2. Programming with Poplar

You can use Poplar library functions to define graph operations and control the execution and profiling of code on the IPU.

Code can be compiled to run on IPU hardware, a simulated IPU Model or the host CPU. Running on an IPU Model or the CPU may be useful for functional testing of simple code when you do not have access to IPU hardware.

The IPU Model is a simulation of the behaviour of the IPU hardware. It does not completely implement every aspect of a real IPU. For example:

  • The IPU Model does not fully support replicated graphs (see Section 2.5, Replicated graphs).

  • The arithmetic results may differ from what would be obtained by using the IPU hardware.

  • Random number generation in the IPU Model is not the same as the hardware. In particular, every simulated tile has the same hard-coded seed (the setSeed() function is a no op). This means all IPU Model codelets will produce the same results every time they are run. Therefore, the IPU Model should not be used to verify any training or accuracy if the graph includes any random number generation.

If you encounter an out of memory error, it may be useful to run on the IPU Model device to debug the problem.

Consider the situation in which the event trace is being used to investigate a graph that creates a tile memory imbalance. In this case, running on the IPU will lead to an out of memory exception before the report is generated. Running on the IPU Model instead of actual hardware will still run out of memory, but the code will run to completion so the report can be generated.

Code running on a CPU device will be faster than the IPU Model, because it does not have the overhead of modelling the IPU hardware. CPU code runs with a single worker thread as if on a single tile on a single IPU. This means you do not need to think about tile allocation or the limited tile memory when initially developing vertex code. Running on a CPU device may also be useful for unit testing of vertices.

Interrogating a CPU device by calling Engine functions such as getBytesPerTile() or getTileClockFrequency() may not return accurate or meaningful results.

If you want to profile your code, you will need to run on either IPU hardware or the IPU Model.

2.1. Poplar programming model

For a more detailed introduction to the IPU architecture and programming model, see the IPU Programmer’s Guide.

A Poplar computation graph defines the input/output relationship between variables and operations. Each variable is a multi-dimensional tensor of typed values and can be distributed across multiple tiles.

Graph representation of variables and processing

Fig. 2.1 Graph representation of variables and processing

The vertices of the graph are the code executed in parallel by the tiles. Each tile executes a sequence of steps, which form a compute set containing one or more vertices.

The edges of the graph define the data that is read and written by the vertices. Each tile only has direct access to the tensor elements that are stored locally.

Each vertex always reads and writes the same tensor elements. In other words, the connections defined by the execution graph are static and cannot be changed at run time. However, the host program can calculate the mapping and graph connectivity at run time when it constructs the execution graph. See Poplar tutorial 7 on the Graphcore GitHub for an example.

The placement of vertices and tensor elements onto tiles is known as the tile mapping.

Mapping tensors and vertices to tiles

Fig. 2.2 Mapping tensors and vertices to tiles

2.2. The structure of a Poplar program

A Poplar program performs the following tasks:

  • Find or create the target device type as a Device representing physical IPU hardware or a simulated IPUModel.

  • Create a Graph object which will define the connections between computation operations and data, and how they are mapped onto the IPUs.

  • Create one or more Program objects which will control the execution of the graph operations.

  • Define the computations to be performed and add them to the Graph and Program objects. You can use the functions defined in Poplar and PopLibs, or you can write your own device code.

  • Create an Engine object, which represents a session on the target device, using the Graph and Program objects.

  • Connect input and output streams to the Engine object, to allow data to be transferred to and from the host.

  • Execute the computation with the Engine object. This will compile your graph code and load it onto the IPU, along with any library functions required, and start execution.

The Poplar and PopLibs libraries also include programs for a wide range of operations on tensor data.

For more detailed descriptions and examples of each of these steps, see the tutorials in the Graphcore GitHub tutorials repository.

2.2.1. Program flow control

A program object can be constructed by combining other program objects in various ways. There are several sub-classes of Program that provide flow control.

The simplest of these is Sequence, which executes a number of sub-programs sequentially. There are also Program classes for executing loops, and for conditional execution.


Looping is supported by the Repeat* classes and the PopLibs counted loop functions, which provide a more flexible interface.

There are two types of repeat programs:

  • Counted: this iterates a fixed number of times (a compile-time constant)

  • While: there are two repeat programs (RepeatWhileTrue and RepeatWhileFalse) which will iterate while the condition is met. Any non-zero value of the predicate is treated as true.

The counted repeat program iterates a fixed number of times. PopLibs provides a more flexible interface with the loop functions in the popops library.

Each of the functions in popops/Loop.hpp returns a program object that implements a “for” loop. There are several versions of these functions that provide varying amounts of control over the loop variables; for example specifying the start, step and end values of the loop counter. The number of iterations can be defined at run time.

The loop count variable can be made available to the program in the body of the loop (unlike loops created with Repeat).

A basic outline for creating a countedForLoop() is shown below:

count = graph.addVariable(poplar::UNSIGNED_INT, {1});
limit = graph.addVariable(poplar::UNSIGNED_INT, {1});
loopBodyProg = Sequence();
popops::mapInPlace(graph, Add(_1, _2), {bodyVar,count}, loopBodyProg,
prog.add(poputil::countedForLoop(graph, count, 0, limit, 1, loopBodyProg,

Conditional execution

The If program is equivalent to an if-then-else: it runs one of two programs depending on the value of a scalar tensor. Any non-zero value of the predicate is treated as true. You can use an empty Sequence for the “else” branch if you just want a simple “if” conditional.

The Switch program runs one of a number of programs depending on the value of a tensor. You can also define a default case for when the value of the tensor does not match one of the switch values.

2.2.2. What happens at run time

When you run your program on the host, the Poplar run-time will compile your graph to create object code for each tile. The code may come from Poplar or PopLibs library functions, or from vertex code you write yourself (see Section 2.8, Device code), and will be linked with any required libraries.

This object will contain:

  1. The control-program code from your graph

  2. Code to manage exchange sequences

  3. Initialised vertex data

  4. The tensor data mapped to that tile

The host program will load the object code onto the target device, which is then ready to execute the program.

2.3. Supported types

2.3.1. Scalar types

The scalar types supported by Poplar are shown in Table 2.1. In addition:

  • By default the char type is signed.

  • long is the same as int.

  • long double is the same as double.

  • The underlying type of an enumerated type is int.

  • Function pointers are the same as data pointers.

  • Only a subset of types are supported as fields of classes derived from the Vertex or MultiVertex classes. See Type for the supported types.

Table 2.1 Scalar data types


Size (bits)

Align (bits)





Character type




Short integer









long long







16-bit IEEE float




32-bit IEEE float




64-bit IEEE float

long double



64-bit IEEE float

void *



Data pointer

2.3.2. Floating point types

The IPU has hardware support for float (32 bit) and half (16 bit) operations.

You can indicate the type of a floating-point literal with the appropriate suffix. This extends the standard C++ for floating point literals.

  • half literals are indicated with _h; for example, 3.14_h

  • float literals are indicated with f or F

  • long double literals are indicated with l or L

  • If a suffix is not specified, the value has the type double


Although the compiler supports double and long double, there is no hardware support for 64-bit floating-point operations and so any calculations will be implemented in software. Therefore, you should be careful to avoid default promotions to double.

Half on the IPU

The IPU instruction set does not support operations on scalar half values, only vectors.

If operations on half values are not vectorised, either explicitly by the user (see Section 2.3.3, Vector types) or the compiler, then they may be promoted to float. The compiler will use vector operations, wherever possible for half types. If not, it will either promote the values to float, or broadcast a single half value to a half2 so it can use vector operations, then discard one half of the vector.

Half on the CPU

For CPU targets, half is, by default, an alias for float and sizeof(half) will be 4.

The parameter accurateHalf can be set to true when creating a CPU or IPU Model target, in which case half will be correctly implemented as 16-bit IEEE floating point. This will be slower, but will produce the same results as the IPU.

Codelets should be written to be generic to the size of half so that changing this setting requires no code changes.

