On Mon, 24 Nov 2014, Tom de Vries wrote:
> On 17-11-14 11:13, Richard Biener wrote:
> > On Sat, 15 Nov 2014, Tom de Vries wrote:
> >
> > > >On 15-11-14 13:14, Tom de Vries wrote:
> > > > > >Hi,
> > > > > >
> > > > > >I'm submitting a patch series with initial support for the oacc
> > > > kernels
> > > > > >directive.
> > > > > >
> > > > > >The patch series uses pass_parallelize_loops to implement
> > > > parallelization of
> > > > > >loops in the oacc kernels region.
> > > > > >
> > > > > >The patch series consists of these 8 patches:
> > > > > >...
> > > > > > 1 Expand oacc kernels after pass_build_ealias
> > > > > > 2 Add pass_oacc_kernels
> > > > > > 3 Add pass_ch_oacc_kernels to pass_oacc_kernels
> > > > > > 4 Add pass_tree_loop_{init,done} to pass_oacc_kernels
> > > > > > 5 Add pass_loop_im to pass_oacc_kernels
> > > > > > 6 Add pass_ccp to pass_oacc_kernels
> > > > > > 7 Add pass_parloops_oacc_kernels to pass_oacc_kernels
> > > > > > 8 Do simple omp lowering for no address taken var
> > > > > >...
> > > >
> > > >This patch lowers integer variables that do not have their address taken
> > > as
> > > >local variable. We use a copy at region entry and exit to copy the value
> > > in
> > > >and out.
> > > >
> > > >In the context of reduction handling in a kernels region, this allows the
> > > >parloops reduction analysis to recognize the reduction, even after oacc
> > > >lowering has been done in pass_lower_omp.
> > > >
> > > >In more detail, without this patch, the omp_data_i load and stores are
> > > >generated in place (in this case, in the loop):
> > > >...
> > > > {
> > > > .omp_data_iD.2201 = &.omp_data_arr.15D.2220;
> > > > {
> > > > unsigned intD.9 iD.2146;
> > > >
> > > > iD.2146 = 0;
> > > > goto <D.2207>;
> > > > <D.2208>:
> > > > D.2216 = .omp_data_iD.2201->cD.2203;
> > > > c.9D.2176 = *D.2216;
> > > > D.2177 = (long unsigned intD.10) iD.2146;
> > > > D.2178 = D.2177 * 4;
> > > > D.2179 = c.9D.2176 + D.2178;
> > > > D.2180 = *D.2179;
> > > > D.2217 = .omp_data_iD.2201->sumD.2205;
> > > > D.2218 = *D.2217;
> > > > D.2217 = .omp_data_iD.2201->sumD.2205;
> > > > D.2219 = D.2180 + D.2218;
> > > > *D.2217 = D.2219;
> > > > iD.2146 = iD.2146 + 1;
> > > > <D.2207>:
> > > > if (iD.2146 <= 524287) goto <D.2208>; else goto
> > > <D.2209>;
> > > > <D.2209>:
> > > > }
> > > >...
> > > >
> > > >With this patch, the omp_data_i load and stores for sum are generated at
> > > entry
> > > >and exit:
> > > >...
> > > > {
> > > > .omp_data_iD.2201 = &.omp_data_arr.15D.2218;
> > > > D.2216 = .omp_data_iD.2201->sumD.2205;
> > > > sumD.2206 = *D.2216;
> > > > {
> > > > unsigned intD.9 iD.2146;
> > > >
> > > > iD.2146 = 0;
> > > > goto <D.2207>;
> > > > <D.2208>:
> > > > D.2217 = .omp_data_iD.2201->cD.2203;
> > > > c.9D.2176 = *D.2217;
> > > > D.2177 = (long unsigned intD.10) iD.2146;
> > > > D.2178 = D.2177 * 4;
> > > > D.2179 = c.9D.2176 + D.2178;
> > > > D.2180 = *D.2179;
> > > > sumD.2206 = D.2180 + sumD.2206;
> > > > iD.2146 = iD.2146 + 1;
> > > > <D.2207>:
> > > > if (iD.2146 <= 524287) goto <D.2208>; else goto
> > > <D.2209>;
> > > > <D.2209>:
> > > > }
> > > > *D.2216 = sumD.2206;
> > > > #pragma omp return
> > > > }
> > > >...
> > > >
> > > >
> > > >So, without the patch the reduction operation looks like this:
> > > >...
> > > > *(.omp_data_iD.2201->sumD.2205) = *(.omp_data_iD.2201->sumD.2205) +
> > > x
> > > >...
> > > >
> > > >And with this patch the reduction operation is simply:
> > > >...
> > > > sumD.2206 = sumD.2206 + x:
> > > >...
> > > >
> > > >OK for trunk?
> > I presume the reason you are trying to do that here is that otherwise
> > it happens too late? What you do is what loop store motion would
> > do.
>
> Richard,
>
> Thanks for the hint. I've built a reduction example:
> ...
> void __attribute__((noinline))
> f (unsigned int *__restrict__ a, unsigned int *__restrict__ sum, unsigned int
> n)
> {
> unsigned int i;
> for (i = 0; i < n; ++i)
> *sum += a[i];
> }...
> and observed that store motion of the *sum store is done by pass_loop_im,
> provided the *sum load is taken out of the the loop by pass_pre first.
That doesn't make much sense. Why is LIM not moving the *sum load?
Ah - if n == 0 the body may not be executed and thus a hoisted load
may trap? I suppose you rather need a loop header copying pass.
> So alternatively, we could use pass_pre and pass_loop_im to achieve the same
> effect.
>
> When trying out adding pass_pre as a part of the pass group pass_oacc_kernels,
> I found that also pass_copyprop was required to get parloops to recognize the
> reduction.
>
> Attached patch adds the pre pass to pass group pass_oacc_kernels.
>
> Bootstrapped and reg-tested in the same way as before.
>
> OK for trunk?
No, I don't think you want this.
Richard.
