Open IanWood1 opened 1 week ago
When investigating https://github.com/iree-org/iree/issues/19131, I ran the input IR on the gpu backend (vs cpu) and it ended up causing a crash in PadDynamicAllocPass.
PadDynamicAllocPass
Here (https://gist.github.com/IanWood1/c2a3ac9cc3801faedead58a958a20b88) is the stacktrace as well as the IR before the failing pass (https://gist.github.com/IanWood1/b8ca10625ffb5efdc1ce1780fab71847).
iree-compile --iree-hal-target-device=hip --iree-hip-target=gfx1100 --iree-dispatch-creation-enable-aggressive-fusion=true quantmm.mlir
2311e0458bfe47a9f9a681c5e1da5dc0dedd7129
Looks like the upper bound is just below INT64_MAX, so it attempts to create a large memref subview
%21:2 = util.assume.int %10<udiv = 4096>, %20<umin = 0, umax = 9007199254740991> : index, index
https://github.com/iree-org/iree/blob/3bb7fd2384e522638994e768c816f39d9af99906/compiler/src/iree/compiler/Codegen/Common/PadDynamicAlloc.cpp#L49-L54
What happened?
When investigating https://github.com/iree-org/iree/issues/19131, I ran the input IR on the gpu backend (vs cpu) and it ended up causing a crash in
PadDynamicAllocPass
.Here (https://gist.github.com/IanWood1/c2a3ac9cc3801faedead58a958a20b88) is the stacktrace as well as the IR before the failing pass (https://gist.github.com/IanWood1/b8ca10625ffb5efdc1ce1780fab71847).
Steps to reproduce your issue
iree-compile --iree-hal-target-device=hip --iree-hip-target=gfx1100 --iree-dispatch-creation-enable-aggressive-fusion=true quantmm.mlir
Version information
2311e0458bfe47a9f9a681c5e1da5dc0dedd7129
Additional context
Looks like the upper bound is just below INT64_MAX, so it attempts to create a large memref subview
https://github.com/iree-org/iree/blob/3bb7fd2384e522638994e768c816f39d9af99906/compiler/src/iree/compiler/Codegen/Common/PadDynamicAlloc.cpp#L49-L54