This is the mail archive of the
mailing list for the GCC project.
Re: [patch/gfortran] Optimizing matmul-transpose combinations
- From: Paul Brook <paul at codesourcery dot com>
- To: fortran at gcc dot gnu dot org
- Cc: Victor Leikehman <LEI at il dot ibm dot com>,gcc-patches at gcc dot gnu dot org
- Date: Tue, 7 Dec 2004 22:20:27 +0000
- Subject: Re: [patch/gfortran] Optimizing matmul-transpose combinations
- Organization: CodeSourcery
- References: <OFAA18B791.F4B29646-ON42256F5C.004A46C8-42256F5C.004BE881@il.ibm.com>
On Tuesday 30 November 2004 13:49, Victor Leikehman wrote:
> This patch follows the previous discussion of this topic.
> As Paul Brook proposed, the preferred
> way to handle matmul-transpose combinations is to
> change the implementation of TRANSPOSE so that it will
> create a new array descriptor with swapped strides.
> There are two parts to this change -- compiler part
> and library part. This patch contains the library
> part, which is just dead code without the accompanying
> change in compiler.
> My experiments with a hacked version of compiler show
> more than a 100% improvement on galgel spec benchmark.
The patch and general strategy look ok.
However it's useless without the compiler bits. I think I'd prefer to wait
until we have the compiler bits before applying the patch.
The other consideration is that we are now in stage3. In theory this means
regression fixes only, though gfortran gets a bit more leeway because it's
new and g77 isn't directly comparable. This is not a correctness issue,
though it's arguable that the poor performance is itself a bug.
I'd need to know how invasive the compiler changes are before deciding if this
is suitable for stage3. It it is too invasive It'll have to wait until 4.1.