tiny.simd.satWidenMulPairwiseAccumulate

1pub fn satWidenMulPairwiseAccumulate(2    comptime DW: type,3    a: anytype,4    b: @TypeOf(a),5    sum: DW.Vector,6) DW.Vector {7    if (comptime DW.Lane != i32 or sourceLane(@TypeOf(a)) != i16 or8        sourceLaneCount(@TypeOf(a)) != DW.lane_count * 2)9    {10        @compileError("satWidenMulPairwiseAccumulate requires paired i16 sources with i32 output");11    }12    var result: DW.Vector = undefined;13    inline for (0..DW.lane_count) |index| {14        const value = @as(i64, sum[index]) +15            @as(i64, a[index * 2]) * @as(i64, b[index * 2]) +16            @as(i64, a[index * 2 + 1]) * @as(i64, b[index * 2 + 1]);17        result[index] = @intCast(std.math.clamp(18            value,19            @as(i64, std.math.minInt(i32)),20            @as(i64, std.math.maxInt(i32)),21        ));22    }23    return result;24}

1pub const satWidenMulPairwiseAccumulate = multiply.satWidenMulPairwiseAccumulate;