public void Wave_Cpu_2x()
{
var kernel = new CpuConfig{Cores = 2}.Configure<WaveKernel>();
var empty = new Matrix<Cell>(202, 202);
var after2s = kernel.Execute(empty);
// todo. validate the result
}
private JitCompiler(CpuConfig config, Type t_kernel, TypeBuilder t_xformed)
{
_config = config;
_t_kernel = t_kernel;
_m_kernel = _t_kernel.GetMethod("RunKernel", BF.All);
ValidateInputParameters();
// todo. think how to support multiple methods: *Dims, *Idxes, synchronization points
// todo. I'd even say that we should embed constants instead of Dims and cache such bodies between calls
// todo. if some dimension is equal to 1, then strip off corresponding loop
var lam = _m_kernel.Decompile();
_hir = lam.Body;
_params = lam.Sig.Syms;
_locals = lam.Body.LocalsRecursive();
_xhir = new Block();
_tids.Add("x", new Local("tid_x", typeof(int)));
_tids.Add("y", new Local("tid_y", typeof(int)));
_tids.Add("z", new Local("tid_z", typeof(int)));
_xhir.Locals.AddElements(_tids.Values);
InferLocalAllocationHints();
ReplicatePrivatelyAllocatedLocals();
LiftLocallyAllocatedLocals();
HoistGloballyAllocatedLocals();
_callsToSyncThreads = _hir.Family().OfType<Eval>().Where(eval =>
{
var m1 = eval.InvokedMethod();
var syncapi_syncThreads = typeof(ISyncApi).GetMethod("SyncThreads");
return m1.Api() == syncapi_syncThreads;
}).ToReadOnly();
TransformBlock(_hir, _xhir, false);
// todo. currently we don't support debuggable IL
// the latter implies correct PDB-mappings and working watch for replicated vars
_config.EmitDebuggableIL.AssertFalse();
_t_xformed = t_xformed;
_t_xformed.AddInterfaceImplementation(typeof(IBlockRunner));
_m_xformed = _t_xformed.DefineMethod("RunBlock", MA.Public, typeof(void), typeof(int3).MkArray());
_m_xformed.DefineParameter(1, ParmA.None, "blockIdx");
CompileTransformedHir();
}
public static void DoCrosscompile(CpuConfig config, Type t_kernel, TypeBuilder t_xformed) { new JitCompiler(config, t_kernel, t_xformed); }
protected CpuConfig(CpuConfig proto) : base(proto)
{
}
protected CpuConfig(CpuConfig proto) : base(proto) {}
public void BigTest_Cpu_2x()
{
var cfg = new CpuConfig{Cores = 2};
1.TimesDo(() => BigTest(cfg));
}
public void MediumTest_Cpu_1x()
{
var cfg = new CpuConfig{Cores = 1};
1.TimesDo(() => MediumTest(cfg));
}
public void SmallTest_Cpu_1x()
{
var cfg = new CpuConfig{Cores = 1};
1.TimesDo(() => SmallTest(cfg));
}
