You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2020/01/24 18:51:26 UTC
[GitHub] [incubator-tvm] alexgl-github opened a new pull request #4775:
conv3d_nchw schedule
alexgl-github opened a new pull request #4775: conv3d_nchw schedule
URL: https://github.com/apache/incubator-tvm/pull/4775
Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from [Reviewers](https://github.com/apache/incubator-tvm/blob/master/CONTRIBUTORS.md#reviewers) by @ them in the pull request thread.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] alexgl-github commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
alexgl-github commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r373155323
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,112 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from tvm import autotvm
+from .. import generic, tag
+from ..nn.conv3d import conv3d, conv3d_ndhwc, conv3d_ncdhw
+from ..generic.nn import schedule_conv3d_ndhwc
+
+@autotvm.register_topi_compute(conv3d, 'cpu', ['direct'])
+def conv3d_x86(cfg, input, filter, strides, padding, dilation, layout='NCDHW', out_dtype=None):
+ if layout == 'NCDHW':
+ return conv3d_ncdhw(input, filter, strides, padding, dilation, out_dtype)
+ elif layout == 'NDHWC':
+ return conv3d_ndhwc(input, filter, strides, padding, dilation, out_dtype)
+
+@autotvm.register_topi_schedule(schedule_conv3d_ndhwc, 'cpu', ['direct'])
+def schedule_conv3d_ndhwc_x86(cfg, outs):
+ """TOPI schedule callback for conv2d
Review comment:
@vinx13 Fixed, thanks.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] vinx13 commented on a change in pull request #4775:
conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
vinx13 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r372586297
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,95 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv2d
+
+ Parameters
+ ----------
+ cfg: ConfigEntity
Review comment:
this argument does not exist
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r370873535
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,112 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from tvm import autotvm
+from .. import generic, tag
+from ..nn.conv3d import conv3d, conv3d_ndhwc, conv3d_ncdhw
+from ..generic.nn import schedule_conv3d_ndhwc
+
+@autotvm.register_topi_compute(conv3d, 'cpu', ['direct'])
+def conv3d_x86(cfg, input, filter, strides, padding, dilation, layout='NCDHW', out_dtype=None):
+ if layout == 'NCDHW':
+ return conv3d_ncdhw(input, filter, strides, padding, dilation, out_dtype)
+ elif layout == 'NDHWC':
+ return conv3d_ndhwc(input, filter, strides, padding, dilation, out_dtype)
+
+@autotvm.register_topi_schedule(schedule_conv3d_ndhwc, 'cpu', ['direct'])
+def schedule_conv3d_ndhwc_x86(cfg, outs):
+ """TOPI schedule callback for conv2d
+
+ Parameters
+ ----------
+ cfg: ConfigEntity
+ The config for this template
+
+ outs: Array of Tensor
+ The computation graph description of conv2d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv2d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def traverse(op):
+ """Traverse operators from computation graph"""
+ # inline all one-to-one-mapping operators except the last stage (output)
+ if tag.is_broadcast(op.tag):
+ if op not in s.outputs:
+ s[op].compute_inline()
+ else: # inject custom schedule
+ if len(op.axis) == 5: # schedule bias + bn + relu
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
Review comment:
Is this applicable?
