Implement subgroup reduction for SYCL RangePolicy parallel_reduce

[masterleinad]

Using subgroups for the reductions should improve the performance for parallel_reduce using SYCL.

[masterleinad]

With 154eff5 (#3940) I'm seeing

n = 128
AXPBY KK: 4.422993e-02 s 6.468522e-02 GB/s
AXPBY SYCL: 2.562794e-02 s 1.116369e-01 GB/s
DOT KK: 6.557975e-02 s 2.908442e-02 GB/s
DOT SYCL: 8.905854e-02 s 2.141680e-02 GB/s
n = 1024
AXPBY KK: 4.248038e-02 s 5.387943e-01 GB/s
AXPBY SYCL: 2.707055e-02 s 8.455012e-01 GB/s
DOT KK: 9.947761e-02 s 1.533892e-01 GB/s
DOT SYCL: 1.804862e-01 s 8.454271e-02 GB/s
n = 8192
AXPBY KK: 5.261190e-02 s 3.480305e+00 GB/s
AXPBY SYCL: 2.776235e-02 s 6.595460e+00 GB/s
DOT KK: 1.428609e-01 s 8.544697e-01 GB/s
DOT SYCL: 2.057009e-01 s 5.934358e-01 GB/s
n = 65536
AXPBY KK: 1.451227e-01 s 1.009383e+01 GB/s
AXPBY SYCL: 2.890844e-02 s 5.067184e+01 GB/s
DOT KK: 2.128002e-01 s 4.589105e+00 GB/s
DOT SYCL: 2.522376e-01 s 3.871598e+00 GB/s
n = 524288
AXPBY KK: 1.962914e-01 s 5.970078e+01 GB/s
AXPBY SYCL: 5.099575e-02 s 2.297986e+02 GB/s
DOT KK: 2.496850e-01 s 3.128943e+01 GB/s
DOT SYCL: 4.468903e-01 s 1.748192e+01 GB/s
n = 4194304
AXPBY KK: 2.972564e-01 s 3.153843e+02 GB/s
AXPBY SYCL: 1.551411e-01 s 6.042888e+02 GB/s
DOT KK: 3.195238e-01 s 1.956036e+02 GB/s
DOT SYCL: 5.425608e-01 s 1.151945e+02 GB/s
n = 33554432
AXPBY KK: 1.244910e+00 s 6.024532e+02 GB/s
AXPBY SYCL: 1.025914e+00 s 7.310552e+02 GB/s
DOT KK: 9.931421e-01 s 5.034526e+02 GB/s
DOT SYCL: 1.252469e+00 s 3.992115e+02 GB/s
n = 268435456
AXPBY KK: 8.666444e+00 s 6.923255e+02 GB/s
AXPBY SYCL: 8.056914e+00 s 7.447020e+02 GB/s
DOT KK: 5.673008e+00 s 7.050933e+02 GB/s
DOT SYCL: 6.063481e+00 s 6.596871e+02 GB/s

on a V100 so we are faster than the native implementation (at least for this simple test case).

[masterleinad]

I am happy with the current status and am looking for reviews.

[crtrott]

Besides the failing test, would it be practical to lift the reduction code for a group and subgroup out of here into its own file and call it from these places. It looks like its the exact same code for the three usecases (minus the copy to global for the team internal reduction). Otherwise this looks pretty clean, and I didn't see a mistake which would explain the failing tests.

[masterleinad]

Retest this please.

[masterleinad]

@crtrott I fixed the failing tests and factored common code out.

[masterleinad]

I'm going to assume that we will merge this more or less as is so I can base other improvements on top of this.

Change rested has been addressed