[30] | 1 | #ifndef NBODY_H_ |
---|
| 2 | #define NBODY_H_ |
---|
| 3 | |
---|
| 4 | #include <CL/cl.h> |
---|
| 5 | |
---|
| 6 | #include <iostream> |
---|
| 7 | #include <stdio.h> |
---|
| 8 | #include <stdlib.h> |
---|
| 9 | #include <assert.h> |
---|
| 10 | #include <string.h> |
---|
| 11 | #include <cmath> |
---|
| 12 | #include <malloc.h> |
---|
| 13 | #include <time.h> |
---|
| 14 | |
---|
| 15 | #include <GL/glut.h> |
---|
| 16 | |
---|
| 17 | #define GROUP_SIZE 64 |
---|
| 18 | #define ITER 5 |
---|
| 19 | |
---|
| 20 | /** |
---|
| 21 | * NBody |
---|
| 22 | * Class implements OpenCL NBody sample |
---|
| 23 | * |
---|
| 24 | */ |
---|
| 25 | |
---|
| 26 | class NBody |
---|
| 27 | { |
---|
| 28 | public: |
---|
| 29 | cl_double setupTime; // Time taken to setup OpenCL resources and building kernel; |
---|
| 30 | cl_double kernelTime; // Time taken to run kernel and read result back; |
---|
| 31 | |
---|
| 32 | size_t maxWorkGroupSize; // Max allowed work-items in a group; |
---|
| 33 | cl_uint maxDimensions; // Max group dimensions allowed; |
---|
| 34 | size_t* maxWorkItemSizes; // Max work-items sizes in each dimensions; |
---|
| 35 | cl_ulong totalLocalMemory; // Max local memory allowed; |
---|
| 36 | cl_ulong usedLocalMemory; // Used local memory; |
---|
| 37 | |
---|
| 38 | cl_float delT; // dT (timestep); |
---|
| 39 | cl_float espSqr; // Softening Factor; |
---|
| 40 | cl_float* initPos; // Initial position; |
---|
| 41 | cl_float* initVel; // Initial velocity; |
---|
| 42 | cl_float* vel; // Output velocity; |
---|
| 43 | cl_float* refPos; // Reference position; |
---|
| 44 | cl_float* refVel; // Reference velocity; |
---|
| 45 | cl_context context; // CL context; |
---|
| 46 | cl_device_id *devices; // CL device list; |
---|
| 47 | cl_mem updatedPos; // Position of partciles; |
---|
| 48 | cl_mem updatedVel; // Velocity of partciles; |
---|
| 49 | cl_command_queue commandQueue; // CL command queue; |
---|
| 50 | cl_program program; // CL program; |
---|
| 51 | cl_kernel kernel; // CL kernel; |
---|
| 52 | |
---|
| 53 | cl_int numParticles; // Number of particles in the system; |
---|
| 54 | const char * filename; // The name of the file that contains the kernel code; |
---|
| 55 | |
---|
| 56 | |
---|
| 57 | private: |
---|
| 58 | |
---|
| 59 | float random(float randMax, float randMin); |
---|
| 60 | int compareArray(const float* mat0, const float* mat1, unsigned int size); |
---|
| 61 | |
---|
| 62 | public: |
---|
| 63 | /** |
---|
| 64 | * Constructor |
---|
| 65 | * Initialize member variables |
---|
| 66 | * @param name name of sample (string) |
---|
| 67 | */ |
---|
| 68 | explicit NBody(std::string name) |
---|
| 69 | { |
---|
| 70 | setupTime = 0; |
---|
| 71 | kernelTime = 0; |
---|
| 72 | delT = 0.005f; |
---|
| 73 | espSqr = 50.0f; |
---|
| 74 | initPos = NULL; |
---|
| 75 | initVel = NULL; |
---|
| 76 | vel = NULL; |
---|
| 77 | refPos = NULL; |
---|
| 78 | refVel = NULL; |
---|
| 79 | devices = NULL; |
---|
| 80 | maxWorkItemSizes = NULL; |
---|
| 81 | filename = "NBody_Kernels.cl"; |
