Bring the remaining apps and test-infrastructure gaps onto CMake, and fix the build/runtime problems uncovered along the way. Apps: - resnet_50, auto_viz, hexagon_dma, and HelloPyTorch each get a CMake build. HelloPyTorch builds its ops with pytorch_wrapper output and synthesizes the pybind extension in CMake (taking include/library paths from torch.utils.cpp_extension), configuring only when Python and PyTorch are found; its CUDA ops are gated behind HELLO_PYTORCH_CUDA. - Remove the simd_op_check app: it linked per-op objects/headers that the correctness test no longer emits (it now JITs via compile_to_callable and emits only assembly), so it can no longer be built. - Drop the stale commented-out nn_ops add_app() placeholder, since the app is no longer in the tree. Benchmarks: - Give the six benchmark apps (bilateral_grid, camera_pipe, lens_blur, local_laplacian, nl_means, stencil_chain) a CMake-native benchmark path: emit a REGISTRATION source, build an <app>_rungen executable, and register an <app>_benchmark CTest. All are labeled benchmark_apps (not plain benchmark, to avoid a regex-match with hexagon_benchmarks) so `ctest -L benchmark_apps` runs them together. Test infrastructure: - Add CMake/CTest support for running the suite under valgrind (via CTest MemCheck and a new Halide_ENABLE_MEMCHECK option) and Intel SDE (via CMAKE_CROSSCOMPILING_EMULATOR). New presets drive both: valgrind, avx512-cannonlake, and avx512-knights-landing. - Fix the hexagon_remote install layout to mirror the committed src/runtime/hexagon_remote/bin/ subtree so the runtime can be rebuilt from source and installed back over the committed blobs. Fixes uncovered during the work: - auto_viz: cast the resampling kernel's tap count to int before using it as an RDom extent; point the auto_viz_demo test's HL_TRACE_FILE at the null device so trace_all output doesn't blow the test timeout under `ctest -V`; and correct the complex/down dispatch entry that pointed at the _up variant (test coverage expanded to all six schedule/direction variants). - hexagon_dma: pin copy_to_device() to HexagonDma, give the YUV chroma planes dynamic dimensionality, pin the Halide libraries to a host target so CI's ambient GPU-testing target isn't inherited, and skip the mock-DMA-driver executables/tests on MSVC. Fixes #8729 Fixes #8732 Fixes #8734 Fixes #8736 Fixes #8740 Fixes #8741 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Uh oh!

Add testing for HelloPyTorch #8729

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions