Awesome

SYCL Reduce Primitive

This is a tiny package implementing what is a giant unmet need in SYCL2020 - proper reductions.

Want to sum a vector coming from every thread in a kernel launch? Want to accumulate a couple different kinds of diagnostic output from a kernel? Too bad. SYCL doesn't have full documentation on how span<> works, and you'll easily get lost writing your own undefined type reducer.

So, instead, try this:

#include <syclreduce/reduce.hpp>

...

namespace SR = syclreduce;

struct FourInts {
    int x[4];
    FourInts() {}
    FourInts(int i) : x{1, i,i,i} {}
	const int& operator[](int i) const { return x[i]; }
	int& operator[](int i) { return x[i]; }
};
struct ReduceFour {
    using T = FourInts; // What we're creating.

    // Initial value for all reduction variables.
    // This is not called often by the library, but
    // is an important part of the mathematics.
    void identity(FourInts &a) const {
        a[0] = 0; a[1] = 0;
        a[2] = ~((int)1<<(sizeof(int)*8-1));
        a[3] =   (int)1<<(sizeof(int)*8-1);
	}

	// How to combine two reduction results.
    // Must be associative.
    // Does not need to be commutative.
    void combine(FourInts &a, const FourInts &b) const {
        a[0] += b[0]; // count
        a[1] += b[1]; // sum
        a[2] = b[2] < a[2] ? b[2] : a[2]; // min
        a[3] = b[3] < a[3] ? a[3] : b[3]; // max
	}
};
SR::Reducer result{ ReduceFour() };

// Like parallel_for, but kernel functions all return
// a ReduceFour by value.
SR::parallel_reduce(cgh, sycl::nd_range({4096, 32}),
			        result, [=](sycl::nd_item<1> it) {
    const size_t tid = it.get_global_id(0);

    return FourInts(tid*31337 % 4792 + 101);
});

FourInts ans = result.get();

The reduction is done internally with the following steps:

Reducing the return results from every work group in parallel, and storing that in a unique position in an internal buffer.
Copying the buffer to the host and reducing over those results.

For a work group size of sycl::nd_range(12,4), this works out like:

out[0] = op(op(0,1), op(2,3)) out[1] = op(op(4,5), op(6,7)) out[2] = op(op(8,9), op(10,11))

result = op( op(out[0], out[1]) , out[2] )

Note that you should be careful about kernel launch sizes using this method. In particular, don't use more than, say, 4x the number of work groups as you have compute units. Otherwise you are wasting memory and storing more intermediate reduction results than you need to.

We're assuming you want already use sycl::nd_range to control launch sizes for this reason.

Installing

This is a header-only package. You can just copy it into an include directory like /usr/local/include/syclreduce/reduce.hpp.

If you also want to test and install the cmake package description, do the following:

mkdir build && cd build
cmake -DCMAKE_CXX_COMPILER=`which syclcc` \
      -DCMAKE_INSTALL_PREFIX=/usr/local \
      ..
make install

Awesome

SYCL Reduce Primitive

Installing

Copyright and License