enable ngraph context in mxnet #350
Conversation
src/ngraph/ngraph_graph.h
Outdated
| @@ -204,7 +201,8 @@ class Graph : public Node { | |||
| context_(context), | |||
| enable_fprop_cache(enable_fprop_cache) { | |||
| fprop_cache = std::make_shared<ngraph::FpropCache>(); | |||
| is_reuse_mem = context.dev_type != mxnet::Context::kNNP; | |||
| is_reuse_mem = context.dev_type == mxnet::Context::CPU().dev_type || | |||
There was a problem hiding this comment.
you can use enum directly, instead of creating default and getting its type. Also, we may want to reuse mem for "ngraph:cpu" (i guess we expect users to use mx.cpu(), but just in case they decide to use ngraph:cpu).
| @@ -459,6 +459,10 @@ void Compiler::DeepCopy(const nnvm::Graph& graph) { | |||
| } | |||
|
|
|||
| bool bad_type(const NodePtr& node) { | |||
| if (!ngraph_context_fallback() && | |||
There was a problem hiding this comment.
we need a test for this functionality, were you able to check by exporting env.
There was a problem hiding this comment.
Hmm, getting an error if I export the env variable, doesn't make much sense, will debug.
mbrookhart
changed the title
src/ngraph/ngraph_context.cc
Outdated
| std::pair<std::string, int32_t> NGraphContextFromDevID(int32_t dev_id) { | ||
| int32_t backend_num = dev_id >> 16; | ||
| int32_t device_num = dev_id - (backend_num << 16); | ||
| std::cout << "NGraphContextfromDevID " << backends.at(backend_num) << " " << device_num << " " << dev_id << std::endl; |
There was a problem hiding this comment.
may have to remove std::cout's.
include/mxnet/base.h
Outdated
| @@ -34,6 +34,8 @@ | |||
| #include <nnvm/op.h> | |||
| #include <nnvm/tuple.h> | |||
| #include <nnvm/symbolic.h> | |||
| // ngraph context | |||
| #include <mxnet/ngraph_context.h> | |||
There was a problem hiding this comment.
ngraph build guard ?
| @@ -150,14 +154,14 @@ struct Context { | |||
| * \return cpu::kDevMask or gpu::kDevMask | |||
| */ | |||
| inline DeviceType dev_mask() const { | |||
| if (dev_type == kCPUPinned || dev_type == kCPUShared || dev_type == kNNP) return kCPU; | |||
| if (dev_type == kCPUPinned || dev_type == kCPUShared || dev_type == kNGraph) return kCPU; | |||
There was a problem hiding this comment.
may give compile error if built without MXNET_USE_NGRAPH
There was a problem hiding this comment.
Made kNGraph always compile to avoid this, but throws an error if python attempts to initialize an nGraph context when mxnet wasn't compiled with nGraph.
|
It didn't build for me @ashokei . |
|
This is the error I saw when I built with 18.04 I built with 16.04 and I got something similar |
|
You need to update your version of nGraph, Adam Proctor broke something a few days ago. |
|
```julia
julia> x = mx.NDArray([1 2; 3 4; 5 6])
3×2 mx.NDArray{Int64,2} @ CPU0:
1 2
3 4
5 6
julia> size(x, 1, 2, 3, 4)
(3, 2, 1, 1)
```
Description
(Brief description on what this PR is about)
Checklist
Essentials
Please feel free to remove inapplicable items for your PR.
Changes
Comments