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] apivovarov commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
apivovarov commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r370922945
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,112 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from tvm import autotvm
+from .. import generic, tag
+from ..nn.conv3d import conv3d, conv3d_ndhwc, conv3d_ncdhw
+from ..generic.nn import schedule_conv3d_ndhwc
+
+@autotvm.register_topi_compute(conv3d, 'cpu', ['direct'])
+def conv3d_x86(cfg, input, filter, strides, padding, dilation, layout='NCDHW', out_dtype=None):
+ if layout == 'NCDHW':
+ return conv3d_ncdhw(input, filter, strides, padding, dilation, out_dtype)
+ elif layout == 'NDHWC':
+ return conv3d_ndhwc(input, filter, strides, padding, dilation, out_dtype)
+
+@autotvm.register_topi_schedule(schedule_conv3d_ndhwc, 'cpu', ['direct'])
+def schedule_conv3d_ndhwc_x86(cfg, outs):
+ """TOPI schedule callback for conv2d
Review comment:
3d
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r370879636
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,112 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from tvm import autotvm
+from .. import generic, tag
+from ..nn.conv3d import conv3d, conv3d_ndhwc, conv3d_ncdhw
+from ..generic.nn import schedule_conv3d_ndhwc
+
+@autotvm.register_topi_compute(conv3d, 'cpu', ['direct'])
+def conv3d_x86(cfg, input, filter, strides, padding, dilation, layout='NCDHW', out_dtype=None):
+ if layout == 'NCDHW':
+ return conv3d_ncdhw(input, filter, strides, padding, dilation, out_dtype)
+ elif layout == 'NDHWC':
+ return conv3d_ndhwc(input, filter, strides, padding, dilation, out_dtype)
+
+@autotvm.register_topi_schedule(schedule_conv3d_ndhwc, 'cpu', ['direct'])
+def schedule_conv3d_ndhwc_x86(cfg, outs):
+ """TOPI schedule callback for conv2d
+
+ Parameters
+ ----------
+ cfg: ConfigEntity
+ The config for this template
+
+ outs: Array of Tensor
+ The computation graph description of conv2d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv2d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def traverse(op):
+ """Traverse operators from computation graph"""
+ # inline all one-to-one-mapping operators except the last stage (output)
+ if tag.is_broadcast(op.tag):
+ if op not in s.outputs:
+ s[op].compute_inline()
+ else: # inject custom schedule
+ if len(op.axis) == 5: # schedule bias + bn + relu
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
+ for tensor in op.input_tensors:
+ if isinstance(tensor.op, tvm.tensor.ComputeOp) and tensor.op not in scheduled_ops:
+ traverse(tensor.op)
+
+ if 'conv3d_ndhwc' in op.tag:
+ conv = op.output(0)
+ kernel = op.input_tensors[1]
+ if isinstance(kernel.op, tvm.tensor.ComputeOp) and "dilate" in kernel.op.tag:
+ s[kernel].compute_inline()
+
+ data = op.input_tensors[0]
+ data_pad = None
+ if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+ data_pad = data
+ data = data_pad.op.input_tensors[0]
+ n_pad, d_pad, h_pad, w_pad, c_pad = data_pad.op.axis
+ pad_fused = s[data_pad].fuse(h_pad, w_pad)
+ s[data_pad].parallel(pad_fused)
+
+ C = conv
+ # data axes
+ n, d, h, w, c = s[C].op.axis
+
+ if True:
+ # tile data h and w
+ ho, wo, hi, wi = s[C].tile(h, w, 2, 2)
+ # kernel axes
+ kd, ky, kx, kc = s[C].op.reduce_axis
+ kxi, kxo = s[C].split(kx, factor=2)
+ kci, kco = s[C].split(kc, factor=2)
+ #
+ s[C].reorder(n, d, ho, wo, hi, wi, c, kxo, kco, kxi, kci)
+ s[C].unroll(kci)
+
+ s[C].vectorize(c)
+ if op != output_op:
+ _, _, _, _, c_out = output_op.axis
+ s[C].compute_at(s[output_op], c_out)
Review comment:
Another lines to think more about - I think compute_at might be better at h axis.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r372133923
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,88 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv2d
+
+ Parameters
+ ----------
+ cfg: ConfigEntity
+ The config for this template
+
+ outs: Array of Tensor
+ The computation graph description of conv2d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv2d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def traverse(op):
+ """Traverse operators from computation graph"""
+ # inline all one-to-one-mapping operators except the last stage (output)
