# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License.  You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied.  See the License for the
# specific language governing permissions and limitations
# under the License.
"""Unittests for tvm.script.parser.tirx"""

import pytest
import tvm_ffi

import tvm.testing
from tvm import ir, tirx
from tvm.script.parser import tirx as T


def test_tir_buffer_proxy():
    buffer_0 = T.Buffer((128, 128), "float32")
    assert (
        isinstance(buffer_0, tirx.Buffer)
        and list(buffer_0.shape) == [128, 128]
        and buffer_0.dtype == ir.PrimType("float32")
    )

    buffer_1 = T.Buffer((64, 64, 64), "int32")
    assert (
        isinstance(buffer_1, tirx.Buffer)
        and list(buffer_1.shape) == [64, 64, 64]
        and buffer_1.dtype == ir.PrimType("int32")
    )


def test_tir_ptr_proxy():
    ptr_0 = T.handle("int32", "global")
    assert (
        isinstance(ptr_0, tirx.Var)
        and isinstance(ptr_0.ty, ir.PointerType)
        and ptr_0.ty.element_type == ir.PrimType("int32")
        and ptr_0.ty.storage_scope == "global"
    )

    ptr_1 = T.handle("float32", "shared")
    assert (
        isinstance(ptr_1, tirx.Var)
        and isinstance(ptr_1.ty, ir.PointerType)
        and ptr_1.ty.element_type == ir.PrimType("float32")
        and ptr_1.ty.storage_scope == "shared"
    )


def test_tir_func_name():
    @T.prim_func(s_tir=True)
    def matmul(a: T.handle, b: T.handle, c: T.handle) -> None:
        A = T.match_buffer(a, [128, 128])
        B = T.match_buffer(b, [128, 128])
        C = T.match_buffer(c, [128, 128])
        for i, j, k in T.grid(128, 128, 128):
            with T.sblock("update"):
                vi, vj, vk = T.axis.remap("SSR", [i, j, k])
                C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

    assert matmul.__name__ == "matmul"
    assert matmul.attrs["global_symbol"] == "matmul"


def test_tir_func_private_attrs():
    @T.prim_func(private=True, s_tir=True)
    def matmul(a: T.handle, b: T.handle, c: T.handle) -> None:
        T.func_attr({"attr": "value"})
        A = T.match_buffer(a, [128, 128])
        B = T.match_buffer(b, [128, 128])
        C = T.match_buffer(c, [128, 128])
        for i, j, k in T.grid(128, 128, 128):
            with T.sblock("update"):
                vi, vj, vk = T.axis.remap("SSR", [i, j, k])
                C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

    assert "global_symbol" not in matmul.attrs


def test_tir_func_private_manual_global_symbol_fail():
    with pytest.raises(tvm.error.DiagnosticError):

        @T.prim_func(private=True, s_tir=True)
        def matmul(a: T.handle, b: T.handle, c: T.handle) -> None:
            T.func_attr({"global_symbol": "matmul"})
            A = T.match_buffer(a, [128, 128])
            B = T.match_buffer(b, [128, 128])
            C = T.match_buffer(c, [128, 128])
            for i, j, k in T.grid(128, 128, 128):
                with T.sblock("update"):
                    vi, vj, vk = T.axis.remap("SSR", [i, j, k])
                    C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

        # should not execute
        assert matmul.__name__ == "matmul"


def test_tir_macro_decorator_signature():
    @T.prim_func(private=True, s_tir=True)
    def evaluate0():
        T.evaluate(0)

    # Ok, no parentheses
    @T.inline
    def func1():
        T.evaluate(0)

    @T.prim_func(private=True, s_tir=True)
    def use1():
        func1()

    tvm.ir.assert_structural_equal(use1, evaluate0)

