On Tue, 14 Sep 2010 09:39:21 -0600 Marcus Daniels <mdani...@lanl.gov> wrote:
> On 9/14/10 8:46 AM, Basile Starynkevitch wrote: > > My current work aims to translate some Gimple into OpenCL source > > code, thus providing GCC with the ability to take advantage of GPU > > running their proprietary OpenCL compilers without asking the user to > > learn OpenCL. > My understanding is that Gimple does not have the notion of data > parallel operations. > For example, in Fortran, array operators are lowered to scalarized form. > OpenCL does have these semantics. kernels enqueued as an NDRanges are > item-by-item data parallel and there are often not any loops expressed > in the kernel itself. And kernels that do have lots of control logic, > and use lots of registers, global memory, etc. will tend not to work > well on GPUs. > > Remember that for geographical & political reasons all my GCC work is > > more "source to source" that "source to machine" oriented. I don't > > have the expertise, and I am not legitimate (internally in my CEA LIST > > organization at least, and also w.r.t. of funding French government > > agencies) to work on anything close to the target processor or silicon > > in GCC. > It seems to me a "source to source" compiler should definitely retain > high level constructs like array operators, DO ALL, OpenMP directives, etc. One can use #pragma-s & builtin-s & attributes for these. This is why I was trying to push the idea of plugin hooks for builtins. > > Marcus -- Basile STARYNKEVITCH http://starynkevitch.net/Basile/ email: basile<at>starynkevitch<dot>net mobile: +33 6 8501 2359 8, rue de la Faiencerie, 92340 Bourg La Reine, France *** opinions {are only mine, sont seulement les miennes} ***
