First version of dispatcher #8713

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Merged

smessmer merged 5 commits into pytorch:master from smessmer:typeid5

Jun 25, 2018

Contributor

smessmer commented Jun 20, 2018 •

edited

Loading

This is a first version of the dispatcher. It uses multi-dispatch by default, but allows the operator schema to override that. It currently only defines unboxed kernel registration and unboxed calling APIs, boxed kernel registration and boxed calling APIs will come as well.

smessmer requested review from Yangqing, dzhulgakov and ezyang

June 20, 2018 21:34

onnxbot added the caffe2 label

Contributor

ezyang commented Jun 20, 2018

@pytorchbot retest this please

Yangqing reviewed

View reviewed changes

caffe2/core/dispatch/DispatchTable.h

    
                  template<class Key_>

                  void emplace(Key_&& key, void* value) {

                    // TODO Locking

                    //std::unique_lock<std::shared_timed_mutex> lock(mutex_);

This comment was marked as off-topic.

Sign in to view

caffe2/core/dispatch/DispatchTable.h

    
                  void erase(const Key& key) {

                    // TODO Locking

                    //std::unique_lock<std::shared_timed_mutex> lock(mutex_);

This comment was marked as off-topic.

Sign in to view

Yangqing reviewed

View reviewed changes

caffe2/core/dispatch/DispatchTable.h Outdated

    
              private:

                  ska::flat_hash_map<Key, void*> map_;

                  // TODO Figure out how to get fast locking in C++11 (use boost::shared_timed_mutex? folly::SharedMutex?)

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

dzhulgakov reviewed

View reviewed changes

caffe2/core/dispatch/DispatchTable.h Outdated

    
                    auto result = map_.emplace(std::forward<Key>(key), value);

                    if (!result.second) {

                      throw std::logic_error("Tried to register conflicting operators to the dispatcher.");

This comment was marked as off-topic.

Sign in to view

caffe2/core/dispatch/DispatchTable.h

    
                  // TODO The current implementation below does not have the correct correctness characteristics

                  // which we need.  It's worth spelling out exactly what we need:

                  //

                  //  - We need LOCK FREE read access to the table (as per the performance benchmark

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

caffe2/core/dispatch/DispatchTable.h Outdated

    
                 * @param func Concrete function implementation to register

                 * @param dispatch_key Dispatch key to implement this function with

                 */

                void registerOp(typename Schema::signature::func_type* func, typename Schema::dispatch::dispatch_key_type dispatch_key) {

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

caffe2/core/dispatch/DispatchTable.h Outdated

    
                  auto dispatch_key = Schema::dispatch::dispatch_key(args...);

                  void* found = ops_.lookup(dispatch_key);

                  if (found == nullptr) {

                    throw std::logic_error("Didn't find operator to dispatch to");

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

caffe2/core/dispatch/KernelRegistration.h Outdated

    
               * @tparam hasKernel Boolean for compile-time checking that a kernel is specified before finalizing the builder

               * @tparam hasDispatchKey Boolean for compile-time checking thhat a dispatch key is specified before finalizing the builder

               */

              template<class OpSchemaDef, bool hasKernel, bool hasDispatchKey>

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

smessmer added 3 commits

June 22, 2018 14:36


          Implement dispatcher prototype

3c02221


          code review

37309e8


          code review comments

b0d9e97

smessmer force-pushed the typeid5 branch from 6df5893 to b0d9e97 Compare

June 22, 2018 21:36

smessmer mentioned this pull request

Implement calling of c10 ops from c2 #8809

Closed

dzhulgakov approved these changes

View reviewed changes

Collaborator

dzhulgakov left a comment

Stamping to unblock - please address comments

caffe2/core/dispatch/DeviceId.h Outdated

    
                      case DeviceTypeId::CUDA: return stream << "DeviceTypeId(CUDA)";

                      case DeviceTypeId::UNDEFINED: return stream << "DeviceTypeId(UNDEFINED)";

                  }

                  throw std::logic_error("Unknown DeviceTypeId");

This comment was marked as off-topic.

Sign in to view

smessmer added 2 commits

June 25, 2018 11:45


          Code review comments

c0d271f


          fixes

ba0e625

smessmer merged commit cca2476 into pytorch:master

smessmer deleted the typeid5 branch

June 25, 2018 20:12

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels