oneAPI简介
Intel 的oneAPI,目的是简化跨CPU、GPU、FPGA、人工智能和其它加速器的各种计算引擎的编程开发。开发人员想达到的目的是一个解决方案,多种架构。2020年5月6月期间发布了Base Kit (Beta)版本。
官网链接
注意可能要Intel账号才能下载。
安装
下载的文件是.exe
结尾的,有图形化安装界面,Next下去即可。最终解压安装完成的效果如图:
2GB左右的安装包,安装完成后大概有14GB的空间占用,如果对于Intel加速库比较熟悉的人应该可以看出MKL,ipp,tbb等库。在此也不得不说Intel的加速库,在正确使用的情况下效果是真的不错。
这里也想说一点,Intel新推出的11代酷睿处理器的核显已经可以赶上入门独显了。而且Intel自家的FPGA产品也是可以支持OpenCL异构计算的。所以Intel迫切的想要退出一种新的解决方案,让开发者不需要过多的了解底层硬件编写语言:比如OpenCL C,Verilog HDL,同样也可以写出高性能代码。
同样新的语言营运而生,DPC++(Data Parallel C++),英特尔在设计DPC++的时候,在语法上和CUDA非常接近,如果程序员对于CUDA非常熟悉的话,那么使用DPC++进行编程应该没有任何问题。本质上还是有C/C++语言基础,看懂代码应该没太大难度。
测试
Intel 给了oneAPI的编程指导,只不过现在还没中文版:
官网指导
你也可以根据自己选的Toolkit和语言,在左侧栏Document出搜寻对应的文档。
根据编程指导里的介绍,Intel把所有的sample code放到了github上:
Github
这里使用的是DPC++ compiler下的vector add,放上代码:
dpc_common.hpp
//==============================================================
// Copyright © 2020 Intel Corporation
//
// SPDX-License-Identifier: MIT
// =============================================================
#ifndef _DP_HPP
#define _DP_HPP
#pragma once
#include
#include
#include
namespace dpc {
// this exception handler with catch async exceptions
static auto exception_handler = [](cl::sycl::exception_list eList) {
for (std::exception_ptr const &e : eList) {
try {
std::rethrow_exception(e);
} catch (std::exception const &e) {
#if _DEBUG
std::cout << "Failure" << std::endl;
#endif
std::terminate();
}
}
};
class queue : public cl::sycl::queue {
// Enable profiling by default
cl::sycl::property_list prop_list =
cl::sycl::property_list{cl::sycl::property::queue::enable_profiling()};
public:
queue()
: cl::sycl::queue(cl::sycl::default_selector{}, exception_handler, prop_list) {}
queue(cl::sycl::device_selector &d)
: cl::sycl::queue(d, exception_handler, prop_list) {}
queue(cl::sycl::device_selector &d, cl::sycl::property_list &p)
: cl::sycl::queue(d, exception_handler, p) {}
};
using Duration = std::chrono::duration;
class Timer {
public:
Timer() : start(std::chrono::steady_clock::now()) {}
Duration elapsed() {
auto now = std::chrono::steady_clock::now();
return std::chrono::duration_cast(now - start);
}
private:
std::chrono::steady_clock::time_point start;
};
}; // namespace dpc
#endif
vector-add-buffers.cpp
//==============================================================
// Vector Add is the equivalent of a Hello, World! sample for data parallel
// programs. Building and running the sample verifies that your development
// environment is setup correctly and demonstrates the use of the core features
// of DPC++. This sample runs on both CPU and GPU (or FPGA). When run, it
// computes on both the CPU and offload device, then compares results. If the
// code executes on both CPU and offload device, the device name and a success
// message are displayed. And, your development environment is setup correctly!
//
// For comprehensive instructions regarding DPC++ Programming, go to
// https://software.intel.com/en-us/oneapi-programming-guide and search based on
// relevant terms noted in the comments.
//
// DPC++ material used in the code sample:
// • A one dimensional array of data.
// • A device queue, buffer, accessor, and kernel.
//==============================================================
// Copyright © 2020 Intel Corporation
//
// SPDX-License-Identifier: MIT
// =============================================================
#include
#include
#include
#include "dpc_common.hpp"
#if FPGA || FPGA_EMULATOR
#include
#endif
using namespace sycl;
// Array type and data size for this example.
constexpr size_t array_size = 10000;
typedef std::array IntArray;
//************************************
// Vector add in DPC++ on device: returns sum in 4th parameter "sum_parallel".
//************************************
void VectorAdd(queue &q, const IntArray &a_array, const IntArray &b_array,
IntArray &sum_parallel) {
// Create the range object for the arrays managed by the buffer.
range<1> num_items{a_array.size()};
// Create buffers that hold the data shared between the host and the devices.
