This is the first integration test for AVX512. Many AVX512
instructions require specialized hardware, so tests like this one cannot be
run natively on most machines. Therefore, AVX512 tests are enabled via a CMake
build parameter. In addition, if the path to Intel SDE is specified, AVX512
instructions are emulated and the test can run on any machine.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
Neat to see AVX512 integration tests!
mlir/CMakeLists.txt | ||
---|---|---|
145 | arch-specific (omit the period) reads a bit better | |
mlir/integration_test/Dialect/Vector/CPU/AVX512/lit.local.cfg | ||
5 | This particular vpsintersection requires tiger lake AVX512_vp2intersect extensions. Do we want to introduce particular ways of specifying the different AVX512_xxx subsets, or do we just assume that AVX512 implies all subsets? (don't need to resolve that in this CL, just wondering) | |
mlir/integration_test/Dialect/Vector/CPU/AVX512/test-vp2intersect-i32.mlir | ||
7 | ust -> must |
mlir/integration_test/Dialect/Vector/CPU/AVX512/lit.local.cfg | ||
---|---|---|
5 | That may be a good idea, depending on how important it is to run the code natively. Maybe for performance experiments. My CPU actually supports a few AVX512 instructions, but not vp2intersect. (With SDE, the test can run on any machine, but it is really slow.) |
mlir/integration_test/Dialect/Vector/CPU/AVX512/test-vp2intersect-i32.mlir | ||
---|---|---|
4 | Nice, thanks for investigating and making this work! |
mlir/integration_test/Dialect/Vector/CPU/AVX512/test-vp2intersect-i32.mlir | ||
---|---|---|
43 | I see this has been submitted, so I am a bit after the fact, but since this deals with the AVX512 dialect, I would have had a small preference for integration_test/Dialect/AVX512/CPU instead of |
arch-specific (omit the period) reads a bit better