Merge AMD-HIP port

[VincentSC]

AMD has just released their port for MXnet to AMD GPUs, to be found here.
Please merge that code, so MXnet works on both NVidia and AMD hardware.

[piiswrong]

I took a quick look but it looks pretty preliminary. Have you verified that its actually working?

[VincentSC]

Nope, I was hinted on new additions to Github and reacted on that. So you could be totally correct on this.
We're going to benchmark Torch7, Caffe and the MXnet in the coming weeks, and then we would've found out. Torch7 and Caffe are full ports for sure, so I assumed MXnet would too - it would explain the lack of build-instructions.

The good part is that you can choose to integrate it gradually, or wait until it's finished.

[bensander]

Hi, this is Ben from AMD. The port is in the early stages of development as you note. Glad to see this issue since we wanted to make sure the community was aware of our plans and get some early feedback before we get too deep into the port. Let me provide a little background on the tools and what we'd like to do, and ask a couple questions:

HIP is an open-source toolkit designed to port CUDA code. After porting, the code can run on AMD ROCm GPU platform or still on the original CUDA (with same performance). HIP runtime and C++ language are familiar to CUDA programmers so the porting process is lightweight and does little damage to the original code. The HIP code is easy to maintain since it is close to CUDA, and provides full C++ support including templates, classes, namespaces, etc. HIP runtime code, headers, and porting tools are maintained on GitHub and updated every 2-4 weeks by AMD developers.

AMD is also developing an open-source optimized machine intelligence library called "MIOpen". This will contain convolution, pooling, and other kernels which have been optimized for AMD GPUs. The MIOpen host runtime API accepts HIP parameters for things like streams, and also defines Tensor and other structures for memory management. MIOpen is currently internal but will be open-sourced shortly.

We'd like to use these tools to enable MXNet to run on the ROCm platform. Overall strategy would be to port the CUDA code to HIP, and add the calls into MIOpen in the appropriate layers so it runs fast. The MXNet port is being developed in the open here.

One early design question we could use some help on: the port currently replaces the CUDA code with HIP code. This was an easy way to start since it requires less initial modification to the MXNet make system - we can just port the .cu files and start building. However, we likely want to instead use a design where the HIP files exist alongside the CUDA ones. What is best way to add a new HIP target to the MXNet code - where should we put the files, and what make changes are required?

Thanks! We look forward to participating in the MXNet community,
-ben

[mli]

I suggest adding a compilation flag USE_HIP. Then in makefile, add the following

ifeq ($(USE_HIP), 1)
	CFLAGS += -DMXNET_USE_HIP=1
	LDFLAGS += -lhip  # if there is a libhip.so
endif

So now we can switch between hip and cuda in source codes by

#if MXNET_USE_HIP == 1
// codes
#endif

[antinucleon]

Also you can modify the CMake building system

On Mon, Jun 5, 2017 at 09:42 Mu Li ***@***.***> wrote: I suggest adding a compilation flag USE_HIP. Then in makefile, add the following ifeq ($(USE_HIP), 1) CFLAGS += -DMXNET_USE_CUDNN=1 LDFLAGS += -lhip # if there is a libhip.so endif So now we can switch between hip and cuda in source codes by #if MXNET_USE_HIP == 1 // codes #endif — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#6257 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABM13n_XhAclw_3LYogGmd7zIjoemavWks5sBDARgaJpZM4NbXQy> .

-- Sent from mobile phone

[sriharikarnam]

Hi,

This is Srihari from AMD:
Above approach is one of the case we considered earlier, but this approach duplicates the code for every runtime API and other CUDA changes.
Need your inputs on how to confine HIP porting changes as a seperate backend alongside CUDA.
For example : Adding a new directory at mxnet/src/hip and confine HIP porting changes to this directory.

[sriharikarnam]

We are thinking of more modular approach, like introducing another flag MXNET_USE_GPU, this would have an inline function defined that will be common for both CUDA and HIP implementations, we would separate the inline function implementations into two different files specific to CUDA and HIP, based on the flag that is defined, the preprocessor would expand the inline function into the corresponding implementation. We think that this approach would help us in future to separate the implentation into different backends(HIP,CUDA,...). Please comment.

[tqchen]

That makes sense.

[yajiedesign]

This issue is closed due to lack of activity in the last 90 days. Feel free to reopen if this is still an active issue. Thanks!

[xiaoxinyin]

Just wonder if MXNet can be used with AMD GPUs now? Thanks!

[resure-tech]

Thank you, Shiwen.

…

On Sat, Oct 14, 2017 at 5:25 PM, Hu Shiwen ***@***.***> wrote: Reopened #6257 <#6257>. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#6257 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AZHmjvca2zRlT8d46yPOnZqa582bCbkxks5ssVEVgaJpZM4NbXQy> .

[sriharikarnam]

@xiaoxinyin This is Work In Progress as a separate package. Very soon we will discuss on refactoring the code and its abstractions(HIP backend) to start merging the code to master.

[szha]

@apache/mxnet-committers: This issue has been inactive for the past 90 days. It has no label and needs triage.

For general "how-to" questions, our user forum (and Chinese version) is a good place to get help.