junrushao1994 commented on a change in pull request #22: URL: https://github.com/apache/tvm-rfcs/pull/22#discussion_r698099022
########## File path: rfcs/0022-tir-non-scalar-constants.md ########## @@ -0,0 +1,107 @@ + +- Feature Name: tir_non_scalar_constants +- Start Date: 2021-06-01 +- RFC PR: https://github.com/apache/tvm-rfcs/pull/22 +- GitHub Issue: TBD + +# 1. Summary + +This RFC proposes how non-scalar constants could be represented in TIR and used by passes in the lowering process. + +# 2. Motivation + +Currently, the non-scalar constants could be represented in Relay (relay.Constant) to be used by relay passes but not in TIR. Therefore, when performing lowering using TIR passes, we have to maintain a side-channel of tir::Var to constant non-scalar data mapping to perform transformations that could use the knowledge where some of the data are constants. + +Few example scenarios as further motivation : + +## Weight compression + +When lowering for accelerators (E.g. : [Arm(R) Ethos(TM)-U NPU](https://github.com/apache/tvm-rfcs/pull/11)), certain operations will need to get tiled to co-optimize performance and memory utilization. Such tiling patterns create slices of weights that need compressing that will end up with varying sizes. Therefore, the knowledge of some tir::Vars refer to constants are critical in the level of TIR to perform this. + +## Memory Planning + +The TIR program has the ability to express both inter and intra operator memory requirement, post-scheduling as explained further by [Unified Static Memory Planning RFC](https://github.com/apache/tvm-rfcs/pull/9). It would be better if the constants could be embedded to the TIR PrimFunc. Moreover, this allows various [target-dependent lowerings](https://github.com/apache/tvm-rfcs/pull/10), to produce TIR PrimFuncs with constants in it. + +## Winograd Constants + +The Winograd transformation (used for fast GEMMs) involves multiplication by a hard-coded constant tensor. This is currently accomplished in TE using a complicated TE compute expression with many nested selects. Being able to directly express a constant tensor here would significantly simplify this code. + + +# 3. Guide-level explanation Review comment: Perhaps it is worthwhile to discuss about the semantics of a TIR constant as well :-) Will the constants be allocated on stack or on heap? Is this designed for small matrices (e.g. the small matrix in winograd), or relatively larger matrices (e.g. the weight that needs prefetching)? How will lowering and code generation be affected? Does it work for GPU and other devices? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
