我是OpenCL的新手,在C / C ++中的背景非常有限。我得到了这个OpenCL程序,它添加了两个向量,并应该弄清楚它是如何工作的。它来自英特尔:
是否正确地说:每个内核使用A中的1个元素和B中的1个元素来计算Z的1个元素?
对我来说,它看起来像是确定设备的数量(num_devices),实质上是将问题大小(N)除以num_devices,以确定每个设备的元素数(n_per_device [])。然后,它使用n_per_device个元素数为每个设备(input_a []和input_b [])创建随机数数组。
然后这些数组由内核使用,在其中执行整个数组的加法并存储为Z。
例如,假设可用设备数为1000,问题大小(N)为1,000,000; n_per_device为1000(由于没有余数,因此对所有设备都是相同的),它将生成1000个input_a和input_b数组,每个数组包含1000个元素。然后内核分别获取一对1000个元素的数组,并将它们加在一起-换句话说,每次执行内核都会添加1000个元素?
我在追踪什么,还是在这里完全错了?
内核是:
// ACL kernel for adding two input vectors
__kernel void vectorAdd(__global const float *x,
__global const float *y,
__global float *restrict z)
{
// get index of the work item
int index = get_global_id(0);
// add the vector elements
z[index] = x[index] + y[index];
}
主机(主)代码是(很长,很抱歉,不确定什么不重要):
///////////////////////////////////////////////////////////////////////////////////
// This host program executes a vector addition kernel to perform:
// C = A + B
// where A, B and C are vectors with N elements.
//
// This host program supports partitioning the problem across multiple OpenCL
// devices if available. If there are M available devices, the problem is
// divided so that each device operates on N/M points. The host program
// assumes that all devices are of the same type (that is, the same binary can
// be used), but the code can be generalized to support different device types
// easily.
//
// Verification is performed against the same computation on the host CPU.
///////////////////////////////////////////////////////////////////////////////////
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include "CL/opencl.h"
#include "AOCL_Utils.h"
using namespace aocl_utils;
// OpenCL runtime configuration
cl_platform_id platform = NULL;
unsigned num_devices = 0;
scoped_array<cl_device_id> device; // num_devices elements
cl_context context = NULL;
scoped_array<cl_command_queue> queue; // num_devices elements
cl_program program = NULL;
scoped_array<cl_kernel> kernel; // num_devices elements
scoped_array<cl_mem> input_a_buf; // num_devices elements
scoped_array<cl_mem> input_b_buf; // num_devices elements
scoped_array<cl_mem> output_buf; // num_devices elements
// Problem data.
const unsigned N = 1000000; // problem size
scoped_array<scoped_aligned_ptr<float> > input_a, input_b; // num_devices elements
scoped_array<scoped_aligned_ptr<float> > output; // num_devices elements
scoped_array<scoped_array<float> > ref_output; // num_devices elements
scoped_array<unsigned> n_per_device; // num_devices elements
// Function prototypes
float rand_float();
bool init_opencl();
void init_problem();
void run();
void cleanup();
// Entry point.
int main() {
// Initialize OpenCL.
if(!init_opencl()) {
return -1;
}
// Initialize the problem data.
// Requires the number of devices to be known.
init_problem();
// Run the kernel.
run();
// Free the resources allocated
cleanup();
return 0;
}
/////// HELPER FUNCTIONS ///////
// Randomly generate a floating-point number between -10 and 10.
float rand_float() {
return float(rand()) / float(RAND_MAX) * 20.0f - 10.0f;
}
// Initializes the OpenCL objects.
bool init_opencl() {
cl_int status;
printf("Initializing OpenCL\n");
if(!setCwdToExeDir()) {
return false;
}
// Get the OpenCL platform.
platform = findPlatform("Altera");
if(platform == NULL) {
printf("ERROR: Unable to find Altera OpenCL platform.\n");
return false;
}
// Query the available OpenCL device.
device.reset(getDevices(platform, CL_DEVICE_TYPE_ALL, &num_devices));
printf("Platform: %s\n", getPlatformName(platform).c_str());
printf("Using %d device(s)\n", num_devices);
for(unsigned i = 0; i < num_devices; ++i) {
printf(" %s\n", getDeviceName(device[i]).c_str());
}
// Create the context.
context = clCreateContext(NULL, num_devices, device, NULL, NULL, &status);
checkError(status, "Failed to create context");
// Create the program for all device. Use the first device as the
// representative device (assuming all device are of the same type).
std::string binary_file = getBoardBinaryFile("vectorAdd", device[0]);
printf("Using AOCX: %s\n", binary_file.c_str());
program = createProgramFromBinary(context, binary_file.c_str(), device, num_devices);
// Build the program that was just created.
