This PR takes ensures that the maxThreadsPerBlock returned by hipFuncGetAttributes is both a multiple of the warp size and that the register usage of the maximum block does not exceed the number of available registers.
Fixes#1662
Fix two errors in hipOccupancyMaxActiveBlocksPerMultiprocessor.
1) Fix a possible segfault if the user passed in a null pointer for
the numBlocks value.
2) Handle the situation when the user is asking for a block size
that is larger than what the target device can hold within a
single block.
* Fix bug in LaunchKernel test
Instead of passing the address of the gpu buffer, pass the address
of the pointer that holds the address of the gpu buffer
* Fix hipLaunchKernel's kernarg buffer construction.
The hipLaunchKernel implementation should rely on ihipModuleLaunchKernel
to construct the kernarg buffer correctly based on kernel metadata.
* Fix a bug in get_functions where the Kernel_descriptor wasn't constructed with the correct kernarg layout information.
* Fix a bug in kernarg layout parsing dealing with kernel without any arg
* teach ihipModuleLaunchKernel to handle kernel without any arg
* Add a more interesting test
This PR is a follow-up on PR# #1698 and it makes two more APIs (hipLaunchCooperativeKernel/hipLaunchCooperativeKernelMultiDevice) inline so that they can work correctly with lazy binding.
[Background] it was found that if lazy linking used for a library that calls hipExtLaunchMultiKernelMultiDevice API then this API can get the wrong program_state object for looking up device kernels leading to a "No device code available" error in this API.
To fix this issue, the API was refactored to be inline and get and pass the correct program_state to an internal hip API to request a multi-device kernel launch.
* [hip] add support for implicit kernel argument for multi-grid sync
* modified code for calculating the prev_sum
* change the impCoopArg type to size_t
* add memory clean up
* launch init_gws and main kernels into two separate loops
* [hip] add initial implementation for hipLaunchCooperativeKernel API
* [hip] use total number of work groups to initialize the GWS resource
* [hip] use only one argument for init_gws kernel
* [hip] use the device associated with the stream for checking the device properties
* Add support for hipFunGetAttribute
* Support NVCC path
* Test using sample module_api_global
* Try fixing CI build failure due to hip_prof_gen scan
* Fix for CI build issue
* Resolve conflict
* Rebase and resolve conflicts with master
* Fix build error
* Fix NVCC path build error
* all thread local access now through single struct
* clean up old commented-out code, more use of GET_TLS()
* fewer calls to GET_TLS by passing tls as a funtion argument
* revert unnecessary change to printf
* fix failing tests due to TLS change
* fix merge conflicts in ihipOccupancyMaxActiveBlocksPerMultiprocessor
* Added support of hipOccupancyMaxActiveBlocksPerMultiprocessor & hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags APIs
* Taking into account of SGPR usage to determine the max active blocks in hipOccupancyMaxActiveBlocksPerMultiprocessor()
* [hip] implement the hipExtLaunchMultiKernelMultiDevice API
* add a guard to check the HCC version for acquire_locked_hsa_queue() API which was introdued in HCC for ROCm 2.5
* modified code based on the requested changes
* changes to lock all streams before launching kernels for each device and unlock them after the dispatches
* check each stream to be valid before starting to lock all the streams
* Implement the hipOccupancyMaxPotentialBlockSize function
* Replaced hipGetDeviceProperties() call by ihipGetDeviceProperties() in ihipOccupancyMaxPotentialBlockSize()
* Add test for hipOccupancyMaxPotentialBlockSize in Module API
* Added extern declaration for ihipGetDeviceProperties() to be accessed inside ihipOccupancyMaxPotentialBlockSize()
* fixed hipOccupancyMaxPotentialBlockSize test build issue
* Fix hipOccupancyMaxPotentialBlockSize dtest
* Add BUILD_CMD in hipOccupancyMaxPotentialBlockSize dtest
* Revert "Add BUILD_CMD in hipOccupancyMaxPotentialBlockSize dtest"
This reverts commit 0480ff56f1441fc515d2c26ce33783e303423938.
* Disable hipOccupancyMaxPotentialBlockSize dtest on NVCC
* move extern declaration of ihipGetDeviceProperties to hip_module.cpp
* Update the limiation of 32 wavefronts per CU and 800/512 SGPRs for VI/pre-VI chips to calculate the occupancy
Appended 48 empty bytes to the kernarg area at runtime. The implicit arguments are enabled primarily for the hostcall services
and it is completely abstracted from the user code. Enabled it for both hip-clang and hip-hcc.
Use the code object manager library to parse the code object metadata. Both
code object v2 and v3 formats are now supported for HCC generated binaries.
* Add ersatz for NVRTC.
* Fix extraneous paren and use correct namespace.
* Use lowerCamelCase (yuck, yuck) consistently.
* Link against FS when building hiprtc lib.
* Correctly mark Manipulators. Fix dual compile.
* Add unit tests. Extend HIT to accept linker options.
* Make sure the HIPRTC library is installed.
* Better logging. Try to auto-detect the target.
* Stop specifying the target explicitly.
* Add missing flavour of `hipModuleLaunchKernel`.
* Program was already destroyed.
* Don't use `--genco`. Fix mangled name trimming.
* Fix HIPRTC breakage due to upstream noise.
* [dtests] Replace RUN -> TEST in hiprtc tests
Change-Id: Ie499e92dfe4e5c94634b1c2b76cf52d241bcfea3
* [hit] Set HIP_PATH to HIP_ROOT_DIR for all tests
Change-Id: Ib0ad1f99bc71c03e363e055dd508a7a4a210680a
* Revert "Revert "Use COMgr to read Kernel Args Metadata (#1006)""
This reverts commit a3d118eaa8.
* Revert "Use COMgr to read Kernel Args Metadata (#1006)"
This reverts commit 8a548bf40b.
* Revert "improve program state commentary"
This reverts commit 7aada87cbd.
* Revert "load program state once per agent"
This reverts commit c9117de8eb.
* start moving function_names() into the hip shared lib
* start moving code_object_blobs to a new "state" object
* Consolidate various program state related static objects into a
single program_state object
* minor clean up
* move more stuffs from functional_grid_launch into program_state
* debug make_kernarg
* moving lookup for kernargs size_align into program_state
* clean up old code for kernarg size and alignment
* update hip_module to use newer api in program_state
* Create public member functions for program_state
* move most program state functions into shared library
* Pass the data buffer size to load_executable
Otherwise, it can't figure what the data size is
just from the char* (since the data is not really a string)
* turning free functions in program state into members of program_state_impl
* change the free function globals() into a member of program_state_impl
* replace the static mutex used for populating globals
* moving associate_code_object_symbols_with_host_allocation into
program_state_impl
* move load_code_object_and_freeze_executable into program_state_impl
* moving executables and functions_names into program_state_impl
* moving kernels() into program_state_impl
* moving functions() into program_state_impl
* move get_kernargs into program_state_impl
* moving kernel_descriptor into program_state_impl
* moving kernargs_size_align calculation into program_state_impl
* Changing the handle to program_state_impl to a pointer
* moving program_state_impl into a separate inline source file
* fixing/cleaning up some header file includes
* moving member function for kernargs_size_align into program_state.cpp
* moving Kernel_descriptor into program_state.inl
* add a new class to manage agent globals
* moving all agent globals processing functions into agent_globals_impl
* load program state once per agent
re-merging PR991 against other program state changes
* fix per-agent program state member initialization
* cache executables based on elf name, isa, and agent.
This avoids program state reloading executables after a shared library is dlopened.
re-merging PR1057 against other program state changes
* protect executables cache by a global mutex
* return ref to executables cache
* adapt PR#981 Make hipModuleGetGlobal be in HIP runtime
jglaser
Joseph Greathouse
Aryan Salmanpour
Rahul Garg
Sarbojit2019
Maneesh Gupta
Siu Chi Chan
satyanveshd
Alex Voicu
searlmc1
Nick Curtis
Jeff Daily
Rahul Garg
wkwchau
cdevadas
Aryan Salmanpour
cdevadas
Laurent Morichetti
Yaxun (Sam) Liu