`omp_is_initial_device` in device code was implemented as a builtin
function in D38968 for a better performance. Therefore there is no chance that
this function will be called to `deviceRTLs`. As we're moving to build `deviceRTLs`
with OpenMP compiler, this function can lead to a compilation error. This patch
just simply removes it.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D95397
This patch makes prep for dropping CUDA when compiling `deviceRTLs`.
CUDA intrinsics are replaced by NVVM intrinsics which refers to code in
`__clang_cuda_intrinsics.h`. We don't want to directly include it because in the
near future we're going to switch to OpenMP and by then the header cannot be
used anymore.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D95327
D95161 removed the option `--libomptarget-nvptx-path`, which is used in
the tests for `libomptarget-nvptx`.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D95293
[libomptarget][nvptx] Replace cuda atomic primitives with clang intrinsics
Tested by diff of IR generated for target_impl.cu before and after. NFC. Part
of removing deviceRTL build time dependency on cuda SDK.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D95294
[libomptarget][cuda] Call v2 functions explicitly
rtl.cpp calls functions like cuMemFree that are replaced by a macro
in cuda.h with cuMemFree_v2. This patch changes the source to use
the v2 names consistently.
See also D95104, D95155 for the idea. Alternatives are to use a mixture,
e.g. call the macro names and explictly dlopen the _v2 names, or to keep
the current status where the symbols are replaced by macros in both files
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D95274
[libomptarget] Build cuda plugin without cuda installed locally
Compiles a new file, `plugins/cuda/dynamic_cuda/cuda.cpp`, to an object file that exposes the same symbols that the plugin presently uses from libcuda. The object file contains dlopen of libcuda and cached dlsym calls. Also provides a cuda.h containing the subset that is used.
This lets the cmake file choose between the system cuda and a dlopen shim, with no changes to rtl.cpp.
The corresponding change to amdgpu is postponed until after a refactor of the plugin to reduce the size of the hsa.h stub required
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D95155
The buckets are initialized in __kmp_dephash_create but when they are extended
the memory is allocated but not NULL'd, potentially leaving some buckets
uninitialized after all entries have been copied into the new allocation.
This commit makes sure the buckets are properly initialized with NULL before
copying the entries.
Differential Revision: https://reviews.llvm.org/D95167
[libomptarget][devicertl] Drop templated atomic functions
The five __kmpc_atomic templates are instantiated a total of seven times.
This change replaces the template with explictly typed functions, which
have the same prototype for amdgcn and nvptx, and implements them with
the same code presently in use.
Rolls in the accepted but not yet landed D95085.
The unsigned long long type can be replaced with uint64_t when replacing
the cuda function. Until then, clang warns on casting a pointer to one to
a pointer to the other.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D95093
Summary:
Prior to D91261 the information checked the OMP_MAP_TARGET_PARAM flag, change this as it has been removed. The INFO macro was changed to accept a flag as input to make conditionally printing information easier.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D95133
Profiling has been recently implemented in libomptarget (D93055). This patch enables time profiling support for libomptarget in libomp, to support profiling of multi-threaded execution of offloaded regions.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D94855
Pretty similar to D95058, this patch added forward declaration for
CUDA atomic functions. We already have definitions with right mangled names in
internal CUDA headers so the forward declaration here can work properly.
Reviewed By: jdoerfert, JonChesterfield
Differential Revision: https://reviews.llvm.org/D95085
Summary:
The custom mapper API did not previously support the mapping names added previously. This means they were not present if a user requested debugging information while using the mapper functions. This adds basic support for passing the mapped names to the runtime library.
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D94806
Once we switch to build deviceRTLs with OpenMP, primitives and CUDA
intrinsics cannot be used directly anymore because `__device__` is not recognized
by OpenMP compiler. To avoid involving all CUDA internal headers we had in `clang`,
we forward declared these functions. Eventually they will be transformed into
right LLVM instrinsics.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D95058
[libomptarget][devicertl][nfc] Simplify target_atomic abstraction
Atomic functions were implemented as a shim around cuda's atomics, with
amdgcn implementing those symbols as a shim around gcc style intrinsics.
This patch folds target_atomic.h into target_impl.h and folds amdgcn.
Further work is likely to be useful here, either changing to openmp's atomic
interface or instantiating the templates on the few used types in order to
move them into a cuda/c++ implementation file. This change is mostly to
group the remaining uses of the cuda api under nvptx' target_impl abstraction.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D95062
[libomptarget][devicertl][nfc] Remove some cuda intrinsics, simplify
Replace __popc, __ffs with clang intrinsics. Move kmpc_impl_min to only file
that uses it and replace template with explictly typed.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D95060
Replaced CUDA builtin vars with LLVM intrinsics such that we don't need
definitions of those intrinsics.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D95013
[libomptarget][devicertl] Wrap source in declare target pragmas
Factored out of D93135 / D94745. C++ and cuda ignore unknown pragmas
so this is a NFC for the current implementation language. Removes noise
from patches for building deviceRTL as openmp.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D95048
The basic design is to create an outer-most parallel team. It is not a regular team because it is only created when the first hidden helper task is encountered, and is only responsible for the execution of hidden helper tasks. We first use `pthread_create` to create a new thread, let's call it the initial and also the main thread of the hidden helper team. This initial thread then initializes a new root, just like what RTL does in initialization. After that, it directly calls `__kmpc_fork_call`. It is like the initial thread encounters a parallel region. The wrapped function for this team is, for main thread, which is the initial thread that we create via `pthread_create` on Linux, waits on a condition variable. The condition variable can only be signaled when RTL is being destroyed. For other work threads, they just do nothing. The reason that main thread needs to wait there is, in current implementation, once the main thread finishes the wrapped function of this team, it starts to free the team which is not what we want.
Two environment variables, `LIBOMP_NUM_HIDDEN_HELPER_THREADS` and `LIBOMP_USE_HIDDEN_HELPER_TASK`, are also set to configure the number of threads and enable/disable this feature. By default, the number of hidden helper threads is 8.
Here are some open issues to be discussed:
1. The main thread goes to sleeping when the initialization is finished. As Andrey mentioned, we might need it to be awaken from time to time to do some stuffs. What kind of update/check should be put here?
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D77609
[libomptarget][nvptx] Reduce calls to cuda header
Remove use of clock_t in favour of a builtin. Drop a preprocessor branch.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D94731
[libomptarget][nvptx][nfc] Move target_impl functions out of header
This removes most of the differences between the two target_impl.h.
Also change name mangling from C to C++ for __kmpc_impl_*_lock.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D94728
`omptarget-nvptx` is still a dependence for `check-libomptarget-nvtpx`
although it has been removed by D94573.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D94725
The comment said CUDA 9 header files use the `nv_weak` attribute which
`clang` is not yet prepared to handle. It's three years ago and now things have
changed. Based on my test, removing the definition doesn't have any problem on
my machine with CUDA 11.1 installed.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D94700
For NVPTX target, OpenMP provides a static library `libomptarget-nvptx`
built by NVCC, and another bitcode `libomptarget-nvptx-sm_{$sm}.bc` generated by
Clang. When compiling an OpenMP program, the `.bc` file will be fed to `clang`
in the second run on the program that compiles the target part. Then the generated
PTX file will be fed to `ptxas` to generate the object file, and finally the driver
invokes `nvlink` to generate the binary, where the static library will be appened
to `nvlink`.
One question is, why do we need two libraries? The only difference is, the static
library contains `omp_data.cu` and the bitcode library doesn't. It's unclear why
they were implemented in this way, but per D94565, there is no issue if we also
include the file into the bitcode library. Therefore, we can safely drop the
static library.
This patch is about the change in OpenMP. The driver will be updated as well if
this patch is accepted.
Reviewed By: jdoerfert, JonChesterfield
Differential Revision: https://reviews.llvm.org/D94573
Restore control of kernel launch tracing to be >= 1 as it was before
export LIBOMPTARGET_KERNEL_TRACE=1
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D94695
Hierarchical barrier is an experimental barrier algorithm that uses aspects
of machine hierarchy to define the barrier tree structure. This patch fixes
offset calculation in hierarchical barrier. The offset is used to store info
on a flag about sleeping threads waiting on a location stored in the flag.
This commit also fixes a potential deadlock in hierarchical barrier when
using infinite blocktime by adjusting the offset value of leaf kids so that
it matches the value of leaf state. It also adds testing of default barriers
with infinite blocktime, and also tests hierarchical barrier algorithm with
both default and infinite blocktime.
Patch by Terry Wilmarth and Nawrin Sultana.
Differential Revision: https://reviews.llvm.org/D94241
Add extra information to the runtime page describing the error messages and add information to the release notes for clang 12.0
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D94562
This change enables volatile use of persistent memory for omp_large_cap_mem*
on supported systems. It depends on libmemkind's support for persistent memory,
and requirements/details can be found at the following url.
https://pmem.io/2020/01/20/memkind-dax-kmem.html
Differential Revision: https://reviews.llvm.org/D94353
Constant static data member can be defined in the class without another
define after the class in C++17. Although it is C++17, Clang can still handle it
even w/o the flag for C++17. Unluckily, GCC cannot handle that.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D94541
[libomptarget][amdgpu][nfc] Fix build on centos
rtl.cpp replaced 224 with a #define from elf.h, but that
doesn't work on a centos 7 build machine with an old elf.h
Reviewed By: ronlieb
Differential Revision: https://reviews.llvm.org/D94528
Some LLVM headers are generated by CMake. Before the installation,
LLVM's headers are distributed everywhere, some of which are in
`${LLVM_SRC_ROOT}/llvm/include/llvm`, and some are in
`${LLVM_BINARY_ROOT}/include/llvm`. After intallation, they're all in
`${LLVM_INSTALLATION_ROOT}/include/llvm`.
OpenMP now depends on LLVM headers. Some headers depend on headers generated
by CMake. When building OpenMP along with LLVM, a.k.a via `LLVM_ENABLE_RUNTIMES`,
we need to tell OpenMP where it can find those headers, especially those still
have not been copied/installed.
Reviewed By: jdoerfert, jhuber6
Differential Revision: https://reviews.llvm.org/D94534
The lifetime of `libomptarget` and its opened plugins are not aligned
and it's hard for `libomptarget` to determine when the plugins are destroyed.
As a result, some issues (see D94256 for details) occur on some platforms.
Actually, if we take target memory as target resources, same as other resources,
such as CUDA streams, in each plugin, then the memory manager should also be in
the plugin. Also considering some platforms may want to opt out the feature, it
makes sense to move the memory manager to plugin, make it a common interface, and
let plguin developers determine whether they need it. This is what this patch does.
CUDA plugin is taken as example to show how to integrate it. In this way, we can
also get a bonus that different thresholds can be set for different platforms.
Reviewed By: jdoerfert, JonChesterfield
Differential Revision: https://reviews.llvm.org/D94379
For now `elf_common.c` is taken as a common part included into
different plugin implementations directly via
`#include "../../common/elf_common.c"`, which is not a best practice. Since it
is simple enough such that we don't need to create a real library for it, we just
take it as a interface library so that other targets can link it directly. Another
advantage of this method is, we don't need to add the folder into header search
path which can potentially pollute the search path.
VE and AMD platforms have not been tested because I don't have target machines.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D94443
For now, `*_STANDALONE_BUILD` is set to ON even if they're built along
with LLVM because of issues mentioned in the comments. This can cause some issues.
For example, if we build OpenMP along with LLVM, we'd like to copy those OpenMP
headers to `<prefix>/lib/clang/<version>/include` such that `clang` can find
those headers without using `-I <prefix>/include` because those headers will be
copied to `<prefix>/include` if it is built standalone.
In this patch, we fixed the dependence issue in OpenMP such that it can be built
correctly even with `OPENMP_STANDALONE_BUILD=OFF`. The issue is in the call to
`add_lit_testsuite`, where `clang` and `clang-resource-headers` are passed as
`DEPENDS`. Since we're building OpenMP along with LLVM, `clang` is set by CMake
to be the C/C++ compiler, therefore these two dependences are no longer needed,
where caused the dependence issue.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93738
Multiple `RTLInfoTy` objects are stored in a list `AllRTLs`. Since
`RTLInfoTy` contains a `std::mutex`, it is by default not a copyable object.
In order to support `AllRTLs.push_back(...)` which is currently used, a customized
copy constructor is provided. Every time we need to add a new data member into
`RTLInfoTy`, we should keep in mind not forgetting to add corresponding assignment
in the copy constructor. In fact, the only use of the copy constructor is to push
the object into the list, we can of course write it in a way that first emplace
a new object back, and then use the reference to the last element. In this way we
don't need the copy constructor anymore. If the element is invalid, we just need
to pop it, and that's what this patch does.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D94361
Fugaku supercomputer is built with the Fujitsu A64FX microprocessor, whose cache line is 256. In current libomp, we only have cache line size 128 for PPC64 and otherwise 64. This patch added the support of cache line 256 for A64FX. It's worth noting that although A64FX is a variant of AArch64, this property is not shared. As a result, in light of UCX source code (392443ab92/src/ucs/arch/aarch64/cpu.c (L17)), we can only determine by checking whether the CPU is FUJITSU A64FX.
Reviewed By: jdoerfert, Hahnfeld
Differential Revision: https://reviews.llvm.org/D93169
Summary:
Currently error messages from the CUDA plugins are only printed to the user if they have debugging enabled. Change this behaviour to always print the messages that result in offloading failure. This improves the error messages by indidcating what happened when the error occurs in the plugin library, such as a segmentation fault on the device.
Reviewed by: jdoerfert
Differential Revision: https://reviews.llvm.org/D94263
Add an example to the OpenMP Documentation on the LIBOMPTARGET_INFO environment variable
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D94246
Wrong LLVM headers might be included if we don't set `include_directories`
to a right place. This will cause a compilation error if LLVM is installed in
system directories.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93737
Currently all built libraries in OpenMP are anywhere if building along
with LLVM. It is not an issue if we don't execute any test. However, almost all
tests for `libomptarget` fails because in the lit configuration, we only set
`<build_dir>/libomptarget` to `LD_LIBRARY_PATH` and `LIBRARY_PATH`. Since those
libraries are everywhere, `clang` can no longer find `libomptarget.so` or those
deviceRTLs anymore.
In this patch, we set a unified path for all built libraries, no matter whether
it is built along with LLVM or not. In this way, our lit configuration can work
propoerly.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93736
Summary:
This patch adds more fine-grained support over which information is output from the libomptarget runtime when run with the environment variable LIBOMPTARGET_INFO set. An extensible set of flags can be used to pick and choose which information the user is interested in.
Reviewers: jdoerfert JonChesterfield grokos
Differential Revision: https://reviews.llvm.org/D93727
[libomptarget][amdgpu] Call into deviceRTL instead of ockl
Amdgpu codegen presently emits a call into ockl. The same functionality
is already present in the deviceRTL. Adds an amdgpu specific entry point
to avoid the dependency. This lets simple openmp code (specifically, that
which doesn't use libm) run without rocm device libraries installed.
Reviewed By: ronlieb
Differential Revision: https://reviews.llvm.org/D93356
This patch partially prepares the runtime source code to be built with
-Wconversion, which should trigger warnings if any implicit conversions
can possibly change a value. For builds done with icc or gcc, all such
warnings are handled in this patch. clang gives a much longer list of
warnings, particularly for sign conversions, which the other compilers
don't report. The -Wconversion flag is commented into cmake files, but
I'm not going to turn it on. If someone thinks it is important, and wants
to fix all the clang warnings, they are welcome to.
Types of changes made here involve either improving the consistency of types
used so that no conversion is needed, or else performing careful explicit
conversions, when we're sure a problem won't arise.
Patch is a combination of changes by Terry Wilmarth and Johnny Peyton.
Differential Revision: https://reviews.llvm.org/D92942
Add support to the OpenMP web pages for environment variables supported
by Libomptarget and their usage.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93723
When setting `LLVM_ENABLE_RUNTIMES`, lower case word should be used;
otherwise, it can cause a CMake error that specific path is not found.
Reviewed By: ye-luo
Differential Revision: https://reviews.llvm.org/D93719
After some issues about building runtimes along with LLVM were fixed,
building an OpenMP offloading capable compiler is pretty simple. This patch updates
the FAQ part in the doc.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93671
If a GPU function is externally reachable we give up trying to find the
(unique) kernel it is called from. This can hinder optimizations. Emit a
remark and explain mitigation strategies.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D93439
This patchs adds CMake variables to add subdirectories and include
directories for libomptarget and explicitly gives the location of source
files.
Differential Revision: https://reviews.llvm.org/D93290
[libomptarget][nfc] Replace static const with enum
Semantically identical. Replaces 0xff... with ~0 to spare counting the f.
Has the advantage that the compiler doesn't need to prove the 4/8 byte
value dead before discarding it, and sidesteps the compilation question
associated with what static means for a single source language.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93328
Introduce new kmp_safe_raii_file_t class with RAII semantics for file
open/close. It is essentially a wrapper around the C-style FILE* object.
This also unifies the way we error report if a file can't be opened.
Differential Revision: https://reviews.llvm.org/D92604
This patch enables serial initialization in the forked child process
to fix unstable runtime behavior when used with Python-based AI tools.
Differential Revision: https://reviews.llvm.org/D93230
[libomptarget][nfc] Remove data_sharing type aliasing
Libomptarget previous used __kmpc_data_sharing_slot to access values of type
__kmpc_data_sharing_{worker,master}_slot_static. This aliasing violation was
benign in practice. The master type has since been removed, so a single type
can be used instead.
This is particularly helpful for the transition to an openmp deviceRTL, as the
c++/openmp compiler for amdgcn currently rejects the flexible array member for
being an incomplete type. Serves the same purpose as abandoned D86324.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93075
This patch introduces a new RTM lock type based on spin lock which is
used for OMP lock with speculative hint on supported architecture.
Differential Revision: https://reviews.llvm.org/D92615
This patch adds new API __kmpc_taskloop_5 to accomadate strict
modifier (introduced in OpenMP 5.1) in num_tasks and grainsize
clause.
Differential Revision: https://reviews.llvm.org/D92352
KMP_AFFINITY=norespect was triggering an error because the underlying
process affinity mask was not updated to include the entire machine.
The Windows documentation states that the thread affinities must be
subsets of the process affinity. This patch also moves the printing
(for KMP_AFFINITY=verbose) of whether the initial mask was respected
out of each topology detection function and to one location where the
initial affinity mask is read.
Differential Revision: https://reviews.llvm.org/D92587
Check pointer returned by strchr, as it can be NULL in case of broken
format of input string. Introduced new function __kmp_str_loc_numbers
for fast parsing of numbers only in the location string.
Also made some cleanup of __kmp_str_loc_init declaration and usage:
- changed type of init_fname parameter to bool;
- changed input from true to false in places where fname is not used.
Differential Revision: https://reviews.llvm.org/D90962
D91692 missed various locations in kmp_gsupport, where the scope for
OMPT_STORE_RETURN_ADDRESS is too narrow, i.e. the scope ends before the OMPT
callback is called in some nested function.
This patch fixes the scoping issue, so that all OMPT tests pass, when the
tests are built with gcc.
Differential Revision: https://reviews.llvm.org/D92121
[libomptarget][amdgpu] Address compiler warnings, drive by fixes
Initialize some variables, remove unused ones.
Changes the debug printing condition to align with the aomp test suite.
Differential Revision: https://reviews.llvm.org/D92559
These changes add support for Intel's umonitor/umwait usage in wait
code, for architectures that support those intrinsic functions. Usage of
umonitor/umwait is off by default, but can be turned on by setting the
KMP_USER_LEVEL_MWAIT environment variable.
Differential Revision: https://reviews.llvm.org/D91189
Added UNLIKELY hint to one-time or rarely executed branches.
This improves performance of the library on some tasking benchmarks.
Differential Revision: https://reviews.llvm.org/D92322
With the change to using shared memory, there were a few problems that need to be fixed.
- The previous filename that was used for SHM only used process id. Given that process is
usually based on 16bit number, this was causing some conflicts on machines. Thus we add
UID to the name to prevent this.
- It appears under some conditions (SIGTERM, etc) the shared memory files were not getting
cleaned up. Added a call to clean up the shm files under those conditions. For this user
needs to set envirable KMP_HANDLE_SIGNALS to true.
Patch by Erdner, Todd <todd.erdner@intel.com>
Differential Revision: https://reviews.llvm.org/D91869
Once __kmp_task_finish is not executed for proxy tasks,
move mutexinoutset dependency code to __kmp_release_deps
which is executed for all task kinds.
Differential Revision: https://reviews.llvm.org/D92326
[libomptarget][cuda] Detect missing symbols in plugin at build time
Passes -z,defs to the linker. Error on unresolved symbol references.
Otherwise, those unresolved symbols present as target code running on the host
as the plugin fails to load. This is significantly harder to debug than a link
time error. Flag matches that passed by amdgcn and ve plugins.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D92143
This is an alternative approach to address inconsistencies pointed out in: D90078
This patch makes sure that the return address is reset, when leaving the scope.
In some cases, I had to move the macro out of an if-statement to have it in the
right scope, in some cases I added an additional block to restrict the scope.
This patch does not handle inconsistencies, which might occur if the return
address is still set when we call into the application.
Test case (repeated_calls.c) provided by @hbae
Differential Revision: https://reviews.llvm.org/D91692
OpenMP 5.1 introduces the new env variable
OMP_TOOL_VERBOSE_INIT=(disabled|stdout|stderr|<filename>) to enable verbose
loading and initialization of OMPT tools.
This env variable helps to understand the cause when loading of a tool fails
(e.g., undefined symbols or dependency not in LD_LIBRARY_PATH)
Output of OMP_TOOL_VERBOSE_INIT is added for OMP_DISPLAY_ENV
Tests for this patch are integrated into the different existing tool loading
tests, making these tests more verbose. An Archer specific verbose test is
integrated into an existing Archer test.
Patch prepared by: Isabel Thärigen
Differential Revision: https://reviews.llvm.org/D91464
Adjusted external reference for Darwin/AARCH64 link compatibility.
Made size directive conditional only if __ELF__ defined.
Patch by Michael_Pique <mpique@icloud.com>
Differential Revision: https://reviews.llvm.org/D88252
This patch is the runtime support for https://reviews.llvm.org/D84192.
In order not to modify the tgt_target_data_update information but still be
able to pass the extra information for non-contiguous map item (offset,
count, and stride for each dimension), this patch overload arg when
the maptype is set as OMP_TGT_MAPTYPE_DESCRIPTOR. The origin arg is for
passing the pointer information, however, the overloaded arg is an
array of descriptor_dim:
```
struct descriptor_dim {
int64_t offset;
int64_t count;
int64_t stride
};
```
and the array size is the dimension size. In addition, since we
have count and stride information in descriptor_dim, we can replace/overload the
arg_size parameter by using dimension size.
Reviewed By: grokos, tianshilei1992
Differential Revision: https://reviews.llvm.org/D82245
Summary:
Add support for passing source locations to libomptarget runtime functions using the ident_t struct present in the rest of the libomp API. This will allow the runtime system to give much more insightful error messages and debugging values.
Reviewers: jdoerfert grokos
Differential Revision: https://reviews.llvm.org/D87946
Summary:
This patch adds basic support for priting the source location and names for the mapped variables. This patch does not support names for custom mappers. This is based on D89802.
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D90172
Summary:
This patch adds support for passing in the original delcaration name in the source file to the libomptarget runtime. This will allow the runtime to provide more intelligent debugging messages. This patch takes the original expression parsed from the OpenMP map / update clause and provides a textual representation if it was explicitly mapped, otherwise it takes the name of the variable declaration as a fallback. The information in passed to the runtime in a global array of strings that matches the existing ident_t source location strings using ";name;filename;column;row;;"
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D89802
This introduces the new `ARCHER_OPTIONS` flag `ignore_serial=0|1` to disable
analysis and logging of memory accesses in the sequential part of the OpenMP
application.
In the sequential part of an OpenMP program no data race is possible, unless
there is non-OpenMP concurrency (such as pthreads, MPI, ...). For the latter
reason, this is not active by default.
Besides reducing the runtime overhead for the sequential part of the program,
this reduces the memory overhead for sequential initialization. In combination
with `flush_shadow=1` this can allow analysis of applications, which run close
to the limit of available memory, but only access smaller parts of shared
memory during each OpenMP parallel region.
A problem for this approach is that Archer only gets active, when the OpenMP
runtime gets initialized, which might be after serial initialization of the
application. In such case, it helps to call for example `omp_get_max_threads()`
at the beginning of main.
Differential Revision: https://reviews.llvm.org/D90473
OpenMP 5.1 adds an extra enum entry for ompt_scope_t, which makes the related
switch statement incomplete.
Also adding cases for newly added barrier variants.
Differential Revision: https://reviews.llvm.org/D90758
Currently the affinity format string has initial value. When users set
the format via OMP_AFFINITY_FORMAT, it will overwrite the format string. However,
when copying the format, the tailing null is missing. As a result, if the user
format string is shorter than default value, the remaining part in the default
value still makes effort. This bug is not exposed because the test case doesn't
check the end of a string. It only checks whether given output "contains" the
check string.
Reviewed By: AndreyChurbanov
Differential Revision: https://reviews.llvm.org/D91309
Summary:
This patch begins to add support for a set of scripts that can be used to get information from OpenMP programs to better describe problems and eventually show the data to the user in formatted output. Right now the only support is forformatting the register and memory usage reports from ptxas and nvlink. This is simply done as a wrapper around clang and clang++.
Reviewers: jdoerfert
DIfferential Revision: https://reviews.llvm.org/D91085
The deadlock/race happens when primary thread gets initz lock and tries to join
the worker thread which waits for the same lock in TLS key destructor.
The patch removes the lock and the code of setting TLS value which needed
the lock. Also removed setting TLS from __kmp_unregister_root_current_thread.
Differential Revision: https://reviews.llvm.org/D90647
This patch allows to pass the OpenMP runtime tests after configuring with
`cmake . -DOPENMP_TEST_FLAGS:STRING="-Werror"`.
The warnings for OMPT tests are addressed in D90752.
Differential Revision: https://reviews.llvm.org/D91280
This doesn't add functionality, but just adds the new types and renames the
master callback to masked callback.
Differential Revision: https://reviews.llvm.org/D90752
Modern Fortran compilers support Fortran 90, so we do not need to use
the source code for Fortran compilers that do not support Fortran 90.
Differential Revision: https://reviews.llvm.org/D90077
This patch is the runtime support for https://reviews.llvm.org/D84192.
In order not to modify the tgt_target_data_update information but still be
able to pass the extra information for non-contiguous map item (offset,
count, and stride for each dimension), this patch overload arg when
the maptype is set as OMP_TGT_MAPTYPE_DESCRIPTOR. The origin arg is for
passing the pointer information, however, the overloaded arg is an
array of descriptor_dim:
```
struct descriptor_dim {
int64_t offset;
int64_t count;
int64_t stride
};
```
and the array size is the dimension size. In addition, since we
have count and stride information in descriptor_dim, we can replace/overload the
arg_size parameter by using dimension size.
Reviewed By: grokos
Differential Revision: https://reviews.llvm.org/D82245
This patch fixes potential division by 0 in case hwloc does not
recognize cores (or architecture has no cores).
Patch by Andrey Churbanov
Differential Revision: https://reviews.llvm.org/D90954
This patch adds the mask and ident_t function to get the
openmp version. It also adds logic to force monotonic:dynamic
behavior when OpenMP version less than 5.0.
The OpenMP version is stored in the format:
major*10+minor e.g., OpenMP 5.0 = 50
Differential Revision: https://reviews.llvm.org/D90632
The macros are used in several places with an if(macro) pattern. This results
in several warnings about extraneous parenteses in equality comparison.
Having the constant at the lhs of the comparison, avoids this warning.
Differential Revision: https://reviews.llvm.org/D90756
There is a non-conforming use of variable-sized array in the test case `parallel_offloading_map.c`. This patch fixed it.
Reviewed By: protze.joachim
Differential Revision: https://reviews.llvm.org/D90642
As reported by @ronlieb, the test shows intermittent fails.
The test failed, if the dependent task was already finished, when the depending
task was to be created. We have other tests to check for the dependences pair.
Since detached tasks are supported by clang and the OpenMP runtime, Archer
must expect to receive the corresponding callbacks.
This patch adds support to interpret the synchronization semantics of
omp_fulfill_event and cleans up the handling of task switches.
Presently, there a number of global variables in libomptarget (devices,
RTLs, tables, mutexes, etc.) that are not placed within a struct. This
patch places them into a struct ``PluginManager``. All of the functions
that act on this data remain free.
Differential Revision: https://reviews.llvm.org/D90519
This adds some initial content as well as structure to the new OpenMP
Sphinx documentation hosted at http://openmp.llvm.org/docs/ .
The content contains some useful links but most pages are still empty.
This uses a "custom" theme which is a copy of the default "agogo" one
with minor modifications to get a nicer table of content in the sidebar.
This way we can also adjust the theme as we go.
Reviewed By: jhuber6, JonChesterfield
Differential Revision: https://reviews.llvm.org/D90256
[AMDGPU] Add __builtin_amdgcn_grid_size
Similar to D76772, loads the data from the dispatch pointer. Marked invariant.
Patch also updates the openmp devicertl to use this builtin.
Reviewed By: yaxunl
Differential Revision: https://reviews.llvm.org/D90251
Summary:
This patch adds basic support for priting the source location and names for the
mapped variables. This patch does not support names for custom mappers. This is
based on D89802. The names information currently will be printed out only in
debug mode or using env LIBOMPTARGET_INFO during execution. But the information
is added when availible to the Device and Private data structures. To get the
information out the code must be built with debug symbols on using -g or
-Rpass=openmp-opt
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D90172
Summary:
This patch adds support for passing in the original delcaration name in the
source file to the libomptarget runtime. This will allow the runtime to provide
more intelligent debugging messages. This patch takes the original expression
parsed from the OpenMP map / update clause and provides a textual
representation if it was explicitly mapped, otherwise it takes the name of the
variable declaration as a fallback. The information in passed to the runtime in
a global array of strings that matches the existing ident_t source location
strings using ";name;filename;column;row;;". See
clang/test/OpenMP/target_map_names.cpp for an example of the generated output
for a given map clause.
Reviewers: jdoervert
Differential Revision: https://reviews.llvm.org/D89802
The implementation of target nowait just wraps the target region into a task. The essential four parameters (base ptr, ptr, size, mapper) are taken as firstprivate such that they will be copied to the private location. When there is no user-defined mapper, the mapper variable will be nullptr. However, it will be still copied to the corresponding place. Therefore, a memcpy will be generated and the source pointer will be nullptr, causing a segmentation fault. The root cause is when calling `emitOffloadingArraysArgument`, the last argument `Options` has a field about whether it requires a task. It only takes depend clause into account. In this patch, the nowait clause is also included.
There're two things that will be done in another patches:
1. target data nowait has not been supported yet. D90099 added the support.
2. When there is no mapper, the mapper array can be nullptr no matter whether it requires outer task or not. It can avoid an unnecessary data copy. This is an optimization that is covered in D90101.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D89844
`size_t` has different width on 32- and 64-bit architecture, but the
computation to floor to power of two assumed it is 64-bit, which can cause an
integer overflow. In this patch, architecture detection is added so that the
operation for 64-bit `size_t`. Thank Luke for reporting the issue.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D89878
[libomptarget] Require LLVM source tree to build libomptarget
This is to permit reliably #including files from the LLVM tree in libomptarget,
as an improvement on the copy and paste that is currently in use. See D87841
for the first example of removing duplication given this new requirement.
The weekly openmp dev call reached consensus on this approach. See also D87841
for some alternatives that were considered. In the future, we may want to
introduce a new top level repo for shared constants, or start using the ADT
library within openmp.
This will break sufficiently exotic build systems, trivial fixes as below.
Building libomptarget as part of the monorepo will continue to work.
If openmp is built separately, it now requires a cmake macro indicating
where to find the LLVM source tree.
If openmp is built separately, without the llvm source tree already on disk,
the build machine will need a copy of a subset of the llvm source tree and
the cmake macro indicating where it is.
Reviewed By: protze.joachim
Differential Revision: https://reviews.llvm.org/D89426
[libomptarget][amdgcn] Refactor memcpy to eliminate maps
Builds on D89776 to remove now dead code.
Reviewed By: pdhaliwal
Differential Revision: https://reviews.llvm.org/D89888
The calls to atmi_memcpy presently determine the direction of copy (host to
device or device to host) by storing pointers in a map during malloc and
looking up the pointers during memcpy. As each call site already knows the
direction, this stash+lookup can be eliminated.
This NFC will be followed by a functional one that deletes those map lookups.
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D89776
Change-Id: I1d9089bc1e56b3a9a30e334735fa07dee1f84990
[libomptarget][amdgcn] Implement missing symbols in deviceRTL
Malloc, wtime are stubs. Malloc needs a hostrpc implementation which is
a work in progress, wtime needs some experimentation to find out the
multiplier to get a time in seconds as documentation is scarce.
Reviewed By: ronlieb
Differential Revision: https://reviews.llvm.org/D89725
This patch fixes a problem whereby the pointee object of a PTR_AND_OBJ entry with a `map(to)` motion clause can be overwritten on the device even if its reference counter is >=1.
Currently, we check the reference counter of the parent struct in order to determine whether the motion clause should be respected, but since the pointee object is not part of the struct, it's got its own reference counter which should be used to enqueue the copy or discard it.
The same behavior has already been implemented in targetDataEnd (omptarget.cpp:539-540), but we somehow missed doing the same in targetDataBegin.
Differential Revision: https://reviews.llvm.org/D89597
[openmp][libomptarget] Include header from LLVM source tree
The change is to the amdgpu plugin so is unlikely to break anything.
The point of contention is whether libomptarget can depend on LLVM.
A community discussion was cautiously not opposed yesterday.
This introduces a compile time dependency on the LLVM source tree, in this case
expressed as skipping the building of the plugin if LLVM_MAIN_INCLUDE_DIR is not
set. One the source files will #include llvm/Frontend/OpenMP/OMPGridValues.h,
instead of copy&pasting the numbers across.
For users that download the monorepo, the llvm tree is already on disk. This will
inconvenience users who download only the openmp source as a tar, as they would
now also have to download (at least a file or two) from the llvm source, if they want
to build the parts of the openmp project that (post this patch) depend on llvm.
There was interest expressed in going further - using llvm tools as part of
building libomp, or linking against llvm libraries. That seems less clear cut
an improvement and worthy of further discussion. This patch seeks only to change
policy to support openmp depending on the llvm source tree. Including in the
other direction, or using libraries / tools etc, are purposefully out of scope.
Reviewers are a best guess at interested parties, please feel free to add others
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D87841
[libomptarget][amdgcn] Implement partial barrier
named_sync is used to coordinate non-spmd kernels. This uses bar.sync on nvptx.
There is no corresponding ISA support on amdgcn, so this is implemented using
shared memory, one word initialized to zero.
Each wave increments the variable by one. Whichever wave is last is responsible
for resetting the variable to zero, at which point it and the others continue.
The race condition on a wave reaching the barrier before another wave has
noticed that it has been released is handled with a generation counter, packed
into the same word.
Uses a shared variable that is not needed on nvptx. Introduces a new hook,
kmpc_impl_target_init, to allow different targets to do extra initialization.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D88602
Summary:
This patch changes the CMake files for Clang and Libomptarget to query the
system for its supported CUDA architecture. This makes it much easier for the
user to build optimal code without needing to set the flags manually. This
relies on the now deprecated FindCUDA method in CMake, but full support for
architecture detection is only availible in CMake >3.18
Reviewers: jdoerfert ye-luo
Subscribers: cfe-commits guansong mgorny openmp-commits sstefan1 yaxunl
Tags: #clang #OpenMP
Differential Revision: https://reviews.llvm.org/D87946
The test disables suppression and therefore sometimes triggers a know false
positive in the openmp runtime. The test should only verify that the env
var is handles as expected.
The worker thread can start execution of the task before creation of the second task
Fixes the spurious failure reported in https://reviews.llvm.org/D61657
Currently, the parser used to tokenize the TSAN_OPTIONS in libomp uses
only spaces as separators, even though TSAN in compiler-rt supports
other separators like ':' or ','.
CTest uses ':' to separate sanitizer options by default.
The documentation for other sanitizers mentions ':' as separator,
but TSAN only lists spaces, which is probably where this mismatch originated.
Patch provided by upsj
Differential Revision: https://reviews.llvm.org/D87144
This patch updates the expected results for the GOMP interface patches: D87267, D87269, and D87271.
The taskwait-depend test is changed to really use taskwait-depend and copied to an task_if0-depend test.
To pass the tests, the handling of the return address was fixed.
Differential Revision: https://reviews.llvm.org/D87680
OpenMP 5.1 defines omp_get_initial_device to return the same value as omp_get_num_devices.
Since this change is also 5.0 compliant, no versioning is needed.
Differential Revision: https://reviews.llvm.org/D88149
[nfc][libomptarget] Drop parameter to named_sync
named_sync has one call site (in sync.cu) where it always passed L1_BARRIER.
Folding this into the call site and dropping the macro is a simplification.
amdgpu doesn't have ptx' bar.sync instruction. A correct implementation of
__kmpc_impl_named_sync in terms of shared memory is much easier if it can
assume that the barrier argument is this constant. Said implementation is left
for a second patch.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D88474
Summary:
Adding a missing directory needed for generating Sphinx documentation without
errors. Directory current contains a placeholder image just to populate the
directory.
It allows customizing MAX_SM for non-flagship GPU and reduces graphic memory usage.
In addition, so far the size is hard-coded up to __CUDA_ARCH__ 700 and is already a hassle for 800.
Introduce MAX_SM for 800 and protect future arch
Reviewed By: JonChesterfield
Differential Revision: https://reviews.llvm.org/D88185
This change introduces the GOMP_taskwait_depend() function. It implements
the OpenMP 5.0 feature of #pragma omp taskwait with depend() clause by
wrapping around __kmpc_omp_wait_deps().
Differential Revision: https://reviews.llvm.org/D87269
Encapsulate GOMP task dependencies in separate class and introduce the
new mutexinoutset dependency type. This separate class allows
future GOMP task APIs easier access to the task dependency functionality
and better ability to propagate new dependency types to all existing GOMP
task APIs which use task dependencies.
Differential Revision: https://reviews.llvm.org/D87267
Implement GOMP_teams_reg() function which enables GOMP support of the
standalone teams construct. The GOMP_parallel* functions were modified
to call __kmp_fork_call() unconditionally so that the teams-specific
code could be reused within __kmp_fork_call() instead of reproduced
inside the GOMP_* functions.
Differential Revision: https://reviews.llvm.org/D87167
If an error code can not be recognized by cuGetErrorString, errStr remains null and causes crashing at DP() printing.
Protect this case.
Reviewed By: jhuber6, tianshilei1992
Differential Revision: https://reviews.llvm.org/D87980
Summary:
Adding support for generated html documentation for OpenMP. Changing
Cmake files to build the documentation and adding the base templates for
future documentation to be added.
Reviewers: jdoerfert
Subscribers: aaron.ballman arphaman guansong mgorny openmp-commits sstefan1 yaxunl
Tags: #OpenMP
Differential Revision: https://reviews.llvm.org/D87797
Summary:
This patch adds additonal support for priting infromation from Libomptarget for
already existing maps and printing the final data mapped on the device at
device destruction.
Reviewers: jdoerfort gkistanova
Subscribers: guansong openmp-commits sstefan1 yaxunl
Tags: #OpenMP
Differential Revision: https://reviews.llvm.org/D87722
Summary:
This patch starts adding support for adding information dumps to libomptarget
and rtl plugins. The information printing is controlled by the
LIBOMPTARGET_INFO environment variable introduced in D86483. The goal of this
patch is to provide the user with additional information about the device
during kernel execution and providing the user with information dumps in the
case of failure. This patch added the ability to dump the pointer mapping table
as well as printing the number of blocks and threads in the cuda RTL.
Reviewers: jdoerfort gkistanova ye-luo
Subscribers: guansong openmp-commits sstefan1 yaxunl ye-luo
Tags: #OpenMP
Differential Revision: https://reviews.llvm.org/D87165
The size of worker_rootS should have been DS_Max_Warp_Number.
This reduces memory usage by deviceRTL on AMDGPU from around 2.3GB
to around 770MB.
