tiny.simd.sumOfMulQuadAccumulate

1pub fn sumOfMulQuadAccumulate(2    comptime DW: type,3    a: anytype,4    b: anytype,5    sum: DW.Vector,6) DW.Vector {7    if (comptime sourceLaneCount(@TypeOf(a)) != DW.lane_count * 4 or8        sourceLaneCount(@TypeOf(b)) != DW.lane_count * 4)9    {10        @compileError("sumOfMulQuadAccumulate requires four source lanes per output lane");11    }12    var result = sum;13    inline for (0..DW.lane_count) |index| {14        inline for (0..4) |offset| {15            result[index] = addLane(DW.Lane, result[index], productLane(16                DW.Lane,17                a[index * 4 + offset],18                b[index * 4 + offset],19            ));20        }21    }22    return result;23}

1pub const sumOfMulQuadAccumulate = multiply.sumOfMulQuadAccumulate;