以下CUDA sample是分別用C++和CUDA實(shí)現(xiàn)的類(lèi)似prior vbox layer的操作，并對(duì)其中使用到的CUDA函數(shù)進(jìn)行了解說(shuō)，各個(gè)文件內(nèi)容如下：

common.hpp:

#ifndef FBC_CUDA_TEST_COMMON_HPP_
#define FBC_CUDA_TEST_COMMON_HPP_#include <typeinfo>
#include<random>
#include <cuda_runtime.h> // For the CUDA runtime routines (prefixed with "cuda_")
#include <device_launch_parameters.h>
#include <opencv2/opencv.hpp>template< typename T >
static inline int check_Cuda(T result, const char * const func, const char * const file, const int line)
{if (result) {fprintf(stderr, "Error CUDA: at %s: %d, error code=%d, func: %s\n", file, line, static_cast<unsigned int>(result), func);cudaDeviceReset(); // Make sure we call CUDA Device Reset before exitingreturn -1;}
}template< typename T >
static inline int check(T result, const char * const func, const char * const file, const int line)
{if (result) {fprintf(stderr, "Error: at %s: %d, error code=%d, func: %s\n", file, line, static_cast<unsigned int>(result), func);return -1;}
}#define checkCudaErrors(val) check_Cuda((val), __FUNCTION__, __FILE__, __LINE__)
#define checkErrors(val) check((val), __FUNCTION__, __FILE__, __LINE__)#define CHECK(x) { \if (x) {} \else { fprintf(stderr, "Check Failed: %s, file: %s, line: %d\n", #x, __FILE__, __LINE__); return -1; } \
}#define PRINT_ERROR_INFO(info) { \fprintf(stderr, "Error: %s, file: %s, func: %s, line: %d\n", #info, __FILE__, __FUNCTION__, __LINE__); \return -1; }#define TIME_START_CPU auto start = std::chrono::steady_clock::now();
#define TIME_END_CPU auto end = std::chrono::steady_clock::now(); \auto duration = std::chrono::duration_cast<std::chrono::nanoseconds>(end - start); \*elapsed_time = duration.count() * 1.0e-6;#define TIME_START_GPU cudaEvent_t start, stop; /* cudaEvent_t: CUDA event types,結(jié)構(gòu)體類(lèi)型, CUDA事件,用于測(cè)量GPU在某個(gè)任務(wù)上花費(fèi)的時(shí)間,CUDA中的事件本質(zhì)上是一個(gè)GPU時(shí)間戳,由于CUDA事件是在GPU上實(shí)現(xiàn)的,因此它們不適于對(duì)同時(shí)包含設(shè)備代碼和主機(jī)代碼的混合代碼計(jì)時(shí) */ \cudaEventCreate(&start); /* 創(chuàng)建一個(gè)事件對(duì)象,異步啟動(dòng) */ \cudaEventCreate(&stop); \cudaEventRecord(start, 0); /* 記錄一個(gè)事件,異步啟動(dòng),start記錄起始時(shí)間 */
#define TIME_END_GPU cudaEventRecord(stop, 0); /* 記錄一個(gè)事件,異步啟動(dòng),stop記錄結(jié)束時(shí)間 */ \cudaEventSynchronize(stop); /* 事件同步,等待一個(gè)事件完成,異步啟動(dòng) */ \cudaEventElapsedTime(elapsed_time, start, stop); /* 計(jì)算兩個(gè)事件之間經(jīng)歷的時(shí)間,單位為毫秒,異步啟動(dòng) */ \cudaEventDestroy(start); /* 銷(xiāo)毀事件對(duì)象,異步啟動(dòng) */ \cudaEventDestroy(stop);#define EPS_ 1.0e-4 // ε(Epsilon),非常小的數(shù)
#define PI 3.1415926535897932f
#define INF 2.e10fstatic inline void generator_random_number(float* data, int length, float a = 0.f, float b = 1.f)
{std::random_device rd; std::mt19937 generator(rd()); // 每次產(chǎn)生不固定的不同的值//std::default_random_engine generator; // 每次產(chǎn)生固定的不同的值std::uniform_real_distribution<float> distribution(a, b);for (int i = 0; i < length; ++i) {data[i] = distribution(generator);}
}template<typename T> // unsigned char, char, int , short
static inline void generator_random_number(T* data, int length, T a = (T)0, T b = (T)1)
{std::random_device rd; std::mt19937 generator(rd()); // 每次產(chǎn)生不固定的不同的值//std::default_random_engine generator; // 每次產(chǎn)生固定的不同的值std::uniform_int_distribution<int> distribution(a, b);for (int i = 0; i < length; ++i) {data[i] = static_cast<T>(distribution(generator));}
}static int save_image(const cv::Mat& mat1, const cv::Mat& mat2, int width, int height, const std::string& name)
{CHECK(mat1.type() == mat2.type());cv::Mat src1, src2, dst;cv::resize(mat1, src1, cv::Size(width / 2, height));cv::resize(mat2, src2, cv::Size(width / 2, height));dst = cv::Mat(height, width / 2 * 2, mat1.type());cv::Mat tmp = dst(cv::Rect(0, 0, width / 2, height));src1.copyTo(tmp);tmp = dst(cv::Rect(width / 2, 0, width / 2, height));src2.copyTo(tmp);cv::imwrite(name, dst);
}template<typename T>
static inline int compare_result(const T* src1, const T* src2, int length)
{CHECK(src1);CHECK(src2);int count{ 0 };for (int i = 0; i < length; ++i) {if (fabs(src1[i] - src2[i]) > EPS_) {if (typeid(float).name() == typeid(T).name() || typeid(double).name() == typeid(T).name())fprintf(stderr, "index: %d, val1: %f, val2: %f\n", i, src1[i], src2[i]);elsefprintf(stderr, "index: %d, val1: %d, val2: %d\n", i, src1[i], src2[i]);++count;}if (count > 100) return -1;}return 0;
}#endif // FBC_CUDA_TEST_COMMON_HPP_

funset.cpp:

#include "funset.hpp"
#include <random>
#include <iostream>
#include <vector>
#include <memory>
#include <string>
#include <algorithm>
#include "common.hpp"int test_layer_prior_vbox()
{std::vector<float> vec1{423.f, 245.f, 1333.f, 1444.f, 123.f, 23.f, 32.f, 66.f};std::vector<float> vec2(vec1[6]);std::vector<float> vec3(4);int length = int(vec1[0] * vec1[1] * vec1[6] * 4 * 2);std::unique_ptr<float[]> data1(new float[length]), data2(new float[length]);std::for_each(data1.get(), data1.get() + length, [](float& n) {n = 0.f; });std::for_each(data2.get(), data2.get() + length, [](float& n) {n = 0.f; });generator_random_number(vec2.data(), vec2.size(), 10.f, 100.f);generator_random_number(vec3.data(), vec3.size(), 1.f, 10.f);float elapsed_time1{ 0.f }, elapsed_time2{ 0.f }; // millisecondsint ret = layer_prior_vbox_cpu(data1.get(), length, vec1, vec2, vec3, &elapsed_time1);if (ret != 0) PRINT_ERROR_INFO(layer_prior_vbox_cpu);ret = layer_prior_vbox_gpu(data2.get(), length, vec1, vec2, vec3, &elapsed_time2);if (ret != 0) PRINT_ERROR_INFO(layer_prior_vbox_gpu);compare_result(data1.get(), data2.get(), length);fprintf(stderr, "test layer prior vbox: cpu run time: %f ms, gpu run time: %f ms\n", elapsed_time1, elapsed_time2);return 0;
}

layer_prior_vbox.cpp:

