xref: /aosp_15_r20/external/eigen/test/gpu_common.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 #ifndef EIGEN_TEST_GPU_COMMON_H
2 #define EIGEN_TEST_GPU_COMMON_H
3 
4 #ifdef EIGEN_USE_HIP
5   #include <hip/hip_runtime.h>
6   #include <hip/hip_runtime_api.h>
7 #else
8   #include <cuda.h>
9   #include <cuda_runtime.h>
10   #include <cuda_runtime_api.h>
11 #endif
12 
13 #include <iostream>
14 
15 #define EIGEN_USE_GPU
16 #include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
17 
18 #if !defined(__CUDACC__) && !defined(__HIPCC__)
19 dim3 threadIdx, blockDim, blockIdx;
20 #endif
21 
22 template<typename Kernel, typename Input, typename Output>
run_on_cpu(const Kernel & ker,int n,const Input & in,Output & out)23 void run_on_cpu(const Kernel& ker, int n, const Input& in, Output& out)
24 {
25   for(int i=0; i<n; i++)
26     ker(i, in.data(), out.data());
27 }
28 
29 
30 template<typename Kernel, typename Input, typename Output>
31 __global__
32 EIGEN_HIP_LAUNCH_BOUNDS_1024
run_on_gpu_meta_kernel(const Kernel ker,int n,const Input * in,Output * out)33 void run_on_gpu_meta_kernel(const Kernel ker, int n, const Input* in, Output* out)
34 {
35   int i = threadIdx.x + blockIdx.x*blockDim.x;
36   if(i<n) {
37     ker(i, in, out);
38   }
39 }
40 
41 
42 template<typename Kernel, typename Input, typename Output>
run_on_gpu(const Kernel & ker,int n,const Input & in,Output & out)43 void run_on_gpu(const Kernel& ker, int n, const Input& in, Output& out)
44 {
45   typename Input::Scalar*  d_in;
46   typename Output::Scalar* d_out;
47   std::ptrdiff_t in_bytes  = in.size()  * sizeof(typename Input::Scalar);
48   std::ptrdiff_t out_bytes = out.size() * sizeof(typename Output::Scalar);
49 
50   gpuMalloc((void**)(&d_in),  in_bytes);
51   gpuMalloc((void**)(&d_out), out_bytes);
52 
53   gpuMemcpy(d_in,  in.data(),  in_bytes,  gpuMemcpyHostToDevice);
54   gpuMemcpy(d_out, out.data(), out_bytes, gpuMemcpyHostToDevice);
55 
56   // Simple and non-optimal 1D mapping assuming n is not too large
57   // That's only for unit testing!
58   dim3 Blocks(128);
59   dim3 Grids( (n+int(Blocks.x)-1)/int(Blocks.x) );
60 
61   gpuDeviceSynchronize();
62 
63 #ifdef EIGEN_USE_HIP
64   hipLaunchKernelGGL(HIP_KERNEL_NAME(run_on_gpu_meta_kernel<Kernel,
65 				     typename std::decay<decltype(*d_in)>::type,
66 				     typename std::decay<decltype(*d_out)>::type>),
67 		     dim3(Grids), dim3(Blocks), 0, 0, ker, n, d_in, d_out);
68 #else
69   run_on_gpu_meta_kernel<<<Grids,Blocks>>>(ker, n, d_in, d_out);
70 #endif
71   // Pre-launch errors.
72   gpuError_t err = gpuGetLastError();
73   if (err != gpuSuccess) {
74     printf("%s: %s\n", gpuGetErrorName(err), gpuGetErrorString(err));
75     gpu_assert(false);
76   }
77 
78   // Kernel execution errors.
