19.12. Tensor locations

class popxl.ExecutionContext

Members:

Normal

AccumulateOuterFragment

WeightsFromHostFragment

WeightsToHostFragment

OptimizerFromHostFragment

Subgraph

AccumulateOuterFragment = <ExecutionContext.AccumulateOuterFragment: 1>
Normal = <ExecutionContext.Normal: 0>
OptimizerFromHostFragment = <ExecutionContext.OptimizerFromHostFragment: 4>
Subgraph = <ExecutionContext.Subgraph: 5>
WeightsFromHostFragment = <ExecutionContext.WeightsFromHostFragment: 2>
WeightsToHostFragment = <ExecutionContext.WeightsToHostFragment: 3>
__init__(self: popart_internal_ir.ExecutionContext, value: int) None
property name
property value
class popxl.ReplicatedTensorSharding(value)

Enum type to specify whether to shard tensors over replicas.

Off = 'Off'

Tensors will not be sharded over replicas.

On = 'On'

Tensors will be sharded over replicas.

class popxl.TensorLocation(storage=TensorStorage.OnChip, replicated_tensor_sharding=ReplicatedTensorSharding.Off)

Class that describes the memory characteristics of one or multiple tensors.

Parameters
Return type

None

__init__(storage=TensorStorage.OnChip, replicated_tensor_sharding=ReplicatedTensorSharding.Off)
Parameters
Return type

None

property replicated_tensor_sharding: popxl.tensorlocation.ReplicatedTensorSharding
property storage: popxl.tensorlocation.TensorStorage
class popxl.TensorStorage(value)

Enum type to specify whether to store tensors on-chp in tile memory, or in Streaming Memory.

OffChip = 'OffChip'

Store the tensor in Streaming Memory.

OnChip = 'OnChip'

Store the tensor in tile memory.

class popxl.TileSet

Members:

Compute

IO

Undefined

N

Compute = <TileSet.Compute: 0>
IO = <TileSet.IO: 1>
N = <TileSet.N: 3>
Undefined = <TileSet.Undefined: 2>
__init__(self: popart_internal_ir.TileSet, value: int) None
property name
property value