#include "funset.hpp"
#include <vector>
#include <chrono>
#include "common.hpp"int layer_prior_vbox_cpu(float* dst, int length, const std::vector<float>& vec1, const std::vector<float>& vec2,const std::vector<float>& vec3, float* elapsed_time)
{TIME_START_CPUint layer_width = (int)vec1[0];int layer_height = (int)vec1[1];int image_width = (int)vec1[2];int image_height = (int)vec1[3];float offset = vec1[4];float step = vec1[5];int num_priors = (int)vec1[6];float width = vec1[7];CHECK(length == layer_width * layer_height * num_priors * 4 * 2);CHECK(vec1.size() == 8);CHECK(vec2.size() == num_priors);CHECK(vec3.size() == 4);float* top_data = dst;int idx = 0;for (int h = 0; h < layer_height; ++h) {for (int w = 0; w < layer_width; ++w) {float center_x = (w + offset) * step;float center_y = (h + offset) * step;for (int s = 0; s < num_priors; ++s) {float box_width = width;float box_height = vec2[s];top_data[idx++] = (center_x - box_width / 2.) / image_width;top_data[idx++] = (center_y - box_height / 2.) / image_height;top_data[idx++] = (center_x + box_width / 2.) / image_width;top_data[idx++] = (center_y + box_height / 2.) / image_height;}}}int len = layer_width * layer_height * num_priors;for (int i = 0; i < len; ++i) {for (int j = 0; j < 4; ++j) {top_data[idx++] = vec3[j];}}TIME_END_CPUreturn 0;
}

layer_prior_vbox.cu:

#include "funset.hpp"
#include <iostream>
#include <memory>
#include <algorithm>
#include <cmath>
#include <cuda_runtime.h> // For the CUDA runtime routines (prefixed with "cuda_")
#include <device_launch_parameters.h>
#include "common.hpp"/* __global__: 函數(shù)類(lèi)型限定符;在設(shè)備上運(yùn)行;在主機(jī)端調(diào)用,計(jì)算能力3.2及以上可以在
設(shè)備端調(diào)用;聲明的函數(shù)的返回值必須是void類(lèi)型;對(duì)此類(lèi)型函數(shù)的調(diào)用是異步的,即在
設(shè)備完全完成它的運(yùn)行之前就返回了;對(duì)此類(lèi)型函數(shù)的調(diào)用必須指定執(zhí)行配置,即用于在
設(shè)備上執(zhí)行函數(shù)時(shí)的grid和block的維度,以及相關(guān)的流(即插入<<<   >>>運(yùn)算符);
a kernel,表示此函數(shù)為內(nèi)核函數(shù)(運(yùn)行在GPU上的CUDA并行計(jì)算函數(shù)稱(chēng)為kernel(內(nèi)核函
數(shù)),內(nèi)核函數(shù)必須通過(guò)__global__函數(shù)類(lèi)型限定符定義);*/
__global__ static void layer_prior_vbox(float* dst, int layer_width, int layer_height, int image_width, int image_height,float offset, float step, int num_priors, float width, const float* height, const float* variance, int channel_size)
{/* gridDim: 內(nèi)置變量,用于描述線(xiàn)程網(wǎng)格的維度,對(duì)于所有線(xiàn)程塊來(lái)說(shuō),這個(gè)變量是一個(gè)常數(shù),用來(lái)保存線(xiàn)程格每一維的大小,即每個(gè)線(xiàn)程格中線(xiàn)程塊的數(shù)量.一個(gè)grid為三維,為dim3類(lèi)型；blockDim: 內(nèi)置變量,用于說(shuō)明每個(gè)block的維度與尺寸.為dim3類(lèi)型,包含了block在三個(gè)維度上的尺寸信息;對(duì)于所有線(xiàn)程塊來(lái)說(shuō),這個(gè)變量是一個(gè)常數(shù),保存的是線(xiàn)程塊中每一維的線(xiàn)程數(shù)量;blockIdx: 內(nèi)置變量,變量中包含的值就是當(dāng)前執(zhí)行設(shè)備代碼的線(xiàn)程塊的索引;用于說(shuō)明當(dāng)前thread所在的block在整個(gè)grid中的位置,blockIdx.x取值范圍是[0,gridDim.x-1],blockIdx.y取值范圍是[0, gridDim.y-1].為uint3類(lèi)型,包含了一個(gè)block在grid中各個(gè)維度上的索引信息;threadIdx: 內(nèi)置變量,變量中包含的值就是當(dāng)前執(zhí)行設(shè)備代碼的線(xiàn)程索引;用于說(shuō)明當(dāng)前thread在block中的位置;如果線(xiàn)程是一維的可獲取threadIdx.x,如果是二維的還可獲取threadIdx.y,如果是三維的還可獲取threadIdx.z;為uint3類(lèi)型,包含了一個(gè)thread在block中各個(gè)維度的索引信息 */int x = threadIdx.x + blockIdx.x * blockDim.x;int y = threadIdx.y + blockIdx.y * blockDim.y;if (x < layer_width && y < layer_height) {float center_x = (x + offset) * step;float center_y = (y + offset) * step;int idx = x * num_priors * 4 + y * (layer_width * num_priors * 4);for (int s = 0; s < num_priors; ++s) {float box_width = width;float box_height = height[s];int idx1 = idx + s * 4;dst[idx1] = (center_x - box_width / 2.) / image_width;dst[idx1 + 1] = (center_y - box_height / 2.) / image_height;dst[idx1 + 2] = (center_x + box_width / 2.) / image_width;dst[idx1 + 3] = (center_y + box_height / 2.) / image_height;int idx2 = channel_size + idx + s * 4;dst[idx2] = variance[0];dst[idx2 + 1] = variance[1];dst[idx2 + 2] = variance[2];dst[idx2 + 3] = variance[3];}}
}int layer_prior_vbox_gpu(float* dst, int length, const std::vector<float>& vec1, const std::vector<float>& vec2,const std::vector<float>& vec3, float* elapsed_time)
{float *dev_dst{ nullptr }, *dev_vec;// cudaMalloc: 在設(shè)備端分配內(nèi)存cudaMalloc(&dev_dst, length * sizeof(float));cudaMalloc(&dev_vec, (vec2.size()+vec3.size()) * sizeof(float));/* cudaMemcpy: 在主機(jī)端和設(shè)備端拷貝數(shù)據(jù),此函數(shù)第四個(gè)參數(shù)僅能是下面之一:(1). cudaMemcpyHostToHost: 拷貝數(shù)據(jù)從主機(jī)端到主機(jī)端(2). cudaMemcpyHostToDevice: 拷貝數(shù)據(jù)從主機(jī)端到設(shè)備端(3). cudaMemcpyDeviceToHost: 拷貝數(shù)據(jù)從設(shè)備端到主機(jī)端(4). cudaMemcpyDeviceToDevice: 拷貝數(shù)據(jù)從設(shè)備端到設(shè)備端(5). cudaMemcpyDefault: 從指針值自動(dòng)推斷拷貝數(shù)據(jù)方向,需要支持統(tǒng)一虛擬尋址(CUDA6.0及以上版本)cudaMemcpy函數(shù)對(duì)于主機(jī)是同步的 */cudaMemcpy(dev_dst, dst, length * sizeof(float), cudaMemcpyHostToDevice);cudaMemcpy(dev_vec, vec2.data(), vec2.size() * sizeof(float), cudaMemcpyHostToDevice);cudaMemcpy(dev_vec + vec2.size(), vec3.data(), vec3.size() * sizeof(float), cudaMemcpyHostToDevice);int layer_width = (int)vec1[0];int layer_height = (int)vec1[1];int image_width = (int)vec1[2];int image_height = (int)vec1[3];float offset = vec1[4];float step = vec1[5];int num_priors = (int)vec1[6];float width = vec1[7];int channel_size = layer_width * layer_height * num_priors * 4;TIME_START_GPU/* dim3: 基于uint3定義的內(nèi)置矢量類(lèi)型，相當(dāng)于由3個(gè)unsigned int類(lèi)型組成的結(jié)構(gòu)體，可表示一個(gè)三維數(shù)組，在定義dim3類(lèi)型變量時(shí)，凡是沒(méi)有賦值的元素都會(huì)被賦予默認(rèn)值1 */// Note：每一個(gè)線(xiàn)程塊支持的最大線(xiàn)程數(shù)量為1024，即threads.x*threads.y必須小于等于1024dim3 threads(32, 32);dim3 blocks((layer_width + 31) / 32, (layer_height + 31) / 32);/* <<< >>>: 為CUDA引入的運(yùn)算符,指定線(xiàn)程網(wǎng)格和線(xiàn)程塊維度等,傳遞執(zhí)行參數(shù)給CUDA編譯器和運(yùn)行時(shí)系統(tǒng),用于說(shuō)明內(nèi)核函數(shù)中的線(xiàn)程數(shù)量,以及線(xiàn)程是如何組織的;尖括號(hào)中這些參數(shù)并不是傳遞給設(shè)備代碼的參數(shù),而是告訴運(yùn)行時(shí)如何啟動(dòng)設(shè)備代碼,傳遞給設(shè)備代碼本身的參數(shù)是放在圓括號(hào)中傳遞的,就像標(biāo)準(zhǔn)的函數(shù)調(diào)用一樣;不同計(jì)算能力的設(shè)備對(duì)線(xiàn)程的總數(shù)和組織方式有不同的約束;必須先為kernel中用到的數(shù)組或變量分配好足夠的空間,再調(diào)用kernel函數(shù),否則在GPU計(jì)算時(shí)會(huì)發(fā)生錯(cuò)誤,例如越界等 ;使用運(yùn)行時(shí)API時(shí),需要在調(diào)用的內(nèi)核函數(shù)名與參數(shù)列表直接以<<<Dg,Db,Ns,S>>>的形式設(shè)置執(zhí)行配置,其中：Dg是一個(gè)dim3型變量,用于設(shè)置grid的維度和各個(gè)維度上的尺寸.設(shè)置好Dg后,grid中將有Dg.x*Dg.y*Dg.z個(gè)block;Db是一個(gè)dim3型變量,用于設(shè)置block的維度和各個(gè)維度上的尺寸.設(shè)置好Db后,每個(gè)block中將有Db.x*Db.y*Db.z個(gè)thread;Ns是一個(gè)size_t型變量,指定各塊為此調(diào)用動(dòng)態(tài)分配的共享存儲(chǔ)器大小,這些動(dòng)態(tài)分配的存儲(chǔ)器可供聲明為外部數(shù)組(extern __shared__)的其他任何變量使用;Ns是一個(gè)可選參數(shù),默認(rèn)值為0;S為cudaStream_t類(lèi)型,用于設(shè)置與內(nèi)核函數(shù)關(guān)聯(lián)的流.S是一個(gè)可選參數(shù),默認(rèn)值0. */// Note: 核函數(shù)不支持傳入?yún)?shù)為vector的data()指針，需要cudaMalloc和cudaMemcpy，因?yàn)関ector是在主機(jī)內(nèi)存中l(wèi)ayer_prior_vbox << <blocks, threads>> >(dev_dst, layer_width, layer_height, image_width, image_height,offset, step, num_priors, width, dev_vec, dev_vec + vec2.size(), channel_size);/* cudaDeviceSynchronize: kernel的啟動(dòng)是異步的, 為了定位它是否出錯(cuò), 一般需要加上cudaDeviceSynchronize函數(shù)進(jìn)行同步; 將會(huì)一直處于阻塞狀態(tài),直到前面所有請(qǐng)求的任務(wù)已經(jīng)被全部執(zhí)行完畢,如果前面執(zhí)行的某個(gè)任務(wù)失敗,將會(huì)返回一個(gè)錯(cuò)誤；當(dāng)程序中有多個(gè)流,并且流之間在某一點(diǎn)需要通信時(shí),那就必須在這一點(diǎn)處加上同步的語(yǔ)句,即cudaDeviceSynchronize；異步啟動(dòng)reference: https://stackoverflow.com/questions/11888772/when-to-call-cudadevicesynchronize */cudaDeviceSynchronize();TIME_END_GPUcudaMemcpy(dst, dev_dst, length * sizeof(float), cudaMemcpyDeviceToHost);// cudaFree: 釋放設(shè)備上由cudaMalloc函數(shù)分配的內(nèi)存cudaFree(dev_dst);cudaFree(dev_vec);return 0;
}

執(zhí)行結(jié)果如下：

GitHub：? https://github.com/fengbingchun/CUDA_Test

總結(jié)

以上是生活随笔為你收集整理的CUDA Samples: approximate prior vbox layer的全部?jī)?nèi)容，希望文章能夠幫你解決所遇到的問(wèn)題。

如果覺(jué)得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。