2.3.3. Vector types

Poplar provides a number of vector types for representing short vectors of most scalar types. The supported vector types are shown in Table 2.2. Only the floating point vectors have direct support in the IPU instruction set.

Table 2.2 Vector data types


Size (bits)

Align (bits)






Vector of 2 char values





Vector of 2 unsigned char values





Vector of 4 char values





Vector of 4 unsigned char values





Vector of 2 int values





Vector of 2 unsigned int values





Vector of 4 int values





Vector of 4 unsigned int values





Vector of 2 long values





Vector of 4 long values





Vector of 2 long long values





Vector of 4 long long values





Vector of 2 short values





Vector of 2 unsigned short values





Vector of 4 short values





Vector of 4 unsigned short values





Vector of 2 float values





Vector of 4 float values





Vector of 2 half values





Vector of 4 half values


The vector types are defined with the vector extensions defined by GCC and Clang, using the vector_size variable attribute.

2.3.4. Structure types

Structure types pack according to the standard rules:

  • Field offsets are aligned according to the field’s type.

  • A structure is aligned according to the maximum alignment of its members.

  • Tail padding is added to make the structure’s size a multiple of its alignment.

2.3.5. Bit fields

The following types may be specified in a bit-field’s declaration: char, short, int, long , long long and enum.

If an enum type has negative values, enum bit-fields are signed. Otherwise, if a signed integer type of the specified width is not able to represent all enum values then enum bit-fields are unsigned. Otherwise, enum bit-fields are signed. All other bit-field types are signed unless explicitly unsigned.

Bit-fields pack from the least significant end of the allocation unit. Each non-zero bit field is allocated at the first available bit offset that allows the bit field to be placed in a properly aligned container of the declared type. Non bit-field members are allocated at the first available offset satisfying their declared type’s size and alignment constraints.

A zero-width bit-field forces padding until the next bit-offset aligned with the bit field’s declared type.

Unnamed bit-fields are allocated space in the same manner as named bit-fields.

A structure is aligned according to each of the bit field’s declared types in addition to the types of any other members. Both zero-width and unnamed bit fields are taken into account when calculating a structure’s alignment.

2.4. Virtual graphs

A graph is created for a target device with a specific number of tiles. It is possible to create a new graph from that, which is a virtual graph for a subset of the tiles. This is effectively a new view onto the parent graph for a virtual target, which has a subset of the real target’s tiles and can be treated like a new graph. You can add vertices and tensors to the virtual sub-graphs. These will also appear in the parent graph.

Any change made to the parent graph, such as adding variables or vertices, may also affect the virtual sub-graph. For example, a variable added to the parent graph will appear in the sub-graph if it is mapped to tiles that are within the subset of tiles in the virtual target.

Virtual graphs can be used to manage the assignment of operations to a subset of the available tiles. This can be used, for example, to implement a pipeline of operations by creating a virtual graph for each stage of the pipeline and adding the operations to be performed on those tiles.

Mapping a pipeline of operations to tiles using virtual graphs

Fig. 2.3 Mapping a pipeline of operations to tiles using virtual graphs

There are several versions of the createVirtualGraph function, which provide different ways of selecting the subset of tiles to include in the virtual target.

2.5. Replicated graphs

You can also create a replicated graph. This effectively creates a number of identical copies, or replicas, of the same graph. Each replica targets a different subset of the available tiles (all subsets are the same size). This may be useful, for example, where the target consists of multiple IPUs and you want to create a replica to run on each IPU (or group of IPUs) in parallel.

Any change made to the replicated graph, such as adding variables or vertices, will affect all the replicas. A variable mapped to tile 0, for example, will have an instance on tile 0 in each of the replicas.

Replicated graphs can be created in two ways:

As an example, imagine you have a graph which targets two IPUs. You can run four copies of it, in parallel, on eight of the IPUs in your system by creating the two-IPU graph and replicating it four times. This can be done using either of the techniques above, each of which has advantages and disadvantages, summarised in the following descriptions.

