The search space for conv2d x86 is defined as
cfg.define_split("tile_ic", in_channel, num_outputs=2)
cfg.define_split("tile_oc", num_filter, num_outputs=2)
cfg.define_split("tile_ow", ow, num_outputs=2, filter=lambda y: y.size[-1] <= 64,
policy="verbose")
if is_kernel_1x1:
cfg.define_knob("tile_oh", [1, 2] if oh > 1 else [1])
else:
cfg.define_knob("unroll_kw", [True, False]
as shown in
# 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.
# pylint: disable=invalid-name,unused-variable,unused-argument,no-member
# pylint: disable=no-value-for-parameter,import-outside-toplevel
"""Conv2D schedule on x86"""
show original
I have seen that “tile_ic” and “tile_oc” are used for the data layout transformation. However, I am not sure where “tile_ow” and “unroll_kw” are used. I also checked nn.con2d_NCHWc in
# 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.
# pylint: disable=invalid-name, unused-variable, too-many-locals
# pylint: disable=unused-argument, redefined-builtin
"""Conv2D operators"""
from __future__ import absolute_import as _abs
show original
but didn’t find anything either.
I would really appreciate any information you can provide on this issue
Thank you for your reply @comaniac . Does that mean that, when regular nn.conv2d is used, tile_ow is not used. And may be used for other type of convolutions such as conv2d_avx_1x1 ?
I think so, because conv2d_avx_1x1 (and some other schedules) shares the same compute function with conv2d.
1 Like
That is interesting, since that would make the search space bigger for
@autotvm.register_topi_compute("conv2d_NCHWc.x86")
def conv2d_NCHWc(cfg, data, kernel, strides, padding, dilation, layout, out_layout, out_dtype):
Since the operation performed for that schedule is nn.conv2d_NCHWc, which does not use tile_ow or unroll_kw.
It’s true, but it should be fine in my opinion, because the search space for X86 workloads are still acceptable in most cases.
1 Like
Hi @comaniac , I have one follow-up question if you are so kind. Do you know why we go from a 4-D to a 6-D vector when it comes to the kernel. I understand the N[C/c]HW[c] transformation, but I am having trouble understanding the one for the kernel (which is not 5-D but 6-D).
Is it perhaps the same idea but with the output filter dimension, since this number is big? followed by a big number in the second fastest dimension given by lower case “c” ?
Because kernel includes the channel of not only input but also output.
1 Like