On entry to SGEMM, parameter 10 had an illegal value

[Sean Duggan]

I'm probably missing something obvious, but I'm trying to get the provided Windows binaries working in a program compiling in visual studio. As a test bed, I have a 4x5 (row-major) matrix multiplied by a 5x4 matrix. As I understand the documentation at http://developer.apple.com/library/mac/DOCUMENTATION/Accelerate/Reference/BLAS_Ref/Reference/reference.html#//apple_ref/c/func/cblas_sgemm, that should be:

cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, 4, 4, 5, 1, &(A[0][0]), 4, &(B[0][0]), 5, 0, &(C[0][0]), 4);

However, when I try that, I get an error message of "On entry to SGEMM, parameter 10 had an illegal value". Looking at the source code, I think that this is the relevant bit that determines the error message:

nrowb = args.k;

  if (transb & 1) nrowb = args.n;

....

if (args.ldb < nrowb)  info = 10;

Above, K is 5, as is ldb. 5 is not less than 5, so I shouldn't be getting this error message. Does anyone have a potential explanation?

[Zhang Xianyi]

Hi Sean,

I think lda and ldb are wrong.

You can try this.

cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, 4, 4, 5, 1, &(A[0][0]), 5, &(B[0][0]), 4, 0, &(C[0][0]), 4);

2013/8/14 Sean Duggan <sean....@gmail.com>

--
You received this message because you are subscribed to the Google Groups "OpenBLAS-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openblas-user...@googlegroups.com.
To post to this group, send email to openbla...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Sean Duggan]

That did the trick. Am I mis-understanding the description of the parameters? It seemed like a 5x4 matrix would have 5 as its value for the first dimension, and it seems like the text above would be more likely to trigger it, providing a value where ldb actually is less than K.

--
Sean Duggan                Sean....@gmail.com
"It's easy to win forgiveness for being wrong; being right is what gets you into real trouble." -Bjarne Stroustrup

[Sean Duggan]

Just to confuse things, it looks like the last parameter is based off of the number of rows to get it to work. Quick question when doing a transpose, do you change any of the parameters other than the last one, to indicate the size of the end matrix? To give a concrete example, without transposition, I have the following without transposing anything:

cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, A.rows(), B.cols(), A.cols(), 1, ptrA, A.cols(), ptrB, B.cols(), 0, ptrC, A.rows());

If I were transposing the first, and  multiplying it  which of the following works?

cblas_sgemm(CblasRowMajor, CblasTrans, CblasNoTrans, A.rows(), B.cols(), A.cols(), 1, ptrA, A.cols(), ptrB, B.cols(), 0, ptrC, A.rows());

cblas_sgemm(CblasRowMajor, CblasTrans, CblasNoTrans, A.cols(), B.cols(), A.cols(), 1, ptrA, A.rows(), ptrB, B.cols(), 0, ptrC, A.rows());

cblas_sgemm(CblasRowMajor, CblasTrans, CblasNoTrans, A.cols(), B.cols(), A.cols(), 1, ptrA, A.cols(), ptrB, B.cols(), 0, ptrC, A.rows());

cblas_sgemm(CblasRowMajor, CblasTrans, CblasNoTrans, A.rows(), B.cols(), A.cols(), 1, ptrA, A.cols(), ptrB, B.cols(), 0, ptrC, A.cols());

cblas_sgemm(CblasRowMajor, CblasTrans, CblasNoTrans, A.cols(), B.cols(), A.cols(), 1, ptrA, A.rows(), ptrB, B.cols(), 0, ptrC, A.cols());

cblas_sgemm(CblasRowMajor, CblasTrans, CblasNoTrans, A.cols(), B.cols(), A.cols(), 1, ptrA, A.cols(), ptrB, B.cols(), 0, ptrC, A.cols());

Essentially, I don't know if I'm supplying the size of the A matrix before or after transpose and whether one needs to tell the function call to handle the final matrix differently.

On Wednesday, August 14, 2013 11:13:51 AM UTC-4, Sean Duggan wrote:

That did the trick. Am I mis-understanding the description of the parameters? It seemed like a 5x4 matrix would have 5 as its value for the first dimension, and it seems like the text above would be more likely to trigger it, providing a value where ldb actually is less than K.

On Wed, Aug 14, 2013 at 10:38 AM, Zhang Xianyi <traits...@gmail.com> wrote:

Hi Sean,

I think lda and ldb are wrong.

You can try this.

cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, 4, 4, 5, 1, &(A[0][0]), 5, &(B[0][0]), 4, 0, &(C[0][0]), 4);

2013/8/14 Sean Duggan <sean....@gmail.com>

I'm probably missing something obvious, but I'm trying to get the provided Windows binaries working in a program compiling in visual studio. As a test bed, I have a 4x5 (row-major) matrix multiplied by a 5x4 matrix. As I understand the documentation at http://developer.apple.com/library/mac/DOCUMENTATION/Accelerate/Reference/BLAS_Ref/Reference/reference.html#//apple_ref/c/func/cblas_sgemm, that should be:

cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, 4, 4, 5, 1, &(A[0][0]), 4, &(B[0][0]), 5, 0, &(C[0][0]), 4);

However, when I try that, I get an error message of "On entry to SGEMM, parameter 10 had an illegal value". Looking at the source code, I think that this is the relevant bit that determines the error message:

nrowb = args.k;

  if (transb & 1) nrowb = args.n;

....

if (args.ldb < nrowb)  info = 10;

Above, K is 5, as is ldb. 5 is not less than 5, so I shouldn't be getting this error message. Does anyone have a potential explanation?

--
You received this message because you are subscribed to the Google Groups "OpenBLAS-users" group.

To unsubscribe from this group and stop receiving emails from it, send an email to openblas-users+unsubscribe@googlegroups.com.
To post to this group, send email to openblas-users@googlegroups.com.

For more options, visit https://groups.google.com/groups/opt_out.

[Sean Duggan]

Alright. I've figured it out. This is what I wound up with for my helper function which takes in our Matrix class and transpose arguments and ejects a result:

  int A_m = (tranA ? A.cols() : A.rows() );

  int A_n = (tranA ? A.rows() : A.cols() );

  int B_m = (tranB ? B.cols() : B.rows() );

  int B_n = (tranB ? B.rows() : B.cols() );

  if (A_n != B_m)

    ISI_THROW(IllegalArgumentException("Mismatch between source matrices sizes"));

  this -> resize(A_m, B_n, 0);

  Float* ptrA = A.el_[0];

  Float* ptrB = B.el_[0];

  Float* ptrC = this->el_[0];

  cblas_sgemm(CblasRowMajor, tranA ? CblasTrans : CblasNoTrans, tranB ? CblasTrans : CblasNoTrans, 

    A_m, B_n, A_n, 1, ptrA, A.cols(), ptrB, B.cols(), 0, ptrC, B_n);

So, in short, the first set of dimensions needs to be converted for the transposed dimensions, the latter set don't, and need to refer to columns when using C++ Row Major arrays.

[Zhang Xianyi]

Sorry for the late reply.

2013/8/17 Sean Duggan <sean....@gmail.com>

Alright. I've figured it out. This is what I wound up with for my helper function which takes in our Matrix class and transpose arguments and ejects a result:

  int A_m = (tranA ? A.cols() : A.rows() );

  int A_n = (tranA ? A.rows() : A.cols() );

  int B_m = (tranB ? B.cols() : B.rows() );

  int B_n = (tranB ? B.rows() : B.cols() );

  if (A_n != B_m)

    ISI_THROW(IllegalArgumentException("Mismatch between source matrices sizes"));

  this -> resize(A_m, B_n, 0);

  Float* ptrA = A.el_[0];

  Float* ptrB = B.el_[0];

  Float* ptrC = this->el_[0];

  cblas_sgemm(CblasRowMajor, tranA ? CblasTrans : CblasNoTrans, tranB ? CblasTrans : CblasNoTrans, 

    A_m, B_n, A_n, 1, ptrA, A.cols(), ptrB, B.cols(), 0, ptrC, B_n);

So, in short, the first set of dimensions needs to be converted for the transposed dimensions, the latter set don't, and need to refer to columns when using C++ Row Major arrays.

Yes, you are right.

From Intel MKL reference, http://software.intel.com/sites/products/documentation/hpc/mkl/mklman/GUID-97718E5C-6E0A-44F0-B2B1-A551F0F164B2.htm

m

INTEGER. Specifies the number of rows of the matrix op(A) and of the matrix C. The value of m must be at least zero.

n

INTEGER. Specifies the number of columns of the matrix op(B) and the number of columns of the matrix C.

k

INTEGER. Specifies the number of columns of the matrix op(A) and the number of rows of the matrix op(B).

The m,n,k need change with transpose.

Xianyi