On 07/10/17 14:09, Rebecca N. Palmer wrote:
The problems with current beignet look to be:
- file timestamps in INPUT_FILES_BLOCK (some of beignet's .h files are
script-generated). This part can be fixed in beignet.
That works (COMMAND touch -d '@$ENV{SOURCE_DATE_EPOCH}'
${OCL_OBJECT_DIR}/in
The problems with current beignet look to be:
- file timestamps in INPUT_FILES_BLOCK (some of beignet's .h files are
script-generated). This part can be fixed in beignet.
- build path captured in ORIGINAL_PCH_DIR. _Might_ be fixable in
beignet by using clang (...) > /path/to/beignet.pch ins
Control: tags -1 upstream patch
Running the affected files through llvm-bcanalyzer says the difference
is in the ordering of OPENCL_EXTENSION_TYPES, and this makes sense as
these are stored in a hash map (i.e. unordered):
http://sources.debian.net/src/llvm-toolchain-5.0/1:5.0%7E%2Brc2-1/clang/
Possible workarounds for beignet:
- Remove the .pch files entirely. This slows down compiling (which is a
run-time operation in OpenCL): 263sec instead of 90sec for the
540-kernel test suite = ~0.3sec extra per kernel compile. At a guess,
that's unlikely but not impossible to matter in pract
Package: clang-4.0
Version: 1:4.0.1-5
Severity: wishlist
User: reproducible-bui...@lists.alioth.debian.org
Usertags: toolchain randomness
Control: affects -1 beignet-opencl-icd liboclgrind-16.10
X-Debbugs-Cc: reproducible-b...@lists.alioth.debian.org
Generating a .pch precompiled header with clan
5 matches
Mail list logo