For the transposition directive in gemv, CblasConjTrans just gives results as CblasTrans.

[shengg]

For blas functions, gemv and gemm, there are parameters to determine how to transpose for the matrix: NoTrans, Trans or ConjTrans. In btas, all of these three parameters are implemented for gemm. However, only NoTrans and Trans are implemented for gemv. If ConjTrans is used, it is considered as Trans. I do not know why.

[shiozaki]

I think complex support is not well implemented (I guess I was the only person who uses complex arithmetics?), so I would expect that it is simply a bug.

[naokin]

That should be a bug.
But, as long as calling CBLAS, I think they work correctly, calling with
TensorView might have a trouble.

On Fri, Jul 18, 2014 at 6:54 AM, Toru Shiozaki notifications@github.com
wrote:

I think complex support is not well implemented (I guess I was the only
person who uses complex arithmetics?), so I would expect that it is simply
a bug.

—
Reply to this email directly or view it on GitHub
#54 (comment).

[evaleev]

@shengg Most contract-like and permute should be assumed to NOT work with
TensorView. Although TensorView meets the TWG.Tensor concept spec (checked
by btas::is_boxtensor), and hence gets accepted by genv() and like, most of
these functions assume "contiguous" view. Assume that most uses of
TensorView will produce wrong results and should be checked.

All, note pull request #53 that addresses the issue of const TensorView not
behaving like TensorConstView. The only way I found this to be solvable is
through a configurable parameter (hence a distinct type of TensorView).
Feel free to comment. test.C has the relevant examples that explain the new
TensorRWView.

On Thu, Jul 17, 2014 at 8:29 PM, Naoki Nakatani notifications@github.com
wrote:

That should be a bug.
But, as long as calling CBLAS, I think they work correctly, calling with
TensorView might have a trouble.

On Fri, Jul 18, 2014 at 6:54 AM, Toru Shiozaki notifications@github.com
wrote:

I think complex support is not well implemented (I guess I was the only
person who uses complex arithmetics?), so I would expect that it is
simply
a bug.

—
Reply to this email directly or view it on GitHub
#54 (comment).

—
Reply to this email directly or view it on GitHub
#54 (comment).

web: valeyev.net

[shiozaki]

The unit test is now broken (with Commit: bde5b24). @shengg Is this related to this issue?

test is a Catch v1.0 b48 host application.
Run with -? for options
-------------------------------------------------------------------------------
Tensor Gemv
  Complex Double Gemv --- ConjTrans
-------------------------------------------------------------------------------
tensor_blas_test.cc:223
...............................................................................

tensor_blas_test.cc:371: FAILED:
  CHECK( res < 1E-10 )
with expansion:
  121.1902981253 < 0.0000000001

===============================================================================
14 test cases - 1 failed (2370 assertions - 1 failed)

Makefile:29: recipe for target 'run' failed
gmake: *** [run] Error 1

[shengg]

Yes.
In gemv_impl.h, there are only TransA == CblasNoTrans and TransA != CblasNoTrans. There is no option to distinguish CblasConjTrans from CblasTrans .

[shengg]

@shiozaki I made a small change. It works now.

[shiozaki]

Thanks - but if _HAS_CBLAS is defined, it still breaks.

[shengg]

I used mkl cblas. When _HAS_CBLAS is defined, several other tests also failed.

[shiozaki]

I do use MKL CBLAS on Mac, and this (tensor_blas_test.cc:371) is the only test that fails.

[shengg]

I should check my MKL or try another version. Maybe, it was not configured well, since MKL in my desktop was never used before.

[shiozaki]

For what it's worth, my MKL is composer_xe_2013_sp1.0.074 and I do in ~/.profile

source /opt/intel/bin/compilervars.sh intel64

[shiozaki]

WAIT! you are right, there are some tests that are broken (I was looking at a wrong build). I will try to look into it as far as I can.