static void AllocateHostMemory(bool bPinGenericMemory, ref int[] pp_a, ref CudaPageLockedHostMemory<int> pp_Aligned_a, int nbytes)
{
Console.Write("> cudaMallocHost() allocating {0:0.00} Mbytes of system memory\n", (float)nbytes / 1048576.0f);
// allocate host memory (pinned is required for achieve asynchronicity)
if (pp_Aligned_a != null)
pp_Aligned_a.Dispose();
pp_Aligned_a = new CudaPageLockedHostMemory<int>(nbytes / sizeof(int));
pp_a = new int[nbytes / sizeof(int)];
}
static void AllocateHostMemory(bool bPinGenericMemory, ref int[] pp_a, ref CudaPageLockedHostMemory <int> pp_Aligned_a, int nbytes)
{
Console.Write("> cudaMallocHost() allocating {0:0.00} Mbytes of system memory\n", (float)nbytes / 1048576.0f);
// allocate host memory (pinned is required for achieve asynchronicity)
if (pp_Aligned_a != null)
{
pp_Aligned_a.Dispose();
}
pp_Aligned_a = new CudaPageLockedHostMemory <int>(nbytes / sizeof(int));
pp_a = new int[nbytes / sizeof(int)];
}
static void Main(string[] args)
{
int cuda_device = 0;
int nstreams = 4; // number of streams for CUDA calls
int nreps = 10; // number of times each experiment is repeated
int n = 16 * 1024 * 1024; // number of ints in the data set
int nbytes = n * sizeof(int); // number of data bytes
dim3 threads, blocks; // kernel launch configuration
float elapsed_time, time_memcpy, time_kernel; // timing variables
float scale_factor = 1.0f;
// allocate generic memory and pin it laster instead of using cudaHostAlloc()
// Untested in C#, so stick to cudaHostAlloc().
bool bPinGenericMemory = false; // we want this to be the default behavior
CUCtxFlags device_sync_method = CUCtxFlags.BlockingSync; // by default we use BlockingSync
int niterations; // number of iterations for the loop inside the kernel
ShrQATest.shrQAStart(args);
Console.WriteLine("[ simpleStreams ]");
foreach (var item in args)
{
if (item.Contains("help"))
{
printHelp();
ShrQATest.shrQAFinishExit(args, ShrQATest.eQAstatus.QA_PASSED);
}
}
bPinGenericMemory = false;
foreach (var item in args)
{
if (item.Contains("use_generic_memory"))
{
bPinGenericMemory = true;
}
}
for (int i = 0; i < args.Length; i++)
{
if (args[i].Contains("sync_method"))
{
int temp = -1;
bool error = false;
if (i < args.Length - 1)
{
error = int.TryParse(args[i + 1], out temp);
switch (temp)
{
case 0:
device_sync_method = CUCtxFlags.SchedAuto;
break;
case 1:
device_sync_method = CUCtxFlags.SchedSpin;
break;
case 2:
device_sync_method = CUCtxFlags.SchedYield;
break;
case 4:
device_sync_method = CUCtxFlags.BlockingSync;
break;
default:
error = true;
break;
}
}
if (!error)
{
Console.Write("Specifying device_sync_method = {0}, setting reps to 100 to demonstrate steady state\n", sDeviceSyncMethod[(int)device_sync_method]);
nreps = 100;
}
else
{
Console.Write("Invalid command line option sync_method=\"{0}\"\n", temp);
ShrQATest.shrQAFinishExit(args, ShrQATest.eQAstatus.QA_FAILED);
}
}
}
int num_devices = CudaContext.GetDeviceCount();
if (0 == num_devices)
{
Console.Write("your system does not have a CUDA capable device, waiving test...\n");
ShrQATest.shrQAFinishExit(args, ShrQATest.eQAstatus.QA_FAILED);
}
cuda_device = CudaContext.GetMaxGflopsDeviceId();
CudaDeviceProperties deviceProp = CudaContext.GetDeviceInfo(cuda_device);
if ((1 == deviceProp.ComputeCapability.Major) && (deviceProp.ComputeCapability.Minor < 1))
{
Console.Write("{0} does not have Compute Capability 1.1 or newer. Reducing workload.\n", deviceProp.DeviceName);
}
if (deviceProp.ComputeCapability.Major >= 2)
{
niterations = 100;
}
else
{
if (deviceProp.ComputeCapability.Minor > 1)
{
niterations = 5;
}
else
{
niterations = 1; // reduced workload for compute capability 1.0 and 1.1
}
}
// Check if GPU can map host memory (Generic Method), if not then we override bPinGenericMemory to be false
// In .net we cannot allocate easily generic aligned memory, so <bPinGenericMemory> is always false in our case...
