Calculating GPU's maximum flops using OpenCL - performance

Question

I am writing a simple OpenCL application, which is going to calculate the maximum experiment FLOPS of a target GPU device. I have decided to keep my cl kernel as simple as possible. Here are my OpenCL kernel and my host code. Kernel code is:
__kernel void flops(__global float *data) {
int gid = get_global_id(0);
double s = data[gid];
data[gid] = s * 0.35;
}
And the host code is:
#include <iostream>
#include <sstream>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include "support.h"
#include "Event.h"
#include "ResultDatabase.h"
#include "OptionParser.h"
#include "ProgressBar.h"
using namespace std;
std::string kernels_folder = "/home/users/saman/shoc/src/opencl/level3/FlopsFolder/";
std::string kernel_file = "flops.cl";
static const char *opts = "-cl-mad-enable -cl-no-signed-zeros "
"-cl-unsafe-math-optimizations -cl-finite-math-only";
cl_program createProgram (cl_context context,
cl_device_id device,
const char* fileName) {
cl_int errNum;
cl_program program;
std::ifstream kernelFile (fileName, std::ios::in);
if (!kernelFile.is_open()) {
std::cerr << "Failed to open file for reading: " << fileName << std::endl;
}
std::ostringstream oss;
oss << kernelFile.rdbuf();
std::string srcStdStr = oss.str();
const char *srcStr = srcStdStr.c_str();
program = clCreateProgramWithSource (context, 1, (const char **)&srcStr,
NULL, &errNum);
CL_CHECK_ERROR(errNum);
errNum = clBuildProgram (program, 0, NULL, NULL, NULL, NULL);
CL_CHECK_ERROR (errNum);
return program;
}
bool createMemObjects (cl_context context, cl_command_queue queue,
cl_mem* memObject,
const int memFloatsSize, float *a) {
cl_int err;
*memObject = clCreateBuffer (context, CL_MEM_READ_WRITE,
memFloatsSize * sizeof(float), NULL, &err);
CL_CHECK_ERROR(err);
if (*memObject == NULL) {
std::cerr << "Error creating memory objects. " << std::endl;
return false;
}
Event evWrite("write");
err = clEnqueueWriteBuffer (queue, *memObject, CL_FALSE, 0, memFloatsSize * sizeof(float),
a, 0, NULL, &evWrite.CLEvent());
CL_CHECK_ERROR(err);
err = clWaitForEvents (1, &evWrite.CLEvent());
CL_CHECK_ERROR(err);
return true;
}
void cleanup (cl_context context, cl_command_queue commandQueue,
cl_program program, cl_kernel kernel, cl_mem memObject) {
if (memObject != NULL)
clReleaseMemObject (memObject);
if (kernel != NULL)
clReleaseKernel (kernel);
if (program != NULL)
clReleaseProgram (program);
}
void addBenchmarkSpecOptions(OptionParser &op) {
}
void RunBenchmark(cl_device_id id,
cl_context ctx,
cl_command_queue queue,
ResultDatabase &resultDB,
OptionParser &op)
{
for (float i = 0.1; i <= 0.2; i+=0.1 ) {
std::cout << "Deploying " << 100*i << "%" << std::endl;
bool verbose = false;
cl_int errNum;
cl_program program = 0;
cl_kernel kernel;
cl_mem memObject = 0;
char maxFloatsStr[128];
char testStr[128];
program = createProgram (ctx, id, (kernels_folder + kernel_file).c_str());
if (program == NULL) {
exit (0);
}
if (verbose) std::cout << "Program created successfully!" << std::endl;
kernel = clCreateKernel (program, "flops", &errNum);
CL_CHECK_ERROR(errNum);
if (verbose) std::cout << "Kernel created successfully!" << std::endl;
// Identify maximum size of the global memory on the device side
cl_long maxAllocSizeBytes = 0;
cl_long maxComputeUnits = 0;
cl_long maxWorkGroupSize = 0;
clGetDeviceInfo (id, CL_DEVICE_MAX_MEM_ALLOC_SIZE,
sizeof(cl_long), &maxAllocSizeBytes, NULL);
clGetDeviceInfo (id, CL_DEVICE_MAX_COMPUTE_UNITS,
sizeof(cl_long), &maxComputeUnits, NULL);
clGetDeviceInfo (id, CL_DEVICE_MAX_WORK_GROUP_SIZE,
sizeof(cl_long), &maxWorkGroupSize, NULL);
// Let's use 80% of this memory for transferring data
cl_long maxFloatsUsageSize = ((maxAllocSizeBytes / 4) * 0.8);
if (verbose) std::cout << "Max floats usage size is " << maxFloatsUsageSize << std::endl;
if (verbose) std::cout << "Max compute unit is " << maxComputeUnits << std::endl;
if (verbose) std::cout << "Max Work Group size is " << maxWorkGroupSize << std::endl;
// Prepare buffer on the host side
float *a = new float[maxFloatsUsageSize];
for (int j = 0; j < maxFloatsUsageSize; j++) {
a[j] = (float) (j % 77);
}
if (verbose) std::cout << "Host buffer been prepared!" << std::endl;
// Creating buffer on the device side
if (!createMemObjects(ctx, queue, &memObject, maxFloatsUsageSize, a)) {
