Summary:
Previously we compiled CUDA device code to PTX assembly and embedded
that asm as text in our host binary. Now we compile to PTX assembly and
then invoke ptxas to assemble the PTX into a cubin file. We gather the
ptx and cubin files for each of our --cuda-gpu-archs and combine them
using fatbinary, and then embed that into the host binary.
Adds two new command-line flags, -Xcuda_ptxas and -Xcuda_fatbinary,
which pass args down to the external tools.
Reviewers: tra, echristo
Subscribers: cfe-commits, jhen
Differential Revision: http://reviews.llvm.org/D16082
llvm-svn: 257809
Summary:
Rename the args to be more human-readable. Among other things, this
lets us get rid of a bunch of comments (e.g. "ensure we don't run the
linker"), greatly shortening these tests.
Also apply consistent formatting and fix some English nits while we're
at it.
Reviewers: tra
Differential Revision: http://reviews.llvm.org/D15975
llvm-svn: 257557
Summary: Trying to make this test a bit more manageable.
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D15974
llvm-svn: 257142
Clang needs to know target triple for both sides of compilation so that
preprocessor macros and target builtins from both sides are available.
This change augments Compilation class to carry information about
toolchains used during different CUDA compilation passes and refactors
BuildActions to use it when it constructs CUDA jobs.
Removed DeviceTriple from CudaHostAction/CudaDeviceAction as it's no
longer needed.
Differential Revision: http://reviews.llvm.org/D13144
llvm-svn: 253385
* Only the last of the --cuda-host-only/--cuda-device-only options has effect.
* CudaHostAction always wraps host-side compilation now.
* Fixed printing of empty action lists.
Differential Revision: http://reviews.llvm.org/D12892
llvm-svn: 248297
Current implementation tries to guess which Action will result in a
job which needs to incorporate device-side GPU binaries. The guessing
was attempting to work around the fact that multiple actions may be
combined into a single compiler invocation. If CudaHostAction ends up
being combined (and thus bypassed during action list traversal) no
device-side actions it pointed to were processed. The guessing worked
for most of the usual cases, but fell apart when external assembler
was used.
This change removes the guessing and makes sure we create and pass
device-side jobs regardless of how the jobs get combined.
* CudaHostAction is always inserted either at Compile phase or the
FinalPhase of current compilation, whichever happens first.
* If selectToolForJob combines CudaHostAction with other actions, it
passes info about CudaHostAction up to the caller
* When it sees that CudaHostAction got combined with other actions
(and hence will never be passed to BuildJobsForActions),
BuildJobsForActions creates device-side jobs the same way they would
be created if CudaHostAction was passed to BuildJobsForActions
directly.
* Added two more test cases to make sure GPU binaries are passed to
correct jobs.
Differential Revision: http://reviews.llvm.org/D11280
llvm-svn: 246174
Host-only cuda compilation does produce valid host object
file and in some cases users do want to proceed on to the linking phase.
The change removes special case that stopped compilation pipeline at
the Assembly phase. Device-side compilation is still stopped early
by the types::getCompilationPhases().
Differential Revision: http://reviews.llvm.org/D11573
llvm-svn: 243478
NOTE: reverts r242077 to reinstate r242058, r242065, 242067
and includes fix for OS X test failures.
- Changed driver pipeline to compile host and device side of CUDA
files and incorporate results of device-side compilation into host
object file.
- Added a test for cuda pipeline creation in clang driver.
New clang options:
--cuda-host-only - Do host-side compilation only.
--cuda-device-only - Do device-side compilation only.
--cuda-gpu-arch=<ARCH> - specify GPU architecture for device-side
compilation. E.g. sm_35, sm_30. Default is sm_20. May be used more
than once in which case one device-compilation will be done per
unique specified GPU architecture.
Differential Revision: http://reviews.llvm.org/D9509
llvm-svn: 242085
The tests were failing on OS X.
Revert "[cuda] Driver changes to compile and stitch together host and device-side CUDA code."
Revert "Fixed regex to properly match '64' in the test case."
Revert "clang/test/Driver/cuda-options.cu REQUIRES clang-driver, at least."
llvm-svn: 242077
- Changed driver pipeline to compile host and device side of CUDA
files and incorporate results of device-side compilation into host
object file.
- Added a test for cuda pipeline creation in clang driver.
New clang options:
--cuda-host-only - Do host-side compilation only.
--cuda-device-only - Do device-side compilation only.
--cuda-gpu-arch=<ARCH> - specify GPU architecture for device-side
compilation. E.g. sm_35, sm_30. Default is sm_20. May be used more
than once in which case one device-compilation will be done per
unique specified GPU architecture.
Differential Revision: http://reviews.llvm.org/D9509
llvm-svn: 242058
Justin Lebar