6. IPU supported operations

Below is a list of currently supported operations that can be executed on IPU hardware. This list will be expanded over time as we add more support. Some overloads and modes of operation for ops are not supported and we’ve tried to list all the caveats but some may have been missed.

6.1. Torch operations

6.1.1. Tensor operations

Many of the tensor operations will be executed before even reaching the IPU so we can consider them supported anyway. Some, like contiguous(), make no sense on a distributed memory system like the IPU so are ignored. There are no constraints on the memory format of how operations should be called other than the constraint that all graph inputs should be contiguous.

We will also create tensor views. However, the aliasing property of views with respect to in-place operations should not be relied on as we may have slightly different view behaviour.

Additionally, some PyTorch operations may be implemented by composition of the listed ops but may not be explicitly listed but are in fact supported.

Creation ops

  • torch.arange

  • tensor.fill

  • torch.full

  • torch.full_like

  • torch.Tensor.new_ones

  • torch.Tensor.new_zeros

  • torch.ones

  • torch.ones_like

  • torch.zeros

  • torch.zeros_like

Indexing, slicing, joining and mutating ops

In PyTorch, slicing a tensor is accessing a subset of the tensor by providing the start and end indices, such as tensor[1:5].

With a PopTorch model, you may take a slice of a tensor only if one of two conditions are met: * The start and end are constants, or can be resolved to be constants (for example, a function of the shape of a tensor which does not change between runs). * The start and end of the slice are related by a constant, for example tensor[x:x+5]. Please note that this will produce different results to PyTorch if the end value exceeds the length of the tensor: PyTorch will output a smaller size tensor but PopTorch will allow the slice to wrap round to the start of the relevant dimension.

PyTorch functions

  • torch.cat

  • torch.chunk

  • torch.gather

  • torch.index_select

  • torch.reshape

  • torch.roll

  • torch.scatter_add

  • torch.stack

  • torch.split

  • torch.squeeze

  • torch.t

  • torch.transpose

  • torch.unbind

  • torch.unsqueeze

  • torch.where

Tensor methods

  • tensor.expand

  • tensor.expand_as

  • tensor.masked_fill

Random samplers

To set the random state, use poptorch.Options.randomSeed

  • torch.bernoulli

  • torch.distributions.Bernoulli

  • torch.randn

  • torch.normal

  • torch.distributions.Normal

  • torch.rand

  • torch.uniform

  • torch.distributions.Uniform

6.1.2. Math operations

Pointwise ops

  • torch.abs

  • torch.acos

  • torch.acosh

  • torch.add

  • torch.addcdiv

  • torch.amax

  • torch.amin

  • torch.asin

  • torch.asinh

  • torch.atan

  • torch.atanh

  • torch.bitwise_not

  • torch.ceil

  • torch.clamp

  • torch.clamp_max

  • torch.clamp_min

  • torch.cos

  • torch.cosh

  • torch.div

  • torch.exp

  • torch.expm1

  • torch.floor

  • torch.floor_divide

  • torch.fmod

  • torch.frac

  • torch.log

  • torch.log10

  • torch.log1p

  • torch.log2

  • torch.mul

  • torch.norm

  • torch.neg

  • torch.pow

  • torch.reciprocal

  • torch.remainder

  • torch.round

  • torch.rsqrt

  • torch.sigmoid

  • torch.sign

  • torch.sin

  • torch.sinh

  • torch.sqrt

  • torch.square

  • torch.sub

  • torch.tan

  • torch.tanh

  • torch.true_divide

  • torch.trunc

Reduction ops

  • torch.argmax

  • torch.argmin

  • torch.mean

  • torch.median

  • torch.prod

  • torch.logsumexp

  • torch.sum

Comparison ops

  • torch.eq

  • torch.ge

  • torch.gt

  • torch.le

  • torch.lt

  • torch.max

  • torch.min

  • torch.ne

  • torch.isnan

    torch.topk only supports sorted=True and largest=True arguments.

  • torch.topk

Other ops

  • torch.cumsum

  • torch.cross

  • torch.meshgrid

  • torch.cartesian_prod

  • torch.tensordot

BLAS and LAPACK Operations

  • torch.addmm

  • torch.matmul

  • torch.bmm

6.2. Torch.nn operations

6.2.1. Containers

torch.nn.Module and torch.nn.Sequential can be passed into our compiler wrappers and just work.

6.2.2. Convolution layers

Conv transpose operations do not yet support dilations.

  • torch.nn.Conv1d

  • torch.nn.Conv2d

  • torch.nn.Conv3d

  • torch.nn.ConvTranspose1d

  • torch.nn.ConvTranspose2d

  • torch.nn.ConvTranspose3d

6.2.3. Pooling layers

Currently the max pool layers do not return the indices so only the variants with return_indices=False are supported.

  • torch.nn.MaxPool1d

  • torch.nn.MaxPool2d

  • torch.nn.MaxPool3d

  • torch.nn.AvgPool1d

  • torch.nn.AvgPool2d

  • torch.nn.AvgPool3d

  • torch.nn.AdaptiveAvgPool1d

  • torch.nn.AdaptiveAvgPool2d

  • torch.nn.AdaptiveAvgPool3d

6.2.4. Padding layers

All padding layers are supported.

  • torch.nn.ReflectionPad1d

  • torch.nn.ReflectionPad2d

  • torch.nn.ReplicationPad1d

  • torch.nn.ReplicationPad2d

  • torch.nn.ReplicationPad3d

  • torch.nn.ZeroPad2d

  • torch.nn.ConstantPad1d

  • torch.nn.ConstantPad2d

  • torch.nn.ConstantPad3d

6.2.5. Activations

  • torch.nn.ELU

  • torch.nn.CELU

  • torch.nn.GELU

  • torch.nn.Hardshrink

  • torch.nn.LeakyReLU

  • torch.nn.LogSoftmax

  • torch.nn.ReLU

  • torch.nn.SELU

  • torch.nn.SiLU

  • torch.nn.Sigmoid

  • torch.nn.Softmax

  • torch.nn.Softplus

  • torch.nn.Softsign

  • torch.nn.Softshrink

  • torch.nn.Tanh

  • torch.nn.PReLU

  • torch.nn.RReLU

  • torch.nn.Hardtanh

  • torch.nn.functional.glu

  • torch.nn.Threshold

6.2.6. Normalization layers

Currently only affine=True is supported as a parameter. That is to say, only the variants with trainable parameters are supported.

  • torch.nn.BatchNorm1d

  • torch.nn.BatchNorm2d

  • torch.nn.BatchNorm3d

  • torch.nn.LayerNorm

  • torch.nn.GroupNorm

  • torch.nn.InstanceNorm1d

  • torch.nn.InstanceNorm2d

  • torch.nn.InstanceNorm3d

6.2.7. Recurrent layers

  • torch.nn.GRU

  • torch.nn.LSTM

6.2.8. Linear layers

  • torch.nn.Identity

  • torch.nn.Linear

  • torch.nn.Bilinear

6.2.9. Dropout

  • torch.nn.dropout

6.2.10. Sparse layers

Embedding and EmbeddingBag are supported with the exception of the padding_idx parameter being unsupported.

  • torch.nn.Embedding

  • torch.nn.EmbeddingBag

  • torch.nn.functional.one_hot

6.2.11. Loss functions

This version supports a limited subset of loss functions. However, we support poptorch.identity_loss() which gives you the ability to implement any arbitrary loss function.

See also

poptorch.identity_loss()

One caveat for the following loss functions is if they are used they will always be included in the back propagation and will always receive a gradient, which is a slight deviation from normal PyTorch operations, where they have to opt in to the gradient pass.

  • torch.nn.L1Loss

  • torch.nn.MSELoss

  • torch.nn.CrossEntropyLoss

  • torch.nn.NLLLoss

  • torch.nn.BCELoss

  • torch.nn.KLDivLoss

  • torch.nn.PoissonNLLLoss

  • torch.nn.HingeEmbeddingLoss

  • torch.nn.BCEWithLogitsLoss

  • torch.nn.SmoothL1Loss

  • torch.nn.SoftMarginLoss

  • torch.nn.CosineEmbeddingLoss

  • torch.nn.MarginRankingLoss

  • torch.nn.TripletMarginLoss

  • torch.nn.CTCLoss

6.2.12. Vision Layers

Only nearest is supported.

  • torch.nn.Upsample

6.3. Float 16 operations

Due to the limitation of PyTorch’s float 16 support on the CPU (used for tracing the model), certain operations may result in the use of float 32 where float 16 would be expected, or float 16 where float 32 would be expected. This is because the model must always be traced with float 16 inputs converted to float 32.

