mozga-intel commented on a change in pull request #20340:
URL: https://github.com/apache/incubator-mxnet/pull/20340#discussion_r649928303
##########
File path: src/operator/tensor/dot.cc
##########
@@ -111,6 +115,34 @@ NNVM_REGISTER_OP(_backward_dot)
.set_attr<FComputeEx>("FComputeEx<cpu>", DotBackwardEx<cpu>)
.add_arguments(DotParam::__FIELDS__());
+#if MXNET_USE_ONEDNN == 1
+static void BatchDotComputeExCPU(const nnvm::NodeAttrs& attrs,
+ const OpContext& ctx,
+ const std::vector<NDArray>& inputs,
+ const std::vector<OpReqType>& req,
+ const std::vector<NDArray>& outputs) {
+ if (SupportMKLDNNBatchDot(inputs, outputs[0])) {
+ MKLDNN_OPCHECK_INIT(false, outputs.size(), inputs, outputs);
+ MKLDNNRun(MKLDNNBatchDotForward, attrs, ctx, inputs, req, outputs);
+ MKLDNN_OPCHECK_RUN(BatchDotForward_<cpu>, attrs, ctx, inputs, req,
outputs);
+ return;
+ }
+ FallBackCompute(BatchDotForward_<cpu>, attrs, ctx, inputs, req, outputs);
+}
+
+inline static bool BatchDotStorageType(const nnvm::NodeAttrs& attrs,
Review comment:
I don't think static does anything useful here. Is this something we
need to account for?
##########
File path: src/operator/nn/mkldnn/mkldnn_batch_dot.cc
##########
@@ -0,0 +1,128 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file mkldnn_batch_dot.cc
+ */
+
+#if MXNET_USE_ONEDNN == 1
+
+#include "./mkldnn_batch_dot-inl.h"
+
+namespace mxnet {
+namespace op {
+
+bool SupportMKLDNNBatchDot(const std::vector<NDArray> &inputs, const NDArray
&output) {
+ return inputs[0].shape().Size() != 0 &&
+ inputs[1].shape().Size() != 0 &&
+ output.shape().Size() != 0 &&
+ (inputs[0].dtype() == mshadow::kFloat32 ||
+ inputs[0].dtype() == mshadow::kBfloat16);
+}
+
+void MKLDNNBatchDotForward(const nnvm::NodeAttrs &attrs,
+ const OpContext &ctx,
+ const std::vector<NDArray> &inputs,
+ const std::vector<OpReqType> &req,
+ const std::vector<NDArray> &outputs) {
+ const DotParam ¶m = nnvm::get<DotParam>(attrs.parsed);
+ MKLDNNBatchDotFwd &fwd = MKLDNNBatchDotFwd::GetCached(param, inputs,
outputs);
+ fwd.Execute(inputs, req, outputs);
+}
+
+MKLDNNBatchDotFwd &MKLDNNBatchDotFwd::GetCached(const DotParam ¶m,
+ const std::vector<NDArray>
&inputs,
+ const std::vector<NDArray>
&outputs) {
+ using batch_dot_fwd_map = std::unordered_map<BatchDotSignature,
MKLDNNBatchDotFwd, OpHash>;
+#if DMLC_CXX11_THREAD_LOCAL
+ static thread_local batch_dot_fwd_map fwds;
+#else
+ static MX_THREAD_LOCAL batch_dot_fwd_map fwds;
+#endif
+
+ BatchDotSignature key(param);
+ key.AddSign(inputs[0]);
+ key.AddSign(inputs[1]);
+ key.AddSign(outputs[0]);
+
+ auto it = fwds.find(key);
+ if (it == fwds.end()) {
+ const MKLDNNBatchDotFwd fwd(param, inputs, outputs);
+ it = AddToCache(&fwds, key, fwd);
+ }
+ return it->second;
+}
+
+MKLDNNBatchDotFwd::MKLDNNBatchDotFwd(const DotParam ¶m,
+ const std::vector<NDArray> &inputs,
+ const std::vector<NDArray> &outputs) {
+ auto shape = inputs[0].shape();
+ auto ndim = shape.ndim();
+ auto bigDim = shape[0];
+ for (size_t i = 1; i < ndim - 2; ++i) {
+ bigDim *= shape[i];
+ }
+
+ auto GetMemoryDesc = [&ndim, &bigDim](const NDArray& tensor, const bool
transpose) {
+ auto shape = tensor.shape();
+ if (transpose) {
+ return mkldnn::memory::desc(mkldnn::memory::dims{bigDim, shape[ndim -
1], shape[ndim - 2]},
+ get_mkldnn_type(tensor.dtype()),
+ mkldnn::memory::format_tag::acb);
+ } else {
+ return mkldnn::memory::desc(mkldnn::memory::dims{bigDim, shape[ndim -
2], shape[ndim - 1]},
+ get_mkldnn_type(tensor.dtype()),
+ mkldnn::memory::format_tag::any);
+ }
+ };
+
+ mkldnn::memory::desc data_md = GetMemoryDesc(inputs[0],
param.transpose_a);
+ mkldnn::memory::desc weights_md = GetMemoryDesc(inputs[1],
param.transpose_b);
+ mkldnn::memory::desc out_md({bigDim, data_md.dims()[1],
weights_md.dims()[2]},
+ get_mkldnn_type(outputs[0].dtype()),
+ mkldnn::memory::format_tag::any);
+ mkldnn::matmul::desc fwd_desc(data_md, weights_md, out_md);
+ fwd_pd = std::make_shared<batch_dot_fwd_pd_t>(fwd_desc,
mxnet::CpuEngine::Get()->get_engine());
+ fwd = std::make_shared<batch_dot_fwd_t>(*fwd_pd);
+}
+
+void MKLDNNBatchDotFwd::Execute(const std::vector<NDArray> &inputs,
+ const std::vector<OpReqType> &req,
+ const std::vector<NDArray> &outputs) {
Review comment:
These spaces seem to be unnecessary. Please keep the same standard.