+ if tag.is_broadcast(op.tag):
+ if op not in s.outputs:
+ s[op].compute_inline()
+ else: # inject custom schedule
+ if len(op.axis) == 5: # schedule bias + bn + relu
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
+ for tensor in op.input_tensors:
+ if isinstance(tensor.op, tvm.tensor.ComputeOp) and tensor.op not in scheduled_ops:
+ traverse(tensor.op)
+
+ if 'conv3d_ndhwc' in op.tag:
+ conv = op.output(0)
+ kernel = op.input_tensors[1]
+ if isinstance(kernel.op, tvm.tensor.ComputeOp) and "dilate" in kernel.op.tag:
+ s[kernel].compute_inline()
+
+ data = op.input_tensors[0]
+ data_pad = None
+ if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+ data_pad = data
+ data = data_pad.op.input_tensors[0]
+ n_pad, d_pad, h_pad, w_pad, c_pad = data_pad.op.axis
+ pad_fused = s[data_pad].fuse(h_pad, w_pad)
+ s[data_pad].parallel(pad_fused)
+ C = conv
Review comment:
Some comments will be helpful here and there to pick it up later.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r372134009
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,88 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv2d
+
+ Parameters
+ ----------
+ cfg: ConfigEntity
+ The config for this template
+
+ outs: Array of Tensor
+ The computation graph description of conv2d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv2d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def traverse(op):
+ """Traverse operators from computation graph"""
+ # inline all one-to-one-mapping operators except the last stage (output)
+ if tag.is_broadcast(op.tag):
+ if op not in s.outputs:
+ s[op].compute_inline()
+ else: # inject custom schedule
+ if len(op.axis) == 5: # schedule bias + bn + relu
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
+ for tensor in op.input_tensors:
+ if isinstance(tensor.op, tvm.tensor.ComputeOp) and tensor.op not in scheduled_ops:
+ traverse(tensor.op)
+
+ if 'conv3d_ndhwc' in op.tag:
+ conv = op.output(0)
+ kernel = op.input_tensors[1]
+ if isinstance(kernel.op, tvm.tensor.ComputeOp) and "dilate" in kernel.op.tag:
+ s[kernel].compute_inline()
+
+ data = op.input_tensors[0]
+ data_pad = None
+ if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+ data_pad = data
+ data = data_pad.op.input_tensors[0]
+ n_pad, d_pad, h_pad, w_pad, c_pad = data_pad.op.axis
Review comment:
Suggest to use _ for unused axes.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] icemelon9 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
icemelon9 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r373715263
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,85 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+from ..util import traverse_inline
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv3d
+
+ Parameters
+ ----------
+ outs: Array of Tensor
+ The computation graph description of conv3d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv3d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def _traverse(op):
+ """Traverse operators from computation graph"""
+ if op in s.outputs and tag.is_broadcast(op.tag) and len(op.axis) == 5:
+ # schedule bias + bn + relu
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
+
+ if 'conv3d_ndhwc' in op.tag:
+ conv = op.output(0)
+ kernel = op.input_tensors[1]
+ # dilation stage
+ if isinstance(kernel.op, tvm.tensor.ComputeOp) and "dilate" in kernel.op.tag:
+ s[kernel].compute_inline()
+
+ # padding stage
+ data = op.input_tensors[0]
+ data_pad = None
+ if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+ # fuse pad h and w
+ data_pad = data
+ data = data_pad.op.input_tensors[0]
+ _, _, h_pad, w_pad, _ = data_pad.op.axis
+ pad_fused = s[data_pad].fuse(h_pad, w_pad)
+ s[data_pad].parallel(pad_fused)
+
+ # compute conv
+ C = conv
+ n, d, h, w, c = s[C].op.axis
+ s[C].vectorize(c)
+ if op != output_op: # fuse bias + bn + activation
+ _, _, _, _, c_out = output_op.axis
+ s[C].compute_at(s[output_op], c_out)
+ else:
+ # fuse batch, depth, height axes
+ fused = s[C].fuse(n, d, h)
+ s[C].parallel(fused)
+
+ scheduled_ops.append(op)
Review comment:
remove this line
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] icemelon9 commented on issue #4775: conv3d_ndhwc
schedule
Posted by GitBox <gi...@apache.org>.
