[libc][math][c23] Enable C23 _Float16 math functions on GPUs #99248
Conversation
|
@llvm/pr-subscribers-libc Author: OverMighty (overmighty) ChangesFull diff: https://github.com/llvm/llvm-project/pull/99248.diff 2 Files Affected:
diff --git a/libc/cmake/modules/CheckCompilerFeatures.cmake b/libc/cmake/modules/CheckCompilerFeatures.cmake
index a6d793d495c45..361c1e710b187 100644
--- a/libc/cmake/modules/CheckCompilerFeatures.cmake
+++ b/libc/cmake/modules/CheckCompilerFeatures.cmake
@@ -15,6 +15,12 @@ set(
# Making sure ALL_COMPILER_FEATURES is sorted.
list(SORT ALL_COMPILER_FEATURES)
+# Compiler features that are unavailable on GPU targets with the in-tree Clang.
+set(
+ CPU_ONLY_COMPILER_FEATURES
+ "float128"
+)
+
# Function to check whether the compiler supports the provided set of features.
# Usage:
# compiler_supports(
@@ -65,13 +71,26 @@ foreach(feature IN LISTS ALL_COMPILER_FEATURES)
set(CMAKE_TRY_COMPILE_TARGET_TYPE EXECUTABLE)
endif()
- try_compile(
- has_feature
- ${CMAKE_CURRENT_BINARY_DIR}/compiler_features
- SOURCES ${LIBC_SOURCE_DIR}/cmake/modules/compiler_features/check_${feature}.cpp
- COMPILE_DEFINITIONS -I${LIBC_SOURCE_DIR} ${compile_options}
- LINK_OPTIONS ${link_options}
- )
+ if(LIBC_TARGET_OS_IS_GPU)
+ # CUDA shouldn't be required to build the libc, only to test it, so we can't
+ # try to build CUDA binaries here. Since GPU builds are always compiled with
+ # the in-tree Clang, we just hardcode which compiler features are available
+ # when targeting GPUs.
+ if(feature IN_LIST CPU_ONLY_COMPILER_FEATURES)
+ set(has_feature FALSE)
+ else()
+ set(has_feature TRUE)
+ endif()
+ else()
+ try_compile(
+ has_feature
+ ${CMAKE_CURRENT_BINARY_DIR}/compiler_features
+ SOURCES ${LIBC_SOURCE_DIR}/cmake/modules/compiler_features/check_${feature}.cpp
+ COMPILE_DEFINITIONS -I${LIBC_SOURCE_DIR} ${compile_options}
+ LINK_OPTIONS ${link_options}
+ )
+ endif()
+
if(has_feature)
list(APPEND AVAILABLE_COMPILER_FEATURES ${feature})
if(${feature} STREQUAL "float16")
diff --git a/libc/config/gpu/entrypoints.txt b/libc/config/gpu/entrypoints.txt
index 63228216c85ec..20c97fbbbadc5 100644
--- a/libc/config/gpu/entrypoints.txt
+++ b/libc/config/gpu/entrypoints.txt
@@ -343,6 +343,77 @@ set(TARGET_LIBM_ENTRYPOINTS
libc.src.math.truncf
)
+if(LIBC_TYPES_HAS_FLOAT16)
+ list(APPEND TARGET_LIBM_ENTRYPOINTS
+ # math.h C23 _Float16 entrypoints
+ libc.src.math.canonicalizef16
+ libc.src.math.ceilf16
+ libc.src.math.copysignf16
+ libc.src.math.f16add
+ libc.src.math.f16addf
+ libc.src.math.f16addl
+ libc.src.math.f16div
+ libc.src.math.f16divf
+ libc.src.math.f16divl
+ libc.src.math.f16fma
+ libc.src.math.f16fmaf
+ libc.src.math.f16fmal
+ libc.src.math.f16sqrt
+ libc.src.math.f16sqrtf
+ libc.src.math.f16sqrtl
+ libc.src.math.f16sub
+ libc.src.math.f16subf
+ libc.src.math.f16subl
+ libc.src.math.fabsf16
+ libc.src.math.fdimf16
+ libc.src.math.floorf16
+ libc.src.math.fmaxf16
+ libc.src.math.fmaximum_mag_numf16
+ libc.src.math.fmaximum_magf16
+ libc.src.math.fmaximum_numf16
+ libc.src.math.fmaximumf16
+ libc.src.math.fminf16
+ libc.src.math.fminimum_mag_numf16
+ libc.src.math.fminimum_magf16
+ libc.src.math.fminimum_numf16
+ libc.src.math.fminimumf16
+ libc.src.math.fmodf16
+ libc.src.math.frexpf16
+ libc.src.math.fromfpf16
+ libc.src.math.fromfpxf16
+ libc.src.math.getpayloadf16
+ libc.src.math.ilogbf16
+ libc.src.math.ldexpf16
+ libc.src.math.llogbf16
+ libc.src.math.llrintf16
+ libc.src.math.llroundf16
+ libc.src.math.logbf16
+ libc.src.math.lrintf16
+ libc.src.math.lroundf16
+ libc.src.math.modff16
+ libc.src.math.nanf16
+ libc.src.math.nearbyintf16
+ libc.src.math.nextafterf16
+ libc.src.math.nextdownf16
+ libc.src.math.nexttowardf16
+ libc.src.math.nextupf16
+ libc.src.math.remainderf16
+ libc.src.math.remquof16
+ libc.src.math.rintf16
+ libc.src.math.roundevenf16
+ libc.src.math.roundf16
+ libc.src.math.scalblnf16
+ libc.src.math.scalbnf16
