merge kernels in existing XCLBIN #418
Conversation
nirvedhmeshram
force-pushed
the
nm_xclbin_merge
branch
from
6791d86 to
aaec27e
Compare
nirvedhmeshram
force-pushed
the
nm_xclbin_merge
branch
from
f5d34a8 to
5f49f35
Compare
nirvedhmeshram
force-pushed
the
nm_xclbin_merge
branch
from
b9c1259 to
34c47e3
Compare
nirvedhmeshram
requested review from
MaheshRavishankar,
yzhang93,
Abhishek-Varma and
jtuyls
as code owners
nirvedhmeshram
force-pushed
the
nm_xclbin_merge
branch
from
2a09422 to
a11e419
Compare
compiler/plugins/target/AMD-AIE/iree-amd-aie/Target/AIETarget.cpp
Outdated
Show resolved
Hide resolved
| @@ -350,11 +368,24 @@ LogicalResult AIETargetBackend::serializeExecutable( | |||
| { | |||
| SmallVector<StringRef> cmdEnvRefs{cmdEnv.begin(), cmdEnv.end()}; | |||
| int result = llvm::sys::ExecuteAndWait(cmdArgs[0], cmdArgs, cmdEnvRefs); | |||
| if (result != 0) | |||
| if (result != 0 && AttemptingMerge) { | |||
There was a problem hiding this comment.
Can we detect before hand if the same kernel appears twice? Just because it'll be better to know why we failed. My concern is that we'll stop being able to merge for some other reason, and not know about it.
There was a problem hiding this comment.
I am not sure I correctly understand the case you are thinking about but if I understand correctly you are thinking some kernel cannot be merged in the existing XCLBIN and if a later kernel is same as the one that couldnt be we will just not be able to merge that either?
If what I interpret is correct then a few reasons why this is not of concern.
- IREE has deduplication passes after dispatch formation so same kernel is never appearing twice.
- The success of the merge is independent of the content in the kernel. The only case that I am aware of when it will fail is when we exceed the number of PDIs that xclbin supports. We have one PDI per kernel for now until we get to super kernels. I believe the PDI limit is 16. So my idea here is the 17th kernel will fail and then start a new xclbin, and the process repeats for the 33rd kernel and so on. And as far as the xclbin utility is concerned all 33 of these kernels (PDIs) could have same content and that fact wouldn't matter. A bit hard to test this though and we arent at that scale anyway now so havent worked on testing that.
Yes that is correct. |
|
On further investigating I realized that with the wheel the CI is using, no actual merging is taking place, you can see this error in the CI test added This causes the merging to fail and it goes through the regular new xclbin per dispatch path. |
This PR uses a new xclbin generation tool from Xilinx/mlir-aie#1561 that allows us to merge kernel PDIs into the same XCLBIN.
A mlir-aie wheel bump is needed to use the tool.
An e2e example which forms multiple (three) dispatches is added.