Q&A
Vitis HLS側のCシミュレーションではだめですか?
Vitisというソフトを用いてハードウェアアクセラレーションを行ったのですが，その比較として高位合成前のC++のコードを実行しようとしました．
用いたコードは以下のものです
krnl_vadd.cpp
1#define BUFFER_SIZE 256 2#define DATA_SIZE 4194304 3#define m_axi_port 128 //変更点はここだけ 4//TRIPCOUNT identifier 5const unsigned int c_len = DATA_SIZE / BUFFER_SIZE; 6const unsigned int c_size = BUFFER_SIZE; 7 8/* 9 Vector Addition Kernel Implementation 10 Arguments: 11 in1 (input) --> Input Vector1 12 in2 (input) --> Input Vector2 13 out_r (output) --> Output Vector 14 size (input) --> Size of Vector in Integer 15*/ 16 17extern "C" { 18void krnl_vadd(const unsigned int *in1, // Read-Only Vector 1 19 const unsigned int *in2, // Read-Only Vector 2 20 unsigned int *out_r, // Output Result 21 int size // Size in integer 22) { 23#pragma HLS INTERFACE m_axi port = in1 offset = slave bundle = gmem 24#pragma HLS INTERFACE m_axi port = in2 offset = slave bundle = gmem 25#pragma HLS INTERFACE m_axi port = out_r offset = slave bundle = gmem 26#pragma HLS INTERFACE s_axilite port = in1 bundle = control 27#pragma HLS INTERFACE s_axilite port = in2 bundle = control 28#pragma HLS INTERFACE s_axilite port = out_r bundle = control 29#pragma HLS INTERFACE s_axilite port = size bundle = control 30#pragma HLS INTERFACE s_axilite port = return bundle = control 31 32 unsigned int v1_buffer[BUFFER_SIZE]; // Local memory to store vector1 33 34 //Per iteration of this loop perform BUFFER_SIZE vector addition 35 for (int i = 0; i < size; i += BUFFER_SIZE) { 36 #pragma HLS LOOP_TRIPCOUNT min=c_len max=c_len 37 int chunk_size = BUFFER_SIZE; 38 //boundary checks 39 if ((i + BUFFER_SIZE) > size) 40 chunk_size = size - i; 41 42 read1: for (int j = 0; j < chunk_size; j++) { 43 #pragma HLS LOOP_TRIPCOUNT min=c_size max=c_size 44 #pragma HLS PIPELINE II=1 45 v1_buffer[j] = in1[i + j]; 46 } 47 48 //Burst reading B and calculating C and Burst writing 49 // to Global memory 50 vadd_writeC: for (int j = 0; j < chunk_size; j++) { 51 #pragma HLS LOOP_TRIPCOUNT min=c_size max=c_size 52 #pragma HLS PIPELINE II=1 53 //perform vector addition 54 out_r[i+j] = v1_buffer[j] + in2[i+j]; 55 } 56 57 } 58} 59} 60
vadd.cpp
1#include <stdlib.h> 2#include <fstream> 3#include <iostream> 4#include "vadd.h" 5 6static const int DATA_SIZE = 4194304; 7 8static const std::string error_message = 9 "Error: Result mismatch:\n" 10 "i = %d CPU result = %d Device result = %d\n"; 11 12int main(int argc, char* argv[]) { 13 14 //TARGET_DEVICE macro needs to be passed from gcc command line 15 if(argc != 2) { 16 std::cout << "Usage: " << argv[0] <<" <xclbin>" << std::endl; 17 return EXIT_FAILURE; 18 } 19 20 char* xclbinFilename = argv[1]; 21 22 // Compute the size of array in bytes 23 size_t size_in_bytes = DATA_SIZE * sizeof(int); 24 25 // Creates a vector of DATA_SIZE elements with an initial value of 10 and 32 26 // using customized allocator for getting buffer alignment to 4k boundary 27 28 std::vector<cl::Device> devices; 29 cl::Device device; 30 std::vector<cl::Platform> platforms; 31 bool found_device = false; 32 33 //traversing all Platforms To find Xilinx Platform and targeted 34 //Device in Xilinx Platform 35 cl::Platform::get(&platforms); 36 for(size_t i = 0; (i < platforms.size() ) & (found_device == false) ;i++){ 37 cl::Platform platform = platforms[i]; 38 std::string platformName = platform.getInfo<CL_PLATFORM_NAME>(); 39 if ( platformName == "Xilinx"){ 40 devices.clear(); 41 platform.getDevices(CL_DEVICE_TYPE_ACCELERATOR, &devices); 42 if (devices.size()){ 43 device = devices[0]; 44 found_device = true; 45 break; 46 } 47 } 48 } 49 if (found_device == false){ 50 std::cout << "Error: Unable to find Target Device " 51 << device.getInfo<CL_DEVICE_NAME>() << std::endl; 52 return EXIT_FAILURE; 53 } 54 55 // Creating Context and Command Queue for selected device 56 cl::Context context(device); 57 cl::CommandQueue q(context, device, CL_QUEUE_PROFILING_ENABLE); 58 59 // Load xclbin 60 std::cout << "Loading: '" << xclbinFilename << "'\n"; 61 std::ifstream bin_file(xclbinFilename, std::ifstream::binary); 62 bin_file.seekg (0, bin_file.end); 63 unsigned nb = bin_file.tellg(); 64 bin_file.seekg (0, bin_file.beg); 65 char *buf = new char [nb]; 66 bin_file.read(buf, nb); 67 68 // Creating Program from Binary File 69 cl::Program::Binaries bins; 70 bins.push_back({buf,nb}); 71 devices.resize(1); 72 cl::Program program(context, devices, bins); 73 74 // This call will get the kernel object from program. A kernel is an 75 // OpenCL function that is executed on the FPGA. 76 cl::Kernel krnl_vector_add(program,"krnl_vadd"); 77 78 // These commands will allocate memory on the Device. The cl::Buffer objects can 79 // be used to reference the memory locations on the device. 80 cl::Buffer buffer_a(context, CL_MEM_READ_ONLY, size_in_bytes); 81 cl::Buffer buffer_b(context, CL_MEM_READ_ONLY, size_in_bytes); 82 cl::Buffer buffer_result(context, CL_MEM_WRITE_ONLY, size_in_bytes); 83 84 //set the kernel Arguments 85 int narg=0; 86 krnl_vector_add.setArg(narg++,buffer_a); 87 krnl_vector_add.setArg(narg++,buffer_b); 88 krnl_vector_add.setArg(narg++,buffer_result); 89 krnl_vector_add.setArg(narg++,DATA_SIZE); 90 91 //We then need to map our OpenCL buffers to get the pointers 92 int *ptr_a = (int *) q.enqueueMapBuffer (buffer_a , CL_TRUE , CL_MAP_WRITE , 0, size_in_bytes); 93 int *ptr_b = (int *) q.enqueueMapBuffer (buffer_b , CL_TRUE , CL_MAP_WRITE , 0, size_in_bytes); 94 int *ptr_result = (int *) q.enqueueMapBuffer (buffer_result , CL_TRUE , CL_MAP_READ , 0, size_in_bytes); 95 96 //setting input data 97 for(int i = 0 ; i< DATA_SIZE; i++){ 98 ptr_a[i] = 10; 99 ptr_b[i] = 20; 100 } 101 102 // Data will be migrated to kernel space 103 q.enqueueMigrateMemObjects({buffer_a,buffer_b},0/* 0 means from host*/); 104 105 //Launch the Kernel 106 q.enqueueTask(krnl_vector_add); 107 108 // The result of the previous kernel execution will need to be retrieved in 109 // order to view the results. This call will transfer the data from FPGA to 110 // source_results vector 111 q.enqueueMigrateMemObjects({buffer_result},CL_MIGRATE_MEM_OBJECT_HOST); 112 113 q.finish(); 114 115 //Verify the result 116 int match = 0; 117 for (int i = 0; i < DATA_SIZE; i++) { 118 int host_result = ptr_a[i] + ptr_b[i]; 119 if (ptr_result[i] != host_result) { 120 printf(error_message.c_str(), i, host_result, ptr_result[i]); 121 match = 1; 122 break; 123 } 124 } 125 126 q.enqueueUnmapMemObject(buffer_a , ptr_a); 127 q.enqueueUnmapMemObject(buffer_b , ptr_b); 128 q.enqueueUnmapMemObject(buffer_result , ptr_result); 129 q.finish(); 130 131 std::cout << "TEST " << (match ? "FAILED" : "PASSED") << std::endl; 132 return (match ? EXIT_FAILURE : EXIT_SUCCESS); 133 134} 135
vadd.h
1#pragma once 2 3#define CL_HPP_CL_1_2_DEFAULT_BUILD 4#define CL_HPP_TARGET_OPENCL_VERSION 120 5#define CL_HPP_MINIMUM_OPENCL_VERSION 120 6#define CL_HPP_ENABLE_PROGRAM_CONSTRUCTION_FROM_ARRAY_COMPATIBILITY 1 7 8#include <CL/cl2.hpp> 9 10//Customized buffer allocation for 4K boundary alignment 11template <typename T> 12struct aligned_allocator 13{ 14 using value_type = T; 15 T* allocate(std::size_t num) 16 { 17 void* ptr = nullptr; 18 if (posix_memalign(&ptr,4096,num*sizeof(T))) 19 throw std::bad_alloc(); 20 return reinterpret_cast<T*>(ptr); 21 } 22 void deallocate(T* p, std::size_t num) 23 { 24 free(p); 25 } 26}; 27
しかし，そのまま実行した場合は
PS C:\cpp\vadd_src> g++ krnl_vadd.cpp vadd.cpp -o vadd In file included from vadd.cpp:47: vadd.h:53:10: fatal error: CL/cl2.hpp: No such file or directory #include <CL/cl2.hpp> ^~~~~~~~~~~~ compilation terminated.