##########
File path: src/operator/nn/mkldnn/mkldnn_batch_dot.cc
##########
@@ -0,0 +1,128 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file mkldnn_batch_dot.cc
+ */
+
+#if MXNET_USE_ONEDNN == 1
+
+#include "./mkldnn_batch_dot-inl.h"
+
+namespace mxnet {
+namespace op {
+
+bool SupportMKLDNNBatchDot(const std::vector<NDArray> &inputs, const NDArray
&output) {
+ return inputs[0].shape().Size() != 0 &&
+ inputs[1].shape().Size() != 0 &&
+ output.shape().Size() != 0 &&
+ (inputs[0].dtype() == mshadow::kFloat32 ||
+ inputs[0].dtype() == mshadow::kBfloat16);
+}
+
+void MKLDNNBatchDotForward(const nnvm::NodeAttrs &attrs,
+ const OpContext &ctx,
+ const std::vector<NDArray> &inputs,
+ const std::vector<OpReqType> &req,
+ const std::vector<NDArray> &outputs) {
+ const DotParam ¶m = nnvm::get<DotParam>(attrs.parsed);
+ MKLDNNBatchDotFwd &fwd = MKLDNNBatchDotFwd::GetCached(param, inputs,
outputs);
+ fwd.Execute(inputs, req, outputs);
+}
+
+MKLDNNBatchDotFwd &MKLDNNBatchDotFwd::GetCached(const DotParam ¶m,
+ const std::vector<NDArray>
&inputs,
+ const std::vector<NDArray>
&outputs) {
+ using batch_dot_fwd_map = std::unordered_map<BatchDotSignature,
MKLDNNBatchDotFwd, OpHash>;
+#if DMLC_CXX11_THREAD_LOCAL
+ static thread_local batch_dot_fwd_map fwds;
+#else
+ static MX_THREAD_LOCAL batch_dot_fwd_map fwds;
+#endif
+
+ BatchDotSignature key(param);
+ key.AddSign(inputs[0]);
+ key.AddSign(inputs[1]);
+ key.AddSign(outputs[0]);
+
+ auto it = fwds.find(key);
+ if (it == fwds.end()) {
+ const MKLDNNBatchDotFwd fwd(param, inputs, outputs);
+ it = AddToCache(&fwds, key, fwd);
+ }
+ return it->second;
+}
+
+MKLDNNBatchDotFwd::MKLDNNBatchDotFwd(const DotParam ¶m,
+ const std::vector<NDArray> &inputs,
+ const std::vector<NDArray> &outputs) {
+ auto shape = inputs[0].shape();
+ auto ndim = shape.ndim();
+ auto bigDim = shape[0];
+ for (size_t i = 1; i < ndim - 2; ++i) {
+ bigDim *= shape[i];
+ }
+
+ auto GetMemoryDesc = [&ndim, &bigDim](const NDArray& tensor, const bool
transpose) {
+ auto shape = tensor.shape();
+ if (transpose) {
+ return mkldnn::memory::desc(mkldnn::memory::dims{bigDim, shape[ndim -
1], shape[ndim - 2]},
+ get_mkldnn_type(tensor.dtype()),
+ mkldnn::memory::format_tag::acb);
+ } else {
+ return mkldnn::memory::desc(mkldnn::memory::dims{bigDim, shape[ndim -
2], shape[ndim - 1]},
+ get_mkldnn_type(tensor.dtype()),
+ mkldnn::memory::format_tag::any);
+ }
+ };
+
+ mkldnn::memory::desc data_md = GetMemoryDesc(inputs[0],
param.transpose_a);
+ mkldnn::memory::desc weights_md = GetMemoryDesc(inputs[1],
param.transpose_b);
+ mkldnn::memory::desc out_md({bigDim, data_md.dims()[1],
weights_md.dims()[2]},
+ get_mkldnn_type(outputs[0].dtype()),
+ mkldnn::memory::format_tag::any);
+ mkldnn::matmul::desc fwd_desc(data_md, weights_md, out_md);
+ fwd_pd = std::make_shared<batch_dot_fwd_pd_t>(fwd_desc,
mxnet::CpuEngine::Get()->get_engine());
+ fwd = std::make_shared<batch_dot_fwd_t>(*fwd_pd);
+}
+
+void MKLDNNBatchDotFwd::Execute(const std::vector<NDArray> &inputs,
+ const std::vector<OpReqType> &req,
+ const std::vector<NDArray> &outputs) {
Review comment:
I agree. Unfortunately, I have difficulty in finding the advantages of
this alignment. Doesn't the code get more unreadable after that?
