tiny.simd.sumsOfShuffledQuadAbsDiff

1pub fn sumsOfShuffledQuadAbsDiff(2    comptime D: type,3    comptime index3: usize,4    comptime index2: usize,5    comptime index1: usize,6    comptime index0: usize,7    a: D.Vector,8    b: D.Vector,9) D.repartition(wideLane(D.Lane)).Vector {10    requireByteLane(D, "sumsOfShuffledQuadAbsDiff");11    if (index0 > 3 or index1 > 3 or index2 > 3 or index3 > 3) {12        @compileError("quad shuffle indices must be less than four");13    }14    const W = wideLane(D.Lane);15    const R = D.repartition(W);16    const selectors = [4]usize{ index0, index1, index2, index3 };17    var shuffled: [D.lane_count]D.Lane = @splat(0);18    inline for (0..(D.lane_count + 15) / 16) |block_index| {19        inline for (0..4) |group| {20            inline for (0..4) |byte| {21                const destination = block_index * 16 + group * 4 + byte;22                const source = block_index * 16 + selectors[group] * 4 + byte;23                if (destination < D.lane_count and source < D.lane_count) {24                    shuffled[destination] = a[source];25                }26            }27        }28    }29    var result: R.Vector = @splat(0);30    inline for (0..R.lane_count) |index| {31        const a_base = (index / 4) * 8 + (index & 3);32        const b_base = (index / 2) * 4;33        if (a_base + 3 < D.lane_count and b_base + 3 < D.lane_count) {34            var total: W = 0;35            inline for (0..4) |offset| {36                total += absoluteDifference(W, shuffled[a_base + offset], b[b_base + offset]);37            }38            result[index] = total;39        }40    }41    return result;42}

1pub const sumsOfShuffledQuadAbsDiff = reduce.sumsOfShuffledQuadAbsDiff;