5.1. Poplar Tutorial 1: Programs and Variables

Before starting this tutorial, take time to familiarise yourself with the IPU’s architecture by reading the IPU Programmer’s Guide. You can learn more about the Poplar programming model in the corresponding section of our documentation: Poplar and PopLibs User Guide: Poplar programming model.

In this tutorial you will:

learn about the structure of Graphcore’s low level C++ Poplar library for programming on the IPU.
learn how graphs, variables and programs can be used to execute computations on the IPU.
learn how streams can be used to exchange data efficiently between the host CPU and the IPU.
complete a small example program which communicates and adds data on the IPU.
optionally you will run this program on the IPU hardware.

A brief summary and a list of additional resources are included at the end this tutorial. Graphcore also provides tutorials using Python deep learning frameworks PyTorch and TensorFlow 2.

Setup

In order to run this tutorial on the IPU you will need to have a Poplar SDK environment enabled (see the Getting Started Guide for your IPU system).

You will also need a C++ toolchain compatible with the C++11 standard, build commands in this tutorial use GCC.

Using tut1_variables/start_here as your working directory, open tut1.cpp in a code editor. The file contains the outline of a C++ program with a main function, some Poplar library headers and the poplar namespace. In the rest of this tutorial, you will be adding code snippets to the main function at the indicated locations.

Graphs, variables and programs

Poplar programs are built of three main components:

a graph, which targets specific hardware devices.
variables, which are part of a graph and store the data on which an IPU can operate.
a program, which controls the operations applied to the graph and to its variables.

Creating the graph

All Poplar programs require a Graph object to construct the computation graph. Graphs are always created for a specific target (where the target is a description of the hardware being targeted, such as an IPU). To obtain the target we need to choose a device.

By default, these Poplar tutorials use a simulated target. As a result, they can run on any machine, even if it has no Graphcore hardware attached. On systems with Graphcore accelerator hardware, the header file poplar/DeviceManager.hpp contains API calls to enumerate and return Device objects for the attached hardware. The simulated devices are created with the IPUModel class, which mimics the functionality of an IPU on the host. The createDevice method creates a new virtual device to work with. Once we have this device, we can create a Graph object to target it.

Add the following code to the body of main:

// Create the IPU Model device
IPUModel ipuModel;
Device device = ipuModel.createDevice();
Target target = device.getTarget();

// Create the Graph object
Graph graph(target);

While the IPUModel provides a convenient way to build and debug Poplar programs without using IPU resources it is not a perfect representation of the hardware. As a result it is preferable to use an IPU if one is available. A description of the limitations of the IPUModel is provided in the Poplar developer guide. Instructions on how to use the hardware with this tutorial example is available in the last section of this tutorial: (Optional) Using the IPU.

Adding variables and mapping them to IPU tiles

Any program running on an IPU needs data to work on. These are defined as variables in the graph.

Add the following code to create the first variable in the program:

// Add variables to the graph
Tensor v1 = graph.addVariable(FLOAT, {4}, "v1");

This adds one vector variable with four elements of type float to the graph. The final string parameter, "v1", is used to identify the data in debugging/profiling tools.

Add three more variables:
- v2: another vector of 4 floats.
- v3: a two-dimensional 4x4 tensor of floats.
- v4: a vector of 10 integers (of type INT).

Note that the return type of addVariable is Tensor. The Tensor type represents data on the device in multi-dimensional tensor form. This type is used to reference the whole variable but, as we will see later, it can also be used to reference partial slices of variables, or data constructed from multiple variables.

Variables must be allocated to tiles. One option is to allocate the whole variable to one tile.

Add the following code:

// Allocate v1 to reside on tile 0
graph.setTileMapping(v1, 0);

Most of the time, programs actually deal with data spread over multiple tiles.

Add the following code:

// Spread v2 over tiles 0..3
for (unsigned i = 0; i < 4; ++i)
  graph.setTileMapping(v2[i], i);

This calls setTileMapping on sub-tensors of the variable v2 to spread it over multiple tiles.

Add code to allocate v3 and v4 to other tiles.

Adding the control program

Now that we have created some variables in the graph, we can create a control program to run on the device. Programs are represented as sub-classes of the Program class. In this example we will use the Sequence sub-class, which represents a number of steps executed sequentially.

Add this declaration:

// Create a control program that is a sequence of steps
program::Sequence prog;

// Debug print the tensor to the host console
prog.add(program::PrintTensor("v1-debug", v1));

Here, the sequence has one step that will perform a debug print (via the host) of the data on the device.

