feat(vpto): support FP8 local contiguous memory lowering - #1116
Conversation
cfc4c63 to
3cf067a
Compare
|
Review note (non-blocking for the intended local-memory path): I validated the LLVM IR produced by this PR for FP8 bisheng --target=hiipu64-hisilicon-cce \
-march=dav-c310-vec \
--cce-aicore-arch=dav-c310-vec \
--cce-aicore-only -O2 -c -x ir <input.ll> -o <output.o>That path compiles successfully. One boundary should probably be made explicit: directly constructed FP8 constants such as arith.constant dense<0.0> : vector<4xf8E5M2>should be treated as unsupported, rather than lowered through the generic typed-carrier pattern. The generic pattern converts the result type while retaining the original typed I suggest adding an early diagnostic/verifier rejection plus a negative test for direct FP8 constants. This does not affect the supported |
3cf067a to
0e9f0ab
Compare
0e9f0ab to
7bd22de
Compare
7bd22de to
47af597
Compare
A3 板测失败
失败用例
|
A3 板测失败详情:PR #1116comm_p2p_binding_variants
comm_p2p
|
A5 板测成功
|
f8E4M3FNandf8E5M2scalar/vector types to native HiVM LLVM FP8 types.llvm.allocaandllvm.getelementptr.