status = clBuildProgram(program, 0, NULL, "", NULL, NULL);
checkError(status, "Failed to build program");
// Create per-device objects.
queue.reset(num_devices);
kernel.reset(num_devices);
n_per_device.reset(num_devices);
input_a_buf.reset(num_devices);
input_b_buf.reset(num_devices);
output_buf.reset(num_devices);
for(unsigned i = 0; i < num_devices; ++i) {
// Command queue.
queue[i] = clCreateCommandQueue(context, device[i], CL_QUEUE_PROFILING_ENABLE, &status);
checkError(status, "Failed to create command queue");
// Kernel.
const char *kernel_name = "vectorAdd";
kernel[i] = clCreateKernel(program, kernel_name, &status);
checkError(status, "Failed to create kernel");
// Determine the number of elements processed by this device.
n_per_device[i] = N / num_devices; // number of elements handled by this device
// Spread out the remainder of the elements over the first
// N % num_devices.
if(i < (N % num_devices)) {
n_per_device[i]++;
}
// Input buffers.
input_a_buf[i] = clCreateBuffer(context, CL_MEM_READ_ONLY,
n_per_device[i] * sizeof(float), NULL, &status);
checkError(status, "Failed to create buffer for input A");
input_b_buf[i] = clCreateBuffer(context, CL_MEM_READ_ONLY,
n_per_device[i] * sizeof(float), NULL, &status);
checkError(status, "Failed to create buffer for input B");
// Output buffer.
output_buf[i] = clCreateBuffer(context, CL_MEM_WRITE_ONLY,
n_per_device[i] * sizeof(float), NULL, &status);
checkError(status, "Failed to create buffer for output");
}
return true;
}
// Initialize the data for the problem. Requires num_devices to be known.
void init_problem() {
if(num_devices == 0) {
checkError(-1, "No devices");
}
input_a.reset(num_devices);
input_b.reset(num_devices);
output.reset(num_devices);
ref_output.reset(num_devices);
// Generate input vectors A and B and the reference output consisting
// of a total of N elements.
// We create separate arrays for each device so that each device has an
// aligned buffer.
for(unsigned i = 0; i < num_devices; ++i) {
input_a[i].reset(n_per_device[i]);
input_b[i].reset(n_per_device[i]);
output[i].reset(n_per_device[i]);
ref_output[i].reset(n_per_device[i]);
for(unsigned j = 0; j < n_per_device[i]; ++j) {
input_a[i][j] = rand_float();
input_b[i][j] = rand_float();
ref_output[i][j] = input_a[i][j] + input_b[i][j];
}
}
}
void run() {
cl_int status;
const double start_time = getCurrentTimestamp();
// Launch the problem for each device.
scoped_array<cl_event> kernel_event(num_devices);
scoped_array<cl_event> finish_event(num_devices);
for(unsigned i = 0; i < num_devices; ++i) {
// Transfer inputs to each device. Each of the host buffers supplied to
// clEnqueueWriteBuffer here is already aligned to ensure that DMA is used
// for the host-to-device transfer.
cl_event write_event[2];
status = clEnqueueWriteBuffer(queue[i], input_a_buf[i], CL_FALSE,
0, n_per_device[i] * sizeof(float), input_a[i], 0, NULL, &write_event[0]);
checkError(status, "Failed to transfer input A");
status = clEnqueueWriteBuffer(queue[i], input_b_buf[i], CL_FALSE,
0, n_per_device[i] * sizeof(float), input_b[i], 0, NULL, &write_event[1]);
checkError(status, "Failed to transfer input B");
// Set kernel arguments.
unsigned argi = 0;
status = clSetKernelArg(kernel[i], argi++, sizeof(cl_mem), &input_a_buf[i]);
checkError(status, "Failed to set argument %d", argi - 1);
status = clSetKernelArg(kernel[i], argi++, sizeof(cl_mem), &input_b_buf[i]);
checkError(status, "Failed to set argument %d", argi - 1);
status = clSetKernelArg(kernel[i], argi++, sizeof(cl_mem), &output_buf[i]);
checkError(status, "Failed to set argument %d", argi - 1);
// Enqueue kernel.
// Use a global work size corresponding to the number of elements to add
// for this device.
//
// We don't specify a local work size and let the runtime choose
// (it'll choose to use one work-group with the same size as the global
// work-size).
