If you're talking of general-purpose computations on GPUs (GPGPU),
then many of the traditional optimizations are indeed useful and
necessary, since the code structure is general and not specialized
graphics code. Optimizations related to performance could be quite
im****tant - strength reduction, induction variable optimizations, code
motion and so on. IPA may not be very im****tant because most of the
compute-intensive "kernels" that run on the GPU are loops with a high
degree of data-parallelism and minimal control-flow branching (like in
function calls).
In fact, NVIDIA's compiler for CUDA uses the high-level optimizer from
the Open64 compiler framework for optimizing the general purpose code
before performing register allocation, scheduling and other
optimizations specific to the underlying GPU.
On Tue, Apr 29, 2008 at 1:03 PM, shrey <shreyas76@[EMAIL PROTECTED]
> wrote:
> Can any one give me some broad hints on what sort of compiler
> optimizations are im****tant for GPU's especially those on mobile
> platforms? Do traditional heavy weight optimizations/analysis like
> IPA, PRE, loop transformations hold any usefulness ?
--
Giridhar
|