Now that we have a graph and a program, we can see what happens when it is deployed on the device. To do this we must first create an Engine object.

Add to the code:

// Create the engine
Engine engine(graph, prog);
engine.load(device);

This object represents the compiled graph and program, which are ready to run on the device.

Add the following code after the engine initialisation to run the control program:

// Run the control program
std::cout << "Running program\n";
engine.run(0);
std::cout << "Program complete\n";

Compiling the poplar executable

The first version of our main function is complete and ready to be compiled.

In a terminal, compile the host program (remembering to link in the Poplar library using the -lpoplar flag):
```
$ g++ --std=c++11 tut1.cpp -lpoplar -o tut1
```
Then run the compiled program:
```
$ ./tut1
```

When the program runs, the debug output prints out uninitialised values, because we allocated a variable in the graph which is never initialised or written to:

v1-debug: [0.0000000 0.0000000 0.0000000 0.0000000]

Initialising variables

One way to initialise data in the graph is to use constant values: unlike variables, constants are set in the graph at compile time.

After the code adding variables to the graph, add the following:

// Add a constant tensor to the graph
Tensor c1 = graph.addConstant<float>(FLOAT, {4}, {1.0, 1.5, 2.0, 2.5});

This line adds a new constant tensor to the graph whose elements have the values shown.

Allocate the data in c1 to tile 0:

// Allocate c1 to tile 0
graph.setTileMapping(c1, 0);

Now add the following to the sequence program, just before the PrintTensor program:

// Add a step to initialise v1 with the constant value in c1
prog.add(program::Copy(c1, v1));

Here we have used a predefined control program called Copy, which copies data between tensors on the device. Copying the constant tensor c1 into the variable v1 will result in v1 containing the same data as c1.

Note that the synchronisation and exchange phases of IPU execution described in the IPU Programmer’s Guide are performed automatically by the Poplar library functions and do not need to be specified explicitly.

If you recompile and run the program you should see the debug print of v1 shows initialised values:

v1-debug: [1.0000000 1.5000000 2.0000000 2.5000000]

Copying can also be used between variables:

After the v1 debug print command, add the following:

// Copy the data in v1 to v2
prog.add(program::Copy(v1, v2));
// Debug print v2
prog.add(program::PrintTensor("v2-debug", v2));

Now running the program will print both v1 and v2 with the same values.

Getting data into and out of the device

Most data to be processed will not be constant, but will come from the host. There are a couple of ways of getting data in and out of the device from the host. The simplest is to create a read or write handle connected to a tensor. This allows the host to transfer data directly to and from that variable.

Add code (before the engine creation instruction) to create read and write handles for the v3 variables:

// Create host read/write handles for v3
graph.createHostWrite("v3-write", v3);
graph.createHostRead("v3-read", v3);

These handles are used after the engine is created.

Add the following code after the engine creation instruction:

// Copy host data via the write handle to v3 on the device
std::vector<float> h3(4 * 4, 0);
engine.writeTensor("v3-write", h3.data(), h3.data() + h3.size());

Here, h3 holds data on the host (initialised to zeros) and the writeTensor call performs a synchronous write over the PCIe bus (simulated in this case) to the tensor on the device. After this call, the values of v3 on the device will be set to zero.

After the call to engine.run(0), add the following:

// Copy v3 back to the host via the read handle
engine.readTensor("v3-read", h3.data(), h3.data() + h3.size());

// Output the copied back values of v3
std::cout << "\nh3 data:\n";
for (unsigned i = 0; i < 4; ++i) {
  std::cout << "  ";
  for (unsigned j = 0; j < 4; ++j) {
    std::cout << h3[i * 4 + j] << " ";
  }
  std::cout << "\n";
}

Here, we are copying device data back to the host and printing it out. When the program is re-compiled and re-run, this prints all zeros (because the program on the device doesn’t modify the v3 variable):

Let’s see what happens when v3 is modified on the device. We will use Copy again, but also start to look at the flexible data referencing capabilities of the Tensor type.

Add the following code to create slices of v1 and v3 immediately after the creation of the host read/write handles for v3:
```
// Copy a slice of v1 into v3
Tensor v1slice = v1.slice(0, 3);
Tensor v3slice = v3.slice({1,1},{2,4});
```

These lines create a new Tensor object that references data in the graph. This does not create new state but just references parts of v1 and v3.