//
// Events are used to ensure that the kernel is not launched until
// the writes to the input buffers have completed.
const size_t global_work_size = n_per_device[i];
printf("Launching for device %d (%d elements)\n", i, global_work_size);
status = clEnqueueNDRangeKernel(queue[i], kernel[i], 1, NULL,
&global_work_size, NULL, 2, write_event, &kernel_event[i]);
checkError(status, "Failed to launch kernel");
// Read the result. This the final operation.
status = clEnqueueReadBuffer(queue[i], output_buf[i], CL_FALSE,
0, n_per_device[i] * sizeof(float), output[i], 1, &kernel_event[i], &finish_event[i]);
// Release local events.
clReleaseEvent(write_event[0]);
clReleaseEvent(write_event[1]);
}
// Wait for all devices to finish.
clWaitForEvents(num_devices, finish_event);
const double end_time = getCurrentTimestamp();
// Wall-clock time taken.
printf("\nTime: %0.3f ms\n", (end_time - start_time) * 1e3);
// Get kernel times using the OpenCL event profiling API.
for(unsigned i = 0; i < num_devices; ++i) {
cl_ulong time_ns = getStartEndTime(kernel_event[i]);
printf("Kernel time (device %d): %0.3f ms\n", i, double(time_ns) * 1e-6);
}
// Release all events.
for(unsigned i = 0; i < num_devices; ++i) {
clReleaseEvent(kernel_event[i]);
clReleaseEvent(finish_event[i]);
}
// Verify results.
bool pass = true;
for(unsigned i = 0; i < num_devices && pass; ++i) {
for(unsigned j = 0; j < n_per_device[i] && pass; ++j) {
if(fabsf(output[i][j] - ref_output[i][j]) > 1.0e-5f) {
printf("Failed verification @ device %d, index %d\nOutput: %f\nReference: %f\n",
i, j, output[i][j], ref_output[i][j]);
pass = false;
}
}
}
printf("\nVerification: %s\n", pass ? "PASS" : "FAIL");
}
// Free the resources allocated during initialization
void cleanup() {
for(unsigned i = 0; i < num_devices; ++i) {
if(kernel && kernel[i]) {
clReleaseKernel(kernel[i]);
}
if(queue && queue[i]) {
clReleaseCommandQueue(queue[i]);
}
if(input_a_buf && input_a_buf[i]) {
clReleaseMemObject(input_a_buf[i]);
}
if(input_b_buf && input_b_buf[i]) {
clReleaseMemObject(input_b_buf[i]);
}
if(output_buf && output_buf[i]) {
clReleaseMemObject(output_buf[i]);
}
}
if(program) {
clReleaseProgram(program);
}
if(context) {
clReleaseContext(context);
}
}
这里有几个子问题,所以让我尝试分别解决。我将在术语上有点书呆子;我这样做并不是为了招架,但希望这可以帮助您更加理解文档,示例等。]
是否正确地说:每个内核使用A中的1个元素和B中的1个元素来计算Z的1个元素?
kernel只是将在OpenCL设备上运行的代码。通常,内核计划运行(使用clEnqueueNDRangeKernel()
)具有多个work-items的内核。仅使用一个工作项,根本不用为OpenCL烦恼。性能优势来自大规模并行处理。无论如何,您引用的语句对于处理该内核的每个工作项是正确的。如果使用1000个工作项运行此内核,则将处理来自A的1000个元素和来自B的1000个元素,以计算Z的1000个元素。故意发生的顺序是不确定的,并且至少要同时处理几组元素。
对我来说,它看起来像是确定设备的数量(num_devices),实质上是将问题大小(N)除以num_devices,以确定每个设备的元素数(n_per_device [])。然后,它使用n_per_device个元素数为每个设备(input_a []和input_b [])创建随机数数组。
是的,在我看来也是如此。
例如,假设可用设备数为1000,
我只想指出,您将从不在系统中拥有这么多的OpenCL设备。单个OpenCL设备的粒度通常为“一个GPU”或“系统中的所有CPU内核”或“一个FPGA加速卡”。
因此,台式机系统上的“正常”设备数量为1个,2个或最多约4个(例如CPU + iGPU +双离散GPU)。装有很多加速卡的大铁可能大约有16个左右。如果您试图在台式机(或小型服务器)应用程序中加速某些代码,通常只会选择一种最适合您的问题的设备并运行该设备。对于除最基本的算法以外的任何事物来说,在异构设备之间平均分配工作负载都是一个难题。
且问题大小(N)为1,000,000; n_per_device为1000(由于没有余数,因此对所有设备都是相同的),它将生成1000个input_a和input_b数组,每个数组包含1000个元素。然后内核分别获取一对1000个元素的数组,并将它们加在一起-
是
换句话说,每次内核执行都会添加1000个元素?
同样,在这里使用“内核”一词不够精确。在您的示例中,您将排队1000个工作项以在1000个设备中的每个设备上执行内核。