---|
| 82 | numParticles = 164; |
---|
| 83 | } |
---|
| 84 | |
---|
| 85 | /** |
---|
| 86 | * Constructor |
---|
| 87 | * Initialize member variables |
---|
| 88 | * @param name name of sample (const char*) |
---|
| 89 | */ |
---|
| 90 | explicit NBody(const char* name) |
---|
| 91 | { |
---|
| 92 | setupTime = 0; |
---|
| 93 | kernelTime = 0; |
---|
| 94 | delT = 0.005f; |
---|
| 95 | espSqr = 50.0f; |
---|
| 96 | initPos = NULL; |
---|
| 97 | initVel = NULL; |
---|
| 98 | vel = NULL; |
---|
| 99 | refPos = NULL; |
---|
| 100 | refVel = NULL; |
---|
| 101 | devices = NULL; |
---|
| 102 | maxWorkItemSizes = NULL; |
---|
| 103 | filename = "NBody_Kernels.cl"; |
---|
| 104 | numParticles = 1564; |
---|
| 105 | } |
---|
| 106 | |
---|
| 107 | ~NBody(); |
---|
| 108 | |
---|
| 109 | /** |
---|
| 110 | * Allocate and initialize host memory array with random values |
---|
| 111 | * @return 1 on success and 0 on failure |
---|
| 112 | */ |
---|
| 113 | int setupNBody(); |
---|
| 114 | |
---|
| 115 | /** |
---|
| 116 | * OpenCL related initialisations. |
---|
| 117 | * Set up Context, Device list, Command Queue, Memory buffers |
---|
| 118 | * Build CL kernel program executable |
---|
| 119 | * @return 1 on success and 0 on failure |
---|
| 120 | */ |
---|
| 121 | int setupCL(); |
---|
| 122 | |
---|
| 123 | /** |
---|
| 124 | * Set values for kernels' arguments |
---|
| 125 | * @return 1 on success and 0 on failure |
---|
| 126 | */ |
---|
| 127 | int setupCLKernels(); |
---|
| 128 | |
---|
| 129 | /** |
---|
| 130 | * Enqueue calls to the kernels |
---|
| 131 | * on to the command queue, wait till end of kernel execution. |
---|
| 132 | * Get kernel start and end time if timing is enabled |
---|
| 133 | * @return 1 on success and 0 on failure |
---|
| 134 | */ |
---|
| 135 | int runCLKernels(); |
---|
| 136 | |
---|
| 137 | /** |
---|
| 138 | * Load a .cl source file as a char* and it will be used |
---|
| 139 | * for as a parameter for creating a program |
---|
| 140 | * @return the char vector |
---|
| 141 | */ |
---|
| 142 | char * load_program_source(const char *filename); |
---|
| 143 | |
---|
| 144 | /** |
---|
| 145 | * Reference CPU implementation of Binomial Option |
---|
| 146 | * for performance comparison |
---|
| 147 | */ |
---|
| 148 | void nBodyCPUReference(); |
---|
| 149 | |
---|
| 150 | /** |
---|
| 151 | * Override from SDKSample. Print sample stats. |
---|
| 152 | */ |
---|
| 153 | void printStats(); |
---|
| 154 | |
---|
| 155 | /** |
---|
| 156 | * Override from SDKSample, adjust width and height |
---|
| 157 | * of execution domain, perform all sample setup |
---|
| 158 | */ |
---|
| 159 | int setup(); |
---|
| 160 | |
---|
| 161 | /** |
---|
| 162 | * Override from SDKSample |
---|
| 163 | * Run OpenCL NBody |
---|
| 164 | */ |
---|
| 165 | int run(); |
---|
| 166 | |
---|
| 167 | /** |
---|
| 168 | * Override from SDKSample |
---|
| 169 | * Cleanup memory allocations |
---|
| 170 | */ |
---|
| 171 | int cleanup(); |
---|
| 172 | }; |
---|
| 173 | |
---|
| 174 | #endif // NBODY_H_ |
---|