Now add this copy program:

prog.add(program::Copy(v1slice, v3slice));

This step copies three elements from v1 into the middle of v3. Re-compile and re-run the program to see the results:

h3 data:
0 0 0
1 1.5 2
0 0 0
0 0

Data streams

During training and inference of machine learning applications, efficiently passing data from the host to the IPU is often critical to enabling high throughput. The most efficient way to get data in and out of the device is to use data streams (see the the Poplar and PopLibs User Guide: data streams for more information). In Poplar, data streams need to be created and explicitly named in the graph; in the code snippets below we add a first-in-first-out (FIFO) input stream, connect it to a memory buffer (a vector of length 30), and we stream chunks of 10 elements of that buffer to the device.

Add the following code to the program definition:

// Add a data stream to fill v4
DataStream inStream = graph.addHostToDeviceFIFO("v4-input-stream", INT, 10);

// Add program steps to copy from the stream
prog.add(program::Copy(inStream, v4));
prog.add(program::PrintTensor("v4-0", v4));
prog.add(program::Copy(inStream, v4));
prog.add(program::PrintTensor("v4-1", v4));

These instructions copy from the input stream to the variable v4 twice. After each copy, v4 holds new data from the host.

After the engine is created, the data streams need to be connected to data on the host. This is achieved with the Engine::connectStream function.

Add the following code after the creation of the engine:

// Create a buffer to hold data to be fed via the data stream
std::vector<int> inData(10 * 3);
for (unsigned i = 0; i < 10 * 3; ++i)
  inData[i] = i;

// Connect the data stream
engine.connectStream("v4-input-stream", &inData[0], &inData[10 * 3]);

Here, we’ve connected the stream to a data buffer on the host, using it as a circular buffer of data. Recompile and run the program again, and you can see that after each copy from the stream, v4 holds new data copied from the host memory buffer:

v4-0: [0 1 2 3 4 5 6 7 8 9]
v4-1: [10 11 12 13 14 15 16 17 18 19]

(Optional) Using the IPU

This section describes how to modify the program to use the IPU hardware. The only changes are needed are related to making sure an IPU is available and acquiring it.

We will create a new file by copying tut1.cpp to tut1_ipu_hardware.cpp and open it in an editor.

Remove the import declaration:
```
#include <poplar/IPUModel.hpp>
```

Add these import declarations:

#include <poplar/DeviceManager.hpp>
#include <algorithm>

Replace the following lines from the start of main:

// Create the IPU Model device
IPUModel ipuModel;
Device device = ipuModel.createDevice();

with this code:

// Create the DeviceManager which is used to discover devices
auto manager = DeviceManager::createDeviceManager();

// Attempt to attach to a single IPU:
auto devices = manager.getDevices(poplar::TargetType::IPU, 1);
std::cout << "Trying to attach to IPU\n";
auto it = std::find_if(devices.begin(), devices.end(), [](Device &device) {
   return device.attach();
});

if (it == devices.end()) {
  std::cerr << "Error attaching to device\n";
  return 1; //EXIT_FAILURE
}

auto device = std::move(*it);
std::cout << "Attached to IPU " << device.getId() << std::endl;

This gets a list of all devices consisting of a single IPU that are attached to the host and tries to attach to each one in turn until successful. This is a useful approach if there are multiple users on the host. It is also possible to get a specific device using its device-manager ID with the getDevice function.

You are now ready to compile the program:

$ g++ --std=c++11 tut1_ipu_hardware.cpp -lpoplar -o tut1_ipu_hardware

Run the program to see the same results.
```
$ ./tut1_ipu_hardware
```

You can make similar modifications to the programs in the other tutorials in order to use the IPU hardware.

Summary

In this tutorial, we learnt how to build a simple application targeting the Graphcore IPU using Poplar. We used the Graph object to map tensors to specific tiles of the IPU and used the Sequence class to define a program with simple operations. Finally, we used data streams to pass data into the device and return results of the operations back to the host CPU process. This process and the classes used in this tutorial are summarised in the Poplar and PopLibs User Guide: Using the Poplar Library.

These three steps form the basis of Poplar applications and will be reused in the next tutorials. In the second tutorial you will learn to use the popops library which streamlines the definition of graphs and programs that include mathematical and tensor operations in Poplar.

To learn more about the programming model of the IPU discussed in this tutorial you may want to consult the IPU Programmer’s Guide or alternatively the Poplar and PopLibs User Guide. For a detailed reference, consult the API documentation. Graphcore also provides tutorials targeted at new users of the IPU using common Python deep learning frameworks PyTorch and TensorFlow 2.

Search help