-
Notifications
You must be signed in to change notification settings - Fork 74.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[AARCH64] Building TF 2.15.0 from sources failed with undefined __Int8x8_t #62490
Comments
I have the same build problem on the Arm64 with CPU: The CPU doesn't have the NEON feature at all. It needs to turn off Arm-neon support at this platform. Any suggestion to turn off? |
Still broken on 2.16.1 |
This seems to work around the issue for me - built on grace hopper, TF 2.16.1 using spack. Change to TF: diff --git a/third_party/absl/workspace.bzl b/third_party/absl/workspace.bzl
index 06f75166ce4b..56d146d65abe 100644
--- a/third_party/absl/workspace.bzl
+++ b/third_party/absl/workspace.bzl
@@ -42,6 +42,7 @@ def repo():
build_file = "//third_party/absl:com_google_absl.BUILD",
system_build_file = "//third_party/absl:system.BUILD",
system_link_files = SYS_LINKS,
+ patch_file = ["//third_party/absl:absl_neon.patch"],
strip_prefix = "abseil-cpp-{commit}".format(commit = ABSL_COMMIT),
urls = tf_mirror_urls("https://github.com/abseil/abseil-cpp/archive/{commit}.tar.gz".format(commit = ABSL_COMMIT)),
) Patch for absl: diff --git a/absl/base/config.h b/absl/base/config.h
index 5fa9f0efe5a4..741e320fe40c 100644
--- a/absl/base/config.h
+++ b/absl/base/config.h
@@ -962,7 +962,7 @@ static_assert(ABSL_INTERNAL_INLINE_NAMESPACE_STR[0] != 'h' ||
// https://llvm.org/docs/CompileCudaWithLLVM.html#detecting-clang-vs-nvcc-from-code
#ifdef ABSL_INTERNAL_HAVE_ARM_NEON
#error ABSL_INTERNAL_HAVE_ARM_NEON cannot be directly set
-#elif defined(__ARM_NEON) && !defined(__CUDA_ARCH__)
+#elif defined(__ARM_NEON) && !defined(__CUDACC__)
#define ABSL_INTERNAL_HAVE_ARM_NEON 1
#endif |
This was fixed in the latest version of absl: abseil/abseil-cpp#1732. Can we bump the vendored copy of absl before the next release? Then we can finally close this issue. |
Issue type
Build/Install
Have you reproduced the bug with TensorFlow Nightly?
No
Source
source
TensorFlow version
2.15.0
Custom code
Yes
OS platform and distribution
RHEL 8
Mobile device
No response
Python version
3.9
Bazel version
6.1.0
GCC/compiler version
GCC 12.3
CUDA/cuDNN version
Cuda 12.2 , cuDNN 8.8.0
GPU model and memory
No response
Current behavior?
Building TF 2.15.0 from sources for aarch64 fails with error like [a]. Note that building TF 2.15.0 from sources for x86_64 worked fine.
[a]
Standalone code to reproduce the issue
Relevant log output
No response
The text was updated successfully, but these errors were encountered: