[Compiler] Rename compiler ROCM target to HIP

[nithinsubbiah]

As a follow-up to #17710, this patch renames the compiler target backend from ROCM to HIP to be consistent with driver.

User-facing changes:
IREE_TARGET_BACKEND_ROCM to IREE_TARGET_BACKEND_HIP
iree-hal-target-backends=rocm to iree-hal-target-backends=hip

[benvanik]

IIRC we didn't want to do this? (I think Lei told me ROCM and HIP were independent? did someone change their mind? :)

[nithinsubbiah]

They are indeed independent. We wanted to be consistent with the backend names and thus I started working on it. But happy to hear thoughts on moving forward
cc: @antiagainst @ScottTodd

[benvanik]

Mostly just thinking about HSA: it will be using the same codegen pipeline but not targeting HIP, so if there's a better name for what it is that'd be ideal ("ROCDL" may have been what Lei mentioned before). There will be a HIP IREE::HAL::TargetDevice and HSA IREE::HAL::TargetDevice that share the same codegen backend.

[stellaraccident]

I'm not sure there's a better answer. Both rocm and hip names refer to software stacked above. If we were naming from scratch, we would name like llvm: "amdgpu". From that perspective, I would rename the runtime driver "amdgpu-hip". That's what I call it in shortfin fwiw.

[qedawkins]

Ok you all type too fast. +1 to amdgpu-hip, or maybe amdgpu-llvm. Either way it would be nice to switch away from "rocm."

[kuhar]

FWIW, the bespoke clang is called hipcc too

[benvanik]

AMDGPU may work, but ROCm as the umbrella also makes sense (HIP and ROCr/HSA are both under the ROCm umbrella) and is equivalent to CUDA.

[benvanik]

(basically, I don't want inconsistency and I don't want the compiler side that targets multiple runtime implementations named after a particular runtime implementation)

[kuhar]

AMDGPU alone is also the ultimate backend for spirv/vulkan with amdvlk. If they ever implement offline shader compilation, we will have a name collision with that.

[qedawkins]

IIUC AMDGPU is the name of the target in llvm right? https://llvm.org/docs/AMDGPUUsage.html

So maybe amdgpu-llvm for clarity? I kind of like the idea of being consistent with the llvm backend we're targeting and also differntiating from spir-v.

[kuhar]

IIUC AMDGPU is the name of the target in llvm right? https://llvm.org/docs/AMDGPUUsage.html

So maybe amdgpu-llvm for clarity? I kind of like the idea of being consistent with the llvm backend we're targeting and also differntiating from spir-v.

amdgpu is the name of the backend, the two targets are amdgcn--amdpal (graphics/vulkan) and amdgcn-amd-hsa (hip/rocm)

[kuhar]

IMO hip and amdgpu-hip are the two least problematic names

[benvanik]

reminder: hip != hsa, and we'll be using the same thing for hsa - in my dreams we end up deleting our hip HAL - hip is problematic :)

(the name here is for us - it doesn't have to match LLVM or anyone else - it's just what we call our target)

[kuhar]

OK, I give up, naming proves hard again 🙃

[benvanik]

yeah, we're in the ROC* name playground that is certifiably inscrutable - there's no good names :P

[nithinsubbiah]

What if we go with amdgpu-rocm for the compiler target thereby encompassing both hip and hsa runtime drivers, and also skirting the potential name clash with vulkan.

Unless someone has a strong reason to not have rocm keyword anywhere

[qedawkins]

Also tangential question, could we have aliased names for target backends? e.g. if I put in hip or hsa as the compiler backend it would alias to amdgpu-rocm or whatever we choose to call it. If so I'd prefer to lean into a more llvm focused name because it makes the C++ very clear that it's going to hand the executables off to llvm.

[qedawkins]

Also tangential question, could we have aliased names for target backends? e.g. if I put in hip or hsa as the compiler backend it would alias to amdgpu-rocm or whatever we choose to call it. If so I'd prefer to lean into a more llvm focused name because it makes the C++ very clear that it's going to hand the executables off to llvm.

My only problem with rocm is that I've never really understood what it is :P

[benvanik]

rocm == cuda (effectively) - an umbrella term for the entire suite of libraries/tools/languages/etc

[stellaraccident]

I think the stem can defensively be amdgpu or rocm. I prefer the first as it is more descriptive of a physical thing that will always be that thing.

[kuhar]

Also tangential question, could we have aliased names for target backends? e.g. if I put in hip or hsa as the compiler backend it would alias to amdgpu-rocm or whatever we choose to call it. If so I'd prefer to lean into a more llvm focused name because it makes the C++ very clear that it's going to hand the executables off to llvm.

the llvm command line flag library supports aliases