wehub-resource-sync
332 lines
11 KiB
Python
332 lines
11 KiB
Python
# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed 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.
|
|
|
|
import unittest
|
|
from unittest import mock
|
|
|
|
import numpy as np
|
|
|
|
import paddle
|
|
import paddle.nn.functional as F
|
|
from paddle.nn.functional.conv import (
|
|
_MEMORY_FORMAT_CHANNELS_LAST,
|
|
_MEMORY_FORMAT_CHANNELS_LAST_3D,
|
|
_MEMORY_FORMAT_CONTIGUOUS,
|
|
_cudnn_conv_suggest_memory_format,
|
|
)
|
|
|
|
|
|
@unittest.skipIf(
|
|
not paddle.is_compiled_with_cuda(),
|
|
"Skipping tests: CUDA is not available.",
|
|
)
|
|
class TestSlowConv2dDilated(unittest.TestCase):
|
|
def setUp(self):
|
|
# Save old flag states
|
|
self.old_flag_acc = paddle.get_flags(
|
|
['FLAGS_use_accuracy_compatible_kernel']
|
|
)
|
|
self.old_flag_disable = paddle.get_flags(['FLAGS_conv2d_disable_cudnn'])
|
|
|
|
self.place = paddle.CUDAPlace(0)
|
|
|
|
np.random.seed(2026)
|
|
paddle.seed(2026)
|
|
|
|
def tearDown(self):
|
|
# Restore flags
|
|
paddle.set_flags(self.old_flag_acc)
|
|
paddle.set_flags(self.old_flag_disable)
|
|
|
|
def _init_data(self, dtype, layout, with_bias):
|
|
groups = 1
|
|
N = 2
|
|
C_in = 4
|
|
C_out = 4
|
|
H, W = 16, 16
|
|
K = 3
|
|
|
|
if layout == "NCHW":
|
|
input_shape = [N, C_in, H, W]
|
|
else: # NHWC
|
|
input_shape = [N, H, W, C_in]
|
|
|
|
weight_shape = [C_out, C_in // groups, K, K]
|
|
|
|
np_x = np.random.randn(*input_shape).astype('float32')
|
|
np_w = np.random.randn(*weight_shape).astype('float32')
|
|
|
|
np_b = None
|
|
if with_bias:
|
|
np_b = np.random.randn(C_out).astype('float32')
|
|
|
|
return np_x, np_w, np_b, groups
|
|
|
|
def _run_op(
|
|
self, np_x, np_w, np_b, dtype, layout, groups, disable_cudnn_flag
|
|
):
|
|
paddle.set_flags({'FLAGS_conv2d_disable_cudnn': disable_cudnn_flag})
|
|
paddle.set_flags({'FLAGS_use_accuracy_compatible_kernel': 1})
|
|
|
|
x = paddle.to_tensor(np_x, place=self.place, dtype=dtype)
|
|
x.stop_gradient = False
|
|
|
|
w = paddle.to_tensor(np_w, place=self.place, dtype=dtype)
|
|
w.stop_gradient = False
|
|
|
|
b = None
|
|
if np_b is not None:
|
|
b = paddle.to_tensor(np_b, place=self.place, dtype=dtype)
|
|
b.stop_gradient = False
|
|
|
|
out = F.conv2d(
|
|
x,
|
|
w,
|
|
b,
|
|
stride=1,
|
|
padding=1,
|
|
dilation=2,
|
|
groups=groups,
|
|
data_format=layout,
|
|
)
|
|
|
|
loss = out.sum()
|
|
loss.backward()
|
|
|
|
return out.numpy()
|
|
|
|
def _check_implementation(self, dtype, layout="NCHW", with_bias=True):
|
|
np_x, np_w, np_b, groups = self._init_data(dtype, layout, with_bias)
|
|
self._run_op(
|
|
np_x, np_w, np_b, dtype, layout, groups, disable_cudnn_flag=1
|
|
)
|
|
|
|
# =================================================================
|
|
# Test Cases for Registered Types
|
|
# =================================================================
|
|
def test_fp16(self):
|
|
self._check_implementation('float16', layout="NCHW", with_bias=True)
|
|
self._check_implementation('float16', layout="NCHW", with_bias=False)
|
|
self._check_implementation('float16', layout="NHWC", with_bias=True)
|
|
