tvm
Functions
tvm::topi::rocm Namespace Reference

Functions

tvm::te::Tensor dense_rocm (const Target &target, const tvm::te::Tensor &data, const tvm::te::Tensor &weight, const tvm::te::Tensor &bias, const DataType &out_dtype)
 Implementation of dense for rocm backend. More...
 
Schedule schedule_dense (const Target &target, const Array< Tensor > &outs)
 Create a rocm schedule for dense. More...
 
Schedule schedule_injective_from_existing (Schedule sch, const Tensor &out)
 Updates an existing schedule for the given injective ops. More...
 
Schedule schedule_injective (const Target &target, const Array< Tensor > &outs)
 Create a rocm schedule for the given output tensors. More...
 
Schedule schedule_pool (const Target &target, const Array< Tensor > &outs)
 Create a rocm schedule for pool. More...
 
Schedule schedule_global_pool (const Target &target, const Array< Tensor > &outs)
 Create a rocm schedule for global_pool. More...
 
Schedule schedule_reduce (const Target &target, Array< Tensor > outs)
 Create a rocm schedule for a reduce operation. More...
 
Schedule schedule_softmax (const Target &target, const Array< Tensor > &outs)
 Create a rocm schedule for the given softmax output tensors. More...
 

Function Documentation

◆ dense_rocm()

tvm::te::Tensor tvm::topi::rocm::dense_rocm ( const Target target,
const tvm::te::Tensor data,
const tvm::te::Tensor weight,
const tvm::te::Tensor bias,
const DataType out_dtype 
)
inline

Implementation of dense for rocm backend.

Parameters
targetThe target device
dataTensor with shape [batch, in_dim]
weightTensor with shape [out_dim, in_dim]
biasTensor with shape [out_dim]. Optional; to omit bias, pass Tensor()
out_dtypeOutput data type. Used for mixed precision.
Returns
Tensor with shape [batch, out_dim]

◆ schedule_dense()

Schedule tvm::topi::rocm::schedule_dense ( const Target target,
const Array< Tensor > &  outs 
)
inline

Create a rocm schedule for dense.

Parameters
targetThe target to generate a schedule for.
outsThe output tensors.
Returns
A schedule for the given ops.

◆ schedule_global_pool()

Schedule tvm::topi::rocm::schedule_global_pool ( const Target target,
const Array< Tensor > &  outs 
)
inline

Create a rocm schedule for global_pool.

Parameters
targetThe target to generate a schedule for.
outsThe output tensors.
Returns
A schedule for the given ops.

◆ schedule_injective()

Schedule tvm::topi::rocm::schedule_injective ( const Target target,
const Array< Tensor > &  outs 
)
inline

Create a rocm schedule for the given output tensors.

Parameters
targetThe target to generate a schedule for.
outsThe output tensors.
Returns
A schedule for the given ops.

◆ schedule_injective_from_existing()

Schedule tvm::topi::rocm::schedule_injective_from_existing ( Schedule  sch,
const Tensor out 
)
inline

Updates an existing schedule for the given injective ops.

Parameters
schThe schedule to update.
outThe tensor representing the injective op.
Returns
The updated schedule.

◆ schedule_pool()

Schedule tvm::topi::rocm::schedule_pool ( const Target target,
const Array< Tensor > &  outs 
)
inline

Create a rocm schedule for pool.

Parameters
targetThe target to generate a schedule for.
outsThe output tensors.
Returns
A schedule for the given ops.

◆ schedule_reduce()

Schedule tvm::topi::rocm::schedule_reduce ( const Target target,
Array< Tensor outs 
)

Create a rocm schedule for a reduce operation.

Parameters
targetThe target to generate a schedule for.
outsThe output tensors.
Returns
A schedule for the given ops.

◆ schedule_softmax()

Schedule tvm::topi::rocm::schedule_softmax ( const Target target,
const Array< Tensor > &  outs 
)
inline

Create a rocm schedule for the given softmax output tensors.

Parameters
targetThe target to generate a schedule for.
outsThe output tensors.
Returns
A schedule for the given ops.