    # Ok, empty parentheses
    @T.inline()
    def func2():
        T.evaluate(0)

    @T.prim_func(private=True, s_tir=True)
    def use2():
        func2()

    tvm.ir.assert_structural_equal(use1, evaluate0)

    with pytest.raises(ValueError):
        # Wrong: non-keyword argument
        @T.inline(True)
        def func3():
            T.evaluate()


def test_tir_macro_signature():
    @T.inline
    def assign(i, *args, t1, **kwargs):
        vi, vj, vk = T.axis.remap("SSR", [i, args[0], args[1]])
        kwargs["t3"][vi, vj] = kwargs["t3"][vi, vj] + t1[vi, vk] * kwargs["t2"][vj, vk]

    @T.prim_func(private=True, s_tir=True)
    def matmul_w_macro(a: T.handle, b: T.handle, c: T.handle) -> None:
        A = T.match_buffer(a, [128, 128])
        B = T.match_buffer(b, [128, 128])
        C = T.match_buffer(c, [128, 128])
        for i, j, k in T.grid(128, 128, 128):
            with T.sblock("update"):
                assign(i, j, k, t1=A, t2=B, t3=C)

    @T.prim_func(private=True, s_tir=True)
    def matmul_no_macro(a: T.handle, b: T.handle, c: T.handle) -> None:
        A = T.match_buffer(a, [128, 128])
        B = T.match_buffer(b, [128, 128])
        C = T.match_buffer(c, [128, 128])
        for i, j, k in T.grid(128, 128, 128):
            with T.sblock("update"):
                vi, vj, vk = T.axis.remap("SSR", [i, j, k])
                C[vi, vj] = C[vi, vj] + A[vi, vk] * B[vj, vk]

    tvm.ir.assert_structural_equal(matmul_no_macro, matmul_w_macro)


def test_tir_macro_hygienic():
    x_value = 128

    @T.inline
    def static_capture(A, B):
        B[()] = A[x_value]

    @T.prim_func(private=True, s_tir=True)
    def use_hygienic(A: T.Buffer((1024,), "int32"), B: T.Buffer((), "int32")) -> None:
        for x_value in T.serial(10):
            static_capture(A, B)

    @T.prim_func(private=True, s_tir=True)
    def expected_hygienic(A: T.Buffer((1024,), "int32"), B: T.Buffer((), "int32")) -> None:
        for x_value in range(10):
            B[()] = A[128]

    tvm.ir.assert_structural_equal(use_hygienic, expected_hygienic)


def test_tir_inline_late_binding():
    """Inline defined inside prim_func uses LEGB late binding:
    it sees the current value of variables from its enclosing scope at call time."""

    @T.prim_func(private=True, s_tir=True)
    def use_late_binding(A: T.Buffer((1024,), "int32"), B: T.Buffer((), "int32")) -> None:
        for x_value in T.serial(10):

            @T.inline
            def capture(A, B):
                B[()] = A[x_value]

            capture(A, B)

    @T.prim_func(private=True, s_tir=True)
    def expected(A: T.Buffer((1024,), "int32"), B: T.Buffer((), "int32")) -> None:
        for x_value in range(10):
            B[()] = A[x_value]

    tvm.ir.assert_structural_equal(use_late_binding, expected)


def test_tir_macro_in_class():
    class Object:
        def __init__(self, x: T.Buffer):
            self.local_x = T.sblock_alloc_buffer(x.shape, x.dtype)

        @T.inline
        def load(self, x: T.Buffer):
            N, M = T.meta_var(self.local_x.shape)
            for i, j in T.grid(N, M):
                with T.sblock("update"):
                    vi, vj = T.axis.remap("SS", [i, j])
                    self.local_x[vi, vj] = x[vi, vj]

    @T.prim_func(private=True, s_tir=True)
    def func_w_macro(a: T.handle):
        A = T.match_buffer(a, [128, 128])
        o1 = T.meta_var(Object(A))
        o1.load(A)
        o2 = T.meta_var(Object(A))
        o2.load(o1.local_x)

