├── CMakeLists.txt
├── gpu_library.cpp
├── gpu_library.cu
└── test.py


/CMakeLists.txt:
--------------------------------------------------------------------------------
 1 | cmake_minimum_required(VERSION 2.8)
 2 | find_package(CUDA)
 3 | find_package(PythonLibs 2.7 REQUIRED)
 4 | 
 5 | include_directories(${PYTHON_INCLUDE_DIRS})
 6 | 
 7 | set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
 8 | 
 9 | cuda_add_library(gpu_library SHARED
10 |   gpu_library.cpp
11 |   gpu_library.cu)
12 | 
13 | target_link_libraries(gpu_library
14 |   ${PYTHON_LIBRARIES}
15 |   cudart)
16 | 
17 | set_target_properties(gpu_library PROPERTIES PREFIX "")
18 | 


--------------------------------------------------------------------------------
/gpu_library.cpp:
--------------------------------------------------------------------------------
 1 | #include <iostream>
 2 | #include <pybind11/pybind11.h>
 3 | #include <pybind11/numpy.h>
 4 | #include <pybind11/stl.h>
 5 | #include <cuda_runtime.h>
 6 | 
 7 | void run_kernel
 8 | (double *vec, double scalar, int num_elements);
 9 | 
10 | void multiply_with_scalar(pybind11::array_t<double> vec, double scalar)
11 | {
12 |   int size = 10;
13 |   double *gpu_ptr;
14 |   cudaError_t error = cudaMalloc(&gpu_ptr, size * sizeof(double));
15 | 
16 |   if (error != cudaSuccess) {
17 |     throw std::runtime_error(cudaGetErrorString(error));
18 |   }
19 |   auto ha = vec.request();
20 | 
21 |   if (ha.ndim != 1) {
22 |     std::stringstream strstr;
23 |     strstr << "ha.ndim != 1" << std::endl;
24 |     strstr << "ha.ndim: " << ha.ndim << std::endl;
25 |     throw std::runtime_error(strstr.str());
26 |   }
27 | 
28 |   double* ptr = reinterpret_cast<double*>(ha.ptr);
29 |   error = cudaMemcpy(gpu_ptr, ptr, size * sizeof(double), cudaMemcpyHostToDevice);
30 |   if (error != cudaSuccess) {
31 |     throw std::runtime_error(cudaGetErrorString(error));
32 |   }
33 | 
34 |   run_kernel(gpu_ptr, scalar, size);
35 | 
36 |   error = cudaMemcpy(ptr, gpu_ptr, size * sizeof(double), cudaMemcpyDeviceToHost);
37 |   if (error != cudaSuccess) {
38 |     throw std::runtime_error(cudaGetErrorString(error));
39 |   }
40 | 
41 |   error = cudaFree(gpu_ptr);
42 |   if (error != cudaSuccess) {
43 |     throw std::runtime_error(cudaGetErrorString(error));
44 |   }
45 | }
46 | 
47 | PYBIND11_MODULE(gpu_library, m)
48 | {
49 |   m.def("multiply_with_scalar", multiply_with_scalar);
50 | }
51 | 


--------------------------------------------------------------------------------
/gpu_library.cu:
--------------------------------------------------------------------------------
 1 | #include <sstream>
 2 | #include <iostream>
 3 | #include <cuda_runtime.h>
 4 | 
 5 | __global__ void kernel
 6 | (double *vec, double scalar, int num_elements)
 7 | {
 8 |   unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
 9 |   if (idx < num_elements) {
10 |     vec[idx] = vec[idx] * scalar;
11 |   }
12 | }
13 | 
14 | 
15 | void run_kernel
16 | (double *vec, double scalar, int num_elements)
17 | {
18 |   dim3 dimBlock(256, 1, 1);
19 |   dim3 dimGrid(ceil((double)num_elements / dimBlock.x));
20 |   
21 |   kernel<<<dimGrid, dimBlock>>>
22 |     (vec, scalar, num_elements);
23 | 
24 |   cudaError_t error = cudaGetLastError();
25 |   if (error != cudaSuccess) {
26 |     std::stringstream strstr;
27 |     strstr << "run_kernel launch failed" << std::endl;
28 |     strstr << "dimBlock: " << dimBlock.x << ", " << dimBlock.y << std::endl;
29 |     strstr << "dimGrid: " << dimGrid.x << ", " << dimGrid.y << std::endl;
30 |     strstr << cudaGetErrorString(error);
31 |     throw strstr.str();
32 |   }
33 | }
34 | 


--------------------------------------------------------------------------------
/test.py:
--------------------------------------------------------------------------------
 1 | import sys
 2 | sys.path.append('./build')
 3 | 
 4 | import gpu_library
 5 | import numpy
 6 | 
 7 | vec = numpy.linspace(0,1,10)
 8 | 
 9 | print("before: ", vec)
10 | gpu_library.multiply_with_scalar(vec, 10)
11 | print("after: ", vec)
12 | 


--------------------------------------------------------------------------------