if (bPinGenericMemory)
{
Console.Write("Device: <{0}> canMapHostMemory: {1}\n", deviceProp.DeviceName, deviceProp.CanMapHostMemory ? "Yes" : "No");
if (deviceProp.CanMapHostMemory == false)
{
Console.Write("Using cudaMallocHost, CUDA device does not support mapping of generic host memory\n");
bPinGenericMemory = false;
}
}
// Anything that is less than 32 Cores will have scaled down workload
scale_factor = Math.Max((32.0f / (ConvertSMVer2Cores(deviceProp.ComputeCapability.Major, deviceProp.ComputeCapability.Minor) * (float)deviceProp.MultiProcessorCount)), 1.0f);
n = (int)Math.Round((float)n / scale_factor);
Console.Write("> CUDA Capable: SM {0}.{1} hardware\n", deviceProp.ComputeCapability.Major, deviceProp.ComputeCapability.Minor);
Console.Write("> {0} Multiprocessor(s) x {1} (Cores/Multiprocessor) = {2} (Cores)\n",
deviceProp.MultiProcessorCount,
ConvertSMVer2Cores(deviceProp.ComputeCapability.Major, deviceProp.ComputeCapability.Minor),
ConvertSMVer2Cores(deviceProp.ComputeCapability.Major, deviceProp.ComputeCapability.Minor) * deviceProp.MultiProcessorCount);
Console.Write("> scale_factor = {0:0.0000}\n", 1.0f / scale_factor);
Console.Write("> array_size = {0}\n\n", n);
// enable use of blocking sync, to reduce CPU usage
Console.Write("> Using CPU/GPU Device Synchronization method ({0})\n", sDeviceSyncMethod[(int)device_sync_method]);
CudaContext ctx;
if (bPinGenericMemory)
{
ctx = new CudaContext(cuda_device, device_sync_method | CUCtxFlags.MapHost);
}
else
{
ctx = new CudaContext(cuda_device, device_sync_method);
}
//Load Kernel image from resources
string resName;
if (IntPtr.Size == 8)
{
resName = "simpleStreams_x64.ptx";
}
else
{
resName = "simpleStreams.ptx";
}
string resNamespace = "simpleStreams";
string resource = resNamespace + "." + resName;
Stream stream = Assembly.GetExecutingAssembly().GetManifestResourceStream(resource);
if (stream == null)
{
throw new ArgumentException("Kernel not found in resources.");
}
CudaKernel init_array = ctx.LoadKernelPTX(stream, "init_array");
// allocate host memory
int c = 5; // value to which the array will be initialized
int[] h_a = null; // pointer to the array data in host memory
CudaPageLockedHostMemory <int> hAligned_a = null; // pointer to the array data in host memory (aligned to MEMORY_ALIGNMENT)
//Note: In .net we have two seperated arrays: One is in managed memory (h_a), the other one in unmanaged memory (hAligned_a).
//In C++ hAligned_a would point somewhere inside the h_a array.
