In the HalideIR's design, DSL components and IR are mixed together.
For example, Call::Halide can containa reference to a function which is
constructed in the tensor expression language.
While this coupled design simplifies certain aspect of the DSL construction,
it prevents the TIR to evolve as a clean standalone IR:
- The additional tensor expression provided in the function is opaque to the IR
and may become obsolete as we transform them.
- The duplication of the information in the DSL tensor and IR makes it hard to
design a stand-alone text format (when there are elements shared in the tensor
expression and normal statements).
This PR aims to clearly de-couple the TIR from high-level DSL structures(tensor expression),
while still provide clear extensions to build DSLs on top of the TIR.
We introduce a DataProducer as a base class for high level tensor expressions objects
that produce data. We then introduce ProducerLoad to replace the Call::Halide usage,
so that the Call node can always be self contained and used for low-level calls.
The high-level tensor expression DSL can still generate a PrimExpr that contains a ProducerLoad.
These PrimExprs contains fragments of information that can be combined together to
generate a low-level TIR PrimFunc.
We also state clearly that DataProducer **should not** appear in any TIR PrimFunc.
Instead, the high-level DSL layer should lowered DataProducers to Buffers and TIR statements
that produces these buffers. We can further provide verifications to validate such invariance.
Changes:
- Introduce DataProducer to serve as a base class for Tensor in tensor expressions.
- Migrate use of Call::Halide to ProducerLoad
- Migrate the other usages of Calls.
We will also create follow-up PRs to migrate the remaining two DSL related IR nodes(Realize/Provide)
to use the DataProducer.
Tianqi Chen
6ae439c8c5