This limitation is much less noticeable when opts.Precision.halfFloatCasting(poptorch.HalfFloatCastingBehavior.HalfUpcastToFloat) has not been set because PopTorch’s default casting functionality is to output a float 16 if any input of the op is float 16. In such situations, any dtype which incorrectly resolves to a float 16 would have been cast to a float 16 in any case.

6.3.1. Casting

The tensor.to(dtype) argument will be ignored if it is torch.float32 because it may refer to one or more float 16 tensors which were converted to float 32 to allow tracing to happen, for example a.to(b.dtype) where b may be a float 16 tensor converted to a float 32 tensor. Once the output of the op or one of its descendants encounters a known float 16 or float 32 input, the type will be resolved to this type.

The following examples show cases where the casting functionality is resolved based on context, correctly or incorrectly:

Listing 6.1 Cases where casting resolves to the correct type
 1class Model(torch.nn.Module):
 2    def forward(self, x, y):
 3        # In spite of "y.dtype" being ignored if it is float32, the dtype used
 4        # for the cast resolves to be the type of y because of the "+ y"
 5        return x.to(y.dtype) + y
 6
 7
 8native_model = Model()
 9
10float16_tensor = torch.tensor([1.0], dtype=torch.float16)
11float32_tensor = torch.tensor([1.0], dtype=torch.float32)
12
13assert native_model(float16_tensor, float16_tensor).dtype == torch.float16
14assert native_model(float16_tensor, float32_tensor).dtype == torch.float32
15assert native_model(float32_tensor, float16_tensor).dtype == torch.float16
16assert native_model(float32_tensor, float32_tensor).dtype == torch.float32
17
18poptorch_model = poptorch.inferenceModel(native_model)
19assert poptorch_model(float16_tensor, float16_tensor).dtype == torch.float16
20
21poptorch_model = poptorch.inferenceModel(native_model)
22assert poptorch_model(float16_tensor, float32_tensor).dtype == torch.float32
23
24poptorch_model = poptorch.inferenceModel(native_model)
25assert poptorch_model(float32_tensor, float16_tensor).dtype == torch.float16
26
27poptorch_model = poptorch.inferenceModel(native_model)
28assert poptorch_model(float32_tensor, float32_tensor).dtype == torch.float32
29
Listing 6.2 Cases where casting resolves to an incorrect type
 1class Model(torch.nn.Module):
 2    def forward(self, x, y):
 3        # torch.float32 is ignored and the type is resolved to be the type of y
 4        return x.to(torch.float32) + y
 5
 6
 7native_model = Model()
 8
 9float16_tensor = torch.tensor([1.0], dtype=torch.float16)
10float32_tensor = torch.tensor([1.0], dtype=torch.float32)
11
12assert native_model(float16_tensor, float16_tensor).dtype == torch.float32
13assert native_model(float32_tensor, float16_tensor).dtype == torch.float32
14
15opts = poptorch.Options()
16opts.Precision.halfFloatCasting(
17    poptorch.HalfFloatCastingBehavior.HalfUpcastToFloat)
18
19# This incorrectly results in a float 16 tensor
20poptorch_model = poptorch.inferenceModel(native_model, opts)
21assert poptorch_model(float16_tensor, float16_tensor).dtype == torch.float16
22
23# This incorrectly results in a float 16 tensor
24poptorch_model = poptorch.inferenceModel(native_model, opts)
25assert poptorch_model(float32_tensor, float16_tensor).dtype == torch.float16

6.3.2. Creation functions

The following functions are affected: * torch.ones * torch.rand * torch.zeros * torch.distributions.uniform.Uniform

The dtype arguments will be ignored because they may refer to float 16 tensors which were converted to float 32 tensors to allow tracing to succeed. Once the output of the op, or its descendant, encounters a known float 16 or float 32 input, the dtype values are resolved to this type.