Reviewed By: JonChesterfield, jdoerfert
Differential Revision: https://reviews.llvm.org/D87084
Summary:
This patch consolidates the error handling and messaging routines to a single
file omptargetmessage. The goal is to simplify the error handling interface
prior to adding more error handling support
Reviewers: jdoerfert grokos ABataev AndreyChurbanov ronlieb JonChesterfield ye-luo tianshilei1992
Subscribers: danielkiss guansong jvesely kerbowa nhaehnle openmp-commits sstefan1 yaxunl
PrivateArgumentManager shall immediately allocate firstprivates if they
are bases for the next parameters and the next paramaters rely on the
fact that the base musst be allocated already.
Differential Revision: https://reviews.llvm.org/D86781
The test command in `private_mapping.c` was set to expect failure by mistake. It is fixed in this patch.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D86758
Summary:
This patch changes the libomptarget runtime to always emit debug messages that
occur before offloading failure. The goal is to provide users with information
about why their application failed in the target region rather than a single
failure message. This is only done in regions that precede offloading failure
so this should not impact runtime performance. if the debug environment
variable is set then the message is forwarded to the debug output as usual.
A new environment variable was added for future use but does nothing in this
current patch. LIBOMPTARGET_INFO will be used to report runtime information to
the user if requrested, such as grid size, SPMD usage, or data mapping. It will
take an integer indicating the level of information verbosity and a value of 0
will disable it.
Reviewers: jdoerfort
Subscribers: guansong sstefan1 yaxunl ye-luo
Tags: #OpenMP
Differential Revision: https://reviews.llvm.org/D86483
In this patch, we pack all small first-private arguments, allocate and transfer them all at once to reduce the number of data transfer which is very expensive.
Let's take the test case as example.
```
int main() {
int data1[3] = {1}, data2[3] = {2}, data3[3] = {3};
int sum[16] = {0};
#pragma omp target teams distribute parallel for map(tofrom: sum) firstprivate(data1, data2, data3)
for (int i = 0; i < 16; ++i) {
for (int j = 0; j < 3; ++j) {
sum[i] += data1[j];
sum[i] += data2[j];
sum[i] += data3[j];
}
}
}
```
Here `data1`, `data2`, and `data3` are three first-private arguments of the target region. In the previous `libomptarget`, it called data allocation and data transfer three times, each of which allocated and transferred 12 bytes. With this patch, it only calls allocation and transfer once. The size is `(12+4)*3=48` where 12 is the size of each array and 4 is the padding to keep the address aligned with 8. It is implemented in this way:
1. First collect all information for those *first*-private arguments. _private_ arguments are not the case because private arguments don't need to be mapped to target device. It just needs a data allocation. With the patch for memory manager, the data allocation could be very cheap, especially for the small size. For each qualified argument, push a place holder pointer `nullptr` to the `vector` for kernel arguments, and we will update them later.
2. After we have all information, create a buffer that can accommodate all arguments plus their paddings. Copy the arguments to the buffer at the right place, i.e. aligned address.
3. Allocate a target memory with the same size as the host buffer, transfer the host buffer to target device, and finally update all place holder pointers in the arguments `vector`.
The reason we only consider small arguments is, the data transfer is asynchronous. Therefore, for the large argument, we could continue to do things on the host side meanwhile, hopefully, the data is also being transferred. The "small" is defined by that the argument size is less than a predefined value. Currently it is 1024. I'm not sure whether it is a good one, and that is an open question. Another question is, do we need to make it configurable via an environment variable?
Reviewed By: ye-luo
Differential Revision: https://reviews.llvm.org/D86307
Instead of copying and pasting the same `#ifdef` expressions in multiple
places, define a type and a pair of macros in `kmp_os.h`, to handle
whether `va_list` is pointer-like or not:
* `kmp_va_list` is the type to use for `__kmp_fork_call()`
* `kmp_va_deref()` dereferences a `va_list`, if necessary
* `kmp_va_addr_of()` takes the address of a `va_list`, if necessary
Also add FreeBSD to the list of OSes that has a non pointer-like
va_list. This can now be easily extended to other OSes too.
Reviewed By: AndreyChurbanov
Differential Revision: https://reviews.llvm.org/D86397
Target memory manager is introduced in this patch which aims to manage target
memory such that they will not be freed immediately when they are not used
because the overhead of memory allocation and free is very large. For CUDA
device, cuMemFree even blocks the context switch on device which affects
concurrent kernel execution.
The memory manager can be taken as a memory pool. It divides the pool into
multiple buckets according to the size such that memory allocation/free
distributed to different buckets will not affect each other.
In this version, we use the exact-equality policy to find a free buffer. This
is an open question: will best-fit work better here? IMO, best-fit is not good
for target memory management because computation on GPU usually requires GBs of
data. Best-fit might lead to a serious waste. For example, there is a free
buffer of size 1960MB, and now we need a buffer of size 1200MB. If best-fit,
the free buffer will be returned, leading to a 760MB waste.
The allocation will happen when there is no free memory left, and the memory
free on device will take place in the following two cases:
1. The program ends. Obviously. However, there is a little problem that plugin
library is destroyed before the memory manager is destroyed, leading to a fact
that the call to target plugin will not succeed.
2. Device is out of memory when we request a new memory. The manager will walk
through all free buffers from the bucket with largest base size, pick up one
buffer, free it, and try to allocate immediately. If it succeeds, it will
return right away rather than freeing all buffers in free list.
Update:
A threshold (8KB by default) is set such that users could control what size of memory
will be managed by the manager. It can also be configured by an environment variable
`LIBOMPTARGET_MEMORY_MANAGER_THRESHOLD`.
Reviewed By: jdoerfert, ye-luo, JonChesterfield
Differential Revision: https://reviews.llvm.org/D81054
This patch contains the following changes:
1. Renamed the function `DeviceTy::data_exchange` to `DeviceTy::dataExchange`;
2. Changed the second argument `DeviceTy DstDev` to `DeviceTy &DstDev`;
3. Renamed the last argument.
Reviewed By: ye-luo
Differential Revision: https://reviews.llvm.org/D86238
Shilei Tian