Support CMAKE_OSX_ARCHITECTURES #8390
Conversation
alexreinking
added
release_notes
| "COMMAND;DEPENDS;EXTRA_OUTPUTS;PARAMS;PLUGINS;TARGETS" | ||
| ) | ||
|
|
||
| ## "hash table" of extra outputs to extensions |
There was a problem hiding this comment.
Can we add 'hlpipe' here too? It should already be hooked up in the generator interface.
There was a problem hiding this comment.
Definitely. Didn't know it was missing!
| OUTPUT_VARIABLE merged_libs_targets) | ||
|
|
||
| find_program(LIPO lipo REQUIRED) | ||
| add_custom_command( |
There was a problem hiding this comment.
Nicely done, and with all the dependencies setup too!
| add_halide_library(adams2019_train_cost_model FROM adams2019_cost_model.generator | ||
| GENERATOR train_cost_model | ||
| FUNCTION_NAME train_cost_model | ||
| TARGETS cmake | ||
| FEATURES[x86-64-osx] avx2 sse41 | ||
| FEATURES[arm-64-osx] arm_dot_prod-arm_fp16 |
There was a problem hiding this comment.
Are these features available on all Apple Silicon?
alexreinking
force-pushed
the
build/osx-universal2
branch
2 times, most recently
from
3f4eb33 to
88701db
Compare
| function(_bundle_static_is_apple_libtool result item) | ||
| function(_bundle_static_is_apple_libtool result_var item) | ||
| _bundle_static_check_output(version_info "${item}" -V) | ||
| if (version_info MATCHES "Apple, Inc.") | ||
| set(result 1 PARENT_SCOPE) | ||
| else () | ||
| set(result 0 PARENT_SCOPE) | ||
| if (NOT version_info MATCHES "Apple, Inc.") | ||
| set(${result_var} 0 PARENT_SCOPE) |
There was a problem hiding this comment.
This is a drive-by fix.
|
Topics for discussion at the dev meeting:
|
|
Finally, we have a clean build! The @derek-gerstmann -- I'd appreciate some help testing offline. |
|
Testing on my MacBook Pro M1 Max and everything appears to work! I agree, it'd be nice to find a testing strategy for both desktop and mobile. |
|
From dev meeting:
|
This is potentially a breaking change for some builds that aren't using add_halide_generator or vcpkg on Windows.
Previously _Halide_try_load_generators was unhygienic with respect to scope. Now it takes a proper argument.
alexreinking
force-pushed
the
build/osx-universal2
branch
from
8396298 to
8df371a
Compare
|
This is ready on my end. @steven-johnson -- we should coordinate adding testing to the buildbots. Should we wait to land this until we have that? |
yes, probably |
alexreinking
changed the title
No, not yet |
|
Landing this with the caveat that it might need adjustment to support IOS cross-compilation via Xcode. At the moment it seems to have not broken existing workflows and has (locally) enabled universal builds on macOS. |
As requested at the last Halide dev meeting, I have drafted an implementation of support for CMAKE_OSX_ARCHITECTURES. This is again quite complicated, in part because I attempted to avoid breaking existing user code.
As it stands, multi-target alternative selection and feature specification are stepping on each others' toes.
I'll write more tomorrow.
Fixes #8134
Breaking suggestion:
TARGETSargument and theHalide_TARGETvariable shall take only the following values:host(passed verbatim to the generator, but resolved for feature selection)cmakeARCH-BITS-OS(a triple)CMAKE_OSX_ARCHITECTURESis set,cmakeexpands to a list of compatible architectures, e.g.arm-64-osx;x86-64-osx. This is reflected inHalide_CMAKE_TARGET.hostexpands to a single triple corresponding to the architecture of thecmakeexecutable itself (this enables compatibility with thearchcommand).FEATURESargument applies a list of features to ALL alternatives on ALL platforms (useful for, e.g.user_context, name mangling, etc.)FEATURES[platform]arguments set up a list of alternatives (e.g.FEATURES[x86-64-linux] "avx2-sse4" "sse4") for a particular triple.What this doesn't resolve is how to do things like
TARGETS cmake-no_bounds_query cmakeeffectively. Perhaps the following adjustments to the above help:FEATURES[platform]toMULTITARGET[platform]MULTITARGET(unadorned) to provide the default value for unspecified[platform]s.Specifying an empty entry in
FEATURES/MULTITARGETor whatever is challenging in CMake because it really wants to drop empty function arguments. We might need to add a dummy feature likenone(either in the actual Generator or just in the build system) that does nothing but isn't the empty string.