// The buffer destructor is responsible to copy the data back to host when it
// goes out of scope.
buffer a_buf(a_array);
buffer b_buf(b_array);
buffer sum_buf(sum_parallel.data(), num_items);
// Submit a command group to the queue by a lambda function that contains the
// data access permission and device computation (kernel).
q.submit([&](handler &h) {
// Create an accessor for each buffer with access permission: read, write or
// read/write. The accessor is a mean to access the memory in the buffer.
auto a = a_buf.get_access(h);
auto b = b_buf.get_access(h);
// The sum_accessor is used to store (with write permission) the sum data.
auto sum = sum_buf.get_access(h);
// Use parallel_for to run vector addition in parallel on device. This
// executes the kernel.
// 1st parameter is the number of work items.
// 2nd parameter is the kernel, a lambda that specifies what to do per
// work item. The parameter of the lambda is the work item id.
// DPC++ supports unnamed lambda kernel by default.
h.parallel_for(num_items, [=](id<1> i) { sum[i] = a[i] + b[i]; });
});
}
//************************************
// Initialize the array from 0 to array_size - 1
//************************************
void InitializeArray(IntArray &a) {
for (size_t i = 0; i < a.size(); i++) a[i] = i;
}
//************************************
// Demonstrate vector add both in sequential on CPU and in parallel on device.
//************************************
int main() {
// Create device selector for the device of your interest.
#if FPGA_EMULATOR
// DPC++ extension: FPGA emulator selector on systems without FPGA card.
intel::fpga_emulator_selector d_selector;
#elif FPGA
// DPC++ extension: FPGA selector on systems with FPGA card.
intel::fpga_selector d_selector;
#else
// The default device selector will select the most performant device.
default_selector d_selector;
#endif
// Create array objects with "array_size" to store the input and output data.
IntArray a, b, sum_sequential, sum_parallel;
// Initialize input arrays with values from 0 to array_size - 1
InitializeArray(a);
InitializeArray(b);
try {
queue q(d_selector, dpc::exception_handler);
// Print out the device information used for the kernel code.
std::cout << "Running on device: "
<< q.get_device().get_info() << "\n";
std::cout << "Vector size: " << a.size() << "\n";
// Vector addition in DPC++
VectorAdd(q, a, b, sum_parallel);
} catch (exception const &e) {
std::cout << "An exception is caught for vector add.\n";
std::terminate();
}
// Compute the sum of two arrays in sequential for validation.
for (size_t i = 0; i < sum_sequential.size(); i++)
sum_sequential[i] = a[i] + b[i];
// Verify that the two arrays are equal.
for (size_t i = 0; i < sum_sequential.size(); i++) {
if (sum_parallel[i] != sum_sequential[i]) {
std::cout << "Vector add failed on device.\n";
return -1;
}
}
int indices[]{0, 1, 2, (a.size() - 1)};
constexpr size_t indices_size = sizeof(indices) / sizeof(int);
// Print out the result of vector add.
for (int i = 0; i < indices_size; i++) {
int j = indices[i];
if (i == indices_size - 1) std::cout << "...\n";
std::cout << "[" << j << "]: " << a[j] << " + " << b[j] << " = "
<< sum_parallel[j] << "\n";
}
std::cout << "Vector add successfully completed on device.\n";
return 0;
}
代码有了,就可以编译看看效果;在这里Intel提供的是非常完整的工具包,所以编译器调试器一应俱全。但是有一点就是没有环境变量,是无法使用这些工具的。Intel提供了环境变量终端,安装完成oneAPI后会提供一个终端:
如上图所示,打开这个终端,它会自动加载环境变量。
而后我们需要做的是切换到源文件目录,手动编译即可。
参考编译命令:
dpcpp -O2 -g -std=C++17 -o vector-add-buffers.exe src/vector-add-buffers.cpp
没有报错即编译成功,编译成功后,产生的文件如下图:
最后,输入.\vector-add-buffers.exe
来运行查看结果:
可以看到用的GPU,两个拥有10000个元素的一维向量相加很快便执行出来了。有多快?就真的和你打印Hello World差不多快,所以源文件开始的说明里面也说了,两个元素个数相同的一维向量的相加,便是并行处理的Hello World。
注意:这个编译过程切不可放进普通终端(cmd/powershell)中进行,因为环境变量的缘故。
针对单个源文件,直接采用命令编译毫无问题,但是若是有很多个cpp和终端设备的工程文件,最好考虑Visual Studio这样的IDE,或者是CMake生成Makefile(比较推荐,跨平台工程常用)。在Github的sample code里面,Intel也给了vs的工程文件,可以直接clone整个工程,打开就可以用。