protected override void CoreRunKernel(IGrid grid, IKernel kernel)
{
// todo. so far we have just this totally uninformative flag
// 1) this solution is better than receiving AppDomain.UnhandledException
// since this way we ensure that we track only threads that we create
// 2) further implementations might include exception unwrapping and multiplexing
// just like TPL/PLINQ does: (soz, can't find the link about AggregatedException)
var crashCount = 0;
var gridDims = new [] { grid.GridDim.Z, grid.GridDim.Y, grid.GridDim.X };
var numBlocks = gridDims.Product();
var workers = 0.UpTo(Config.Cores - 1).Select(i => new Thread(() =>
{
var chunkSize = (int)Math.Ceiling(numBlocks * 1.0 / Config.Cores);
var start = i * chunkSize;
var end = Math.Min((i + 1) * chunkSize, (int)numBlocks) - 1;
start.UpTo(end).ForEach(j =>
{
var dimSizes = gridDims.Scanrae(1, (curr, dim, _) => curr * dim).ToReadOnly();
var indices = dimSizes.SkipLast(1).Scanrbi(j, (curr, dimSize, _) => curr % dimSize, (curr, dimSize, _) => curr / dimSize, (curr, _) => curr).ToReadOnly();
var blid = new int3(indices[2], indices[1], indices[0]);
try
{
var blockRunner = kernel.AssertCast <IBlockRunner>();
blockRunner.RunBlock(blid);
}
catch (Exception ex)
{
Interlocked.Increment(ref crashCount);
throw new KernelThreadException(kernel, GridDim, blid, BlockDim, null, Thread.CurrentThread.Name, ex);
}
});
// todo. also mess with thread affinities to ensure maximally possible CPU load
})
{
Name = "Conflux CPU runtime worker thread #" + i
}).ToReadOnly();
workers.ForEach(w => w.Start());
workers.ForEach(w => w.Join());
(crashCount == 0).AssertTrue();
}