79   err = gpuDeviceSynchronize();
80   if (err != gpuSuccess) {
81     printf("%s: %s\n", gpuGetErrorName(err), gpuGetErrorString(err));
82     gpu_assert(false);
83   }
84 
85 
86   // check inputs have not been modified
87   gpuMemcpy(const_cast<typename Input::Scalar*>(in.data()),  d_in,  in_bytes,  gpuMemcpyDeviceToHost);
88   gpuMemcpy(out.data(), d_out, out_bytes, gpuMemcpyDeviceToHost);
89 
90   gpuFree(d_in);
91   gpuFree(d_out);
92 }
93 
94 
95 template<typename Kernel, typename Input, typename Output>
run_and_compare_to_gpu(const Kernel & ker,int n,const Input & in,Output & out)96 void run_and_compare_to_gpu(const Kernel& ker, int n, const Input& in, Output& out)
97 {
98   Input  in_ref,  in_gpu;
99   Output out_ref, out_gpu;
100   #if !defined(EIGEN_GPU_COMPILE_PHASE)
101   in_ref = in_gpu = in;
102   out_ref = out_gpu = out;
103   #else
104   EIGEN_UNUSED_VARIABLE(in);
105   EIGEN_UNUSED_VARIABLE(out);
106   #endif
107   run_on_cpu (ker, n, in_ref,  out_ref);
108   run_on_gpu(ker, n, in_gpu, out_gpu);
109   #if !defined(EIGEN_GPU_COMPILE_PHASE)
110   VERIFY_IS_APPROX(in_ref, in_gpu);
111   VERIFY_IS_APPROX(out_ref, out_gpu);
112   #endif
113 }
114 
115 struct compile_time_device_info {
116   EIGEN_DEVICE_FUNC
operatorcompile_time_device_info117   void operator()(int i, const int* /*in*/, int* info) const
118   {
119     if (i == 0) {
120       EIGEN_UNUSED_VARIABLE(info)
121       #if defined(__CUDA_ARCH__)
122       info[0] = int(__CUDA_ARCH__ +0);
123       #endif
124       #if defined(EIGEN_HIP_DEVICE_COMPILE)
125       info[1] = int(EIGEN_HIP_DEVICE_COMPILE +0);
126       #endif
127     }
128   }
129 };
130 
ei_test_init_gpu()131 void ei_test_init_gpu()
132 {
133   int device = 0;
134   gpuDeviceProp_t deviceProp;
135   gpuGetDeviceProperties(&deviceProp, device);
136 
137   ArrayXi dummy(1), info(10);
138   info = -1;
139   run_on_gpu(compile_time_device_info(),10,dummy,info);
140 
141 
142   std::cout << "GPU compile-time info:\n";
143 
144   #ifdef EIGEN_CUDACC
145   std::cout << "  EIGEN_CUDACC:                 " << int(EIGEN_CUDACC) << "\n";
146   #endif
147 
148   #ifdef EIGEN_CUDA_SDK_VER
149   std::cout << "  EIGEN_CUDA_SDK_VER:             " << int(EIGEN_CUDA_SDK_VER) << "\n";
150   #endif
151 
152   #ifdef EIGEN_COMP_NVCC
153   std::cout << "  EIGEN_COMP_NVCC:             " << int(EIGEN_COMP_NVCC) << "\n";
154   #endif
155 
156   #ifdef EIGEN_HIPCC
157   std::cout << "  EIGEN_HIPCC:                 " << int(EIGEN_HIPCC) << "\n";
158   #endif
159 
160   std::cout << "  EIGEN_CUDA_ARCH:             " << info[0] << "\n";
161   std::cout << "  EIGEN_HIP_DEVICE_COMPILE:    " << info[1] << "\n";
162 
163   std::cout << "GPU device info:\n";
164   std::cout << "  name:                        " << deviceProp.name << "\n";
165   std::cout << "  capability:                  " << deviceProp.major << "." << deviceProp.minor << "\n";
166   std::cout << "  multiProcessorCount:         " << deviceProp.multiProcessorCount << "\n";
167   std::cout << "  maxThreadsPerMultiProcessor: " << deviceProp.maxThreadsPerMultiProcessor << "\n";
168   std::cout << "  warpSize:                    " << deviceProp.warpSize << "\n";
169   std::cout << "  regsPerBlock:                " << deviceProp.regsPerBlock << "\n";
170   std::cout << "  concurrentKernels:           " << deviceProp.concurrentKernels << "\n";
171   std::cout << "  clockRate:                   " << deviceProp.clockRate << "\n";
172   std::cout << "  canMapHostMemory:            " << deviceProp.canMapHostMemory << "\n";
173   std::cout << "  computeMode:                 " << deviceProp.computeMode << "\n";
174 }
175 
176 #endif // EIGEN_TEST_GPU_COMMON_H
177