This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
rsp push and rsp pull for comm device, used in kvstore('device') #8732
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -34,6 +34,7 @@ | |
| #include "gradient_compression.h" | ||
| #include "../ndarray/ndarray_function.h" | ||
| #include "../operator/tensor/sparse_retain-inl.h" | ||
| #include "./utils.h" | ||
| namespace mxnet { | ||
| namespace kvstore { | ||
| /** | ||
|
|
@@ -176,17 +177,17 @@ class CommCPU : public Comm { | |
| reduce[i] = buf.copy_buf[i]; | ||
| const_vars[i] = reduce[i].var(); | ||
| } | ||
| NDArray result = buf.merged; | ||
| Resource rsc = ResourceManager::Get()->Request(result.ctx(), | ||
| ResourceRequest(ResourceRequest::kTempSpace)); | ||
| Engine::Get()->PushAsync( | ||
| [reduce, result, rsc, this](RunContext rctx, Engine::CallbackOnComplete on_complete) { | ||
| NDArray out = result; | ||
| is_serial_push_? | ||
| ReduceSumCPUExSerial(reduce, &out) | ||
| : mxnet::ndarray::ElementwiseSum(rctx.get_stream<cpu>(), rsc, reduce, &out); | ||
| on_complete(); | ||
| }, Context::CPU(), const_vars, {result.var(), rsc.var}, | ||
| FnProperty::kCPUPrioritized, priority, PROFILER_MESSAGE("KVStoreReduce")); | ||
| } | ||
|
|
||
|
|
@@ -491,11 +492,7 @@ class CommDevice : public Comm { | |
|
|
||
| void Init(int key, const NDArrayStorageType stype, const TShape& shape, | ||
| int dtype = mshadow::kFloat32) override { | ||
| sorted_key_attrs_.push_back(std::make_tuple(key, shape, dtype, stype)); | ||
|
There was a problem hiding this comment. Using |
||
| } | ||
|
|
||
| void InitBuffersAndComm(const std::vector<NDArray>& src) { | ||
|
|
@@ -528,26 +525,42 @@ class CommDevice : public Comm { | |
| InitBuffersAndComm(src); | ||
| auto& buf = merge_buf_[key]; | ||
| std::vector<NDArray> reduce(src.size()); | ||
|
|
||
| const NDArrayStorageType stype = buf.merged.storage_type(); | ||
| if (stype == kDefaultStorage) { | ||
| CopyFromTo(src[0], &(buf.merged), priority); | ||
| reduce[0] = buf.merged; | ||
|
|
||
| if (buf.copy_buf.empty()) { | ||
| // TODO(mli) this results in large device memory usage for huge ndarray, | ||
| // such as the largest fullc in VGG. consider to do segment reduce with | ||
| // NDArray.Slice or gpu direct memory access. for the latter, we need to | ||
| // remove some ctx check, and also it reduces 20% perf | ||
| buf.copy_buf.resize(src.size()-1); | ||
| for (size_t i = 0; i < src.size()-1; ++i) { | ||
| buf.copy_buf[i] = NDArray( | ||
| buf.merged.shape(), buf.merged.ctx(), false, buf.merged.dtype()); | ||
| } | ||
| } | ||
| for (size_t i = 0; i < src.size()-1; ++i) { | ||
| CopyFromTo(src[i+1], &(buf.copy_buf[i]), priority); | ||
| reduce[i+1] = buf.copy_buf[i]; | ||
| } | ||
| } else { | ||
| if (buf.copy_buf.empty()) { | ||
| buf.copy_buf.resize(src.size()); | ||
| for (size_t j = 0; j < src.size(); ++j) { | ||
| buf.copy_buf[j] = NDArray( | ||
| buf.merged.storage_type(), buf.merged.shape(), buf.merged.ctx(), | ||
| true, buf.merged.dtype()); | ||
| } | ||
| } | ||
| for (size_t i = 0; i < src.size(); ++i) { | ||
| CopyFromTo(src[i], &(buf.copy_buf[i]), priority); | ||
| reduce[i] = buf.copy_buf[i]; | ||
| } | ||
| } | ||
| ElementwiseSum(reduce, &buf.merged, priority); | ||
| return buf.merged; | ||
| } | ||
|
|
||
|
|
@@ -621,7 +634,62 @@ class CommDevice : public Comm { | |
| const std::vector<std::pair<NDArray*, NDArray>>& dst, | ||
| const bool use_copy, | ||
| const int priority) override { | ||
| CHECK_EQ(src.storage_type(), kRowSparseStorage) | ||
| << "BroadcastRowSparse expects row-sparse src NDArray"; | ||
|
|
||
| // whether the indices are the same | ||
|
There was a problem hiding this comment. Is code is duplicated in |
||
| const bool is_same_rowid = CheckSameRowid(dst); | ||
| for (size_t i = 0; i < dst.size(); ++i) { | ||
| // the result can be copied to other devices without invoking sparse retain operator | ||
