Lunderberg commented on code in PR #13569:
URL: https://github.com/apache/tvm/pull/13569#discussion_r1082618777
##########
include/tvm/ir/function.h:
##########
@@ -63,6 +63,13 @@ enum class CallingConv : int {
* - Implementation: defined by device runtime(e.g. runtime/cuda)
*/
kDeviceKernelLaunch = 2,
+ /*!
+ * \brief For functions only called by other functions within the same IR
module.
+ *
+ * Indicates that various signature transformations (e.g. those provided by
the
+ * MakePackedAPI pass) are not desired.
+ */
+ kIntraModule = 3,
Review Comment:
I like the specification of the intent here, rather than just stating
"disable packed API" or something equivalent. I think I'd expand on the
statement that all callers of this function may be assumed to be within the
current module, specifically that this function provides no guarantees as to
function signature to end-users.
That way, it makes it clear that this isn't just a statement of the current
behavior, but also what sort of changes may be made in the future. (e.g. If a
parameter of an internal function is unused, it may be stripped from the
internal function and removed from all callsites, but this wouldn't be legal
for a user-facing function)
##########
tests/python/contrib/test_hexagon/test_call_tir.py:
##########
@@ -0,0 +1,248 @@
+# 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.
+
+""" Various tests related to the (WIP) support for having
+one PrimFunc call another PrimFunc within the same IRModule.
+"""
+
+from typing import List
+import pytest
+import numpy as np
+
+import tvm
+import tvm.testing
+import tvm.script
+from tvm.script import tir as T
+
+from tvm.contrib.hexagon.session import Session
+from tvm.contrib.hexagon import allocate_hexagon_array
+from .infrastructure import get_hexagon_target
+
+# NOTE(cconvey): These pylint warnings should be re-enabled as TVM's pylint
configuration matures.
+# pylint: disable=missing-function-docstring,no-self-argument,invalid-name
+# pylint: disable=redefined-outer-name,missing-class-docstring
+
+#
--------------------------------------------------------------------------------------------------
+# Test parameters
+#
--------------------------------------------------------------------------------------------------
+
+# The shape of the original (unsplit) tensors.
+# We assume that each shape describes a non-empty 2D tensor.
+original_shape = tvm.testing.parameter(
+ # degenerate cases...
+ [1, 1],
+ [1, 2],
+ [2, 1],
+ [2, 2],
+ # arbitrary, provided for variety
+ [5, 3],
+ [3, 5],
+)
+
+# This dtype is arbitrary, but it must match the dtype that's hardcoded into
the
+# callee's function signature. E.g., 'a_data: T.Ptr[T.int8]'.
+#
+# Hopefully over time we'll find a way to soften this limitation, at least for
+# some approaches to PrimFunc-to-PrimFunc calls.
+dtype = tvm.testing.parameter("int8")
+
+#
--------------------------------------------------------------------------------------------------
+# Helper functions / definitions...
+#
--------------------------------------------------------------------------------------------------
+
+HEXAGON_TARGET_ = get_hexagon_target("v69")
+
+ENTRY_PRIMFUNC_NAME_ = "main"
+
+
+def get_reference_input_tensor_(shape: list, dtype: str) -> np.array:
+ assert len(shape) == 2
+
+ a = np.ndarray(shape, dtype=dtype)
+ np_dtype = a.dtype
+
+ if np_dtype.kind in ["i", "u"]:
+ # We allow overflow for integer types because it tends to be
well-behaved
+ # and well-understood...
+ min_value = np.iinfo(np_dtype).min
+ max_value = np.iinfo(np_dtype).max
+
+ next_value = min_value
+
+ for i in range(shape[0]):
+ for j in range(shape[1]):
+ a[i, j] = next_value
+ next_value += 1
+
+ elif np_dtype.kind == "f":
+ # NOTE: For simplicity, we avoid test data that that require
+ # well-defined behavior on floating-point overflow.
+ # But it may be reasonable to test that in the future.
+ min_value = np.finfo(np_dtype).min
+ max_value = np.finfo(np_dtype).max
+
+ min_input_value = min_value / 2.0 + 1
+ max_input_value = max_value / 2.0 - 2
+ delta = (max_input_value - min_input_value) / (shape[0] * shape[1])
+
+ next_value = min_input_value
+
+ for i in range(shape[0]):
+ for j in range(shape[1]):
+ a[i, j] = next_value
+ next_value += delta
+
+ else:
+ assert False, f"Unexpected data type: {np_dtype}"
+
+ return a
+
+
+def get_reference_output_tensor_(shape: list, dtype: str) -> np.array:
Review Comment:
Instead of a function, should this be a fixture instead?
##########
tests/python/contrib/test_hexagon/test_call_tir.py:
##########
@@ -0,0 +1,248 @@
+# 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.
+
+""" Various tests related to the (WIP) support for having
+one PrimFunc call another PrimFunc within the same IRModule.
+"""
+
+from typing import List
+import pytest
+import numpy as np
+
+import tvm
+import tvm.testing
+import tvm.script
+from tvm.script import tir as T
+
+from tvm.contrib.hexagon.session import Session
+from tvm.contrib.hexagon import allocate_hexagon_array
+from .infrastructure import get_hexagon_target
+
+# NOTE(cconvey): These pylint warnings should be re-enabled as TVM's pylint
configuration matures.
+# pylint: disable=missing-function-docstring,no-self-argument,invalid-name
+# pylint: disable=redefined-outer-name,missing-class-docstring
+
+#
--------------------------------------------------------------------------------------------------
+# Test parameters
+#
--------------------------------------------------------------------------------------------------
+
+# The shape of the original (unsplit) tensors.
+# We assume that each shape describes a non-empty 2D tensor.
+original_shape = tvm.testing.parameter(
+ # degenerate cases...
+ [1, 1],
+ [1, 2],
+ [2, 1],
+ [2, 2],
+ # arbitrary, provided for variety
+ [5, 3],
+ [3, 5],
+)
+
+# This dtype is arbitrary, but it must match the dtype that's hardcoded into
the
+# callee's function signature. E.g., 'a_data: T.Ptr[T.int8]'.
+#
+# Hopefully over time we'll find a way to soften this limitation, at least for
+# some approaches to PrimFunc-to-PrimFunc calls.
+dtype = tvm.testing.parameter("int8")
+
+#
--------------------------------------------------------------------------------------------------
+# Helper functions / definitions...
+#
--------------------------------------------------------------------------------------------------
+
+HEXAGON_TARGET_ = get_hexagon_target("v69")
+
+ENTRY_PRIMFUNC_NAME_ = "main"
+
+
+def get_reference_input_tensor_(shape: list, dtype: str) -> np.array:
+ assert len(shape) == 2
+
+ a = np.ndarray(shape, dtype=dtype)
+ np_dtype = a.dtype
+
+ if np_dtype.kind in ["i", "u"]:
+ # We allow overflow for integer types because it tends to be
well-behaved
+ # and well-understood...
+ min_value = np.iinfo(np_dtype).min
+ max_value = np.iinfo(np_dtype).max
+
+ next_value = min_value
+
+ for i in range(shape[0]):
+ for j in range(shape[1]):
+ a[i, j] = next_value
+ next_value += 1
+
+ elif np_dtype.kind == "f":
+ # NOTE: For simplicity, we avoid test data that that require
+ # well-defined behavior on floating-point overflow.
+ # But it may be reasonable to test that in the future.
+ min_value = np.finfo(np_dtype).min
+ max_value = np.finfo(np_dtype).max
+
+ min_input_value = min_value / 2.0 + 1
Review Comment:
As an alternative, what about either of the following two methods?
```python
# Pseudo-random values on [-1,1], seeded to always get same values
np.random.seed(0)
np.random.uniform(-1,1, size=shape).astype(dtype)
# Sequential values on range [-1,1]
N = shape[0]*shape[1]
2*(np.arange(N).reshape(shape).astype(dtype) / N - 0.5)
```
##########
tests/python/contrib/test_hexagon/test_call_tir.py:
##########
@@ -0,0 +1,248 @@
+# 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.