2.5.1. Replicating an existing graph

Replicating an existing graph

Fig. 2.4 Replicating an existing graph

We can start by creating a graph for eight IPUs, and then creating a replicated graph from that:

// Create a graph for 'target' which has 8 IPUs
Graph g  = Graph(target);
// Create 4 replicas each of which targets 2 IPUs
Graph rg = g.createReplicatedGraph(4);

Any changes, such as adding code or variables, made to the replica rg will be duplicated over all four replicas.

However, you can still do things with the original “parent” graph g that do not affect all the replicas. For example, a variable or an operation can be added to the parent graph and mapped to only one IPU. This will only be present on the replica that targets that IPU. It is also possible to access a variable that exists on all the replicas as a single tensor, using the getNonReplicatedTensor function. This adds an extra dimension to the variable to represent the mapping across the replicas.

This approach provides more flexibility but means that the graph of each replica needs to be compiled separately. This can make it slower to build the program.

2.5.2. Creating a replicated graph

Creating a replicated graph

Fig. 2.5 Creating a replicated graph

In this case, we start by creating a replicated graph using the graph constructor:

// Create a graph with 4 replicas for each 2 IPUs
Graph rg = Graph(target, replication_factor(4));

We can add variables and vertices to this graph as usual. These additions will be applied to every replica. This graph only exists as a replica, with no parent graph that can be used to make modifications differently to each replica. Therefore, as all the replicas are guaranteed to be identical, the graph only needs to be compiled once. Copies of the object code are then loaded onto each of the pairs of IPUs when the program runs. Each instance of the replica is given a unique ID at load time; this can be used to identify it in functions such as crossReplicaCopy.

Any functions that rely on the existence of a parent, such as getTopLevelGraph or getNonReplicatedTensor, will fail.

2.6. Data streams and remote buffers

Memory external to the IPU can be accessed in two ways. Data streams enable the IPU to transfer data to and from host memory. Remote buffers enable the IPU to store data in external (off-chip) memory.

2.6.1. Data streams

Data streams are used for communication between the host and the IPU device. The data transfers are controlled by the IPU.

Each stream is a unidirectional communication from the host to the device, or from the device to the host. A stream is defined to transfer a specific number of elements of a given type. This means the buffer storage required by the stream is known (the size of the data elements times the number of elements).

The Poplar graph compiler will merge multiple stream transfers into a single transfer (up to the limits described in stream_buffer_size_limit).

Device-side streams

A stream object, represented by the DataStream class, is created and added to a graph using the addHostToDeviceFIFO() or addDeviceToHostFIFO() functions. The stream is defined to have:

  • A name for the stream

  • The type of data to be transferred

  • The number of elements to be transferred

A host-to-device stream can also have a replication mode, if it is connected to a replicated graph. This defines whether a single stream will send the same data to all the replicated graphs (broadcast mode) or there will be a stream per replica.

Stream data transfer is done with a Copy program which copies data from the stream to a tensor, or from a tensor to the stream.

Host-side stream access

On the host side, a data stream is connected to a buffer allocated in memory. You connect a buffer to the stream using the connectStream() function of an Engine object. This can, optionally, be implemented as a circular buffer to support more flexible transfers.

In order to synchronise with the data transfers from the IPU, a callback is connected to the stream using the connectStreamToCallback() function. Callback implementations are derived from the StreamCallback interface and have a pointer to the stream buffer as an argument.

  • For a device-to-host transfer, the callback function will be called when the transfer is complete so that the host can read the data from the buffer.

  • For a host-to-device stream, by default the callback function will be called immediately before the IPU transfers the buffer contents to device memory. The host-side code should populate the stream buffer and then return.

Stream buffer size limit

The IPU has a memory address translation table which defines the external memory address range it can access. As a result, there is a maximum buffer size for data transferred by a stream. This limit is 128 MB per stream copy operation for Mk1, and 256 MB for Mk2. More data can be transferred by a sequence of copies, separated by sync operations, so that the buffer memory can be reused for each transfer.