| // if the indices are the same | ||
| if (is_same_rowid && i != 0) { | ||
| CopyFromTo(*dst[0].first, dst[i].first, priority); | ||
| continue; | ||
| } | ||
|
|
||
| NDArray* out = dst[i].first; | ||
| NDArray row_id = dst[i].second; | ||
| if (use_copy) { | ||
| CopyFromTo(src, out, priority); | ||
| } else { | ||
| CHECK_EQ(out->storage_type(), kRowSparseStorage) | ||
| << "BroadcastRowSparse expects row_sparse dst NDArray"; | ||
|
|
||
| const bool is_diff_ctx = out->ctx() != src.ctx(); | ||
|
There was a problem hiding this comment. Are we assuming src is always on GPU? There was a problem hiding this comment. src is not assumed to be on gpu. Actually src is always on cpu. As you can see in https://github.com/apache/incubator-mxnet/blob/master/src/kvstore/kvstore_local.h#L233, src is There was a problem hiding this comment. That's true at the beginning. But as soon as you push some gradients on GPU, it copies the weight from pinned_ctx to GPU. See There was a problem hiding this comment. Nonetheless, I think performing sparse retain before the copy makes more sense since the source array is usually very large. |
||
| NDArray out_gpu = is_diff_ctx? NDArray(kRowSparseStorage, out->shape(), | ||
| src.ctx(), true, out->dtype(), out->aux_types()) : *out; | ||
|
|
||
| CHECK_EQ(row_id.ctx(), src.ctx()) | ||
| << "row_id and src are expected to be on the same context"; | ||
|
|
||
| Engine::Get()->PushAsync([=](RunContext rctx, Engine::CallbackOnComplete on_complete) { | ||
| NDArray temp = out_gpu; | ||
| const TBlob& indices = row_id.data(); | ||
| switch (temp.ctx().dev_mask()) { | ||
| case cpu::kDevMask: { | ||
| mxnet::common::SparseRetainOpForwardRspWrapper<cpu>(rctx.get_stream<cpu>(), | ||
| src, indices, kWriteTo, &temp); | ||
| break; | ||
| } | ||
| #if MXNET_USE_CUDA | ||
| case gpu::kDevMask: { | ||
| mxnet::common::SparseRetainOpForwardRspWrapper<gpu>(rctx.get_stream<gpu>(), | ||
| src, indices, kWriteTo, &temp); | ||
| // wait for GPU operations to complete | ||
| rctx.get_stream<gpu>()->Wait(); | ||
| break; | ||
| } | ||
|
There was a problem hiding this comment. is |
||
| #endif | ||
| default: LOG(FATAL) << MXNET_GPU_NOT_ENABLED_ERROR; | ||
| } | ||
| on_complete(); | ||
| }, out_gpu.ctx(), {src.var(), row_id.var()}, {out_gpu.var()}, | ||
| FnProperty::kNormal, priority, PROFILER_MESSAGE("KVStoreSparseRetain")); | ||
| if (is_diff_ctx) { | ||
| CopyFromTo(out_gpu, out, priority); | ||
| } | ||
| } | ||
| } | ||
| } | ||
|
|
||
| private: | ||
|
|
@@ -667,7 +735,7 @@ class CommDevice : public Comm { | |
| #endif | ||
| } | ||
|
|
||
| using KeyAttrs = std::tuple<int, TShape, int, NDArrayStorageType>; | ||
| // try to allocate buff on device evenly | ||
| void InitMergeBuffer(const std::vector<Context>& devs) { | ||
| std::sort(sorted_key_attrs_.begin(), sorted_key_attrs_.end(), []( | ||
|
|
@@ -681,8 +749,9 @@ class CommDevice : public Comm { | |
| } | ||
| for (size_t i = 0; i < sorted_key_attrs_.size(); ++i) { | ||
| int key = std::get<0>(sorted_key_attrs_[i]); | ||
|
|
||
| TShape shape = std::get<1>(sorted_key_attrs_[i]); | ||
|
|
||
| int type = std::get<2>(sorted_key_attrs_[i]); | ||
|
|
||
| const NDArrayStorageType stype = std::get<3>(sorted_key_attrs_[i]); | ||
| auto& buf = merge_buf_[key]; | ||
| Context ctx; | ||
| size_t min_size = std::numeric_limits<size_t>::max(); | ||
|
|
@@ -693,8 +762,12 @@ class CommDevice : public Comm { | |
| min_size = size; | ||
| } | ||
| } | ||
| if (stype == kDefaultStorage) { | ||
| buf.merged = NDArray(shape, ctx, false, type); | ||
| } else { | ||
| buf.merged = NDArray(stype, shape, ctx, true, type); | ||
| } | ||
| ctx_info[ctx.dev_id].second += shape.Size(); | ||
| } | ||
| inited_ = true; | ||
| } | ||
|
|
||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Let's make sure functions in
.hare documented. Should add some description forCastStorageDispatchtoo...