public void cudaMalloc_cudaFree_cudaMemcpy_test_cudaMemset_test()
{
int length = 1024 * 2;
byte testValue = 5;
byte[] test = new byte[length];
byte[] result = new byte[length];
IntPtr d_ptr = IntPtr.Zero;
var size = length * sizeof(byte);
var status = CudaRuntimeApi.cudaMalloc(ref d_ptr, (ulong)size);
Assert.AreEqual(status, cudaError.cudaSuccess);
Console.WriteLine($"ptr : {d_ptr}");
GCHandle gchTest = GCHandle.Alloc(test, GCHandleType.Pinned);
GCHandle gchResult = GCHandle.Alloc(result, GCHandleType.Pinned);
IntPtr h_ptrTest = Marshal.UnsafeAddrOfPinnedArrayElement(test, 0);
IntPtr h_ptrResult = Marshal.UnsafeAddrOfPinnedArrayElement(result, 0);
status = CudaRuntimeApi.cudaMemcpy(d_ptr, h_ptrTest, (ulong)size, cudaMemcpyKind.HostToDevice);
Assert.AreEqual(status, cudaError.cudaSuccess);
status = CudaRuntimeApi.cudaMemset(d_ptr, testValue, (ulong)size);
Assert.AreEqual(status, cudaError.cudaSuccess);
status = CudaRuntimeApi.cudaMemcpy(h_ptrResult, d_ptr, (ulong)size, cudaMemcpyKind.DeviceToHost);
Assert.AreEqual(status, cudaError.cudaSuccess);
for (int i = 0; i < length; i++)
{
Assert.AreEqual(result[i], testValue);
}
status = CudaRuntimeApi.cudaFree(d_ptr);
Assert.AreEqual(status, cudaError.cudaSuccess);
gchTest.Free();
gchResult.Free();
}
public void cublasCgemm_test()
{
int devCount = 0;
var status = CudaRuntimeApi.cudaGetDeviceCount(ref devCount);
Assert.AreEqual(cudaError.cudaSuccess, status);
int devId = 0;
status = CudaRuntimeApi.cudaSetDevice(devId);
var handle = IntPtr.Zero;
var cblasStatus = Cublas_api.cublasCreate_v2(ref handle);
Random rand = new Random();
int rows_a = rand.Next(2, 10);
int cols_a = rand.Next(2, 10);
int rows_b = cols_a;
int cols_b = rand.Next(2, 10);
int rows_c = rows_a;
int cols_c = cols_b;
var A = new float2[rows_a * cols_a];
var B = new float2[rows_b * cols_b];
var C = new float2[rows_c * cols_c];
var resultC = new float2[rows_c * cols_c];
var cA = new Complex32[rows_a * cols_a];
var cB = new Complex32[rows_b * cols_b];
var cC = new Complex32[rows_c * cols_c];
var cResultC = new Complex32[rows_c * cols_c];
for (int i = 0; i < A.Length; i++)
{
var real = Convert.ToSingle(rand.Next(0, 10));
var imag = Convert.ToSingle(rand.Next(0, 10));
A[i] = new float2(real, imag);
cA[i] = new Complex32(real, imag);
}
for (int i = 0; i < B.Length; i++)
{
var real = Convert.ToSingle(rand.Next(0, 10));
var imag = Convert.ToSingle(rand.Next(0, 10));
B[i] = new float2(real, imag);
cB[i] = new Complex32(real, imag);
}
for (int i = 0; i < C.Length; i++)
{
var real = Convert.ToSingle(rand.Next(0, 10));
var imag = Convert.ToSingle(rand.Next(0, 10));
C[i] = new float2(real, imag);
cC[i] = new Complex32(real, imag);
}
var alphaReal = Convert.ToSingle(rand.Next(0, 10));
var alphaImag = Convert.ToSingle(rand.Next(0, 10));
var alpha = new float2(alphaReal, alphaImag);
var cAlpha = new Complex32(alphaReal, alphaImag);
var betaReal = Convert.ToSingle(rand.Next(0, 10));
var betaImag = Convert.ToSingle(rand.Next(0, 10));
var beta = new float2(betaReal, betaImag);
var cBeta = new Complex32(betaReal, betaImag);
var d_a = IntPtr.Zero;
var d_b = IntPtr.Zero;
var d_c = IntPtr.Zero;
status = CudaRuntimeApi.cudaMalloc(ref d_a, (ulong)(A.Length * Marshal.SizeOf(typeof(float2))));
status = CudaRuntimeApi.cudaMalloc(ref d_b, (ulong)(B.Length * Marshal.SizeOf(typeof(float2))));
status = CudaRuntimeApi.cudaMalloc(ref d_c, (ulong)(C.Length * Marshal.SizeOf(typeof(float2))));
var gch_a = GCHandle.Alloc(A, GCHandleType.Pinned);
var gch_b = GCHandle.Alloc(B, GCHandleType.Pinned);
var gch_c = GCHandle.Alloc(C, GCHandleType.Pinned);
var gch_resultC = GCHandle.Alloc(resultC, GCHandleType.Pinned);
var h_a = Marshal.UnsafeAddrOfPinnedArrayElement(A, 0);
var h_b = Marshal.UnsafeAddrOfPinnedArrayElement(B, 0);
var h_c = Marshal.UnsafeAddrOfPinnedArrayElement(C, 0);
var h_resultC = Marshal.UnsafeAddrOfPinnedArrayElement(resultC, 0);
status = CudaRuntimeApi.cudaMemcpy(d_a, h_a, (ulong)(A.Length * Marshal.SizeOf(typeof(float2))), cudaMemcpyKind.HostToDevice);
status = CudaRuntimeApi.cudaMemcpy(d_b, h_b, (ulong)(B.Length * Marshal.SizeOf(typeof(float2))), cudaMemcpyKind.HostToDevice);
status = CudaRuntimeApi.cudaMemcpy(d_c, h_c, (ulong)(C.Length * Marshal.SizeOf(typeof(float2))), cudaMemcpyKind.HostToDevice);
cblasStatus = Cublas_api.cublasCgemm_v2(
handle,
cublasOperation_t.CUBLAS_OP_N,
cublasOperation_t.CUBLAS_OP_N,
rows_a,
cols_b,
cols_a,
ref alpha,
d_a,
rows_a,
d_b,
rows_b,
ref beta,
d_c,
rows_c
);
status = CudaRuntimeApi.cudaMemcpy(h_resultC, d_c, (ulong)(resultC.Length * Marshal.SizeOf(typeof(float2))), cudaMemcpyKind.DeviceToHost);
for (int i = 0; i < rows_c * cols_c; i++)
{
cResultC[i] = new Complex32(resultC[i].X, resultC[i].Y);
}
var mResultC = Matrix <Complex32> .Build.Dense(rows_c, cols_c, cResultC);
var mA = Matrix <Complex32> .Build.Dense(rows_a, cols_a, cA);
var mB = Matrix <Complex32> .Build.Dense(rows_b, cols_b, cB);
var mExpectedC = Matrix <Complex32> .Build.Dense(rows_c, cols_c, cC).Clone();
mExpectedC = cAlpha * mA * mB + cBeta * mExpectedC;
Complex32[] expected = mExpectedC.ToColumnMajorArray();
Console.WriteLine("alpha : {0}, beta : {1}", alpha, beta);
Console.WriteLine("A");
Console.WriteLine(mA.ToString());
Console.WriteLine();
Console.WriteLine("B");
Console.WriteLine(mB.ToString());
Console.WriteLine();
Console.WriteLine("resultC");
Console.WriteLine(mResultC.ToString());
Console.WriteLine();
Console.WriteLine("expectedC");
Console.WriteLine(mExpectedC.ToString());
for (int i = 0; i < C.Length; i++)
{
Assert.AreEqual(expected[i], cResultC[i]);
}
cblasStatus = Cublas_api.cublasDestroy_v2(handle);
status = CudaRuntimeApi.cudaFree(d_a);
status = CudaRuntimeApi.cudaFree(d_b);
status = CudaRuntimeApi.cudaFree(d_c);
gch_a.Free();
gch_b.Free();
gch_c.Free();
}
public void cublasSgemm_test()
{
int devCount = 0;
var status = CudaRuntimeApi.cudaGetDeviceCount(ref devCount);
Assert.AreEqual(cudaError.cudaSuccess, status);
int devId = 0;
status = CudaRuntimeApi.cudaSetDevice(devId);
var handle = IntPtr.Zero;
