[Runtime] Support arbitrary bitwidth integer in LLVM execution engine

[chhzh123]

Description

This PR ports the PR of HeteroCL #493 to support arbitrary bitwidth integers in LLVM execution engine.

Proposed Solutions

This PR constructs arrays with arbitrary bitwidth integers by creating customized numpy struct data type, packing and unpacking the results before and after the execution. Different signless integers have been tested. Unsigned types will be added in the next PR, since it still needs more annotations in the IR builder.

Examples

See the following example. Casting and (un)wrapping can be done automatically by the execution engine, so users have no need to worry about the data types, but just directly pass in the numpy arrays.

def test_arbitrary_bitwidth_gemm_alloc_output():
    M, N, K = 4, 4, 4
    T_IN, T_OUT = Int(5), Int(14)

    def gemm(A: T_IN[M, K], B: T_IN[K, N]) -> T_OUT[M, N]:
        C: T_OUT[M, N] = 0
        for i, j, k in allo.grid(M, N, K, name="C"):
            C[i, j] += A[i, k] * B[k, j]
        return C

    s = allo.customize(gemm)
    mod = s.build()
    np_A = np.random.randint(-4, 4, size=(M, K)).astype(np.int32)
    np_B = np.random.randint(-4, 4, size=(K, N)).astype(np.int32)
    np_C = np.matmul(np_A, np_B)
    np_C_allo = mod(np_A, np_B)
    np.testing.assert_allclose(np_C, np_C_allo, rtol=1e-5)

Checklist

• PR's title starts with a category (e.g. [Bugfix], [IR], [Builder], etc)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage (It would be better to provide ~2 different test cases to test the robustness of your code)
• Code is well-documented