AllocateHostMemory(bPinGenericMemory, ref h_a, ref hAligned_a, nbytes);
Console.Write("\nStarting Test\n");
// allocate device memory
CudaDeviceVariable <int> d_c = c; //using new implicit cast to allocate memory and asign value
CudaDeviceVariable <int> d_a = new CudaDeviceVariable <int>(nbytes / sizeof(int));
CudaStream[] streams = new CudaStream[nstreams];
for (int i = 0; i < nstreams; i++)
{
streams[i] = new CudaStream();
}
// create CUDA event handles
// use blocking sync
CudaEvent start_event, stop_event;
CUEventFlags eventflags = ((device_sync_method == CUCtxFlags.BlockingSync) ? CUEventFlags.BlockingSync : CUEventFlags.Default);
start_event = new CudaEvent(eventflags);
stop_event = new CudaEvent(eventflags);
// time memcopy from device
start_event.Record(); // record in stream-0, to ensure that all previous CUDA calls have completed
hAligned_a.AsyncCopyToDevice(d_a, streams[0].Stream);
stop_event.Record();
stop_event.Synchronize(); // block until the event is actually recorded
time_memcpy = CudaEvent.ElapsedTime(start_event, stop_event);
Console.Write("memcopy:\t{0:0.00}\n", time_memcpy);
// time kernel
threads = new dim3(512, 1);
blocks = new dim3(n / (int)threads.x, 1);
start_event.Record();
init_array.BlockDimensions = threads;
init_array.GridDimensions = blocks;
init_array.RunAsync(streams[0].Stream, d_a.DevicePointer, d_c.DevicePointer, niterations);
stop_event.Record();
stop_event.Synchronize();
time_kernel = CudaEvent.ElapsedTime(start_event, stop_event);
Console.Write("kernel:\t\t{0:0.00}\n", time_kernel);
//////////////////////////////////////////////////////////////////////
// time non-streamed execution for reference
threads = new dim3(512, 1);
blocks = new dim3(n / (int)threads.x, 1);
start_event.Record();
for (int k = 0; k < nreps; k++)
{
init_array.BlockDimensions = threads;
init_array.GridDimensions = blocks;
init_array.Run(d_a.DevicePointer, d_c.DevicePointer, niterations);
hAligned_a.SynchronCopyToHost(d_a);
}
stop_event.Record();
stop_event.Synchronize();
elapsed_time = CudaEvent.ElapsedTime(start_event, stop_event);
Console.Write("non-streamed:\t{0:0.00} ({1:00} expected)\n", elapsed_time / nreps, time_kernel + time_memcpy);
//////////////////////////////////////////////////////////////////////
// time execution with nstreams streams
threads = new dim3(512, 1);
blocks = new dim3(n / (int)(nstreams * threads.x), 1);
byte[] memset = new byte[nbytes]; // set host memory bits to all 1s, for testing correctness
for (int i = 0; i < nbytes; i++)
{
memset[i] = 255;
}
System.Runtime.InteropServices.Marshal.Copy(memset, 0, hAligned_a.PinnedHostPointer, nbytes);
d_a.Memset(0); // set device memory to all 0s, for testing correctness
start_event.Record();
for (int k = 0; k < nreps; k++)
{
init_array.BlockDimensions = threads;
init_array.GridDimensions = blocks;
// asynchronously launch nstreams kernels, each operating on its own portion of data
for (int i = 0; i < nstreams; i++)
{
init_array.RunAsync(streams[i].Stream, d_a.DevicePointer + i * n / nstreams * sizeof(int), d_c.DevicePointer, niterations);
}
// asynchronously launch nstreams memcopies. Note that memcopy in stream x will only
// commence executing when all previous CUDA calls in stream x have completed
for (int i = 0; i < nstreams; i++)
{
hAligned_a.AsyncCopyFromDevice(d_a, i * n / nstreams * sizeof(int), i * n / nstreams * sizeof(int), nbytes / nstreams, streams[i].Stream);
}
}
stop_event.Record();
stop_event.Synchronize();
elapsed_time = CudaEvent.ElapsedTime(start_event, stop_event);
Console.Write("{0} streams:\t{1:0.00} ({2:0.00} expected with compute capability 1.1 or later)\n", nstreams, elapsed_time / nreps, time_kernel + time_memcpy / nstreams);
// check whether the output is correct
Console.Write("-------------------------------\n");
//We can directly access data in hAligned_a using the [] operator, but copying
//data first to h_a is faster.
System.Runtime.InteropServices.Marshal.Copy(hAligned_a.PinnedHostPointer, h_a, 0, nbytes / sizeof(int));
bool bResults = correct_data(h_a, n, c * nreps * niterations);
// release resources
for (int i = 0; i < nstreams; i++)
{
streams[i].Dispose();
}
start_event.Dispose();
stop_event.Dispose();
hAligned_a.Dispose();
d_a.Dispose();
d_c.Dispose();
CudaContext.ProfilerStop();
ctx.Dispose();
Console.ReadKey();
ShrQATest.shrQAFinishExit(args, bResults ? ShrQATest.eQAstatus.QA_PASSED : ShrQATest.eQAstatus.QA_FAILED);
}
