[GitHub] [tvm] adstraw commented on a change in pull request #8960: [Hexagon] Add contrib tests for blocked conv2d and maxpool2d

[GitBox <gi...@apache.org>]

adstraw commented on a change in pull request #8960:
URL: https://github.com/apache/tvm/pull/8960#discussion_r704845672



##########
File path: tests/python/contrib/test_hexagon/test_conv2d_blocked.py
##########
@@ -0,0 +1,474 @@
+# 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.
+
+import sys
+
+import tvm
+from tvm import te
+from tvm import topi
+from tvm.topi import testing
+from .infrastructure import ceildiv
+from .infrastructure import get_packed_layout
+from .infrastructure import build_and_run
+
+import numpy as np
+import pytest
+
+
+def conv2d_logical(
+    shape_nhwc,
+    shape_kcrs,
+    kernel_size,
+    stride,
+    padding,
+    dtype,
+    storage_scope="global",
+):
+    """
+    Conv2d TE wherein both input activation and filter tensors
+    are defined with their logical NHWC/OIHW shapes, respectively.
+    The packed physical layout for the activation and filter are:
+      Activation: nhwc8h8w32c
+      Filter: oihw8i32o4i
+    """
+    assert kernel_size == tuple(shape_kcrs[2:])
+
+    block_shape = 8, 8, 32

Review comment:
       Might be better to add a function e.g. `get_block_shape` here even if we don't think we will ever change these numbers.  Much easier to change / generalize later if needed.

##########
File path: tests/python/contrib/test_hexagon/conftest.py
##########
@@ -0,0 +1,46 @@
+# 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.
+
+""" Hexagon testing fixtures used to deduce testing argument
+    values from testing parameters """
+
+import tvm
+from .infrastructure import ceildiv
+
+
+@tvm.testing.fixture
+def shape_nhwc(batch, in_channel, in_size):
+    return (batch, in_size, in_size, in_channel)
+
+
+@tvm.testing.fixture
+def shape_oihw(out_channel, in_channel, kernel):
+    return (out_channel, in_channel, kernel, kernel)
+
+
+@tvm.testing.fixture
+def shape_oihw8i32o4i(out_channel, in_channel, kernel):
+    out_factor, in_factor, in_second_factor = 32, 32, 4
+    return (
+        int(ceildiv(out_channel, out_factor)),
+        int(ceildiv(in_channel, in_factor)),
+        kernel,
+        kernel,
+        in_factor // in_second_factor,

Review comment:
       in_first_factor

##########
File path: tests/python/contrib/test_hexagon/infrastructure.py
##########
@@ -0,0 +1,57 @@
+# 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.
+
+""" Hexagon testing infrastructure """
+
+import tvm
+import numpy
+
+
+def ceildiv(o, d):
+    return tvm.tir.floordiv(o + d - 1, d)
+
+
+def get_packed_layout(logical_shape_nhwc, block_shape, packed_C=True):
+    shape = [logical_shape_nhwc[0]]
+    off_h, off_w, off_c = block_shape
+    shape.append(ceildiv(logical_shape_nhwc[1], off_h))
+    shape.append(ceildiv(logical_shape_nhwc[2], off_w))
+    if packed_C:
+        shape.append(ceildiv(logical_shape_nhwc[3], off_c))
+        shape.extend(block_shape)
+    else:
+        shape.extend([off_h, off_w, logical_shape_nhwc[-1]])

Review comment:
       Wondering why `-1` was used here instead of `3`.  Concerned about the size of `logical_shape_nhwc` being != 4?  Should this be an assert?

##########
File path: tests/python/contrib/test_hexagon/infrastructure.py
##########
@@ -0,0 +1,57 @@
+# 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.
+
+""" Hexagon testing infrastructure """
+
+import tvm
+import numpy
+
+
+def ceildiv(o, d):
+    return tvm.tir.floordiv(o + d - 1, d)
+
+
+def get_packed_layout(logical_shape_nhwc, block_shape, packed_C=True):
+    shape = [logical_shape_nhwc[0]]
+    off_h, off_w, off_c = block_shape
+    shape.append(ceildiv(logical_shape_nhwc[1], off_h))
+    shape.append(ceildiv(logical_shape_nhwc[2], off_w))
+    if packed_C:
+        shape.append(ceildiv(logical_shape_nhwc[3], off_c))
+        shape.extend(block_shape)
+    else:
+        shape.extend([off_h, off_w, logical_shape_nhwc[-1]])
+    return shape
+
+
+def build_and_run(inputs, func, target, target_host, *args, **kwargs):
+    s, placeholders, binds = func(*args, **kwargs)

Review comment:
       `s` is for schedule right?  I would prefer a more verbose name here but perhaps that's because I am not used to this naming convention.

##########
File path: tests/python/contrib/test_hexagon/test_conv2d_blocked.py
##########
@@ -0,0 +1,474 @@
+# 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.
+
+import sys
+
+import tvm
+from tvm import te
+from tvm import topi
+from tvm.topi import testing
+from .infrastructure import ceildiv
+from .infrastructure import get_packed_layout
+from .infrastructure import build_and_run
+
+import numpy as np
+import pytest
+
+
+def conv2d_logical(
+    shape_nhwc,
+    shape_kcrs,

Review comment:
       I agree with @csullivan that KCRS is better from a code perspective and avoids name conflicts.  However, OIHW is used throughout this PR and the linkage that KCRS == OIHW is not exceedingly clear.  Perhaps that could be made more explicit?

##########
File path: tests/python/contrib/test_hexagon/test_conv2d_blocked.py
##########
@@ -0,0 +1,474 @@
+# 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.
+
+import sys
+
+import tvm
+from tvm import te
+from tvm import topi
+from tvm.topi import testing
+from .infrastructure import ceildiv
+from .infrastructure import get_packed_layout
+from .infrastructure import build_and_run
+
+import numpy as np
+import pytest
+
+
+def conv2d_logical(
+    shape_nhwc,
+    shape_kcrs,
+    kernel_size,
+    stride,
+    padding,
+    dtype,
+    storage_scope="global",
+):
+    """
+    Conv2d TE wherein both input activation and filter tensors
+    are defined with their logical NHWC/OIHW shapes, respectively.
+    The packed physical layout for the activation and filter are:
+      Activation: nhwc8h8w32c
+      Filter: oihw8i32o4i
+    """
+    assert kernel_size == tuple(shape_kcrs[2:])
+
+    block_shape = 8, 8, 32
+    block_H, block_W, block_C = block_shape
+    shape = get_packed_layout(shape_nhwc, block_shape)
+    logical_output_shape = (

Review comment:
       General comment (or perhaps a question) : I am a little surprised that this sort of shape calculation is explicitly stated here.  This seems to be a pretty general "what is the logical shape of my convolution?" computation.  Should this not be in library code somewhere?  Perhaps this is by convention?  Comment / question applies to other shape / pad / etc. calculations below, I think.

##########
File path: tests/python/contrib/test_hexagon/conftest.py
##########
@@ -0,0 +1,46 @@
+# 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.
+
+""" Hexagon testing fixtures used to deduce testing argument
+    values from testing parameters """
+
+import tvm
+from .infrastructure import ceildiv
+
+
+@tvm.testing.fixture
+def shape_nhwc(batch, in_channel, in_size):
+    return (batch, in_size, in_size, in_channel)
+
+
+@tvm.testing.fixture
+def shape_oihw(out_channel, in_channel, kernel):
+    return (out_channel, in_channel, kernel, kernel)
+
+
+@tvm.testing.fixture
+def shape_oihw8i32o4i(out_channel, in_channel, kernel):
+    out_factor, in_factor, in_second_factor = 32, 32, 4

Review comment:
       I think it would be more clear match the order and numbers in the layout (8, 32, 4) with the factors as follows:
   
   ```
   in_first_factor, out_factor, in_second_factor = 8, 32, 4
   in_factor = in_first_factor * in_second_factor
   ```
   




-- 
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: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org