+
+""" Various tests related to the (WIP) support for having
+one PrimFunc call another PrimFunc within the same IRModule.
+"""
+
+from typing import List
+import pytest
+import numpy as np
+
+import tvm
+import tvm.testing
+import tvm.script
+from tvm.script import tir as T
+
+from tvm.contrib.hexagon.session import Session
+from tvm.contrib.hexagon import allocate_hexagon_array
+from .infrastructure import get_hexagon_target
+
+# NOTE(cconvey): These pylint warnings should be re-enabled as TVM's pylint
configuration matures.
+# pylint: disable=missing-function-docstring,no-self-argument,invalid-name
+# pylint: disable=redefined-outer-name,missing-class-docstring
+
+#
--------------------------------------------------------------------------------------------------
+# Test parameters
+#
--------------------------------------------------------------------------------------------------
+
+# The shape of the original (unsplit) tensors.
+# We assume that each shape describes a non-empty 2D tensor.
+original_shape = tvm.testing.parameter(
+ # degenerate cases...
+ [1, 1],
+ [1, 2],
+ [2, 1],
+ [2, 2],
+ # arbitrary, provided for variety
+ [5, 3],
+ [3, 5],
+)
+
+# This dtype is arbitrary, but it must match the dtype that's hardcoded into
the
+# callee's function signature. E.g., 'a_data: T.Ptr[T.int8]'.
+#
+# Hopefully over time we'll find a way to soften this limitation, at least for
+# some approaches to PrimFunc-to-PrimFunc calls.
+dtype = tvm.testing.parameter("int8")
+
+#
--------------------------------------------------------------------------------------------------
+# Helper functions / definitions...
+#
--------------------------------------------------------------------------------------------------
+
+HEXAGON_TARGET_ = get_hexagon_target("v69")
+
+ENTRY_PRIMFUNC_NAME_ = "main"
+
+
+def get_reference_input_tensor_(shape: list, dtype: str) -> np.array:
+ assert len(shape) == 2
+
+ a = np.ndarray(shape, dtype=dtype)
+ np_dtype = a.dtype
+
+ if np_dtype.kind in ["i", "u"]:
+ # We allow overflow for integer types because it tends to be
well-behaved
+ # and well-understood...
+ min_value = np.iinfo(np_dtype).min
+ max_value = np.iinfo(np_dtype).max
+
+ next_value = min_value
+
+ for i in range(shape[0]):
+ for j in range(shape[1]):
+ a[i, j] = next_value
+ next_value += 1
+
+ elif np_dtype.kind == "f":
+ # NOTE: For simplicity, we avoid test data that that require
+ # well-defined behavior on floating-point overflow.
+ # But it may be reasonable to test that in the future.
+ min_value = np.finfo(np_dtype).min
+ max_value = np.finfo(np_dtype).max
+
+ min_input_value = min_value / 2.0 + 1
Review Comment:
The `min_value` for floating point numbers is the smallest positive number
that can be represented, so `min_value / 2.0` will always evaluate to zero.
Similarly, numbers that are anywhere close to `max_value` will create a very
large amount of rounding error, and probably won't be useful test inputs.
##########
tests/python/contrib/test_hexagon/test_call_tir.py:
##########
@@ -0,0 +1,248 @@
+# 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.
+
+""" Various tests related to the (WIP) support for having
+one PrimFunc call another PrimFunc within the same IRModule.
+"""
+
+from typing import List
+import pytest
+import numpy as np
+
+import tvm
+import tvm.testing
+import tvm.script
+from tvm.script import tir as T
+
+from tvm.contrib.hexagon.session import Session
+from tvm.contrib.hexagon import allocate_hexagon_array
+from .infrastructure import get_hexagon_target
+
+# NOTE(cconvey): These pylint warnings should be re-enabled as TVM's pylint
configuration matures.
+# pylint: disable=missing-function-docstring,no-self-argument,invalid-name
+# pylint: disable=redefined-outer-name,missing-class-docstring
+
+#
--------------------------------------------------------------------------------------------------
+# Test parameters
+#
--------------------------------------------------------------------------------------------------
+
+# The shape of the original (unsplit) tensors.