exit (0);
}
errNum = clSetKernelArg (kernel, 0, sizeof(cl_mem), &memObject);
CL_CHECK_ERROR(errNum);
size_t wg_size, wg_multiple;
cl_ulong local_mem, private_usage, local_usage;
errNum = clGetKernelWorkGroupInfo (kernel, id,
CL_KERNEL_WORK_GROUP_SIZE,
sizeof (wg_size), &wg_size, NULL);
CL_CHECK_ERROR (errNum);
errNum = clGetKernelWorkGroupInfo (kernel, id,
CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE,
sizeof (wg_multiple), &wg_multiple, NULL);
CL_CHECK_ERROR (errNum);
errNum = clGetKernelWorkGroupInfo (kernel, id,
CL_KERNEL_LOCAL_MEM_SIZE,
sizeof (local_usage), &local_usage, NULL);
CL_CHECK_ERROR (errNum);
errNum = clGetKernelWorkGroupInfo (kernel, id,
CL_KERNEL_PRIVATE_MEM_SIZE,
sizeof (private_usage), &private_usage, NULL);
CL_CHECK_ERROR (errNum);
if (verbose) std::cout << "Work Group size is " << wg_size << std::endl;
if (verbose) std::cout << "Preferred Work Group size is " << wg_multiple << std::endl;
if (verbose) std::cout << "Local memory size is " << local_usage << std::endl;
if (verbose) std::cout << "Private memory size is " << private_usage << std::endl;
size_t globalWorkSize[1] = {maxFloatsUsageSize};
size_t localWorkSize[1] = {1};
Event evKernel("flops");
errNum = clEnqueueNDRangeKernel (queue, kernel, 1, NULL,
globalWorkSize, localWorkSize,
0, NULL, &evKernel.CLEvent());
CL_CHECK_ERROR (errNum);
if (verbose) cout << "Waiting for execution to finish ";
errNum = clWaitForEvents(1, &evKernel.CLEvent());
CL_CHECK_ERROR(errNum);
evKernel.FillTimingInfo();
if (verbose) cout << "Kernel execution terminated successfully!" << std::endl;
delete[] a;
sprintf (maxFloatsStr, "Size: %d", maxFloatsUsageSize);
sprintf (testStr, "Flops: %f\% Memory", 100*i);
double flopCount = maxFloatsUsageSize * 16000;
double gflop = flopCount / (double)(evKernel.SubmitEndRuntime());
resultDB.AddResult (testStr, maxFloatsStr, "GFLOPS", gflop);
// Now it's time to read back the data
a = new float[maxFloatsUsageSize];
errNum = clEnqueueReadBuffer(queue, memObject, CL_TRUE, 0, maxFloatsUsageSize*sizeof(float), a, 0, NULL, NULL);
CL_CHECK_ERROR(errNum);
if (verbose) {
for (int j = 0; j < 10; j++) {
std::cout << a[j] << " ";
}
}
delete[] a;
if (memObject != NULL)
clReleaseMemObject (memObject);
if (program != NULL)
clReleaseProgram (program);
if (kernel != NULL)
clReleaseKernel (kernel);
}
std::cout << "Program executed successfully!" << std::endl;
}
Explaining the code, in the kernel code I actually do a single floating point operation, which means every single task will do on FOPS. In the host code, I first retrieve the maximum global memory size of the GPU, allocate portion of it (for loop define how much of it), then push the data and kernel execution into it. I will measure the execution time of clEnqueueNDRangeKernel and then calculate the GFLOPS of application. In my current implementation, no matter what is the size of cl_mem, I get around 0.28 GFLOPS of performance, which is much less than the advertised power. I assume I do specific things inefficiently here. Or in general my method for calculating the GPU performance is not right. Does anyone can tell my what kind of changes should I make into the code?

Answer 1

With local group size of 1, you are wasting 31/32 of the resources (thus you can have 1/32 of the peak performance at most). You need local group size of at least 32 (and is multiple of 32) to fully utilize computation resources and 64 to achieve 100% occupancy (100% occupancy is not necessary though).
Memory access has high latency and low bandwidth. Your kernel will always be waiting for memory controllers if other things are right. You need do more arithmetic operations to make the ALU's busy.
You need read the document first and make use of the Visual Profiler. In the previous two parts I just want to tell that things are stranger than you thought. But more strange things are waiting.
You can achieve peak performance eaily on CPU with assembly language (By doing only independent arithmetic operations. If you write such code in C it will simply be dropped by the compiler). NVidia only provides us an IL interface called PTX, and I'm not sure if compiler will optimize it. And you can only use PTX in CUDA I think.
edit: It seems that compiler will optimize unused PTX code away, at least in inline assembers.