##########
File path: src/operator/nn/mkldnn/mkldnn_batch_dot.cc
##########
@@ -0,0 +1,128 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file mkldnn_batch_dot.cc
+ */
+
+#if MXNET_USE_ONEDNN == 1
+
+#include "./mkldnn_batch_dot-inl.h"
+
+namespace mxnet {
+namespace op {
+
+bool SupportMKLDNNBatchDot(const std::vector<NDArray> &inputs, const NDArray
&output) {
+ return inputs[0].shape().Size() != 0 &&
+ inputs[1].shape().Size() != 0 &&
+ output.shape().Size() != 0 &&
+ (inputs[0].dtype() == mshadow::kFloat32 ||
+ inputs[0].dtype() == mshadow::kBfloat16);
+}
+
+void MKLDNNBatchDotForward(const nnvm::NodeAttrs &attrs,
+ const OpContext &ctx,
+ const std::vector<NDArray> &inputs,
+ const std::vector<OpReqType> &req,
+ const std::vector<NDArray> &outputs) {
+ const DotParam ¶m = nnvm::get<DotParam>(attrs.parsed);
+ MKLDNNBatchDotFwd &fwd = MKLDNNBatchDotFwd::GetCached(param, inputs,
outputs);
+ fwd.Execute(inputs, req, outputs);
+}
+
+MKLDNNBatchDotFwd &MKLDNNBatchDotFwd::GetCached(const DotParam ¶m,
+ const std::vector<NDArray>
&inputs,
+ const std::vector<NDArray>
&outputs) {
+ using batch_dot_fwd_map = std::unordered_map<BatchDotSignature,
MKLDNNBatchDotFwd, OpHash>;
+#if DMLC_CXX11_THREAD_LOCAL
+ static thread_local batch_dot_fwd_map fwds;
+#else
+ static MX_THREAD_LOCAL batch_dot_fwd_map fwds;
+#endif
+
+ BatchDotSignature key(param);
+ key.AddSign(inputs[0]);
+ key.AddSign(inputs[1]);
+ key.AddSign(outputs[0]);
+
+ auto it = fwds.find(key);
+ if (it == fwds.end()) {
+ const MKLDNNBatchDotFwd fwd(param, inputs, outputs);
+ it = AddToCache(&fwds, key, fwd);
+ }
+ return it->second;
+}
+
+MKLDNNBatchDotFwd::MKLDNNBatchDotFwd(const DotParam ¶m,
+ const std::vector<NDArray> &inputs,
+ const std::vector<NDArray> &outputs) {
+ auto shape = inputs[0].shape();
+ auto ndim = shape.ndim();
+ auto bigDim = shape[0];
+ for (size_t i = 1; i < ndim - 2; ++i) {
+ bigDim *= shape[i];
+ }
+
+ auto GetMemoryDesc = [&ndim, &bigDim](const NDArray& tensor, const bool
transpose) {
+ auto shape = tensor.shape();
+ if (transpose) {
+ return mkldnn::memory::desc(mkldnn::memory::dims{bigDim, shape[ndim -
1], shape[ndim - 2]},
+ get_mkldnn_type(tensor.dtype()),
+ mkldnn::memory::format_tag::acb);
+ } else {
+ return mkldnn::memory::desc(mkldnn::memory::dims{bigDim, shape[ndim -
2], shape[ndim - 1]},
+ get_mkldnn_type(tensor.dtype()),
+ mkldnn::memory::format_tag::any);
+ }
+ };
+
+ mkldnn::memory::desc data_md = GetMemoryDesc(inputs[0],
param.transpose_a);
+ mkldnn::memory::desc weights_md = GetMemoryDesc(inputs[1],
param.transpose_b);
+ mkldnn::memory::desc out_md({bigDim, data_md.dims()[1],
weights_md.dims()[2]},
+ get_mkldnn_type(outputs[0].dtype()),
+ mkldnn::memory::format_tag::any);
+ mkldnn::matmul::desc fwd_desc(data_md, weights_md, out_md);
+ fwd_pd = std::make_shared<batch_dot_fwd_pd_t>(fwd_desc,
mxnet::CpuEngine::Get()->get_engine());
+ fwd = std::make_shared<batch_dot_fwd_t>(*fwd_pd);
+}
+
+void MKLDNNBatchDotFwd::Execute(const std::vector<NDArray> &inputs,
+ const std::vector<OpReqType> &req,
+ const std::vector<NDArray> &outputs) {
Review comment:
Sure! Please run auto-code formatter to keep the same standard.
--
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:
[email protected]
