public static void IlGpu(CudaAccelerator gpu, Real[] result, Real[] left, Real[] right, int n)
{
using (var cudaResult = gpu.Allocate(result))
using (var cudaLeft = gpu.Allocate(left))
using (var cudaRight = gpu.Allocate(right))
{
using var blas = new CuBlas(gpu, CuBlasAPIVersion.V11);
var timer = Stopwatch.StartNew();
blas.Gemm(
CuBlasOperation.NonTranspose,
CuBlasOperation.NonTranspose,
n, n, n,
1, cudaLeft.View, n,
cudaRight.View, n, 0,
cudaResult.View, n);
gpu.Synchronize();
PrintPerformance(timer, "MatrixMultiplication.IlGpu.cuBLAS", n, n, n);
cudaResult.CopyTo(result, 0, 0, result.Length);
}
}
public static unsafe cublasContext *blas_handle()
{
if (!cublasInit)
{
cublasContext *handle;
SafeCall(CuBlas.cublasCreate_v2(&handle));
cublasInit = true;
cublasHandle = handle;
}
return(cublasHandle);
}
static void Main()
{
const int DataSize = 1024;
const CuBlasAPIVersion CuBlasVersion = CuBlasAPIVersion.V10;
using (var context = new Context())
{
// Enable algorithms library
context.EnableAlgorithms();
// Check for Cuda support
foreach (var acceleratorId in CudaAccelerator.CudaAccelerators)
{
using (var accelerator = new CudaAccelerator(context, acceleratorId))
{
Console.WriteLine($"Performing operations on {accelerator}");
var buf = accelerator.Allocate <float>(DataSize);
var buf2 = accelerator.Allocate <float>(DataSize);
accelerator.Initialize(accelerator.DefaultStream, buf, 1.0f);
accelerator.Initialize(accelerator.DefaultStream, buf2.View, 1.0f);
// Initialize the CuBlas library using manual pointer mode handling
// (default behavior)
using (var blas = new CuBlas(accelerator, CuBlasVersion))
{
// Set pointer mode to Host to enable data transfer to CPU memory
blas.PointerMode = CuBlasPointerMode.Host;
float output = blas.Nrm2(buf);
// Set pointer mode to Device to enable data transfer to GPU memory
blas.PointerMode = CuBlasPointerMode.Device;
blas.Nrm2(buf, buf2);
// Use pointer mode scopes to recover the previous pointer mode
using (var scope = blas.BeginPointerScope(CuBlasPointerMode.Host))
{
float output2 = blas.Nrm2(buf);
}
}
// Initialize the CuBlas<T> library using custom pointer mode handlers
using (var blas = new CuBlas <CuBlasPointerModeHandlers.AutomaticMode>(accelerator, CuBlasVersion))
{
// Automatic transfer to host
float output = blas.Nrm2(buf);
// Automatic transfer to device
blas.Nrm2(buf, buf2);
}
}
}
}
}
static void Main()
{
const int DataSize = 1024;
using var context = Context.Create(builder => builder.Cuda().EnableAlgorithms());
// Check for Cuda support
foreach (var device in context.GetCudaDevices())
{
using var accelerator = device.CreateCudaAccelerator(context);
Console.WriteLine($"Performing operations on {accelerator}");
var buf = accelerator.Allocate1D <float>(DataSize);
var buf2 = accelerator.Allocate1D <float>(DataSize);
accelerator.Initialize(accelerator.DefaultStream, buf.View, 1.0f);
accelerator.Initialize(accelerator.DefaultStream, buf2.View, 1.0f);
// Initialize the CuBlas library using manual pointer mode handling
// (default behavior)
using (var blas = new CuBlas(accelerator))
{
// Set pointer mode to Host to enable data transfer to CPU memory
blas.PointerMode = CuBlasPointerMode.Host;
float output = blas.Nrm2(buf.View.AsGeneral());
// Set pointer mode to Device to enable data transfer to GPU memory
blas.PointerMode = CuBlasPointerMode.Device;
blas.Nrm2(buf.View.AsGeneral(), buf2.View);
// Use pointer mode scopes to recover the previous pointer mode
using var scope = blas.BeginPointerScope(CuBlasPointerMode.Host);
float output2 = blas.Nrm2(buf.View.AsGeneral());
}
// Initialize the CuBlas<T> library using custom pointer mode handlers
using (var blas = new CuBlas <CuBlasPointerModeHandlers.AutomaticMode>(accelerator))
{
// Automatic transfer to host
float output = blas.Nrm2(buf.View.AsGeneral());
// Automatic transfer to device
blas.Nrm2(buf.View.AsGeneral(), buf2.View);
}
}
}
public static unsafe void gemm_ongpu(int ta, int tb, int m, int n, int k, float alpha,
float[] a, int lda,
float[] b, int ldb,
float beta,
float[] c, int ldc)
{
using (var gpuA = Gpu.Default.AllocateDevice(a.ToArray()))
using (var gpuB = Gpu.Default.AllocateDevice(b.ToArray()))
using (var gpuC = Gpu.Default.AllocateDevice(c.ToArray()))
{
var handle = CudaUtils.blas_handle();
CudaUtils.SafeCall(CuBlas.cublasSgemm_v2(handle,
(tb != 0 ? cublasOperation_t.CUBLAS_OP_T : cublasOperation_t.CUBLAS_OP_N),
(ta != 0 ? cublasOperation_t.CUBLAS_OP_T : cublasOperation_t.CUBLAS_OP_N), n, m, k, &alpha, (float *)gpuB.Handle, ldb,
(float *)gpuA.Handle, lda, &beta, (float *)gpuC.Handle, ldc));
a = Gpu.CopyToHost(gpuA);
b = Gpu.CopyToHost(gpuB);
c = Gpu.CopyToHost(gpuC);
}
}
