[UKernel] Add ukernel to be compiled through peano #1097
Conversation
jtuyls
requested review from
makslevental,
nirvedhmeshram,
newling,
MaheshRavishankar,
yzhang93 and
Abhishek-Varma
as code owners
compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/XCLBinGen.cpp
Outdated
Show resolved
Hide resolved
compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/XCLBinGen.cpp
Outdated
Show resolved
Hide resolved
compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/test/aie_elf_files_gen_test.cpp
Show resolved
Hide resolved
Co-authored-by: Abhishek Varma <[email protected]>
| args.emplace_back("-o"); | ||
| args.emplace_back(outputFile); | ||
| if (verbose) args.emplace_back("-v"); | ||
| if (failed(runTool((peanoDir / "bin" / "clang").string(), args, verbose))) { |
There was a problem hiding this comment.
Wow, you can do this with peano? I must play around with this, see what the .ll generated looks like.
There was a problem hiding this comment.
Indeed, this is seems quite useful for experimentation and speed of dev. However, I have run into a couple of issues already that I think you have seen as well with vectorization:
- O0 results in compilation errors
- Sometimes the stack size usage blows up and this results in incorrect results if it's larger than 1024 and I needed to manually increase it: . Ideally, we should know the stack size usage before allocating a stack buffer...
Also, it should work with the high level AIE_API as well, but then you need to include those header files and I think they are only available inside Vitis. But with Vitis installed on our CI machines, that should work as well in theory, although I haven't tried it yet.
There was a problem hiding this comment.
Stack overflow, I suspect that's what is causing the current numerical issues I'm observing Xilinx/llvm-aie#342 -- thanks for this pointer, very well timed!
Adds a path to compile ukernels through peano. The compilation time is a lot faster this way, potentially at a decrease in performance, but potentially ukernels could be written in a way that enables good performance via peano as well.