This is an automated email from the ASF dual-hosted git repository.
kparzysz pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new 832c7674fe [hexagon][testing] nonrandom tests (#12164)
832c7674fe is described below
commit 832c7674fef385887a2fb99b8530736a40dfc820
Author: Christian Convey <[email protected]>
AuthorDate: Sat Jul 23 12:45:45 2022 -0400
[hexagon][testing] nonrandom tests (#12164)
Add support for populating unit-test input tensors
with values other than random.
---
tests/python/contrib/test_hexagon/pytest_util.py | 66 ++++++++++++++++++++++
.../test_hexagon/topi/test_avg_pool2d_slice.py | 30 ++++++++--
2 files changed, 92 insertions(+), 4 deletions(-)
diff --git a/tests/python/contrib/test_hexagon/pytest_util.py
b/tests/python/contrib/test_hexagon/pytest_util.py
index fb28ebeb68..0264fc1bc2 100644
--- a/tests/python/contrib/test_hexagon/pytest_util.py
+++ b/tests/python/contrib/test_hexagon/pytest_util.py
@@ -59,6 +59,22 @@ def get_test_id(*test_params, test_param_descs:
List[Optional[str]] = None) -> s
val_str = "F"
need_prefix_separator = True
+ elif type(param_val) == TensorContentConstant:
+ val_str = f"const({param_val.elem_value})"
+ need_prefix_separator = True
+
+ elif type(param_val) == TensorContentDtypeMin:
+ val_str = "min"
+ need_prefix_separator = True
+
+ elif type(param_val) == TensorContentDtypeMax:
+ val_str = "max"
+ need_prefix_separator = True
+
+ elif type(param_val) == TensorContentRandom:
+ val_str = "random"
+ need_prefix_separator = True
+
else:
val_str = str(param_val)
need_prefix_separator = True
@@ -91,3 +107,53 @@ def get_multitest_ids(
get_test_id(*single_test_param_list, test_param_descs=param_descs)
for single_test_param_list in multitest_params_list
]
+
+
+def get_numpy_dtype_info(np_dtype_name: str) -> Union[np.finfo, np.iinfo]:
+ """
+ Return an appropriate 'np.iinfo' or 'np.finfo' object corresponding to
+ the specified dtype.
+ """
+ np_dtype = np.dtype(np_dtype_name)
+ kind = np_dtype.kind
+
+ if kind == "f":
+ return np.finfo(np_dtype_name)
+ elif kind == "i":
+ return np.iinfo(np_dtype_name)
+ else:
+ raise TypeError(
+ f"np_dtype_name ({np_dtype_name}) must indicate some
floating-point or integral data type"
+ )
+
+
+TensorContentConstant = collections.namedtuple("TensorContentConstant",
["elem_value"])
+TensorContentRandom = collections.namedtuple("TensorContentRandom", [])
+TensorContentDtypeMin = collections.namedtuple("TensorContentDtypeMin", [])
+TensorContentDtypeMax = collections.namedtuple("TensorContentDtypeMax", [])
+
+
+def create_populated_numpy_ndarray(
+ input_shape: Union[list, tuple], dtype: str, input_tensor_populator
+) -> np.ndarray:
+ """
+ Create a numpy tensor with the specified shape, dtype, and content.
+ """
+ itp = input_tensor_populator # just for brevity
+
+ if type(itp) == TensorContentConstant:
+ return np.full(tuple(input_shape), itp.elem_value, dtype=dtype)
+
+ elif type(itp) == TensorContentDtypeMin:
+ info = get_numpy_dtype_info(dtype)
+ return np.full(tuple(input_shape), info.min, dtype=dtype)
+
+ elif type(itp) == TensorContentDtypeMax:
+ info = get_numpy_dtype_info(dtype)
+ return np.full(tuple(input_shape), info.max, dtype=dtype)
+
+ elif type(itp) == TensorContentRandom:
+ return np.random.random(input_shape).astype(dtype)
+
+ else:
+ raise ValueError(f"Unexpected input_tensor_populator type:
{type(itp)}")
diff --git a/tests/python/contrib/test_hexagon/topi/test_avg_pool2d_slice.py
b/tests/python/contrib/test_hexagon/topi/test_avg_pool2d_slice.py
index 34e9b751b9..af60e0f2e0 100644
--- a/tests/python/contrib/test_hexagon/topi/test_avg_pool2d_slice.py
+++ b/tests/python/contrib/test_hexagon/topi/test_avg_pool2d_slice.py
@@ -18,7 +18,6 @@
import pytest
import numpy as np
from typing import *
-import collections
from tvm import te
import tvm.testing
@@ -27,7 +26,14 @@ from tvm.contrib.hexagon.build import HexagonLauncher
from tvm.contrib.hexagon.session import Session
import tvm.topi.hexagon.slice_ops as sl
from ..infrastructure import allocate_hexagon_array, transform_numpy
-from ..pytest_util import get_multitest_ids
+from ..pytest_util import (
+ get_multitest_ids,
+ create_populated_numpy_ndarray,
+ TensorContentConstant,
+ TensorContentRandom,
+ TensorContentDtypeMin,
+ TensorContentDtypeMax,
+)
input_layout = tvm.testing.parameter(
@@ -36,8 +42,8 @@ input_layout = tvm.testing.parameter(
@tvm.testing.fixture
-def input_np(input_shape, dtype):
- return np.random.random(input_shape).astype(dtype)
+def input_np(input_shape, dtype: str, input_tensor_populator):
+ return create_populated_numpy_ndarray(input_shape, dtype,
input_tensor_populator)
@tvm.testing.fixture
@@ -61,6 +67,7 @@ class TestAvgPool2dSlice:
"cnt_padded", # count_include_pad
"out_layout", # output_layout
None, # dtype
+ None, # input_tensor_populator
]
_multitest_params = [
@@ -74,6 +81,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 16, 16, 32],
@@ -85,6 +93,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 8, 8, 32],
@@ -96,6 +105,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
# Test non-one stride and dilation
(
@@ -108,6 +118,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 8, 8, 32],
@@ -119,6 +130,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 8, 8, 32],
@@ -130,6 +142,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
# Test non-zero padding
(
@@ -142,6 +155,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 8, 8, 32],
@@ -153,6 +167,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 8, 8, 32],
@@ -164,6 +179,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 8, 8, 32],
@@ -175,6 +191,7 @@ class TestAvgPool2dSlice:
True,
"nhwc-8h2w32c2w-2d",
"float16",
+ TensorContentRandom(),
),
# Test n11c-1024c-2d layout which will require input and output to
have different layout
(
@@ -187,6 +204,7 @@ class TestAvgPool2dSlice:
True,
"n11c-1024c-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 1, 1, 2048],
@@ -198,6 +216,7 @@ class TestAvgPool2dSlice:
True,
"n11c-1024c-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 1, 1, 2048],
@@ -209,6 +228,7 @@ class TestAvgPool2dSlice:
True,
"n11c-1024c-2d",
"float16",
+ TensorContentRandom(),
),
(
[1, 1, 1, 2048],
@@ -220,6 +240,7 @@ class TestAvgPool2dSlice:
True,
"n11c-1024c-2d",
"float16",
+ TensorContentRandom(),
),
]
@@ -236,6 +257,7 @@ class TestAvgPool2dSlice:
count_include_pad,
output_layout,
dtype,
+ input_tensor_populator,
) = tvm.testing.parameters(*_multitest_params, ids=_param_ids)
@tvm.testing.fixture
