[API] Enhancement on Data Movement APIs - Githubissues

cornell-zhang / heterocl

HeteroCL: A Multi-Paradigm Programming Infrastructure for Software-Defined Heterogeneous Computing

https://cornell-zhang.github.io/heterocl/

Apache License 2.0

322 stars 92 forks source link

[API] Enhancement on Data Movement APIs #321

Closed hecmay closed 3 years ago

hecmay commented 3 years ago

Enhance .to primitive

[x] Improve explicit loop unrolling to support 2D PE array
[x] Add more test cases for host-device graph partitioning algorithms
[x] Automatic memory coalescing for repetitive memory access when building the PE
[x] AutoSA integration done. tested on GEMM example.
[x] Connect AutoSA with .to (consider packing, SA layout inference e.t.c)
[ ] Automatic IO optimization for the given target device (e.g. packing, burst size, and memory device selection)

Enhance HCL runtime and build APIs

[x] Support streaming interface for HCL module
[x] Convert array/scalar values to hcl.asarray automatically in TVM packed function call
[x] Added dockerfile for AutoSA+SODA+HCL

Bugs fixing

[x] The VHLS codegen generates A[0] for a scalar instead of A
[x] Support unrolling parent stage with attaching sub-stages (the sub-stage will be copied across the unrolled PEs)
[x] Support flattened and multi-dimensional array in the host code generator

hecmay commented 3 years ago

@zzzDavid Please see this example for streaming between HCL modules: https://github.com/Hecmay/heterocl/blob/fix/tests/test_schedule_systolic.py#L132-L156

I updated the API to be user-friendly and intuitive:

# Stream module add's B port to module mul's B port
s.to(kernel.add.B, kernel.mul.B)

The underlying IR pass will check the legality of the primitive: it has to be one-read-one-write.