public static void Invoke(TSCudaContext context, CudaContext cudaContext, byte[] ptx, string baseName, params object[] args)
{
ThrowIfAnyTensorInvalid(args);
cudaContext.SetCurrent();
CudaDeviceProperties deviceInfo = context.DeviceInfoForContext(cudaContext);
IEnumerable <Tensor> allTensors = args.OfType <Tensor>();
Tensor firstTensor = allTensors.First();
long elementCount = firstTensor.ElementCount();
ApplySpecialization spec = new ApplySpecialization(allTensors.ToArray());
ConvertTensorArgs.Convert(cudaContext, spec.Use32BitIndices, args);
ManagedCuda.VectorTypes.dim3 block = ApplyUtils.GetApplyBlock();
ManagedCuda.VectorTypes.dim3 grid = ApplyUtils.GetApplyGrid(deviceInfo, elementCount);
string fullKernelName = PermutationGenerator.GetMangledName(baseName, spec);
CudaKernel kernel = context.KernelCache.Get(cudaContext, ptx, fullKernelName);
kernel.GridDimensions = grid;
kernel.BlockDimensions = block;
kernel.RunAsync(CUstream.NullStream, args);
}
public static void InvokeReduceAll(TSCudaContext context, CudaContext cudaContext, byte[] ptx, string kernelName, dim3 grid, dim3 block, uint smemSize, ApplySpecialization spec, params object[] args)
{
ConvertTensorArgs.Convert(cudaContext, spec.Use32BitIndices, args);
CudaKernel kernel = context.KernelCache.Get(cudaContext, ptx, kernelName);
kernel.GridDimensions = grid;
kernel.BlockDimensions = block;
kernel.DynamicSharedMemory = smemSize;
kernel.Run(args);
}
private void Invoke(TSCudaContext context, CudaContext cudaContext, string kernelName, dim3 grid, dim3 block, uint smemSize, CUstream stream, bool index32, params object[] args)
{
ConvertTensorArgs.Convert(cudaContext, index32, args);
byte[] ptx = GetPtx(context.Compiler);
CudaKernel kernel = context.KernelCache.Get(cudaContext, ptx, kernelName);
kernel.GridDimensions = grid;
kernel.BlockDimensions = block;
kernel.DynamicSharedMemory = smemSize;
kernel.RunAsync(stream, args);
}
/// <summary>
/// Invokes the specified context.
/// </summary>
/// <param name="context">The context.</param>
/// <param name="cudaContext">The cuda context.</param>
/// <param name="ptx">The PTX.</param>
/// <param name="baseName">Name of the base.</param>
/// <param name="args">The arguments.</param>
public static void Invoke(TSCudaContext context, CudaContext cudaContext, byte[] ptx, string baseName, params object[] args)
{
ThrowIfAnyTensorInvalid(args);
var deviceInfo = context.DeviceInfoForContext(cudaContext);
var allTensors = args.OfType <NDArray>();
var firstTensor = allTensors.First();
var elementCount = firstTensor.ElementCount();
var spec = new ApplySpecialization(allTensors.ToArray());
ConvertTensorArgs.Convert(cudaContext, spec.Use32BitIndices, args);
var block = ApplyUtils.GetApplyBlock();
var grid = ApplyUtils.GetApplyGrid(deviceInfo, elementCount);
var fullKernelName = PermutationGenerator.GetMangledName(baseName, spec);
var kernel = context.KernelCache.Get(cudaContext, ptx, fullKernelName);
kernel.GridDimensions = grid;
kernel.BlockDimensions = block;
kernel.RunAsync(CUstream.NullStream, args);
}