icemelon9 commented on issue #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#issuecomment-580976944
Thanks @alexgl-github . This is now merged.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on issue #4775: conv3d_ndhwc
schedule
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on issue #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#issuecomment-579539128
@vinx13 @kevinthesun for further reviews.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] alexgl-github commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
alexgl-github commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r373714466
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,92 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv3d
+
+ Parameters
+ ----------
+ outs: Array of Tensor
+ The computation graph description of conv3d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv3d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def traverse(op):
+ """Traverse operators from computation graph"""
+ # inline all one-to-one-mapping operators except the last stage (output)
+ if tag.is_broadcast(op.tag):
+ if op not in s.outputs:
+ s[op].compute_inline()
+ else: # inject custom schedule
+ if len(op.axis) == 5:
+ # schedule bias + bn + activation
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
+ for tensor in op.input_tensors:
+ if isinstance(tensor.op, tvm.tensor.ComputeOp) and tensor.op not in scheduled_ops:
+ traverse(tensor.op)
+
+ if 'conv3d_ndhwc' in op.tag:
+ conv = op.output(0)
+ kernel = op.input_tensors[1]
+ # dilation stage
+ if isinstance(kernel.op, tvm.tensor.ComputeOp) and "dilate" in kernel.op.tag:
+ s[kernel].compute_inline()
+
+ # padding stage
+ data = op.input_tensors[0]
+ data_pad = None
+ if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+ # fuse pad h and w
+ data_pad = data
+ data = data_pad.op.input_tensors[0]
+ _, _, h_pad, w_pad, _ = data_pad.op.axis
+ pad_fused = s[data_pad].fuse(h_pad, w_pad)
+ s[data_pad].parallel(pad_fused)
+
+ # compute conv
+ C = conv
+ n, d, h, w, c = s[C].op.axis
+ s[C].vectorize(c)
+ if op != output_op: # fuse bias + bn + activation
+ _, _, _, _, c_out = output_op.axis
+ s[C].compute_at(s[output_op], c_out)
+ else:
+ # fuse batch, depth, height axes
+ fused = s[C].fuse(n, d, h)
+ s[C].parallel(fused)
+
+ scheduled_ops.append(op)
+
+ traverse(output_op)
Review comment:
@icemelon9 Changed to traverse_inline, thanks
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] vinx13 commented on a change in pull request #4775:
conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
vinx13 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r372740925
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,112 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from tvm import autotvm
+from .. import generic, tag
+from ..nn.conv3d import conv3d, conv3d_ndhwc, conv3d_ncdhw
+from ..generic.nn import schedule_conv3d_ndhwc
+
+@autotvm.register_topi_compute(conv3d, 'cpu', ['direct'])
+def conv3d_x86(cfg, input, filter, strides, padding, dilation, layout='NCDHW', out_dtype=None):
+ if layout == 'NCDHW':
+ return conv3d_ncdhw(input, filter, strides, padding, dilation, out_dtype)
+ elif layout == 'NDHWC':
+ return conv3d_ndhwc(input, filter, strides, padding, dilation, out_dtype)
+
+@autotvm.register_topi_schedule(schedule_conv3d_ndhwc, 'cpu', ['direct'])
+def schedule_conv3d_ndhwc_x86(cfg, outs):
+ """TOPI schedule callback for conv2d
Review comment:
@alexgl-github this comment hasn't been resolved
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] anijain2305 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
anijain2305 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r372133805
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,88 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv2d
+
+ Parameters
+ ----------
+ cfg: ConfigEntity
+ The config for this template
+
+ outs: Array of Tensor
+ The computation graph description of conv2d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv2d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def traverse(op):
+ """Traverse operators from computation graph"""
+ # inline all one-to-one-mapping operators except the last stage (output)
+ if tag.is_broadcast(op.tag):
+ if op not in s.outputs:
+ s[op].compute_inline()
+ else: # inject custom schedule
+ if len(op.axis) == 5: # schedule bias + bn + relu
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
+ for tensor in op.input_tensors:
+ if isinstance(tensor.op, tvm.tensor.ComputeOp) and tensor.op not in scheduled_ops:
+ traverse(tensor.op)
+
+ if 'conv3d_ndhwc' in op.tag:
+ conv = op.output(0)
+ kernel = op.input_tensors[1]
+ if isinstance(kernel.op, tvm.tensor.ComputeOp) and "dilate" in kernel.op.tag:
+ s[kernel].compute_inline()
+
+ data = op.input_tensors[0]
+ data_pad = None
+ if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+ data_pad = data
+ data = data_pad.op.input_tensors[0]
+ n_pad, d_pad, h_pad, w_pad, c_pad = data_pad.op.axis
+ pad_fused = s[data_pad].fuse(h_pad, w_pad)
+ s[data_pad].parallel(pad_fused)
+ C = conv
+ n, d, h, w, c = s[C].op.axis
+ s[C].vectorize(c)
+ if op != output_op:
+ n_out, d_out, h_out, w_out, c_out = output_op.axis
Review comment:
Suggest to use _ for unused axes.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] alexgl-github commented on issue #4775:
conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
alexgl-github commented on issue #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#issuecomment-579823237
> @alexgl-github Thanks for the contribution.