というエラーが発生（CL/cl2.hppがない）してしまいます．
CL/cl2.hppを用いずにこのプログラムを動かす方法はありますでしょうか．
また，実行にかかった時間なども知りたいのでその方法も教えていただけると助かります．
回答2件
0
ご回答ありがとうございます。
試してみた結果、以下のようなエラーとなりました。
vadd.cpp: In function ‘int main(int, char**)’: vadd.cpp:71:10: error: ‘vector’ is not a member of ‘std’ std::vector<cl::Device> devices; ^~~~~~ vadd.cpp:71:17: error: ‘cl’ has not been declared std::vector<cl::Device> devices; ^~ vadd.cpp:71:29: error: ‘devices’ was not declared in this scope std::vector<cl::Device> devices; ^~~~~~~ vadd.cpp:72:5: error: ‘cl’ has not been declared cl::Device device; ^~ vadd.cpp:73:10: error: ‘vector’ is not a member of ‘std’ std::vector<cl::Platform> platforms; ^~~~~~ vadd.cpp:73:17: error: ‘cl’ has not been declared std::vector<cl::Platform> platforms; ^~ vadd.cpp:73:31: error: ‘platforms’ was not declared in this scope std::vector<cl::Platform> platforms; ^~~~~~~~~ vadd.cpp:78:5: error: ‘cl’ has not been declared cl::Platform::get(&platforms); ^~ vadd.cpp:80:9: error: ‘cl’ has not been declared cl::Platform platform = platforms[i]; ^~ vadd.cpp:81:36: error: ‘platform’ was not declared in this scope std::string platformName = platform.getInfo<CL_PLATFORM_NAME>(); ^~~~~~~~ vadd.cpp:81:36: note: suggested alternative: ‘platformName’ std::string platformName = platform.getInfo<CL_PLATFORM_NAME>(); ^~~~~~~~ platformName vadd.cpp:81:53: error: ‘CL_PLATFORM_NAME’ was not declared in this scope std::string platformName = platform.getInfo<CL_PLATFORM_NAME>(); ^~~~~~~~~~~~~~~~ vadd.cpp:81:71: error: expected primary-expression before ‘)’ token std::string platformName = platform.getInfo<CL_PLATFORM_NAME>(); ^ vadd.cpp:84:33: error: ‘CL_DEVICE_TYPE_ACCELERATOR’ was not declared in this scope platform.getDevices(CL_DEVICE_TYPE_ACCELERATOR, &devices); ^~~~~~~~~~~~~~~~~~~~~~~~~~ vadd.cpp:86:7: error: ‘device’ was not declared in this scope device = devices[0]; ^~~~~~ vadd.cpp:86:7: note: suggested alternative: ‘dysize’ device = devices[0]; ^~~~~~ dysize vadd.cpp:94:15: error: ‘device’ was not declared in this scope << device.getInfo<CL_DEVICE_NAME>() << std::endl; ^~~~~~ vadd.cpp:94:15: note: suggested alternative: ‘dysize’ << device.getInfo<CL_DEVICE_NAME>() << std::endl; ^~~~~~ dysize vadd.cpp:94:30: error: ‘CL_DEVICE_NAME’ was not declared in this scope << device.getInfo<CL_DEVICE_NAME>() << std::endl; ^~~~~~~~~~~~~~ vadd.cpp:94:30: note: suggested alternative: ‘__LC_NAME’ << device.getInfo<CL_DEVICE_NAME>() << std::endl; ^~~~~~~~~~~~~~ __LC_NAME vadd.cpp:94:46: error: expected primary-expression before ‘)’ token << device.getInfo<CL_DEVICE_NAME>() << std::endl; ^ vadd.cpp:99:5: error: ‘cl’ has not been declared cl::Context context(device); ^~ vadd.cpp:100:5: error: ‘cl’ has not been declared cl::CommandQueue q(context, device, CL_QUEUE_PROFILING_ENABLE); ^~ vadd.cpp:112:5: error: ‘cl’ has not been declared cl::Program::Binaries bins; ^~ vadd.cpp:113:5: error: ‘bins’ was not declared in this scope bins.push_back({buf,nb}); ^~~~ vadd.cpp:113:5: note: suggested alternative: ‘uint’ bins.push_back({buf,nb}); ^~~~ uint vadd.cpp:115:5: error: ‘cl’ has not been declared cl::Program program(context, devices, bins); ^~ vadd.cpp:119:5: error: ‘cl’ has not been declared cl::Kernel krnl_vector_add(program,"krnl_vadd"); ^~ vadd.cpp:123:5: error: ‘cl’ has not been declared cl::Buffer buffer_a(context, CL_MEM_READ_ONLY, size_in_bytes); ^~ vadd.cpp:124:5: error: ‘cl’ has not been declared cl::Buffer buffer_b(context, CL_MEM_READ_ONLY, size_in_bytes); ^~ vadd.cpp:125:5: error: ‘cl’ has not been declared cl::Buffer buffer_result(context, CL_MEM_WRITE_ONLY, size_in_bytes); ^~ vadd.cpp:129:5: error: ‘krnl_vector_add’ was not declared in this scope krnl_vector_add.setArg(narg++,buffer_a); ^~~~~~~~~~~~~~~ vadd.cpp:129:35: error: ‘buffer_a’ was not declared in this scope krnl_vector_add.setArg(narg++,buffer_a); ^~~~~~~~ vadd.cpp:130:35: error: ‘buffer_b’ was not declared in this scope krnl_vector_add.setArg(narg++,buffer_b); ^~~~~~~~ vadd.cpp:131:35: error: ‘buffer_result’ was not declared in this scope krnl_vector_add.setArg(narg++,buffer_result); ^~~~~~~~~~~~~ vadd.cpp:131:35: note: suggested alternative: ‘__codecvt_result’ krnl_vector_add.setArg(narg++,buffer_result); ^~~~~~~~~~~~~ __codecvt_result vadd.cpp:135:26: error: ‘q’ was not declared in this scope int *ptr_a = (int *) q.enqueueMapBuffer (buffer_a , CL_TRUE , CL_MAP_WRITE , 0, size_in_bytes); ^ vadd.cpp:135:57: error: ‘CL_TRUE’ was not declared in this scope int *ptr_a = (int *) q.enqueueMapBuffer (buffer_a , CL_TRUE , CL_MAP_WRITE , 0, size_in_bytes); ^~~~~~~ vadd.cpp:135:67: error: ‘CL_MAP_WRITE’ was not declared in this scope int *ptr_a = (int *) q.enqueueMapBuffer (buffer_a , CL_TRUE , CL_MAP_WRITE , 0, size_in_bytes); ^~~~~~~~~~~~ vadd.cpp:137:77: error: ‘CL_MAP_READ’ was not declared in this scope *ptr_result = (int *) q.enqueueMapBuffer (buffer_result , CL_TRUE , CL_MAP_READ , 0, size_in_bytes); ^~~~~~~~~~~ vadd.cpp:137:77: note: suggested alternative: ‘CLONE_THREAD’ *ptr_result = (int *) q.enqueueMapBuffer (buffer_result , CL_TRUE , CL_MAP_READ , 0, size_in_bytes); ^~~~~~~~~~~ CLONE_THREAD vadd.cpp:154:48: error: ‘CL_MIGRATE_MEM_OBJECT_HOST’ was not declared in this scope q.enqueueMigrateMemObjects({buffer_result},CL_MIGRATE_MEM_OBJECT_HOST);
投稿2022/02/01 05:16
総合スコア4
0
[回答ではありません。が...]
CL/cl2.hppを用いずにこのプログラムを動かす方法はありますでしょうか
#include <CL/cl2.hpp>
を
#include <cstdlib>
#include <new>
に差し替えてもダメですかね?
投稿2022/02/01 00:37
総合スコア16614
ご回答ありがとうございます。
回答欄に実行結果を載せましたので見ていただけたら幸いです。
となると、「CL/cl2.hppを用いずにこのプログラムを動かす方法はありますでしょうか」のコタエは NO です。
...ところで OpenCL はインストールされていますか?
あなたの回答
tips
プレビュー
