[spirv] [tracking] SPIR-V lowering

[antiagainst]

This is a tracking bug for implementing lowering towards the SPIR-V dialect. It will be periodically updated to provide the current status. Anybody interested in helping is very welcome to pick up tasks here! :)

Common utilities

• Target environment (version, extensions, capabilities)
• Shader interface ABI (descriptor, storage class, etc.) and its lowering
• Entry point ABI (dispatch configuration, etc.) and its lowering
• SPIR-V module and resource variable creation
• Layout decoration utilities
• Recommended pass pipeline/recipe
• Simple end-to-end testing infrastructure

From GPU dialect

• Kernel function
• Non-kernel functions
• Block/grid dim
• Thread/block id

From Std dialect

• Std to spv type converter
  • Memref's with dynamic shapes
• constant
  • scalar/vector
  • composite
• load, store
• Binary ops
  • addi, subi, muli
  • divis, remis
  • diviu, remiu
  • float cases
• comparison ops
  • int eq, ne
  • int slt, sle, sgt, sge
  • int ult, ule, ugt, uge
  • float eq, ne
  • float slt, sle, sgt, sge
  • float ult, ule, ugt, uge
• Logical ops
  • select
• Bit ops
• Cast ops
• return
• br
• cond_br
• call

From Linalg dialect

• Pass pipeline to go from linalg to spv

[denis0x0D]

@antiagainst can you please clarify on "Layout decoration utilities", I thought they exist, do I miss something? Thanks.

[antiagainst]

Oh, sorry I thought I ticked it... Fixed now. Yes we can have a few more things to do there like supporting more rules but I don't think they are of high priority. They can be added later when needed. I would consider the general utility is there for this reason.

Also note that this is a coarse summary right now and expect this to expand. :) We also plan to expand the SPIR-V.md doc to explain more over the lowering internals and instructions on adding an new op lowering. Would highly appreciate your help (as always) on completing the lowering! :)

[denis0x0D]

Yes we can have a few more things to do there like supporting more rules but I don't think they are of high priority.

Thanks for mention this, I completely forget about different rules for layout, sorry about it.

Also note that this is a coarse summary right now and expect this to expand. :) We also plan to expand the SPIR-V.md doc to explain more over the lowering internals and instructions on adding an new op lowering. Would highly appreciate your help (as always) on completing the lowering! :)

Sounds great, I would like to help, if it possible. Thanks!

[antiagainst]

Feel free to take whatever you are interested and not done. Please create an issue; I'll link it here. Thanks! :)

[MaheshRavishankar]

Updated to add non-kernel functions and dynamic shapes as not completed items.
I plan to focus on dynamic shapes starting next year.

[denis0x0D]

@MaheshRavishankar as far as I understood the lowering from std->spv mostly done by you, in this case can you please share the task which I can start from and you are not working on it? Thanks.

[MaheshRavishankar]

@denis0x0D : Thanks for picking up some of these tasks. For Std to SPIR-V all except call operation are fairly low hanging fruit. I am not working on this right now. Feel free to pick up any that make sense to you. Thanks!

[denis0x0D]

@MaheshRavishankar nice, thanks!

[denis0x0D]

@antiagainst @MaheshRavishankar sorry for delay, is it ok if I'll take some bit and cast op lowering from std to spv? Thanks.

[antiagainst]

Sure, please go ahead!

[denis0x0D]

@antiagainst @MaheshRavishankar can you please clarify on lowering for std.and and std.or, it seems like they have bitwise semantics? Do I miss something? Thanks.
https://github.com/llvm/llvm-project/blob/master/mlir/lib/Conversion/StandardToSPIRV/StandardToSPIRV.td#L23

[antiagainst]

Oops, yes indeed as shown by llvm/llvm-project@a8a5c06. Good catch! This is a bug that should be fixed. Would you mind to send a PR for it? Thanks!

[MaheshRavishankar]

What's the bug here. The lowering is only supported for bool. So that should be fine right?

[denis0x0D]

@MaheshRavishankar oh sorry, you are right, I missed that Bitwise works for SPV_Integer

[antiagainst]

Oops again. Yes @MaheshRavishankar is correct. I commented too quickly...