You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2020/06/24 19:42:46 UTC
[GitHub] [incubator-tvm] t-vi opened a new pull request #5920: Two small fixes to AMDCPU codegen for LLVM 10+ and ROCm 3.5+
t-vi opened a new pull request #5920:
URL: https://github.com/apache/incubator-tvm/pull/5920
- LLVM 10+ got stricter and fails hard when Aligning to 0 bytes, so we don't.
- ROCm 3.5 deprecates code object v2. So if we have the new ROCm, we use the (default) v3.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] t-vi commented on pull request #5920: Two small fixes to AMDCPU codegen for LLVM 10+ and ROCm 3.5+
Posted by GitBox <gi...@apache.org>.
t-vi commented on pull request #5920:
URL: https://github.com/apache/incubator-tvm/pull/5920#issuecomment-649657507
@tqchen Yeah, so the background to this is that the recent release of ROCm 3.5 brings rather sweeping changes (changing the compiler backend for the HIP compilation among other things).
My conclusion from this would be
- I'd have some sympathy for people not switching immediately (I was one of them 2 days ago),
- In half a year, serious users will have switched. Then we can then just bump the requirement to ROCm >= 3.5 then (IMHO).
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] tqchen commented on pull request #5920: Two small fixes to AMDCPU codegen for LLVM 10+ and ROCm 3.5+
Posted by GitBox <gi...@apache.org>.
tqchen commented on pull request #5920:
URL: https://github.com/apache/incubator-tvm/pull/5920#issuecomment-649618356
NOTE: using runtime detection of rocm features will only work if we are building on the same machine and won't work for cross compilation. While it is OK for now, let us keep that in mind and once we land https://discuss.tvm.ai/t/rfc-tvm-target-specification/6844, we might want to allow user to explicitly specify the attr and only use auto detect if the attr is not specified(or march=native is used)
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] t-vi commented on pull request #5920: Two small fixes to AMDCPU codegen for LLVM 10+ and ROCm 3.5+
Posted by GitBox <gi...@apache.org>.
t-vi commented on pull request #5920:
URL: https://github.com/apache/incubator-tvm/pull/5920#issuecomment-649546779
Sorry, another small thing:
- I'm throwing in an addition of kMaxRegistersPerBlock for ROCm. This was introduced for CUDA in #5898.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] t-vi commented on pull request #5920: Two small fixes to AMDCPU codegen for LLVM 10+ and ROCm 3.5+
Posted by GitBox <gi...@apache.org>.
t-vi commented on pull request #5920:
URL: https://github.com/apache/incubator-tvm/pull/5920#issuecomment-649501459
@masahi This is needed for the latest ROCm and also recent LLVM.
Without it, the execution will fail with "symbol not found error", which probably is hard to decypher.
I'm not particularly fond of needing to go through the device API, but I haven't found a better way, as the codegen wants to be compilable without having HIP installed.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] tqchen merged pull request #5920: Two small fixes to AMDCPU codegen for LLVM 10+ and ROCm 3.5+
Posted by GitBox <gi...@apache.org>.
tqchen merged pull request #5920:
URL: https://github.com/apache/incubator-tvm/pull/5920
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org