To support std::complex and some other standard C/C++ functions in HIP device code,
they need to be forced to be __host__ __device__ functions by pragmas. This is done
by some clang standard C++ wrapper headers which are shared between cuda-clang and hip-Clang.
For these standard C++ wapper headers to work properly, specific include path order
has to be enforced:
clang C++ wrapper include path
standard C++ include path
clang include path
Also, these C++ wrapper headers require device version of some standard C/C++ functions
must be declared before including them. This needs to be done by including a default
header which declares or defines these device functions. The default header is always
included before any other headers are included by users.
This patch adds the the default header and include path for HIP.
Differential Revision: https://reviews.llvm.org/D81176
I didn't realize HIP was a distinct offloading kind, so the subtarget
was looking for -march, which isn't correct for HIP. We also have the
possibility of different denormal defaults in the case of multiple
offload targets, so we need to thread the JobAction through the target
hook.
Generate PTX using newer versions of PTX and allow using sm_80 with CUDA-11.
None of the new features of CUDA-10.2+ have been implemented yet, so using these
versions will still produce a warning.
Differential Revision: https://reviews.llvm.org/D77670
Instead of hardcoding individual GPU mappings in multiple functions, keep them
all in one table and use it to look up the mappings.
We also don't care about 'virtual' architecture much, so the API is trimmed down
down to a simpler GPU->Virtual arch name lookup.
Differential Revision: https://reviews.llvm.org/D77665
The argument after -Xarch_device will be added to the arguments for CUDA/HIP
device compilation and will be removed for host compilation.
The argument after -Xarch_host will be added to the arguments for CUDA/HIP
host compilation and will be removed for device compilation.
Differential Revision: https://reviews.llvm.org/D76520
Extract common code to a function. To prepare for
adding an option for CUDA/HIP host and device only
option.
Differential Revision: https://reviews.llvm.org/D76455
This fixes an issue with clang issuing a warning about unknown CUDA SDK if it's
detected during non-CUDA compilation.
Differential Revision: https://reviews.llvm.org/D76030
AMDGPU and x86 at least both have separate controls for whether
denormal results are flushed on output, and for whether denormals are
implicitly treated as 0 as an input. The current DAGCombiner use only
really cares about the input treatment of denormals.
This is how it should've been and brings it more in line with
std::string_view. There should be no functional change here.
This is mostly mechanical from a custom clang-tidy check, with a lot of
manual fixups. It uncovers a lot of minor inefficiencies.
This doesn't actually modify StringRef yet, I'll do that in a follow-up.
This makes clang somewhat forward-compatible with new CUDA releases
without having to patch it for every minor release without adding
any new function.
If an unknown version is found, clang issues a warning (can be disabled
with -Wno-cuda-unknown-version) and assumes that it has detected
the latest known version. CUDA releases are usually supersets
of older ones feature-wise, so it should be sufficient to keep
released clang versions working with minor CUDA updates without
having to upgrade clang, too.
Differential Revision: https://reviews.llvm.org/D73231
Currently there are 4 different mechanisms for controlling denormal
flushing behavior, and about as many equivalent frontend controls.
- AMDGPU uses the fp32-denormals and fp64-f16-denormals subtarget features
- NVPTX uses the nvptx-f32ftz attribute
- ARM directly uses the denormal-fp-math attribute
- Other targets indirectly use denormal-fp-math in one DAGCombine
- cl-denorms-are-zero has a corresponding denorms-are-zero attribute
AMDGPU wants a distinct control for f32 flushing from f16/f64, and as
far as I can tell the same is true for NVPTX (based on the attribute
name).
Work on consolidating these into the denormal-fp-math attribute, and a
new type specific denormal-fp-math-f32 variant. Only ARM seems to
support the two different flush modes, so this is overkill for the
other use cases. Ideally we would error on the unsupported
positive-zero mode on other targets from somewhere.
Move the logic for selecting the flush mode into the compiler driver,
instead of handling it in cc1. denormal-fp-math/denormal-fp-math-f32
are now both cc1 flags, but denormal-fp-math-f32 is not yet exposed as
a user flag.
-cl-denorms-are-zero, -fcuda-flush-denormals-to-zero and
-fno-cuda-flush-denormals-to-zero will be mapped to
-fp-denormal-math-f32=ieee or preserve-sign rather than the old
attributes.
Stop emitting the denorms-are-zero attribute for the OpenCL flag. It
has no in-tree users. The meaning would also be target dependent, such
as the AMDGPU choice to treat this as only meaning allow flushing of
f32 and not f16 or f64. The naming is also potentially confusing,
since DAZ in other contexts refers to instructions implicitly treating
input denormals as zero, not necessarily flushing output denormals to
zero.
This also does not attempt to change the behavior for the current
attribute. The LangRef now states that the default is ieee behavior,
but this is inaccurate for the current implementation. The clang
handling is slightly hacky to avoid touching the existing
denormal-fp-math uses. Fixing this will be left for a future patch.
AMDGPU is still using the subtarget feature to control the denormal
mode, but the new attribute are now emitted. A future change will
switch this and remove the subtarget features.
Skip distro detection when we're not running on Linux, or when the target triple is not Linux. This saves a few OS calls for each invocation of clang.exe.
Differential Revision: https://reviews.llvm.org/D70467
This patch removes the remaining part of the OpenMP offload linker scripts which was used for inserting device binaries into the output linked binary. Device binaries are now inserted into the host binary with a help of the wrapper bit-code file which contains device binaries as data. Wrapper bit-code file is dynamically created by the clang driver with a help of new tool clang-offload-wrapper which takes device binaries as input and produces bit-code file with required contents. Wrapper bit-code is then compiled to an object and resulting object is appended to the host linking by the clang driver.
This is the second part of the patch for eliminating OpenMP linker script (please see https://reviews.llvm.org/D64943).
Differential Revision: https://reviews.llvm.org/D68166
llvm-svn: 374219
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
Differential revision: https://reviews.llvm.org/D66259
llvm-svn: 368942
Summary:
The bug was reported on the OpenMP-dev list:
.../obj-release/lib/clang/9.0.0/include/__clang_cuda_intrinsics.h:173:35: error: '__nvvm_shfl_sync_idx_i32' needs target feature ptx60|ptx61|ptx63|ptx64
__MAKE_SYNC_SHUFFLES(__shfl_sync, __nvvm_shfl_sync_idx_i32,
This problem occurs when trying to compile a .cu file that requires a newer ptx version (>ptx60 in this case) than ptx42.
Reviewers: tra, ABataev, caomhin
Reviewed By: tra
Subscribers: jdoerfert, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D61474
llvm-svn: 359910
CUDA 10.1 tools deprecated some command line options.
fatbinary no longer needs --cuda.
Differential Revision: https://reviews.llvm.org/D61470
llvm-svn: 359838
These builtins provide access to the new integer and
sub-integer variants of MMA (matrix multiply-accumulate) instructions
provided by CUDA-10.x on sm_75 (AKA Turing) GPUs.
Also added a feature for PTX 6.4. While Clang/LLVM does not generate
any PTX instructions that need it, we still need to pass it through to
ptxas in order to be able to compile code that uses the new 'mma'
instruction as inline assembly (e.g used by NVIDIA's CUTLASS library
https://github.com/NVIDIA/cutlass/blob/master/cutlass/arch/mma.h#L101)
Differential Revision: https://reviews.llvm.org/D60279
llvm-svn: 359248
..and use it to control that parts of CUDA compilation
that depend on the specific version of CUDA SDK.
This patch has a placeholder for a 'new launch API' support
which is in a separate patch. The list will be further
extended in the upcoming patch to support CUDA-10.1.
Differential Revision: https://reviews.llvm.org/D57487
llvm-svn: 352798
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Summary:
Added support for the -gline-directives-only option + fixed logic of the
debug info for CUDA devices. If optimization level is O0, then options
--[no-]cuda-noopt-device-debug do not affect the debug info level. If
the optimization level is >O0, debug info options are used +
--no-cuda-noopt-device-debug is used or no --cuda-noopt-device-debug is
used, the optimization level for the device code is kept and the
emission of the debug directives is used.
If the opt level is > O0, debug info is requested +
--cuda-noopt-device-debug option is used, the optimization is disabled
for the device code + required debug info is emitted.
Reviewers: tra, echristo
Subscribers: aprantl, guansong, JDevlieghere, cfe-commits
Differential Revision: https://reviews.llvm.org/D51554
llvm-svn: 348930
This patch moves the virtual file system form clang to llvm so it can be
used by more projects.
Concretely the patch:
- Moves VirtualFileSystem.{h|cpp} from clang/Basic to llvm/Support.
- Moves the corresponding unit test from clang to llvm.
- Moves the vfs namespace from clang::vfs to llvm::vfs.
- Formats the lines affected by this change, mostly this is the result of
the added llvm namespace.
RFC on the mailing list:
http://lists.llvm.org/pipermail/llvm-dev/2018-October/126657.html
Differential revision: https://reviews.llvm.org/D52783
llvm-svn: 344140
This patch renames -f{no-}cuda-rdc to -f{no-}gpu-rdc and keeps the original
options as aliases. When -fgpu-rdc is off,
clang will assume the device code in each translation unit does not call
external functions except those in the device library, therefore it is possible
to compile the device code in each translation unit to self-contained kernels
and embed them in the host object, so that the host object behaves like
usual host object which can be linked by lld.
The benefits of this feature is: 1. allow users to create static libraries which
can be linked by host linker; 2. amortized device code linking time.
This patch modifies HIP action builder to insert actions for linking device
code and generating HIP fatbin, and pass HIP fatbin to host backend action.
It extracts code for constructing command for generating HIP fatbin as
a function so that it can be reused by early finalization. It also modifies
codegen of HIP host constructor functions to embed the device fatbin
when it is available.
Differential Revision: https://reviews.llvm.org/D52377
llvm-svn: 343611
When looking for the bclib Clang considered the default library
path first while it preferred directories in LIBRARY_PATH when
constructing the invocation of nvlink. The latter actually makes
more sense because during development it allows using a non-default
runtime library. So change the search for the bclib to start
looking in directories given by LIBRARY_PATH.
Additionally add a new option --libomptarget-nvptx-path= which
will be searched first. This will be handy for testing purposes.
Differential Revision: https://reviews.llvm.org/D51686
llvm-svn: 343230
Summary:
Some targets support only default set of the debug options and do not
support additional debug options, like NVPTX target. Patch introduced
virtual function supportsDebugInfoOptions() that can be overloaded
by the toolchain, checks if the target supports some debug
options and emits warning when an unsupported debug option is
found.
Reviewers: echristo
Subscribers: aprantl, JDevlieghere, cfe-commits
Differential Revision: https://reviews.llvm.org/D49148
llvm-svn: 338155
The option enables use of 32-bit pointers for accessing
const/local/shared memory. The feature is disabled by default.
Differential Revision: https://reviews.llvm.org/D46148
llvm-svn: 331938
Currently we always include PTX into the fatbin along
with the GPU code.It about doubles the size of the GPU binary
we need to carry in the executable. These options allow control
inclusion of PTX into GPU binary.
This patch does not change the defaults, though we may consider
making no-PTX the default in the future.
Differential Revision: https://reviews.llvm.org/D45495
llvm-svn: 329737
Found via codespell -q 3 -I ../clang-whitelist.txt
Where whitelist consists of:
archtype
cas
classs
checkk
compres
definit
frome
iff
inteval
ith
lod
methode
nd
optin
ot
pres
statics
te
thru
Patch by luzpaz! (This is a subset of D44188 that applies cleanly with a few
files that have dubious fixes reverted.)
Differential revision: https://reviews.llvm.org/D44188
llvm-svn: 329399
Summary: This patch adds an additional flag to the OpenMP device offloading toolchain to link in the runtime library bitcode.
Reviewers: Hahnfeld, ABataev, carlo.bertolli, caomhin, grokos, hfinkel
Reviewed By: ABataev, grokos
Subscribers: jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D43197
llvm-svn: 327460
Summary: This patch adds an additional flag to the OpenMP device offloading toolchain to link in the runtime library bitcode.
Reviewers: Hahnfeld, ABataev, carlo.bertolli, caomhin, grokos, hfinkel
Reviewed By: ABataev, grokos
Subscribers: jholewinski, guansong, cfe-commits
Differential Revision: https://reviews.llvm.org/D43197
llvm-svn: 327438
As a first step, pass '-c/--compile-only' to ptxas so that it
doesn't complain about references to external function. This
will successfully generate object files, but they won't work
at runtime because the registration routines need to adapted.
Differential Revision: https://reviews.llvm.org/D42921
llvm-svn: 324878
If the CUDA toolkit is not installed to its default locations
in /usr/local/cuda, the user is forced to specify --cuda-path.
This is tedious and the driver can be smarter if well-known tools
(like ptxas) can already be found in the PATH environment variable.
Add option --cuda-path-ignore-env if the user wants to ignore
set environment variables. Also use it in the tests to make sure
the driver always finds the same CUDA installation, regardless
of the user's environment.
Differential Revision: https://reviews.llvm.org/D42642
llvm-svn: 323848
Clang can use CUDA-9.1 now, though new APIs (are not implemented yet.
The major change is that headers in CUDA-9.1 went through substantial
changes that started in CUDA-9.0 which required substantial changes
in the cuda compatibility headers provided by clang.
There are two major issues:
* CUDA SDK no longer provides declarations for libdevice functions.
* A lot of device-side functions have become nvcc's builtins and
CUDA headers no longer contain their implementations.
This patch changes the way CUDA headers are handled if we compile
with CUDA 9.x. Both 9.0 and 9.1 are affected.
* Clang provides its own declarations of libdevice functions.
* For CUDA-9.x clang now provides implementation of device-side
'standard library' functions using libdevice.
This patch should not affect compilation with CUDA-8. There may be
some observable differences for CUDA-9.0, though they are not expected
to affect functionality.
Tested: CUDA test-suite tests for all supported combinations of:
CUDA: 7.0,7.5,8.0,9.0,9.1
GPU: sm_20, sm_35, sm_60, sm_70
Differential Revision: https://reviews.llvm.org/D42513
llvm-svn: 323713
../tools/clang/lib/Driver/ToolChains/Cuda.cpp:80:18: error: reference to non-static member function must be called; did you mean to call it with no arguments?
if (Distro(D.getVFS).IsDebian())
~~^~~~~~
()
llvm-svn: 319322
This was previously done in some places, but for example not for
bundling so that single object compilation with -c failed. In
addition cubin was used for all file types during unbundling which
is incorrect for assembly files that are passed to ptxas.
Tighten up the tests so that we can't regress in that area.
Differential Revision: https://reviews.llvm.org/D40250
llvm-svn: 318763
Summary:
CUDA 9's minimum sm is sm_30.
Ideally we should also make sm_30 the default when compiling with CUDA
9, but that seems harder than it should be.
Subscribers: sanjoy
Differential Revision: https://reviews.llvm.org/D39109
llvm-svn: 316611
For the shuffle instructions in reductions we need at least sm_30
but the user may want to customize the default architecture.
Differential Revision: https://reviews.llvm.org/D38883
llvm-svn: 315996
If the user passes -nocudalib, we can live without it being present.
Simplify the code by just checking whether LibDeviceMap is empty.
Differential Revision: https://reviews.llvm.org/D38901
llvm-svn: 315902
Summary: If we only use the compiler front-end, do not throw an error about the cuda device library not being found. This allows the front-end to be run on systems where no Cuda installation is found.
Reviewers: Hahnfeld, ABataev, carlo.bertolli, caomhin, tra
Reviewed By: tra
Subscribers: hfinkel, tra, cfe-commits
Differential Revision: https://reviews.llvm.org/D37914
llvm-svn: 314217
Summary: Enable the -nocudalib flag for the OpenMP device offloading toolchain as well. Currently it can only be used for the CUDA toolchain.
Reviewers: Hahnfeld, ABataev, carlo.bertolli, caomhin, hfinkel, tra
Reviewed By: tra
Subscribers: hfinkel, cfe-commits
Differential Revision: https://reviews.llvm.org/D37913
llvm-svn: 314164
Summary: When composing the output file name, the path to the file is being dropped. The full path is required.
Reviewers: Hahnfeld, ABataev, caomhin, carlo.bertolli, hfinkel, tra
Reviewed By: tra
Subscribers: hfinkel, tra, cfe-commits
Differential Revision: https://reviews.llvm.org/D37912
llvm-svn: 314156
Summary: If we only use the compiler front-end, do not throw an error about the cuda device library not being found. This allows the front-end to be run on systems where no Cuda installation is found.
Reviewers: Hahnfeld, ABataev, carlo.bertolli, caomhin, tra
Reviewed By: tra
Subscribers: hfinkel, tra, cfe-commits
Differential Revision: https://reviews.llvm.org/D37914
llvm-svn: 314150
For now CUDA-9 is not included in the list of CUDA versions clang
searches for, so the path to CUDA-9 must be explicitly passed
via --cuda-path=.
On LLVM side NVPTX added sm_70 GPU type which bumps required
PTX version to 6.0, but otherwise is equivalent to sm_62 at the moment.
Differential Revision: https://reviews.llvm.org/D37576
llvm-svn: 312734
Create a separate test file to contain all tests for OpenMP
offloading to GPUs.
Make libdevice checking more robust by accounting for
the case in which no libdevice is found.
This changes are in connrection with diff: D29660
llvm-svn: 310718
Summary: Invoking the compiler inside a script causes the clang-offload-bundler executable to not be found. This patch enables the lookup for executables in the driver directory where the clang-offload-bundler resides.
Reviewers: hfinkel, carlo.bertolli, arpith-jacob, ABataev, caomhin
Reviewed By: hfinkel
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D36537
llvm-svn: 310513
Summary:
This flag "--fopenmp-ptx=" enables the overwriting of the default PTX version used for GPU offloaded OpenMP target regions: "+ptx42".
Reviewers: arpith-jacob, caomhin, carlo.bertolli, ABataev, Hahnfeld, jlebar, hfinkel, tstellar
Reviewed By: ABataev
Subscribers: rengolin, cfe-commits
Differential Revision: https://reviews.llvm.org/D29660
llvm-svn: 310489
Summary: Previously we have added the "-c" flag which gets passed to PTXAS by default to generate relocatable OpenMP target code by default. This set of flags exposes control over this behaviour.
Reviewers: arpith-jacob, caomhin, carlo.bertolli, ABataev, Hahnfeld, jlebar, hfinkel, tstellar
Reviewed By: ABataev
Subscribers: Hahnfeld, rengolin, cfe-commits
Differential Revision: https://reviews.llvm.org/D29659
llvm-svn: 310484
The commit r310291 introduced the failure. r310332 was a test fix commit and
r310300 was a followup commit. I reverted these two to avoid merge conflicts
when reverting.
The 'openmp-offload.c' test is failing on Darwin because the following
run lines:
// RUN: touch %t1.o
// RUN: touch %t2.o
// RUN: %clang -### -no-canonical-prefixes -fopenmp=libomp -fopenmp-targets=nvptx64-nvidia-cuda -save-temps -no-canonical-prefixes %t1.o %t2.o 2>&1 \
// RUN: | FileCheck -check-prefix=CHK-TWOCUBIN %s
trigger the following assertion:
Driver.cpp:3418:
assert(CachedResults.find(ActionTC) != CachedResults.end() &&
"Result does not exist??");
llvm-svn: 310345
Summary: When device offloading is enabled and the device is an NVIDIA GPU, OpenMP target regions must be compiled with relocation enabled by passing the "-c" flag to the PTXAS invocation.
Reviewers: arpith-jacob, caomhin, carlo.bertolli, ABataev, Hahnfeld, jlebar, hfinkel, tstellar
Reviewed By: Hahnfeld
Subscribers: Hahnfeld, rengolin, mkuron, cfe-commits
Differential Revision: https://reviews.llvm.org/D29642
llvm-svn: 310300
Summary: When compiling code being offloaded by OpenMP to an NVIDIA GPU, pass the -v to PTXAS if it was passed to the CLANG driver.
Reviewers: arpith-jacob, caomhin, carlo.bertolli, ABataev, jlebar, hfinkel, tstellar
Reviewed By: jlebar
Subscribers: Hahnfeld, rengolin, cfe-commits
Differential Revision: https://reviews.llvm.org/D29644
llvm-svn: 310295
Summary:
OpenMP has the ability to offload target regions to devices which may have different architectures.
A new -fopenmp-target-arch flag is introduced to specify the device architecture.
In this patch I use the new flag to specify the compute capability of the underlying NVIDIA architecture for the OpenMP offloading CUDA tool chain.
Only a host-offloading test is provided since full device offloading capability will only be available when [[ https://reviews.llvm.org/D29654 | D29654 ]] lands.
Reviewers: hfinkel, Hahnfeld, carlo.bertolli, caomhin, ABataev
Reviewed By: hfinkel
Subscribers: guansong, cfe-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D34784
llvm-svn: 310263
Summary: Pass the type of the device offloading when building the tool chain for a particular target architecture. This is required when supporting multiple tool chains that target a single device type. In our particular use case, the OpenMP and CUDA tool chains will use the same ```addClangTargetOptions ``` method. This enables the reuse of common options and ensures control over options only supported by a particular tool chain.
Reviewers: arpith-jacob, caomhin, carlo.bertolli, ABataev, jlebar, hfinkel, tstellar, Hahnfeld
Reviewed By: hfinkel
Subscribers: jgravelle-google, aheejin, rengolin, jfb, dschuff, sbc100, cfe-commits
Differential Revision: https://reviews.llvm.org/D29647
llvm-svn: 307272
Summary:
(This is a move-only refactoring patch. There are no functionality changes.)
This patch splits apart the Clang driver's tool and toolchain implementation
files. Each target platform toolchain is moved to its own file, along with the
closest-related tools. Each target platform toolchain has separate headers and
implementation files, so the hierarchy of classes is unchanged.
There are some remaining shared free functions, mostly from Tools.cpp. Several
of these move to their own architecture-specific files, similar to r296056. Some
of them are only used by a single target platform; since the tools and
toolchains are now together, some helpers now live in a platform-specific file.
The balance are helpers related to manipulating argument lists, so they are now
in a new file pair, CommonArgs.h and .cpp.
I've tried to cluster the code logically, which is fairly straightforward for
most of the target platforms and shared architectures. I think I've made
reasonable choices for these, as well as the various shared helpers; but of
course, I'm happy to hear feedback in the review.
There are some particular things I don't like about this patch, but haven't been
able to find a better overall solution. The first is the proliferation of files:
there are several files that are tiny because the toolchain is not very
different from its base (usually the Gnu tools/toolchain). I think this is
mostly a reflection of the true complexity, though, so it may not be "fixable"
in any reasonable sense. The second thing I don't like are the includes like
"../Something.h". I've avoided this largely by clustering into the current file
structure. However, a few of these includes remain, and in those cases it
doesn't make sense to me to sink an existing file any deeper.
Reviewers: rsmith, mehdi_amini, compnerd, rnk, javed.absar
Subscribers: emaste, jfb, danalbert, srhines, dschuff, jyknight, nemanjai, nhaehnle, mgorny, cfe-commits
Differential Revision: https://reviews.llvm.org/D30372
llvm-svn: 297250
Yaxun (Sam) Liu