+ libc.src.math.setpayloadf16
+ libc.src.math.setpayloadsigf16
+ libc.src.math.totalorderf16
+ libc.src.math.totalordermagf16
+ libc.src.math.truncf16
+ libc.src.math.ufromfpf16
+ libc.src.math.ufromfpxf16
+ )
+endif()
+
set(TARGET_LLVMLIBC_ENTRYPOINTS
${TARGET_LIBC_ENTRYPOINTS}
${TARGET_LIBM_ENTRYPOINTS}
|
|
@jhuber6 I built the tests for AMDGPU, but only ran them on NVIDIA. Could you please try to run them on AMD for me, or should I try to run them on my other laptop? |
|
Oh, they actually don't use builtins. I forgot to add an |
|
Should we try to run some benchmarks before enabling builtins on GPUs? |
If the builitin is supported it's pretty much always going to be faster than the bit-twidding generic implementation. Maybe there's some exceptions for things like |
sqrt is always correct. The intrinsics should all be lowered correctly. If not, it's a bug that should be addressed in the backend and not worked around in a library. I'm not sure if the sin/cos cases pass conformance, OpenCL only just finally added half tests to the conformance suite so we can check. The f16 cases are easier because it's a lot easier to upcast to float and back |
| @@ -111,7 +111,7 @@ add_entrypoint_object( | |||
| DEPENDS | |||
| libc.src.__support.macros.properties.types | |||
| libc.src.__support.FPUtil.nearest_integer_operations | |||
| libc.src.__support.macros.properties.architectures | |||
| libc.src.__support.macros.properties.cpu_features | |||
There was a problem hiding this comment.
Do you mind running the bazel builds to see if these extra dependencies need to be added to bazel overlay?
There was a problem hiding this comment.
Bazel build seems to work:
~/projects/llvm-project/utils/bazel libc-math-float16-gpu ➜ bazel build --config=generic_clang @llvm-project//libc/...
INFO: Analyzed 897 targets (68 packages loaded, 6253 targets configured).
INFO: Found 897 targets...
INFO: Elapsed time: 166.512s, Critical Path: 151.57s
INFO: 3876 processes: 2132 internal, 1744 linux-sandbox.
INFO: Build completed successfully, 3876 total actions
~/projects/llvm-project/utils/bazel libc-math-float16-gpu ➜ bazel test --config=generic_clang @llvm-project//libc/...
INFO: Analyzed 897 targets (0 packages loaded, 0 targets configured).
INFO: Found 659 targets and 238 test targets...
INFO: Elapsed time: 5.366s, Critical Path: 4.26s
INFO: 239 processes: 476 linux-sandbox.
INFO: Build completed successfully, 239 total actions
@llvm-project//libc/test/src/__support:arg_list_test PASSED in 0.0s
...
Executed 238 out of 238 tests: 238 tests pass.
There were tests whose specified size is too big. Use the --test_verbose_timeout_warnings command line option to see which ones these are.
~/projects/llvm-project/utils/bazel libc-math-float16-gpu ➜
Why are only some math functions listed in https://github.com/llvm/llvm-project/blob/main/utils/bazel/llvm-project-overlay/libc/BUILD.bazel?
There was a problem hiding this comment.
B/c I've been lazy / forget to update it, especially since there are quite some work need to be done to add smoke tests to the lists. I planned to update it all at once at some point in late July, early August.
|
The PR looks good to me. Let's wait for @jhuber6 to do the confirmation / approval. |
|
#98972 was merged yesterday so maybe I should enable these functions in this PR too while at it, unless you would prefer another PR as this one is already approved. |
Enable ROUND_OPT flag and builtin usage for _Float16 functions on GPUs.
Disable narrowing functions that take `long double` arguments.
Enable f16mul and f16mulf.
overmighty
force-pushed
the
libc-math-float16-gpu
branch
from
a0de91e to
3c7f767
Compare
|
Rebased to enable |
|
@jhuber6 Is this still okay? I just disabled the functions that use |
Yes, so long as there's no long double stuff it should still work. If not we can turn it off, but it worked when I last tested it. |
No description provided.