Skip to content

Commit

Permalink
Next prototype of the framework integration
Browse files Browse the repository at this point in the history
  • Loading branch information
makortel committed Aug 6, 2018
1 parent c2aba96 commit 741388b
Show file tree
Hide file tree
Showing 31 changed files with 1,867 additions and 0 deletions.
2 changes: 2 additions & 0 deletions HeterogeneousCore/CUDACore/BuildFile.xml
Original file line number Diff line number Diff line change
@@ -1,10 +1,12 @@
<use name="FWCore/Concurrency"/>
<use name="FWCore/Framework"/>
<use name="FWCore/ServiceRegistry"/>
<use name="FWCore/ParameterSet"/>
<use name="DataFormats/Provenance"/>
<use name="HeterogeneousCore/CUDAServices"/>
<use name="cuda"/>
<use name="cuda-api-wrappers"/>
<use name="root"/>

<export>
<lib name="1"/>
Expand Down
283 changes: 283 additions & 0 deletions HeterogeneousCore/CUDACore/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,283 @@
# Next iteration of the prototype for CMSSW interface to heterogeneous algorithms

## Introduction

The current prototype with `HeterogeneousEDProducer` and
`HeterogeneousProduct` is documented [here](../Producer/README.md).
The main differences wrt. that are
* Split device-specific code to different EDProducers
* Plug components together in the configuration

This page documents the CUDA integration, and discusses briefly on how
to extend to other devices. It will be extended if/when it gets
deployed and `HeterogeneousEDProducer` retired.

## Sub-packages
* [`CUDACore`](#cuda-integration) CUDA-specific core components
* [`CUDAServices`](../CUDAServices) Various edm::Services related to CUDA
* [`CUDAUtilities`](../CUDAUtilities) Various utilities for CUDA kernel code

# CUDA integration

## Choosing device

The device choosing logic is split to an EDProducer, an EDFilter, and
use of Paths in the configuration.

First, a `CUDADeviceChooser` EDProducer is run. It has the logic to
device whether the following chain of EDModules should run on a CUDA
device or not, and if yes, on which CUDA device. If it decides "yes",
it produces a `CUDAToken`, which contains the device id and a CUDA
stream. If it decides "no", it does not produce anything.

Next step is a `CUDADeviceFilter` EDFilter. It checks whether the
`CUDADeviceChooser` produced a product or not. If "yes", it returns
`true`, and if "no", it returns `false`.

Finally, the pieces need to be put together in the configuration. The
`CUDADeviceChooser` can be "anywhere", but the `CUDADeviceFilter`
should be the first module on a `cms.Path`, followed by the CUDA
EDProducers (in the future it may become sufficient to have only the
first EDProducer of a chain in the `Path`).
```python
process.fooCUDADevice = cms.EDProducer("CUDADeviceChooser")
process.fooCUDADeviceFilter = cms.EDFilter("CUDADeviceFilter",
src = cms.InputTag("fooCUDADevice")
)
process.fooCUDA = cms.EDProducer("FooProducerCUDA")
process.fooPathCUDA = cms.Path(
process.fooCUDADeviceFilter + process.fooCUDA
)
process.fooTask = cms.Task(
process.fooDevice
)
```

## Data model

The GPU data can be a single pointer to device data, or a class/struct
containing such pointers (among other stuff). When putting the data to
event, the data is wrapped to `CUDA<T>` template, which holds
* the GPU data
* must be movable, but no other restrictions (except need to be able to generate ROOT dictionaries from it)
* the current device where the data was produced, and the CUDA stream the data was produced with
* [CUDA event for synchronization between multiple CUDA streams](#synchronizing-between-cuda-streams)

Note that the `CUDA<T>` wrapper can be constructed only with
`CUDAScopedContext::wrap()`, and the data `T` can be obtained from it
only with `CUDAScopedContext::get()`, as described further below.

## CUDA EDProducer

### Class declaration

For time being (may disappear in the future) a CUDA producer should
inherit from `CUDAStreamEDProducer<...>`. The template parameters are
the usual
[stream producer extensions](https://twiki.cern.ch/twiki/bin/view/CMSPublic/FWMultithreadedFrameworkStreamModuleInterface#Template_Arguments).
Note that contrary to `HeterogeneousEDProducer`, the `ExternalWork`
extension is **not** implied.

```cpp
#include "HeterogeneousCore/CUDACore/interface/CUDAStreamEDProducer.h"
class FooProducerCUDA: public CUDAStreamEDProducer<> {
...
```
### Memory allocation
The only effect of the `CUDAStreamEDProducer` base class is that
`beginStream(edm::StreamID)` is replaced with
`beginStreamCUDA(edm::StreamID)`. This is done in order to set the
current CUDA device before the user code starts. **If the algorithm
has to allocate memory buffers for the duration of the whole job, the
recommended place is here.** Note that a CUDA stream is not passed to
the user code. If a CUDA stream is really needed, the developer should
create+synchronize it by him/herself. (although if this appears to be
common practice, we should try to provide the situation somehow)
### Setting the current device
A CUDA producer should read either `CUDAToken` (from
`CUDADeviceChooser`) or one or more `CUDA<T>` products. Then, in the
`acquire()`/`produce()`, it should construct `CUDAScopedContext` from
one of them
```cpp
// From CUDAToken
edm::Handle<CUDAToken> htoken;
iEvent.getByToken(srcToken_, htoken);
auto ctx = CUDAScopedContext(*htoken);
/// From CUDA<T>
edm::Handle<CUDA<GPUClusters> > handle;
iEvent.getByToken(srctoken_, handle);
auto ctx = CUDAScopedContext(*handle);
```