var cblasStatus = Cublas_api.cublasCreate_v2(ref handle);
Random rand = new Random();
int rows_a = rand.Next(2, 10);
int cols_a = rand.Next(2, 10);
int rows_b = cols_a;
int cols_b = rand.Next(2, 10);
int rows_c = rows_a;
int cols_c = cols_b;
float alpha = 1.0F;
float beta = 0.0F;
float[] A = new float[rows_a * cols_a];
float[] B = new float[rows_b * cols_b];
float[] C = new float[rows_c * cols_c];
float[] resultC = new float[rows_c * cols_c];
for (int i = 0; i < A.Length; i++)
{
A[i] = Convert.ToSingle(rand.Next(0, 10));
}
for (int i = 0; i < B.Length; i++)
{
B[i] = Convert.ToSingle(rand.Next(0, 10));
}
var d_a = IntPtr.Zero;
var d_b = IntPtr.Zero;
var d_c = IntPtr.Zero;
status = CudaRuntimeApi.cudaMalloc(ref d_a, (ulong)A.Length * sizeof(float));
status = CudaRuntimeApi.cudaMalloc(ref d_b, (ulong)B.Length * sizeof(float));
status = CudaRuntimeApi.cudaMalloc(ref d_c, (ulong)C.Length * sizeof(float));
var gch_a = GCHandle.Alloc(A, GCHandleType.Pinned);
var gch_b = GCHandle.Alloc(B, GCHandleType.Pinned);
var gch_c = GCHandle.Alloc(C, GCHandleType.Pinned);
var gch_resultC = GCHandle.Alloc(resultC, GCHandleType.Pinned);
var h_a = Marshal.UnsafeAddrOfPinnedArrayElement(A, 0);
var h_b = Marshal.UnsafeAddrOfPinnedArrayElement(B, 0);
var h_c = Marshal.UnsafeAddrOfPinnedArrayElement(C, 0);
var h_resultC = Marshal.UnsafeAddrOfPinnedArrayElement(resultC, 0);
status = CudaRuntimeApi.cudaMemcpy(d_a, h_a, (ulong)A.Length * sizeof(float), cudaMemcpyKind.HostToDevice);
status = CudaRuntimeApi.cudaMemcpy(d_b, h_b, (ulong)B.Length * sizeof(float), cudaMemcpyKind.HostToDevice);
status = CudaRuntimeApi.cudaMemcpy(d_c, h_c, (ulong)C.Length * sizeof(float), cudaMemcpyKind.HostToDevice);
cblasStatus = Cublas_api.cublasSgemm_v2(
handle,
cublasOperation_t.CUBLAS_OP_N,
cublasOperation_t.CUBLAS_OP_N,
rows_a,
cols_b,
cols_a,
ref alpha,
d_a,
rows_a,
d_b,
rows_b,
ref beta,
d_c,
rows_c
);
status = CudaRuntimeApi.cudaMemcpy(h_resultC, d_c, (ulong)C.Length * sizeof(float), cudaMemcpyKind.DeviceToHost);
var mResultC = Matrix <float> .Build.Dense(rows_c, cols_c, resultC);
var mA = Matrix <float> .Build.Dense(rows_a, cols_a, A);
var mB = Matrix <float> .Build.Dense(rows_b, cols_b, B);
var mExpectedC = Matrix <float> .Build.Dense(rows_c, cols_c, C).Clone();
mExpectedC = alpha * mA * mB + beta * mExpectedC;
float[] expected = mExpectedC.ToColumnMajorArray();
Console.WriteLine("alpha : {0}, beta : {1}", alpha, beta);
Console.WriteLine("A");
Console.WriteLine(mA.ToString());
Console.WriteLine();
Console.WriteLine("B");
Console.WriteLine(mB.ToString());
Console.WriteLine();
Console.WriteLine("resultC");
Console.WriteLine(mResultC.ToString());
Console.WriteLine();
Console.WriteLine("expectedC");
Console.WriteLine(mExpectedC.ToString());
for (int i = 0; i < C.Length; i++)
{
Assert.AreEqual(expected[i], resultC[i]);
}
cblasStatus = Cublas_api.cublasDestroy_v2(handle);
status = CudaRuntimeApi.cudaFree(d_a);
status = CudaRuntimeApi.cudaFree(d_b);
status = CudaRuntimeApi.cudaFree(d_c);
gch_a.Free();
gch_b.Free();
gch_c.Free();
}