To support stream copies that may exceed the buffer capacity, Poplar will split larger stream copies into a sequence of stream copies with distinct sync IDs. The size at which a stream copy is split is determined by splitLimit option. You can specify this when creating the data stream with addHostToDeviceFIFO() and addDeviceToHostFIFO(). This limit is smaller than the maximum stream copy transfer limit to account for internal Poplar streams which are added.

Note that if buffers are not overlapped in host memory (see the Engine option exchange.streamBufferOverlap) then stream splitting is not possible. This is because a buffer is allocated exclusively to a data stream, independently of sync IDs, and so nothing can be done to reduce the buffer usage.

Each IPU has its own translation table. So, if there are multiple IPUs, this limit applies to each IPU individually.

2.6.2. Optimising host data transfers

There are several things you can do to optimise the use of data streams to and from the host. These are described below.


You can specify that the the IPU should call the callback function as early as possible (for example, immediately after it releases the stream buffer from a previous transfer). The host is then able to fill the buffer in advance of the transfer, meaning the IPU spends less time waiting for the host.

This mode of operation, known as prefetch, is enabled by setting the Engine option exchange.enablePrefetch to “true” when the engine object is created.


Prefetch is only possible if the address range of the stream’s data buffer does not overlap with another stream’s buffer (this may be done to optimise memory use).

This means that the Engine option exchange.streamBufferOverlap must be set to either “HostRearrangeOnly” or “None”. The first of these is most useful as the performance of streams that are being rearranged is often less important. Setting the option to “None” may use too much memory.

The callback function returns a value that indicates if the buffer was filled.

If there is data available to fill the buffer, the callback function should return Result::Success. The device code will then call the complete() callback when it has transferred the data.

Otherwise, if data is not available (either because it is the end of the stream, or the data is not ready yet), then the callback returns Result::NotAvailable.


Data streams also supports multibuffering. Multibuffering means that a single data stream uses many buffers, instead of one. Having many buffers allows better I/O performance by getting the IPU and the host to work concurently, on different buffers. The number of buffers used by a stream is called its depth. Fig. 2.6 shows how multibuffering using a buffering depth of 2 improves IPU to host communication:


Fig. 2.6 Host and IPU communication with and without multibuffering

You can enable multibuffering for a data stream by setting its depth using the option bufferingDepth in the OptionFlags passed to the addHostToDeviceFIFO() or addDeviceToHostFIFO() functions.

2.6.3. Remote memory buffers

The IPU can also access off-chip Streaming Memory as a remote buffer. This may be host memory or external memory associated with the IPU system. This is not used for transferring data to the host, but just for data storage by the IPU program. Accesses to and from the remote buffer are much faster than data streams because they are entirely driven by the IPU during program execution and do not need to interact with the host.

A RemoteBuffer object is created and added to the graph with the addRemoteBuffer() function of the graph object. Data transfers to and from the remote buffer are performed using a Copy program; this copies data from the buffer to a tensor, or from a tensor to the buffer.

The data type and size of the remote buffer are defined when it is created. The definition of the buffer and the parameters to the Copy program allow for very flexible addressing.

You can think of the remote buffer containing a number of “rows” of data to be transferred. (These rows do not need to correspond to the structure the tensor being transferred or the organisation of the data in the buffer, but are just a way of managing data transfers.)

The size of each row and the number of rows are parameters to addRemoteBuffer() when the buffer is created:

  • numElements is the number of data items in each row

  • repeats is the number of rows in the buffer

Each Copy program can copy one or more rows of data to or from the remote buffer The rows to be copied are specified by the offset parameter to the Copy program. The number of offsets specifies the number of rows to copy. This cannot be larger than repeats. However, each row is implemented as a separate transfer.

Remote buffer restrictions

  • All copies to and from a given RemoteBuffer must be from the same IPU (each remote buffer is associated with exactly one IPU)

  • The size of a single transfer can not exceed 256 MB. In other words, each row must be less than 256 MB. The size of the remote buffer can be larger than 256 MB, depending on the value of the repeats parameter.