The following examples show cases where the type output differs from PyTorch:

Listing 6.3 Type resolution when using torch.zeros
 1## torch.ones and zeros
 2class Model(torch.nn.Module):
 3    def forward(self, x):
 4        # dtype is ignored, however the type is resolved to be the type of x
 5        return torch.zeros((2, 3, 4), dtype=torch.float32) + x
 6
 7
 8native_model = Model()
 9
10float16_tensor = torch.tensor([1.0], dtype=torch.float16)
11float32_tensor = torch.tensor([1.0], dtype=torch.float32)
12
13# The native model always yields a float32 tensor
14assert native_model(float16_tensor).dtype == torch.float32
15assert native_model(float32_tensor).dtype == torch.float32
16
17opts = poptorch.Options()
18opts.Precision.halfFloatCasting(
19    poptorch.HalfFloatCastingBehavior.HalfUpcastToFloat)
20
21# The poptorch model will resolve to the type of x
22poptorch_model = poptorch.inferenceModel(native_model, opts)
23assert poptorch_model(float16_tensor).dtype == torch.float16
24
25poptorch_model = poptorch.inferenceModel(native_model, opts)
26assert poptorch_model(float32_tensor).dtype == torch.float32
27
Listing 6.4 Type resolution when using torch.rand
 1## torch.rand
 2class Model(torch.nn.Module):
 3    def forward(self, x):
 4        # dtype is ignored, however the type is resolved to be the type of x
 5        return torch.rand((2, 3, 4), dtype=torch.float32) + x
 6
 7
 8native_model = Model()
 9
10float16_tensor = torch.tensor([1.0], dtype=torch.float16)
11float32_tensor = torch.tensor([1.0], dtype=torch.float32)
12
13opts = poptorch.Options()
14opts.Precision.halfFloatCasting(
15    poptorch.HalfFloatCastingBehavior.HalfUpcastToFloat)
16
17# The native model always yields a float32 tensor
18assert native_model(float16_tensor).dtype == torch.float32
19assert native_model(float32_tensor).dtype == torch.float32
20
21# The poptorch model will resolve to the type of x
22poptorch_model = poptorch.inferenceModel(native_model, opts)
23assert poptorch_model(float16_tensor).dtype == torch.float16
24
25poptorch_model = poptorch.inferenceModel(native_model, opts)
26assert poptorch_model(float32_tensor).dtype == torch.float32
27
Listing 6.5 Type resolution when using torch.distributions.uniform.Uniform
 1## torch.distributions.uniform.Uniform
 2class Model(torch.nn.Module):
 3    def forward(self, x):
 4        # dtype is ignored, however the type is resolved to be the type of x
 5        ud = torch.distributions.uniform.Uniform(
 6            torch.tensor([0.0], dtype=torch.float16),
 7            torch.tensor([1.0], dtype=torch.float32))
 8        return ud.sample((10, 10, 1000)) + x
 9
10
11native_model = Model()
12
13float16_tensor = torch.tensor([1.0], dtype=torch.float16)
14float32_tensor = torch.tensor([1.0], dtype=torch.float32)
15
16# The native model always yields a float32 tensor
17assert native_model(float16_tensor).dtype == torch.float32
18assert native_model(float32_tensor).dtype == torch.float32
19
20opts = poptorch.Options()
21opts.Precision.halfFloatCasting(
22    poptorch.HalfFloatCastingBehavior.HalfUpcastToFloat)
23
24# The poptorch model will resolve to the type of x
25poptorch_model = poptorch.inferenceModel(native_model, opts)
26assert poptorch_model(float16_tensor).dtype == torch.float16
27
28poptorch_model = poptorch.inferenceModel(native_model, opts)
29assert poptorch_model(float32_tensor).dtype == torch.float32

6.3.3. Normalization

Some normalization layers require the computation of running statistics - mean and variance. These tensors will be computed as float32 even though the inputs to the operator can be float16. This behaviour has been chosen to strike a balance between performance and numerical accuracy.

The following operators are affected: * torch.nn.BatchNorm1d * torch.nn.BatchNorm2d * torch.nn.BatchNorm3d

The type of running statistics computations may be controlled via opts.Precision.runningStatisticsAlwaysFloat(bool). For example, in the script below, mean and variance computations will be performed in half precision:

Listing 6.6 Controlling type of running mean and variance computations
1model = torch.nn.Sequential()
2model.add_module('lin', torch.nn.Linear(16, 16))
3model.add_module('bn', torch.nn.BatchNorm1d(16))
4model.float()
5
6opts = poptorch.Options()
7opts.Precision.runningStatisticsAlwaysFloat(False)
8poptorch_model = poptorch.inferenceModel(model, opts)