>
> Almost every PR needs to have a test case in Apache TVM project. In this case, it can be a test case that uses this schedule and compares accuracy with some golden reference.
There's already topi conv3d_ndhwc test using this schedule:
https://github.com/apache/incubator-tvm/blob/1b8522e475e9a889b80f069a83928cafa3502a74/topi/tests/python/test_topi_conv3d_ndhwc.py#L60
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] icemelon9 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
icemelon9 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r373643228
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,92 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv3d
+
+ Parameters
+ ----------
+ outs: Array of Tensor
+ The computation graph description of conv3d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv3d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
+
+ def traverse(op):
+ """Traverse operators from computation graph"""
+ # inline all one-to-one-mapping operators except the last stage (output)
+ if tag.is_broadcast(op.tag):
+ if op not in s.outputs:
+ s[op].compute_inline()
+ else: # inject custom schedule
+ if len(op.axis) == 5:
+ # schedule bias + bn + activation
+ n, d, h, w, c = op.axis
+ fused = s[op].fuse(n, d, h, w)
+ s[op].parallel(fused)
+ s[op].vectorize(c)
+ for tensor in op.input_tensors:
+ if isinstance(tensor.op, tvm.tensor.ComputeOp) and tensor.op not in scheduled_ops:
+ traverse(tensor.op)
+
+ if 'conv3d_ndhwc' in op.tag:
+ conv = op.output(0)
+ kernel = op.input_tensors[1]
+ # dilation stage
+ if isinstance(kernel.op, tvm.tensor.ComputeOp) and "dilate" in kernel.op.tag:
+ s[kernel].compute_inline()
+
+ # padding stage
+ data = op.input_tensors[0]
+ data_pad = None
+ if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+ # fuse pad h and w
+ data_pad = data
+ data = data_pad.op.input_tensors[0]
+ _, _, h_pad, w_pad, _ = data_pad.op.axis
+ pad_fused = s[data_pad].fuse(h_pad, w_pad)
+ s[data_pad].parallel(pad_fused)
+
+ # compute conv
+ C = conv
+ n, d, h, w, c = s[C].op.axis
+ s[C].vectorize(c)
+ if op != output_op: # fuse bias + bn + activation
+ _, _, _, _, c_out = output_op.axis
+ s[C].compute_at(s[output_op], c_out)
+ else:
+ # fuse batch, depth, height axes
+ fused = s[C].fuse(n, d, h)
+ s[C].parallel(fused)
+
+ scheduled_ops.append(op)
+
+ traverse(output_op)
Review comment:
You can use traverse_inline function to avoid redundant inlining broadcast ops.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] icemelon9 commented on a change in pull request
#4775: conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
icemelon9 commented on a change in pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#discussion_r373715242
##########
File path: topi/python/topi/x86/conv3d.py
##########
@@ -0,0 +1,85 @@
+# 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, no-else-return
+"""Conv3D operators"""
+import tvm
+from .. import generic, tag
+from ..util import traverse_inline
+
+@generic.schedule_conv3d_ndhwc.register("cpu")
+def schedule_conv3d_ndhwc(outs):
+ """TOPI schedule callback for conv3d
+
+ Parameters
+ ----------
+ outs: Array of Tensor
+ The computation graph description of conv3d
+ in the format of an array of tensors.
+
+ Returns
+ -------
+ s: Schedule
+ The computation schedule for conv3d.
+ """
+ s = tvm.create_schedule([x.op for x in outs])
+ output_op = outs[0].op
+ scheduled_ops = []
Review comment:
Remove this line since it's useless now
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] alexgl-github commented on issue #4775:
conv3d_ndhwc schedule
Posted by GitBox <gi...@apache.org>.
alexgl-github commented on issue #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775#issuecomment-579536676
@apivovarov
@anijain2305
Could you please review again?
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services
[GitHub] [incubator-tvm] icemelon9 merged pull request #4775: conv3d_ndhwc
schedule
Posted by GitBox <gi...@apache.org>.
icemelon9 merged pull request #4775: conv3d_ndhwc schedule
URL: https://github.com/apache/incubator-tvm/pull/4775
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services