英特尔MKL错误：进入cblas_zgemm时参数9不正确（C＃Linux）

Question

我使用以下命令构建了custom MKL library (2019 Update 2) for Windows（10）和Linux（Ubuntu 18.04）：

nmake libintel64 MKLROOT="C:\Program Files (x86)\IntelSWTools\compilers_and_libraries\windows\mkl" name=win\intel64\custom_mkl interface="lp64"

和

make libintel64 MKLROOT="opt/intel/mkl" name=linux/intel64/custom_mkl interface=lp64

我正在使用C＃的DllImport来调用cblas_zgemm

[DllImport(DLLName, CallingConvention = CallingConvention.Cdecl, ExactSpelling = true, SetLastError = false)]
internal static extern void cblas_zgemm(
    int Order, int TransA, int TransB, int M, int N, int K,
    ComplexDouble alpha, [In] ComplexDouble[,] A, int lda, [In] ComplexDouble[,] B, int ldb,
    ComplexDouble beta, [In, Out] ComplexDouble[,] C, int ldc);

见https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm

其中ComplexDouble定义为：

public struct ComplexDouble
{
    public double real;
    public double imag;

    public static implicit operator ComplexDouble(double d)
    {
        return new ComplexDouble() { real = d };
    }
}

我已经为矩阵乘法定义了以下静态方法：

public static ComplexDouble[,] Dot(ComplexDouble[,] a, ComplexDouble[,] b)
{
    int n1 = a.GetLength(0);
    int n2 = a.GetLength(1);
    int n3 = b.GetLength(0);
    int n4 = b.GetLength(1);
    if (n2 != n3) throw new System.Exception("Inner matrix dimensions must agree");
    int Order = 101; // row-major arrays
    int TransA = 111; // trans='N'
    int TransB = 111; // trans='N'
    int M = n1, N = n4, K = n2;
    int lda = K, ldb = N, ldc = N;
    ComplexDouble alpha = 1, beta = 0;
    ComplexDouble[,] c = new ComplexDouble[n1, n4];
    _mkl.cblas_zgemm(Order, TransA, TransB, M, N, K, alpha, a, lda, b, ldb, beta, c, ldc);
    return c;
}

最后，我得到了以下测试代码，这些代码在Windows下运行良好，但在Linux上运行失败（参数9在进入cblas_zgemm时不正确）。

ComplexDouble[,] A = new ComplexDouble[,] { { 1, 2, 3 }, { 4, 5, 6 } };
ComplexDouble[,] B = new ComplexDouble[,] { { 0, 1, 0, 1 }, { 1, 0, 0, 1 }, { 1, 0, 1, 0 } };
ComplexDouble[,] C = MKL.Dot(A, B);

我做了以下观察：

• 它适用于Windows
• 它在Linux（Mono）上失败，参数9在输入cblas_zgemm或Null Reference时不正确。
• 如果我用cblas_dgemm和double重复实验，它可以在Windows和Linux上运行。

Answer 1

仔细看看cblas_zgemm

void cblas_zgemm (
    const CBLAS_LAYOUT Layout, const CBLAS_TRANSPOSE transa, const CBLAS_TRANSPOSE transb, const MKL_INT m, const MKL_INT n, const MKL_INT k, 
    const void *alpha, const void *a, const MKL_INT lda, const void *b, const MKL_INT ldb, 
    const void *beta, void *c, const MKL_INT ldc);

表明参数alpha和beta必须通过引用传递。所以DllImport必须看起来像那样：

[DllImport(DLLName, CallingConvention = CallingConvention.Cdecl, ExactSpelling = true, SetLastError = false)]
internal static extern void cblas_zgemm(
    int Order, int TransA, int TransB, int M, int N, int K,
    ref ComplexDouble alpha, [In] ComplexDouble[,] A, int lda, [In] ComplexDouble[,] B, int ldb,
    ref ComplexDouble beta, [In, Out] ComplexDouble[,] C, int ldc);

我不知何故错过了因为cblas_dgemm

void cblas_dgemm (
    const CBLAS_LAYOUT Layout, const CBLAS_TRANSPOSE transa, const CBLAS_TRANSPOSE transb, const MKL_INT m, const MKL_INT n, const MKL_INT k, 
    const double alpha, const double *a, const MKL_INT lda, const double *b, const MKL_INT ldb, 
    const double beta, double *c, const MKL_INT ldc);

参数alpha和beta按值传递。

[DllImport(DLLName, CallingConvention = CallingConvention.Cdecl, ExactSpelling = true, SetLastError = false)]
internal static extern void cblas_dgemm(
    int Order, int TransA, int TransB, int M, int N, int K,
    double alpha, [In] double[,] A, int lda, [In] double[,] B, int ldb,
    double beta, [In, Out] double[,] C, int ldc);

就像那个cblas_zgemm适用于Windows和Linux。我仍然不明白为什么它使用ref在Windows上工作？