    @T.prim_func(private=True, s_tir=True)
    def func_no_macro(a: T.handle):
        A = T.match_buffer(a, [128, 128])
        local_a = T.sblock_alloc_buffer([128, 128])
        for i, j in T.grid(128, 128):
            with T.sblock("update"):
                vi, vj = T.axis.remap("SS", [i, j])
                local_a[vi, vj] = A[vi, vj]
        local_b = T.sblock_alloc_buffer([128, 128])
        for i, j in T.grid(128, 128):
            with T.sblock("update"):
                vi, vj = T.axis.remap("SS", [i, j])
                local_b[vi, vj] = local_a[vi, vj]

    tvm.ir.assert_structural_equal(func_no_macro, func_w_macro)


def test_tir_starred_expression():
    dims = (128, 128)

    @T.prim_func(private=True, s_tir=True)
    def starred(a: T.handle) -> None:
        A = T.match_buffer(a, [128, *dims], "int32")
        for i, j, k in T.grid(128, *dims):
            A[i, j, k] = T.int32(1)

    @T.prim_func(private=True, s_tir=True)
    def non_starred(a: T.handle) -> None:
        A = T.match_buffer(a, [128, 128, 128], "int32")
        for i, j, k in T.grid(128, 128, 128):
            A[i, j, k] = T.int32(1)

    tvm.ir.assert_structural_equal(starred, non_starred)


def test_tir_starred_shape_expression():
    dims = (128, 128)

    @T.prim_func(private=True, s_tir=True)
    def starred(a: T.handle) -> None:
        A = T.match_buffer(a, [128, *dims], "int32")
        for i, j, k in T.grid(*A.shape):
            A[i, j, k] = T.int32(1)

    @T.prim_func(private=True, s_tir=True)
    def non_starred(a: T.handle) -> None:
        A = T.match_buffer(a, [128, 128, 128], "int32")
        for i, j, k in T.grid(128, 128, 128):
            A[i, j, k] = T.int32(1)

    tvm.ir.assert_structural_equal(starred, non_starred)


def test_tir_dynamic_for_loop():
    dims = (128, 128)

    @T.prim_func(private=True, s_tir=True)
    def starred(a: T.handle) -> None:
        A = T.match_buffer(a, [128, *dims], "int32")
        for iters in T.grid(*A.shape):
            A[iters] = T.int32(1)

    @T.prim_func(private=True, s_tir=True)
    def non_starred(a: T.handle) -> None:
        A = T.match_buffer(a, [128, 128, 128], "int32")
        for i, j, k in T.grid(128, 128, 128):
            A[i, j, k] = T.int32(1)

    tvm.ir.assert_structural_equal(starred, non_starred)


def test_tir_starred_for_loop():
    dims = (128, 128)

    @T.prim_func(private=True, s_tir=True)
    def starred(a: T.handle, b: T.handle):
        A = T.match_buffer(a, [*dims, 128], "int32")
        B = T.match_buffer(b, dims, "int32")
        for *spatial, reduction in T.grid(*A.shape):
            with T.sblock("reduce"):
                with T.init():
                    B[spatial] = T.int32(0)
                B[spatial] = B[spatial] + A[(*spatial, reduction)]

    @T.prim_func(private=True, s_tir=True)
    def non_starred(a: T.handle, b: T.handle):
        A = T.match_buffer(a, [128, 128, 128], "int32")
        B = T.match_buffer(b, [128, 128], "int32")
        for i, j, k in T.grid(128, 128, 128):
            with T.sblock("reduce"):
                with T.init():
                    B[i, j] = T.int32(0)
                B[i, j] = B[i, j] + A[i, j, k]

    tvm.ir.assert_structural_equal(starred, non_starred)


def test_tir_loop_steps():
    N = T.Var("N", "int32")

    @T.prim_func(private=True, s_tir=True)
    def loop_with_steps(
        A: T.Buffer((N,)), B: T.Buffer((N,)), C: T.Buffer((N,)), tid: T.int32, v: T.int32
    ):
        for i in T.serial(tid, N, step=2):
            C[i] = A[i] + B[i]
        for i in T.unroll(tid, N, step=3):
            C[i] = A[i] + B[i]
        for i in T.vectorized(tid, N, step=4):
            C[i] = A[i] + B[i]
        for i in T.parallel(tid, N, step=5):
            C[i] = A[i] + B[i]
        for i in T.serial(tid, N, step=v):
            C[i] = A[i] + B[i]

    stmts = loop_with_steps.body.seq
    assert stmts[0].step == 2
    assert stmts[1].step == 3
    assert stmts[2].step == 4
    assert stmts[3].step == 5
    assert stmts[4].step.name == "v"


def test_tir_empty_tuple_index():
    @T.inline
    def bar(val):
        T.evaluate(val)

    @T.prim_func(private=True, s_tir=True)
    def func_with_empty_tuple(A: T.Buffer((), "int32"), B: T.Buffer((), "int32")):
        bar(val=A[()])

    @T.prim_func(private=True, s_tir=True)
    def expected(A: T.Buffer((), "int32"), B: T.Buffer((), "int32")):
        T.evaluate(A[()])

    tvm.ir.assert_structural_equal(func_with_empty_tuple, expected)


def test_tir_builtin_expression():
    dims = (128, 128)

    @T.prim_func(private=True, s_tir=True)
    def with_builtin(a: T.handle) -> None:
        A = T.match_buffer(a, [len(dims), *dims], "int32")
        for i, j, k in T.grid(*A.shape):
            A[i, j, k] = T.int32(1 + len(A.shape))

    @T.prim_func(private=True, s_tir=True)
    def evaluated(A: T.Buffer((2, 128, 128), "int32")):
        for i, j, k in T.grid(2, 128, 128):
            A[i, j, k] = 4

    tvm.ir.assert_structural_equal(with_builtin, evaluated)


def test_thread_binding_dtype():
    @T.prim_func(private=True, s_tir=True)
    def func(A: T.Buffer((128, 128)), B: T.Buffer((128, 128))):
        for i in T.thread_binding(T.int64(128), "threadIdx.x"):
            for j in T.thread_binding(128, "threadIdx.y"):
                B[i, j] = A[i, j]

    loop_i = func.body
    loop_j = loop_i.body
    assert loop_i.loop_var.ty.dtype == "int64"
    assert loop_i.thread_binding.var.ty.dtype == "int64"
    assert loop_j.loop_var.ty.dtype == "int32"
    assert loop_j.thread_binding.var.ty.dtype == "int32"


def test_inferred_ty_with_prim_args():
    """A PrimFunc may have inferred Type"""

    @T.prim_func(s_tir=True)
    def func(M: T.int32, N: T.int32) -> T.int32:
        T.ret(M * N)

    expected = tvm.relax.FuncType(
        [
            tvm.ir.PrimType("int32"),
            tvm.ir.PrimType("int32"),
        ],
        tvm.ir.PrimType("int32"),
        purity=True,
    )
    tvm.ir.assert_structural_equal(func.ty, expected)


def test_inferred_ty_with_buffer_args():
    """PrimFunc buffer arguments are inferred as R.Tensor"""

    @T.prim_func(s_tir=True)
    def func(A: T.Buffer([16, 16], "float32"), B: T.Buffer([256], "int32")) -> T.float32:
        T.ret(T.float32(42.0))

    expected = tvm.relax.FuncType(
        [
            tvm.relax.TensorType([16, 16], "float32"),
            tvm.relax.TensorType([256], "int32"),
        ],
        tvm.ir.PrimType("float32"),
        purity=True,
    )
    tvm.ir.assert_structural_equal(func.ty, expected)


def test_inferred_ty_with_internal_allocation():
    """A pure function may still write to internal allocations.

    Whether a function writes to internal allocations is not a visible
    effect, and does not impact the purity of a function.
    """

    @T.prim_func(s_tir=True)
    def func(A: T.Buffer([16, 16], "float32")) -> T.float32:
        Sum = T.decl_buffer([], "float32")
        Sum[()] = 0.0
        for i, j in T.grid(16, 16):
            Sum[()] = Sum[()] + A[i, j]

        T.ret(Sum[()])

    expected = tvm.relax.FuncType(
        [
            tvm.relax.TensorType([16, 16], "float32"),
        ],
        tvm.ir.PrimType("float32"),
        purity=True,
    )
    tvm.ir.assert_structural_equal(func.ty, expected)


def test_inferred_ty_with_output_buffer():
    """A pure function may not write to an argument buffer

    If an argument buffer is written to, the function must be impure.
    """

    @T.prim_func(s_tir=True)
    def func(A: T.Buffer(16, "float32"), B: T.Buffer(16, "float32")):
        for i in range(16):
            B[i] = A[i]

    expected = tvm.relax.FuncType(
        [
            tvm.relax.TensorType([16], "float32"),
            tvm.relax.TensorType([16], "float32"),
        ],
        tvm.relax.TupleType([]),
        purity=False,
    )
    tvm.ir.assert_structural_equal(func.ty, expected)


def test_inferred_ty_with_dynamic_buffer():
    """The inferred Type may contain dynamic shapes"""

    @T.prim_func(s_tir=True)
    def func(a_handle: T.handle, b_handle: T.handle):
        M = T.int64()
        N = T.int64()
        A = T.match_buffer(a_handle, [M, N], "float32")
        B = T.match_buffer(b_handle, [M * N], "float32")
        for i, j in T.grid(M, N):
            B[i * N + j] = A[i, j]

    M = tvm.tirx.Var("M", "int64")
    N = tvm.tirx.Var("N", "int64")
    expected = tvm.relax.FuncType(
        [
            tvm.relax.TensorType([M, N], "float32"),
            tvm.relax.TensorType([M * N], "float32"),
        ],
        tvm.relax.TupleType([]),
        purity=False,
    )
    tvm.ir.assert_structural_equal(func.ty, expected)


def test_reinterpret_nop():
    """Test builtin reinterpret op"""

    @T.prim_func(s_tir=True)
    def func(A: T.Buffer((32,), "float32"), B: T.Buffer((32,), "float32")) -> None:
        T.func_attr({"global_symbol": "main"})
        for i in T.serial(0, 32):
            with T.sblock():
                vi = T.axis.remap("S", [i])
                B[vi] = T.reinterpret("float32", A[vi])

    @T.prim_func(s_tir=True)
    def expected(A: T.Buffer((32,), "float32"), B: T.Buffer((32,), "float32")) -> None:
        T.func_attr({"global_symbol": "main"})
        for i in T.serial(0, 32):
            with T.sblock():
                vi = T.axis.remap("S", [i])
                B[vi] = A[vi]

    tvm.ir.assert_structural_equal(func, expected)


def test_launch_thread_i64():
    """Test launching thread with int64"""

    @T.prim_func(s_tir=True)
    def func() -> None:
        blockIdx_x = T.launch_thread("blockIdx.x", T.int64(1))
        if blockIdx_x == T.int64(0):
            T.evaluate(T.int64(0))
        else:
            T.evaluate(T.int64(1))

    assert func.body.node.dom.min.ty.dtype == "int64"
    assert func.body.node.dom.extent.ty.dtype == "int64"


def test_deterministic_branch():
    """Test deterministic branch"""

    def create_func(predicate: bool):
        @T.prim_func(private=True, s_tir=True)
        def func() -> None:
            if predicate:
                T.evaluate(0)
            else:
                T.evaluate(1)

        return func

    def create_expected(value):
        @T.prim_func(private=True, s_tir=True)
        def expected() -> None:
            T.evaluate(value)

        return expected

    tvm.ir.assert_structural_equal(create_func(True), create_expected(0))
    tvm.ir.assert_structural_equal(create_func(False), create_expected(1))


def test_block_annotation_merge():
    def _to_dict(anno: tvm_ffi.container.Map):
        result = {}
        for k, v in anno.items():
            result[k] = _to_dict(v) if isinstance(v, tvm_ffi.container.Map) else v
        return result

    @T.prim_func(s_tir=True)
    def func0():
        with T.sblock():
            T.sblock_attr({"key1": "block1"})
            T.sblock_attr({"key2": "block2"})
            T.evaluate(0)

    assert _to_dict(func0.body.block.annotations) == {"key1": "block1", "key2": "block2"}