`CUDAScopedContext` works in the RAII way and does the following
* Sets the current device (for the scope) from `CUDAToken`/`CUDA<T>`
* Gives access to the CUDA stream the algorithm should use to queue asynchronous work
* Calls `edm::WaitingTaskWithArenaHolder::doneWaiting()` when necessary
* [Synchronizes between CUDA streams if necessary](#synchronizing-between-cuda-streams)
* Needed to get/put `CUDA<T>` from/to the event

In case of multiple input products, from possibly different CUDA
streams and/or CUDA devices, this approach gives the developer full
control in which of them the kernels of the algorithm should be run.

### Getting input

The real product (`T`) can be obtained from `CUDA<T>` only with the
help of `CUDAScopedContext`.

```cpp
edm::Handle<CUDA<GPUClusters> > hclus;
iEvent.getByToken(srctoken_, hclus);
GPUClusters const& clus = ctx.get(*hclus);
```

This step is needed to
* check that the data are on the same CUDA device
* if not, throw an exception (with unified memory could prefetch instead)
* if the CUDA streams are different, synchronize between them

### Calling the CUDA kernels

There is nothing special, except the CUDA stream can be obtained from
the `CUDAScopedContext`

```cpp
gpuAlgo.makeClustersAsync(..., ctx.stream());
```

### Putting output

The GPU data needs to be wrapped to `CUDA<T>` template with `CUDAScopedContest.wrap()`

```cpp
GPUClusters clusters = gpuAlgo.makeClustersAsync(..., ctx.stream());
std::unique_ptr<CUDA<GPUClusters> > ret = ctx.wrap(clusters);
iEvent.put(std::move(ret));

// or with one line
iEvent.put(ctx.wrap(gpuAlgo.makeClustersAsync(ctx.stream())));
```

This step is needed to
* store the current device and CUDA stream into `CUDA<T>`
* record the CUDA event needed for CUDA stream synchronization

### `ExternalWork` extension

Everything above works both with and without `ExternalWork`.

Without `ExternalWork` the `EDProducer`s act similar to TBB
flowgraph's "streaming node". I.e. they just queue more asynchronous
work in their `produce()`.

The `ExternalWork` is needed when one would otherwise call
`cudeStreamSynchronize()`, e.g. transferring something to CPU needed
for downstream DQM, or to queue more asynchronous work. With
`ExternalWork` an `acquire()` method needs to be implemented that gets
an `edm::WaitingTaskWithArenaHolder` parameter. The
`WaitingTaskWithArenaHolder` should then be passed to the constructor
of `CUDAScopedContext` along

```cpp
void acquire(..., edm::WaitingTaskWithArenaHolder waitingTaskHolder) {
edm::Handle<CUDA<GPUClusters> > handle;
iEvent.getByToken(token_, handle);
auto ctx = CUDAScopedContext(*handle, std::move(waitingTaskHolder)); // can also copy instead of move if waitingTaskHolder is needed for something else as well
...
```
When constructed this way, `CUDAScopedContext` registers a callback
function to the CUDA stream in its destructor to call
`waitingTaskHolder.doneWaiting()`.
A GPU->GPU producer needs a `CUDAScopedContext` also in its
`produce()`. Currently the best way is to read the input again in
`produce()` and construct the `CUDAScopedContext` from there. This
point will be improved.
### Transferring GPU data to CPU
The GPU->CPU data transfer needs synchronization to ensure the CPU
memory to have all data before putting that to the event. This means
the `ExternalWork` needs to be used along
* In `acquire()`
* (allocate CPU memory buffers)
* Queue all GPU->CPU transfers asynchronously
* In `produce()`
* If needed, read additional CPU products (e.g. from `edm::Ref`s)
* Reformat data back to legacy data formats
* Note: `CUDAScopedContext` is **not** needed in in `produce()`
### Synchronizing between CUDA streams
In case the producer needs input data that were produced in two (or
more) CUDA streams, these streams have to be synchronized (since CMSSW
framework no longer guarantees the synchronization as was the case
with `HeterogeneousEDProducer`). Here this synchronization is achieved
with CUDA events.
Each `CUDA<T>` constains also a CUDA event object. The call to
`CUDAScopedContext::wrap()` will *record* the event in the CUDA stream.
This means that when all work queued to the CUDA stream up to that
point has been finished, the CUDA event becomes *occurred*. Then, in
`CUDAScopedContext::get()`, if the `CUDA<T>` to get from has a
different CUDA stream than the `CUDAScopedContext`,
`cudaStreamWaitEvent(stream, event)` is called. This means that all
subsequent work queued to the CUDA stream will wait for the CUDA event
to become occurred. Therefore this subsequent work can assume that the
to-be-getted CUDA product exists.
## Configuration
```python
process.fooCPU = cms.EDProducer("FooProducer") # legacy CPU
process.fooCUDADevice = cms.EDProducer("CUDADeviceChooser")
process.fooCUDADeviceFilter = cms.EDFilter("CUDADeviceFilter",
src = cms.InputTag("fooCUDADevice")
)
process.fooCUDA = cms.EDProducer("FooProducerCUDA")
process.fooFromCUDA = cms.EDProducer("FooProducerCUDAtoCPU", src="fooCUDA")
process.foo = cms.EDProducer("FooProducerFallback",
src = cms.VInputTag("fooFromCUDA", "fooCPU")
)
process.fooPathCUDA = cms.Path(
process.fooCUDADeviceFilter + process.fooCUDA
)
process.fooPathCPU = cms.Path(
~process.fooCUDADeviceFilter + process.fooCPU
)
process.fooTask = cms.Task(
process.fooDevice,
process.fooFromCUDA,
process.foo
)
...
```
For a more complete example, see [here](test/testCUDA_cfg.py).

# Extension to other devices

The C++ side extends in a straightforward way. One has to add classes
similar to `CUDAToken`, `CUDA<T>`, and `CUDAScopedContext`. Of course,
much depends on the exact details. The python configuration side
extends as well, one "just" has to add more modules there. Also the
device choosing logic is also extendable
```python
process.fooCUDADevice = ...
process.fooFPGADevice = ...
process.fooPathCUDA = cms.Path(
process.fooCUDADeviceFilter + ...
)
process.fooPathFPGA = cms.Path(
~process.fooCUDADeviceFilter + process.fooFPGADeviceFilter + ...
)
process.fooPathCPU = cms.Path(
~process.fooCUDADeviceFilter + ~process.fooFPGADeviceFilter + ...
)
```
77 changes: 77 additions & 0 deletions HeterogeneousCore/CUDACore/interface/CUDA.h
Original file line number Diff line number Diff line change
@@ -0,0 +1,77 @@
#ifndef HeterogeneousCore_CUDACore_CUDA_h
#define HeterogeneousCore_CUDACore_CUDA_h

#include <optional>

#include <cuda/api_wrappers.h>

/**
* The purpose of this class is to wrap CUDA data to edm::Event in a
* way which forces correct use of various utilities.
*
* The non-default construction has to be done with CUDAScopedContext
* (in order to properly register the CUDA event).
*
* The default constructor is needed only for the ROOT dictionary generation.
*
* The CUDA event is in practice needed only for stream-stream
* synchronization, but someone with long-enough lifetime has to own
* it. Here is a somewhat natural place. If overhead is too much, we
* can e.g. make CUDAService own them (creating them on demand) and
* use them only where synchronization between streams is needed.
*/
template <typename T>
class CUDA {
public:
CUDA() = default; // Needed only for ROOT dictionary generation

CUDA(const CUDA&) = delete;
CUDA& operator=(const CUDA&) = delete;
CUDA(CUDA&&) = default;
CUDA& operator=(CUDA&&) = default;

bool isValid() const { return streamEvent_.get() != nullptr; }

int device() const { return device_; }

const cuda::stream_t<>& stream() const { return streamEvent_->stream; }
cuda::stream_t<>& stream() { return streamEvent_->stream; }

const cuda::event_t& event() const { return streamEvent_->event; }
cuda::event_t& event() { return streamEvent_->event; }

private:
friend class CUDAScopedContext;
friend class TestCUDA;

template <typename TokenOrContext>
explicit CUDA(T data, const TokenOrContext& token):
streamEvent_(std::make_unique<StreamEvent>(token)),
data_(std::move(data)),
device_(token.device())
{}

// Using unique_ptr to support the default constructor. Tried
// std::optional, but cuda::stream_t and cuda::event_t have their
// move assignment operators deleted. Use a struct to save one
// memory allocation.
public: // need to be public for ROOT dicrionary generation?
struct StreamEvent {
template <typename TokenOrContext>
explicit StreamEvent(const TokenOrContext& token):
stream(token.stream()),
event(cuda::event::create(token.device(),
cuda::event::sync_by_busy_waiting, // default; we should try to avoid explicit synchronization, so maybe the value doesn't matter much?
cuda::event::dont_record_timings)) // it should be a bit faster to ignore timings
{}

cuda::stream_t<> stream; // stream_t is just a handle, the real CUDA stream is owned by CUDAToken (with long-enough life time)
cuda::event_t event;
};
private:
std::unique_ptr<StreamEvent> streamEvent_;
T data_;
int device_ = -1;
};

#endif
Loading

0 comments on commit 741388b

Please sign in to comment.