  • On IPU-M2000 and IPU-POD systems, each IPU may use up to one quarter of the Streaming Memory of the IPU-M2000. However, each IPU reserves 256 MB from this for data streams. For example, with 32 GB attached to an IPU-M2000, each of the four IPUs have (8 GB - 256 MB) available to be used for remote buffers.

  • By default IPUs are limited to (16 GB - 256 MB) of remote buffers each, regardless of the size of the Streaming Memory. If using more than 16 GB per IPU, the engine option target.extendedMemory must be set to “true”. This has a slight performance impact for copies so is disabled by default.

2.8. Device code

Each vertex of the graph is associated with some device code. This can come from a library function or you can write your own as a codelet. Codelets are specified as a class that inherits from the poplar::Vertex type. For example:

#include <poplar/Vertex.hpp>

using namespace poplar;

class AdderVertex : public Vertex {
  Input<float> x;
  Input<float> y;
  Output<float> sum;

  bool compute() {
    *sum = x + y;
    return true;

The Input and Output fields connect the vertex to the tensor data that it reads and writes. An Input field should not be written and an Output field should not be read; the results are undefined. If you need a field that is read and written, then it should be defined as InOut.

These fields have begin, end, operator[] and operator* methods so they can be iterated over and accessed like other C++ containers. For Input fields all of these methods are const.

The Output field can be successfully updated even if the corresponding tensor is on another tile. This is because the data is not transferred to the destination tile until the compute is complete. However, reading an Output field is not guaranteed to return the expected value. If you need to both write to and read from a field, then it should be declared as an InOut type.

The types used in vertex code are described in the runtime API section of the Poplar and PopLibs API Reference.

You can add a codelet to your graph by using the Graph::addCodelets function. This will load the source file and compile the codelet when the host program runs. See the adder example provided with the Poplar distribution.

You can also pass compilation options (for example “-O3”). The code is compiled for both the host and for the IPU so the program can be run on IPU hardware or on the host.

There are a couple of predefined macros that may be useful when writing vertex code. __POPC__ is defined when code is compiled by the codelet compiler. The macro __IPU__ is defined when code is being compiled for the IPU (rather than the host).

Graphcore’s tutorials repository includes an introductory tutorial on creating codelets and compute sets with Poplar.

You can also write codelets in assembly language for the IPU. See Section 9, Writing vertices in assembly for more information.

2.8.1. Stack allocation

When C++ functions are compiled, the compiler is usually able to determine the stack required. This is not possible if you use recursion, function calls via pointers or variable-length arrays (array variables that have a size that is not a compile time constant).

If you must use these techniques, then you must explicitly specify the stack used by your functions. Macros are provided for this purpose. These are defined in the Poplar header file StackSizeDefs.hpp. See the runtime API section of the Poplar API Reference for more information.

  • DEF_STACK_USAGE size function

    This defines the total stack usage (in bytes) for the function specified and any functions that it calls. This means that Poplar will not traverse the call graph of the function to determine the total stack usage of the function.

    If you use recursion, this macro must be used to specify the total stack usage of the recursive function itself, taking into account the maximum depth of the recursion and any other functions that can be called.


    This defines a list of other functions that may be called via pointers. Note that this creates a maintainability problem as the macro use must be updated every time the code changes its use of function pointers.

2.8.2. Pre-compiling codelets

There is a command line tool to pre-compile codelets. This reduces loading time, and allows you to check for errors before running the host program.

The codelet compiler, popc, takes your source code as input and creates a graph program object file (conventionally, with a .gp file extension). For example:

$ popc codelets.cpp -o codelets.gp

This object file can be added to your graph in the same way as source codelets, using the same Graph::addCodelets function. See the adder_popc example provided with the Poplar distribution.

The general form of the popc command is:

$ popc [options] <input file> -o <output file>

The command takes several command line options. Most are similar to any other C compiler. For example:


Add a macro definition


Add a directory to the include search path


Enable debugging


Set the optimization level (n = 0 to 3)

For a full list of options, use the --help option.