private static string GetOpenCLSourceHeader(OpenCLNet.Platform platform, OpenCLNet.Device device, HlGraphEntry KernelGraphEntry)
{
StringBuilder result = new System.Text.StringBuilder();
result.AppendLine("// BEGIN GENERATED OpenCL");
setExtensionIfAvailable(result, device, "cl_amd_fp64");
setExtensionIfAvailable(result, device, "cl_khr_fp64");
setExtensionIfAvailable(result, device, "cl_khr_global_int32_base_atomics");
setExtensionIfAvailable(result, device, "cl_khr_global_int32_extended_atomics");
setExtensionIfAvailable(result, device, "cl_khr_local_int32_base_atomics");
setExtensionIfAvailable(result, device, "cl_khr_local_int32_extended_atomics");
if (KernelGraphEntry.HlGraph.RandomStateLocation != null)
{
result.AppendLine();
result.AppendLine("// Source: http://www.doc.ic.ac.uk/~dt10/research/rngs-gpu-mwc64x.html");
result.AppendLine("uint MWC64X(uint2 *state)");
result.AppendLine("{");
result.AppendLine(" enum{ A=4294883355U };");
result.AppendLine(" uint x=(*state).x, c=(*state).y; // Unpack the state");
result.AppendLine(" uint res=x^c; // Calculate the result");
result.AppendLine(" uint hi=mul_hi(x,A); // Step the RNG");
result.AppendLine(" x=x*A+c;");
result.AppendLine(" c=hi+(x<c);");
result.AppendLine(" *state=(uint2)(x,c); // Pack the state back up");
result.AppendLine(" return res; // Return the next result");
result.AppendLine("}");
}
return(result.ToString());
}
private static string GetOpenCLSourceFooter(OpenCLNet.Platform platform, OpenCLNet.Device device)
{
StringBuilder result = new System.Text.StringBuilder();
result.AppendLine();
result.AppendLine("// END GENERATED OpenCL");
return(result.ToString());
}
internal static void CallOpenCLNet(int[] WorkSize, HlGraphEntry CacheEntry, InvokeContext ctx, HighLevel.HlGraph HLgraph, OpenCLNet.Device device)
{
// We can invoke the kernel using the arguments from ctx now :)
if (device == null)
{
device = GetFirstGpu();
if (device == null)
{
device = GetFirstCpu();
}
if (device == null)
{
throw new ArgumentNullException("device");
}
}
OpenCLNet.Platform Platform = device.Platform;
OpenCLNet.Context context;
OpenCLNet.Program program;
lock (CacheEntry)
{
context = CacheEntry.Context;
if (context == null)
{
IntPtr[] properties = new IntPtr[]
{
new IntPtr((long)OpenCLNet.ContextProperties.PLATFORM), Platform.PlatformID,
IntPtr.Zero,
};
context = CacheEntry.Context = Platform.CreateContext(properties, new OpenCLNet.Device[] { device }, null, IntPtr.Zero);
}
program = CacheEntry.Program;
if (program == null)
{
StringBuilder source = new StringBuilder();
source.Append(GetOpenCLSourceHeader(Platform, device, CacheEntry));
source.Append(CacheEntry.Source);
source.Append(GetOpenCLSourceFooter(Platform, device));
program = context.CreateProgramWithSource(source.ToString());
try
{
program.Build();
}
catch (Exception ex)
{
string err = program.GetBuildLog(device);
throw new Exception(err, ex);
}
CacheEntry.Program = program;
}
}
using (CallContext CallContext = new CallContext(context, device, OpenCLNet.CommandQueueProperties.PROFILING_ENABLE, program.CreateKernel(HLgraph.MethodName)))
{
OpenCLNet.CommandQueue commandQueue = CallContext.CommandQueue;
for (int i = 0; i < ctx.Arguments.Count; i++)
{
ctx.Arguments[i].WriteToKernel(CallContext, i);
}
//OpenCLNet.Event StartEvent, EndEvent;
//commandQueue.EnqueueMarker(out StartEvent);
IntPtr[] GlobalWorkSize = new IntPtr[WorkSize.Length];
for (int i = 0; i < WorkSize.Length; i++)
{
GlobalWorkSize[i] = new IntPtr(WorkSize[i]);
}
commandQueue.EnqueueNDRangeKernel(CallContext.Kernel, (uint)GlobalWorkSize.Length, null, GlobalWorkSize, null);
for (int i = 0; i < ctx.Arguments.Count; i++)
{
ctx.Arguments[i].ReadFromKernel(CallContext, i);
}
commandQueue.Finish();
//commandQueue.EnqueueMarker(out EndEvent);
//commandQueue.Finish();
//ulong StartTime, EndTime;
//StartEvent.GetEventProfilingInfo(OpenCLNet.ProfilingInfo.QUEUED, out StartTime);
//EndEvent.GetEventProfilingInfo(OpenCLNet.ProfilingInfo.END, out EndTime);
}
}