self._check_implementation('float16', layout="NHWC", with_bias=False)
|
|
|
|
def test_fp32(self):
|
|
self._check_implementation('float32', layout="NCHW", with_bias=True)
|
|
self._check_implementation('float32', layout="NCHW", with_bias=False)
|
|
self._check_implementation('float32', layout="NHWC", with_bias=True)
|
|
self._check_implementation('float32', layout="NHWC", with_bias=False)
|
|
|
|
|
|
@unittest.skipIf(
|
|
not paddle.is_compiled_with_cuda(),
|
|
"Skipping tests: CUDA is not available.",
|
|
)
|
|
class TestSlowConv3dDilated(unittest.TestCase):
|
|
def setUp(self):
|
|
# Save old flag states
|
|
self.old_flag_acc = paddle.get_flags(
|
|
['FLAGS_use_accuracy_compatible_kernel']
|
|
)
|
|
self.old_flag_disable = paddle.get_flags(['FLAGS_conv3d_disable_cudnn'])
|
|
|
|
self.place = paddle.CUDAPlace(0)
|
|
|
|
np.random.seed(2026)
|
|
paddle.seed(2026)
|
|
|
|
def tearDown(self):
|
|
# Restore flags
|
|
paddle.set_flags(self.old_flag_acc)
|
|
paddle.set_flags(self.old_flag_disable)
|
|
|
|
def _init_data(self, dtype, layout, with_bias):
|
|
groups = 1
|
|
N = 2
|
|
C_in = 4
|
|
C_out = 4
|
|
D, H, W = 8, 8, 8
|
|
K = 3
|
|
|
|
if layout == "NCDHW":
|
|
input_shape = [N, C_in, D, H, W]
|
|
else: # NDHWC
|
|
input_shape = [N, D, H, W, C_in]
|
|
|
|
weight_shape = [C_out, C_in // groups, K, K, K]
|
|
|
|
np_x = np.random.randn(*input_shape).astype('float32')
|
|
np_w = np.random.randn(*weight_shape).astype('float32')
|
|
|
|
np_b = None
|
|
if with_bias:
|
|
np_b = np.random.randn(C_out).astype('float32')
|
|
|
|
return np_x, np_w, np_b, groups
|
|
|
|
def _run_op(
|
|
self, np_x, np_w, np_b, dtype, layout, groups, disable_cudnn_flag
|
|
):
|
|
paddle.set_flags({'FLAGS_conv3d_disable_cudnn': disable_cudnn_flag})
|
|
paddle.set_flags({'FLAGS_use_accuracy_compatible_kernel': 1})
|
|
|
|
x = paddle.to_tensor(np_x, place=self.place, dtype=dtype)
|
|
x.stop_gradient = False
|
|
|
|
w = paddle.to_tensor(np_w, place=self.place, dtype=dtype)
|
|
w.stop_gradient = False
|
|
|
|
b = None
|
|
if np_b is not None:
|
|
b = paddle.to_tensor(np_b, place=self.place, dtype=dtype)
|
|
b.stop_gradient = False
|
|
|
|
out = F.conv3d(
|
|
x,
|
|
w,
|
|
b,
|
|
stride=1,
|
|
padding=1,
|
|
dilation=2,
|
|
groups=groups,
|
|
data_format=layout,
|
|
)
|
|
|
|
loss = out.sum()
|
|
loss.backward()
|
|
|
|
return out.numpy()
|
|
|
|
def _check_implementation(self, dtype, layout="NCDHW", with_bias=True):
|
|
np_x, np_w, np_b, groups = self._init_data(dtype, layout, with_bias)
|
|
self._run_op(
|
|
np_x, np_w, np_b, dtype, layout, groups, disable_cudnn_flag=1
|
|
)
|
|
|
|
# =================================================================
|
|
# Test Cases for Registered Types
|
|
# =================================================================
|
|
def test_fp16(self):
|
|
self._check_implementation('float16', layout="NCDHW", with_bias=True)
|
|
self._check_implementation('float16', layout="NCDHW", with_bias=False)
|
|
self._check_implementation('float16', layout="NDHWC", with_bias=True)
|
|
self._check_implementation('float16', layout="NDHWC", with_bias=False)
|
|
|
|
def test_fp32(self):
|
|
self._check_implementation('float32', layout="NCDHW", with_bias=True)