+# We assume that each shape describes a non-empty 2D tensor.
+original_shape = tvm.testing.parameter(
+ # degenerate cases...
+ [1, 1],
+ [1, 2],
+ [2, 1],
+ [2, 2],
+ # arbitrary, provided for variety
+ [5, 3],
+ [3, 5],
+)
+
+# This dtype is arbitrary, but it must match the dtype that's hardcoded into
the
+# callee's function signature. E.g., 'a_data: T.Ptr[T.int8]'.
+#
+# Hopefully over time we'll find a way to soften this limitation, at least for
+# some approaches to PrimFunc-to-PrimFunc calls.
+dtype = tvm.testing.parameter("int8")
+
+#
--------------------------------------------------------------------------------------------------
+# Helper functions / definitions...
+#
--------------------------------------------------------------------------------------------------
+
+HEXAGON_TARGET_ = get_hexagon_target("v69")
+
+ENTRY_PRIMFUNC_NAME_ = "main"
+
+
+def get_reference_input_tensor_(shape: list, dtype: str) -> np.array:
+ assert len(shape) == 2
+
+ a = np.ndarray(shape, dtype=dtype)
+ np_dtype = a.dtype
+
+ if np_dtype.kind in ["i", "u"]:
+ # We allow overflow for integer types because it tends to be
well-behaved
+ # and well-understood...
+ min_value = np.iinfo(np_dtype).min
+ max_value = np.iinfo(np_dtype).max
+
+ next_value = min_value
Review Comment:
Numpy has a couple functions that may be useful here. The array could be
generated as `np.arange(shape[0]*shape[1]).reshape(shape).astype(np_dtype)`.
For integer types, this would start at zero and wrap at the maximum value of
each int.
##########
tests/python/contrib/test_hexagon/test_call_tir.py:
##########
@@ -0,0 +1,248 @@
+# 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.
+
+""" Various tests related to the (WIP) support for having
+one PrimFunc call another PrimFunc within the same IRModule.
+"""
+
+from typing import List
+import pytest
+import numpy as np
+
+import tvm
+import tvm.testing
+import tvm.script
+from tvm.script import tir as T
+
+from tvm.contrib.hexagon.session import Session
+from tvm.contrib.hexagon import allocate_hexagon_array
+from .infrastructure import get_hexagon_target
+
+# NOTE(cconvey): These pylint warnings should be re-enabled as TVM's pylint
configuration matures.
+# pylint: disable=missing-function-docstring,no-self-argument,invalid-name
+# pylint: disable=redefined-outer-name,missing-class-docstring
+
+#
--------------------------------------------------------------------------------------------------
+# Test parameters
+#
--------------------------------------------------------------------------------------------------
+
+# The shape of the original (unsplit) tensors.
+# We assume that each shape describes a non-empty 2D tensor.
+original_shape = tvm.testing.parameter(
+ # degenerate cases...
+ [1, 1],
+ [1, 2],
+ [2, 1],
+ [2, 2],
+ # arbitrary, provided for variety
+ [5, 3],
+ [3, 5],
+)
+
+# This dtype is arbitrary, but it must match the dtype that's hardcoded into
the
+# callee's function signature. E.g., 'a_data: T.Ptr[T.int8]'.
+#
+# Hopefully over time we'll find a way to soften this limitation, at least for
+# some approaches to PrimFunc-to-PrimFunc calls.
+dtype = tvm.testing.parameter("int8")
+
+#
--------------------------------------------------------------------------------------------------
+# Helper functions / definitions...
+#
--------------------------------------------------------------------------------------------------
+
+HEXAGON_TARGET_ = get_hexagon_target("v69")
+
+ENTRY_PRIMFUNC_NAME_ = "main"
+
+
+def get_reference_input_tensor_(shape: list, dtype: str) -> np.array:
Review Comment:
Instead of a function, should this be a fixture instead? That way, it could
be cached across multiple tests that use the same shape and dtype.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
