A hash calculated via FNV-1a algorithm is introduced in ihipModule_t, the
internal of hipModule_t. The hash is used by HIP module APIs such as
- read_agent_global_from_module
to determine whether the agent-scope globals for a module have been iterated.
This commit fixes one issue that applications which load / unload modules
frequently would occasionally fail. After deep investigation of the issue it
turns out the old implementation in read_agent_global_from_module uses
hipModule_t as the key, which is not robust enough, as hipModule_t instances
are allocated dynamically so there are cases that one memory address may be
used by multiple hipModule_t instances. The real solution is to introduce a
uniquely identifiable hash for the code object associated with the HIP module.
And that's the rationale behind this commit.
hipify-clang could be built by Visual Studio 14 2015 or higher.
[Reason]
hipify-clang contains c++14 features which are not supported even by the latest VS 2013 and the Platform Toolset "Visual C++ Compiler Nov 2013 CTP".
- Improved install instructions for AMD-hcc and NVIDIA-nvcc.
- Removed redundant section related to HCC-Native-GCN-ISA.
Change-Id: I33a9e71d95df51dbdcade4a48834840146bdbe6f
Header files in those dirs are an integral part of clang, which includes wrapper headers for CUDA.
There is no need in checking those dirs.
[Reasons]
1. clang will not take into account empty or irrelevant include dirs;
2. Packages for different clang versions will have different header files in the above dirs.
[Reasons]
+ We don't compile kernel code at least for now as HIP kernel syntax is almost equal CUDA's;
+ clang always includes PTX in its binaries, so e.g. a binary compiled with --cuda-gpu-arch= would be forwards-compatible with e.g. sm_35 GPUs.
+ Due to latest Windows SDK implementation based on c++14 features, '-std+c++14' is a must;
+ Doesn't affect older versions of MSVC 2017, 2015 (the latest supported is MSVC 2015);
+ On linux still '-std=c++11' is a must;
+ Exclude '-std=' from lit config.
[Reasons]
+ File is opened further by clang
+ There is a regression with LLVM 5.x, when system error occurs: temporary file couldn't be renamed - permission denied.
Option -o-dir for output directory:
+ if not specified source file(s) directory is used;
+ creates the directory if the directory doesn't exist (only one level in a tree);
+ if -o and -o-dir both are specified the hipified file is placed to "-o-dir" + "-o";
+ reports an error in case of a wrong directory specified, in case of necessity of creating a tree of subfolders, or in case of a filename specified.
1. Option -temp-dir for temporary directory:
+ if not specified system temp is used;
+ creates the directory if the directory doesn't exist (only one level in a tree);
+ reports an error in case of a wrong directory specified, in case of necessity of creating a tree of subfolders, or in case of a filename specified.
2. Option -save-temps for preserving temporary files:
+ if specified temporary files are not being deleted from system temps and user temps as well.
3. Work with files in terms of calculated absolute paths by collapsing all '.' and '..' patterns, resolving symlinks and expanding '~' expression to the user's home directory:
+ to produce correct include paths;
+ to avoid possible errors on file routines.
+ Files are not being placed near original source before hipification anymore, system tmp is used for them now.
+ Additional include (-I) to absolute original source dir is added
+ Partially implements #812
+ Tested on windows/linux, single/multiple inputs, with and without includes to user header files
+ Just do load_config for lit.site.cfg earlier in lit.cfg to get initialized cuda_version
+ Remove lit.cfg.in for CMake variables substitution
+ Return back lit.cfg
+ Exclude CUDA 9.0 tests if CUDA < 9.0
+ Report CUDA version before running tests
+ Add lit.cfg.in for CMake variables substitution to now generated lit.cfg
+ Exclude lit.cfg
+ Exclude CUDA 9.0 tests if CUDA < 9.0
+ Report CUDA version before running tests
+ Option '-roc' is added to translate to ROCm instead of HIP
+ Update hipification mechanism accordingly
+ Populate all translation maps with additional field for ROCm identifiers
+ There are 4 degrees of CUDA support now: full support, only by HIP, only by ROCm, unsupported
Evgeny Mankov
Wen-Heng (Jack) Chung
Maneesh Gupta
emankov
Sameer Sahasrabuddhe
Alex Voicu
Yaxun Sam Liu
Rahul Garg
Aaron Enye Shi