Manuel Drehwald

Results 98 issues of Manuel Drehwald

Down to 1.2k LoC of IR. Wip ```.cpp #include #include #include class SuperGeometryStatistics3D2 { private: std::map foo; double bar; public: void update() { typename std::map::const_iterator iter; for (iter = foo.begin();...

To test Enzyme for higher orders and improve our docs I tried to follow the[ jax cookbook](https://jax.readthedocs.io/en/latest/notebooks/autodiff_cookbook.html). It fails however pretty early due to Eigen introducing some integer operation which...

enhancement

Not sure if worth investigating, but while writing a test i just noticed this failure: ```.sh FAIL: Enzyme :: Enzyme/ReverseMode/ompsqloopPostOpt.ll (670 of 677) ******************** TEST 'Enzyme :: Enzyme/ReverseMode/ompsqloopPostOpt.ll' FAILED ********************...

While trying to re(pro)duce another `cannot deal with ptr that isn't arg` issue for OpenLB I ran into this assertion (using undef-fn-inactive, strict-aliasing=0, release-ta). How interesting is this to minimize,...

Apparently Tim's update didn't fix all of them @wsmoses. When compiling OpenLB while also setting globals-default-inactive, (beside of the mandatory strict-aliasing=0 and relaxing TA) the compilation succeeds (although with likely...

Just some deduplications / simplifications to make it easier to generate the same code using tg.

[x] emit missing flags for some args [x] update tests [x] adjust d_alpha update (not a pure blas call) [] add new tests for axpy and scal

I do have a trivial but huge program which takes 7s to compile using clang-14.04 with -O3 and lto for the normal pass. using __enzyme_autodiff together with -O3 and lto...

bug
needs more information

Motivated by https://discourse.llvm.org/t/cast-x-is-broken-implications-and-proposal-to-address/63033 I just build enzyme on top of llvm-head. Enzyme would compile fine and ninja check-enzyme only has 6 extra tests failing (on top of the old 6...

enhancement