|
|
self._check_implementation('float32', layout="NCDHW", with_bias=False)
|
|
self._check_implementation('float32', layout="NDHWC", with_bias=True)
|
|
self._check_implementation('float32', layout="NDHWC", with_bias=False)
|
|
|
|
|
|
@unittest.skipIf(
|
|
not paddle.is_compiled_with_cuda(), "CUDA is required for coverage test"
|
|
)
|
|
class TestCudnnConvCoverage(unittest.TestCase):
|
|
def setUp(self):
|
|
self.place = paddle.CUDAPlace(0)
|
|
|
|
def test_cudnn_conv_suggest_memory_format_coverage(self):
|
|
x_fp32 = paddle.randn([1, 3, 16, 16], dtype='float32')
|
|
w_fp32_4d = paddle.randn([3, 3, 3, 3], dtype='float32')
|
|
w_fp32_5d = paddle.randn([3, 3, 3, 3, 3], dtype='float32')
|
|
x_fp64 = paddle.cast(x_fp32, 'float64')
|
|
w_fp64_4d = paddle.cast(w_fp32_4d, 'float64')
|
|
|
|
self.assertEqual(
|
|
_cudnn_conv_suggest_memory_format(x_fp64, w_fp32_4d),
|
|
_MEMORY_FORMAT_CONTIGUOUS,
|
|
)
|
|
self.assertEqual(
|
|
_cudnn_conv_suggest_memory_format(x_fp32, w_fp64_4d),
|
|
_MEMORY_FORMAT_CONTIGUOUS,
|
|
)
|
|
|
|
with mock.patch(
|
|
'paddle.nn.functional.conv.get_cudnn_version', return_value=8500
|
|
):
|
|
self.assertEqual(
|
|
_cudnn_conv_suggest_memory_format(
|
|
x_fp32, w_fp32_4d, data_format="NHWC"
|
|
),
|
|
_MEMORY_FORMAT_CHANNELS_LAST,
|
|
)
|
|
|
|
self.assertEqual(
|
|
_cudnn_conv_suggest_memory_format(
|
|
x_fp32, w_fp32_5d, data_format="NHWC"
|
|
),
|
|
_MEMORY_FORMAT_CHANNELS_LAST_3D,
|
|
)
|
|
|
|
self.assertEqual(
|
|
_cudnn_conv_suggest_memory_format(
|
|
x_fp32, w_fp32_4d, data_format="NCHW"
|
|
),
|
|
_MEMORY_FORMAT_CONTIGUOUS,
|
|
)
|
|
|
|
with mock.patch(
|
|
'paddle.nn.functional.conv.get_cudnn_version', return_value=7000
|
|
):
|
|
self.assertEqual(
|
|
_cudnn_conv_suggest_memory_format(
|
|
x_fp32, w_fp32_4d, data_format="NHWC"
|
|
),
|
|
_MEMORY_FORMAT_CONTIGUOUS,
|
|
)
|
|
|
|
def test_is_cudnn_supported_coverage(self):
|
|
from paddle.nn.functional.conv import _is_cudnn_supported
|
|
|
|
x_gpu_fp16 = paddle.randn([1, 3, 8, 8, 8], dtype='float16').to(
|
|
self.place
|
|
)
|
|
x_cpu = paddle.randn([1, 3, 8, 8, 8], dtype='float16').cpu()
|
|
|
|
self.assertFalse(_is_cudnn_supported(x_gpu_fp16, None, "NCDHW", False))
|
|
|
|
self.assertFalse(_is_cudnn_supported(x_cpu, None, "NCDHW", True))
|
|
|
|
w_trivial = paddle.randn([3, 3, 1, 1, 1], dtype='float16')
|
|
w_non_trivial = paddle.randn([3, 3, 3, 3, 3], dtype='float16')
|
|
|
|
with mock.patch(
|
|
'paddle.nn.functional.conv.get_cudnn_version', return_value=91000
|
|
):
|
|
self.assertFalse(
|
|
_is_cudnn_supported(x_gpu_fp16, w_non_trivial, "NCDHW", True)
|
|
)
|
|
|
|
self.assertTrue(
|
|
_is_cudnn_supported(x_gpu_fp16, w_trivial, "NCDHW", True)
|
|
)
|
|
|
|
x_gpu_fp32 = paddle.randn([1, 3, 8, 8, 8], dtype='float32').to(
|
|
self.place
|
|
)
|
|
self.assertTrue(
|
|
_is_cudnn_supported(x_gpu_fp32, w_non_trivial, "NCDHW", True)
|
|
)
|
|
|
|
with mock.patch(
|
|
'paddle.nn.functional.conv.get_cudnn_version', return_value=85000
|
|
):
|
|
self.assertTrue(
|
|
_is_cudnn_supported(x_gpu_fp16, w_non_trivial, "NCDHW", True)
|
|
)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|