    @T.prim_func(s_tir=True)
    def func1():
        with T.sblock():
            T.sblock_attr({"key": {"key1": "block1"}})
            T.sblock_attr({"key": {"key2": "block2"}})
            T.evaluate(0)

    assert _to_dict(func1.body.block.annotations) == {"key": {"key1": "block1", "key2": "block2"}}

    @T.prim_func(s_tir=True)
    def func2():
        with T.sblock():
            T.sblock_attr({"key1": "block1"})
            T.sblock_attr({"key1": "block1"})
            T.evaluate(0)

    assert _to_dict(func2.body.block.annotations) == {"key1": "block1"}

    with pytest.raises(RuntimeError):

        @T.prim_func(s_tir=True)
        def func3():
            with T.sblock():
                T.sblock_attr({"key1": "block1"})
                T.sblock_attr({"key1": "block2"})
                T.evaluate(0)


def test_alloc_inside_block():
    @T.prim_func(private=True, s_tir=True)
    def func() -> None:
        with T.sblock():
            A = T.sblock_alloc_buffer([10], "float32")
            for i in T.serial(0, 10):
                B = T.sblock_alloc_buffer([10], "float32")
                for j in T.serial(0, 10):
                    B[j] = T.float32(j)
                    A[i] += B[j]

    @T.prim_func(private=True, s_tir=True)
    def expected() -> None:
        with T.sblock():
            A = T.sblock_alloc_buffer([10], "float32")
            B = T.sblock_alloc_buffer([10], "float32")
            for i, j in T.grid(10, 10):
                B[j] = T.float32(j)
                A[i] += B[j]

    tvm.ir.assert_structural_equal(func, expected)


def test_tir_macro_block_name_suffix():
    @T.inline
    def operation(A, idx):
        with T.sblock("op"):
            v = T.axis.remap("S", [idx])
            A[v] = A[v] * T.float32(2)

    @T.prim_func(private=True, s_tir=True)
    def func_w_macro(a: T.handle) -> None:
        A = T.match_buffer(a, [10])
        for i in T.serial(0, 10):
            operation(A, i)
            operation(A, i)
            operation(A, i)

    @T.prim_func(private=True, s_tir=True)
    def expected(a: T.handle) -> None:
        A = T.match_buffer(a, [10])
        for i in T.serial(0, 10):
            with T.sblock("op"):
                v = T.axis.remap("S", [i])
                A[v] = A[v] * T.float32(2)
            with T.sblock("op_1"):
                v = T.axis.remap("S", [i])
                A[v] = A[v] * T.float32(2)
            with T.sblock("op_2"):
                v = T.axis.remap("S", [i])
                A[v] = A[v] * T.float32(2)

    tvm.ir.assert_structural_equal(func_w_macro, expected)


def test_ifexp():
    @T.prim_func(private=True, s_tir=True)
    def func(A: T.buffer((128, 128), "float32")):
        for i, j in T.grid(128, 128):
            A[i, j] = i if i < j else j

    @T.prim_func(private=True, s_tir=True)
    def expected(A: T.buffer((128, 128), "float32")):
        for i, j in T.grid(128, 128):
            A[i, j] = T.if_then_else(i < j, i, j)

    tvm.ir.assert_structural_equal(func, expected)


def test_sequence_compare():
    @T.prim_func(private=True, s_tir=True)
    def tir_func(A: T.Buffer((128, 128), "float32")):
        for i, j in T.grid(128, 128):
            if 0 < i < 128 and 0 < j < 128:
                A[i, j] = 1
            else:
                A[i, j] = 0

    @T.prim_func(private=True, s_tir=True)
    def expected(A: T.buffer((128, 128), "float32")):
        for i, j in T.grid(128, 128):
            if (0 < i and i < 128) and (0 < j and j < 128):
                A[i, j] = 1
            else:
                A[i, j] = 0

    tvm.ir.assert_structural_equal(tir_func, expected)


if __name__ == "__main__":
    tvm.testing.main()