scheduler redesign #101
Merged
scheduler redesign #101
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
The MemoryManager is supposed to be a single point of all allocations in SYCL RT. Signed-off-by: Vlad Romanov <vlad.romanov@intel.com>
Do not allocate memory on host in buffer_impl constructors when user's ptr is provided. This is possible because the ownership of that memory is given to the constructed SYCL buffer for the duration of it' lifetime. The memory still need to be allocated in constructor that takes iterators as they can point to non-consecutive memory(need consecutive memory to interact with OpenCL). This may improve performance as we avoid several memory allocation/copy. As required by the Specification the host memory now always allocated using allocator provided by user. Moved methods that allocate, fill and copy memory to separate class. Copy method was improved to support device to host and host to device partial(aka CopyRect) version of copies. So, now the buffer_impl class is decoupled from OpenCL. Fix possible bug, now constructor from shared_ptr stores a copy of shared_ptr in lambda to avoid deallocation of memory until buffer is destructed or set_final_data is invoked to override destination for updating host memory(aka copy back). Signed-off-by: Vlad Romanov <vlad.romanov@intel.com>
Modified handler class so it creates special CG class. CG class represents different types of command groups, such as kernel execution, fill and so on. This is needed to decouple Scheduler from template parameters introduced by lambda, various ranges and arguments. Signed-off-by: Vlad Romanov <vlad.romanov@intel.com>
…ass. The Requirement class is added as base non-templated class for accessor. This class describes command group's requirement to memory object and is used to build dependency graph in Scheduler. Signed-off-by: Vlad Romanov <vlad.romanov@intel.com>
This patch introduces new scheduler which is enabled when SCHEDULER_20 macro is set(set by default). The new scheduler is based on accessor rather on buffer, so it will support images. Also now there are two commands that implements moving memory to other contexts instead of one - alloca memory and memcpy. There is new command - release memory which will be ran during the sycl::buffer or sycl::image object destruction and releases memory instances for the memory object. Signed-off-by: Vlad Romanov <vlad.romanov@intel.com>
This patch fixes compilation of sycl application with default compiler(without specifying -fsycl). To do this default KernelInfo implementation is added + check that skips integration header processing if specialization for KernelInfo(which is supposed to come from integraion header) is not available. Signed-off-by: Vlad Romanov <vlad.romanov@intel.com>
romanovvlad
force-pushed
the
private/vromanov/SchedulerRedesign
branch
from
824bfa7
to
d75ca01
Compare
vladimirlaz
approved these changes
Apr 23, 2019
| cl_int MapMemObject::enqueueImp() { | ||
| std::vector<cl_event> RawEvents = | ||
| Command::prepareEvents(detail::getSyclObjImpl(MQueue->get_context())); | ||
| assert(MDstReq.getNumOfDims() == 1); |
There was a problem hiding this comment.
Debug build is broken by this line of code. Looks like it should be
Suggested change
| assert(MDstReq.getNumOfDims() == 1); | |
| assert(MDstReq.MDims == 1); |
There was a problem hiding this comment.
Yes, you are right. Thanks.
vmaksimo
pushed a commit
to vmaksimo/llvm
that referenced
this pull request
Feb 19, 2021
CONFLICT (content): Merge conflict in clang/utils/TableGen/ClangProgModelBuiltinEmitter.cpp
steffenlarsen
pushed a commit
to steffenlarsen/llvm
that referenced
this pull request
May 21, 2021
CONFLICT (content): Merge conflict in llvm/include/llvm/LinkAllPasses.h
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Plan for next commits: