpuError_t err = gpuGetLastError();
if (err != gpuSuccess) {
printf("%s: %s\n", gpuGetErrorName(err), gpuGetErrorString(err));
gpu_assert(false);
}
// Kernel execution errors.
err = gpuDeviceSynchronize();
if (err != gpuSuccess) {
printf("%s: %s\n", gpuGetErrorName(err), gpuGetErrorString(err));
gpu_assert(false);
}
// check inputs have not been modified
gpuMemcpy(const_cast<typename Input::Scalar*>(in.data()), d_in, in_bytes, gpuMemcpyDeviceToHost);
gpuMemcpy(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost);
gpuFree(d_in);
gpuFree(d_out);
}
template<typename Kernel, typename Input, typename Output>
void run_and_compare_to_gpu(const Kernel& ker, int n, const Input& in, Output& out)
{
Input in_ref, in_gpu;
Output out_ref, out_gpu;
#if !defined(EIGEN_GPU_COMPILE_PHASE)
in_ref = in_gpu = in;
out_ref = out_gpu = out;
#else
EIGEN_UNUSED_VARIABLE(in);
EIGEN_UNUSED_VARIABLE(out);
#endif
run_on_cpu (ker, n, in_ref, out_ref);
run_on_gpu(ker, n, in_gpu, out_gpu);
#if !defined(EIGEN_GPU_COMPILE_PHASE)
VERIFY_IS_APPROX(in_ref, in_gpu);
VERIFY_IS_APPROX(out_ref, out_gpu);
#endif
}
struct compile_time_device_info {
EIGEN_DEVICE_FUNC
void operator()(int i, const int* /*in*/, int* info) const
{
if (i == 0) {
EIGEN_UNUSED_VARIABLE(info)
#if defined(__CUDA_ARCH__)
info[0] = int(__CUDA_ARCH__ +0);
#endif
#if defined(EIGEN_HIP_DEVICE_COMPILE)
info[1] = int(EIGEN_HIP_DEVICE_COMPILE +0);
#endif
}
}
};
void ei_test_init_gpu()
{
int device = 0;
gpuDeviceProp_t deviceProp;
gpuGetDeviceProperties(&deviceProp, device);
ArrayXi dummy(1), info(10);
info = -1;
run_on_gpu(compile_time_device_info(),10,dummy,info);
std::cout << "GPU compile-time info:\n";
#ifdef EIGEN_CUDACC
std::cout << " EIGEN_CUDACC: " << int(EIGEN_CUDACC) << "\n";
#endif
#ifdef EIGEN_CUDA_SDK_VER
std::cout << " EIGEN_CUDA_SDK_VER: " << int(EIGEN_CUDA_SDK_VER) << "\n";
#endif
#ifdef EIGEN_COMP_NVCC
std::cout << " EIGEN_COMP_NVCC: " << int(EIGEN_COMP_NVCC) << "\n";
#endif
#ifdef EIGEN_HIPCC
std::cout << " EIGEN_HIPCC: " << int(EIGEN_HIPCC) << "\n";
#endif
std::cout << " EIGEN_CUDA_ARCH: " << info[0] << "\n";
std::cout << " EIGEN_HIP_DEVICE_COMPILE: " << info[1] << "\n";
std::cout << "GPU device info:\n";
std::cout << " name: " << deviceProp.name << "\n";
std::cout << " capability: " << deviceProp.major << "." << deviceProp.minor << "\n";
std::cout << " multiProcessorCount: " << deviceProp.multiProcessorCount << "\n";
std::cout << " maxThreadsPerMultiProcessor: " << deviceProp.maxThreadsPerMultiProcessor << "\n";
std::cout << " warpSize: " << deviceProp.warpSize << "\n";
std::cout << " regsPerBlock: " << deviceProp.regsPerBlock << "\n";
std::cout << " concurrentKernels: " << deviceProp.concurrentKernels << "\n";
std::cout << " clockRate: " << deviceProp.clockRate << "\n";
std::cout << " canMapHostMemory: " << deviceProp.canMapHostMemory << "\n";
std::cout << " computeMode: " << deviceProp.computeMode << "\n";
}
#endif // EIGEN_TEST_GPU_COMMON_H
|
