hcc_detail/math_functions.h used to include hcc_detail/hip_runtime.h.
Removing it has caused regression in TensorFlow 1.8.
Put it back for backward compatibiliity.
SWDEV-145570 - [HIP] - Fix some issues in hip runtime
- Set stream for event
- Free mem needs to be reported in bytes but runtime backends reports in Kb
ReviewBoardURL = http://ocltc.amd.com/reviews/r/15586/diff/
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_memory.cpp#40 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#15 edit
SWDEV-145570 - [HIP] - hipHccModuleLaunchKernel needs to be a c++ sym
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_hcc.map.in#9 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#14 edit
SWDEV-145570 - [HIP] Refactored some g_* stuff
Refactored g_functions into a platform state.
Added a _vars for registered variables.
Added an execution stack similar to Hcc-clang.
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_platform.cpp#14 edit
SWDEV-145570 - [HIP] refactor hipMemcpy* functions to correctly handle copies using prepinned memory
The current implementation of hipMemcpy functions picks the copy type based on a flag that the user passes. However, one can use the hipMemcpyHostToDevice/hipMemcpyDeviceToHost flag in a combination with prepinned memory. By using the WriteMemoryCommand/ReadMemoyCommand in this case, we will pin the same host memory twice. This is fine on PAL/Linux, since pinning the same VA range is a noop, but this will start failing once we switch to using device memory with HIP/VDI/HSA.
The solution is to ignore the hipMemcpyKind flag and let the runtime decide what kind of copy is best to do. Except for the case when hipMemcpyHostToHost is passed, since both host pointers may be prepinned.
ReviewBoardURL = http://ocltc.amd.com/reviews/r/15482/diff/
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_memory.cpp#37 edit
When shareWithAll memory (e.g., host memory) is allocated, set appId
in hc::AmPointerInfo to -1 to indicate that this memory is not mapped
to any device. Peer checking in ihipStream_t::canSeeMemory is not
necessary if memory is shared with all devices. Thus, it is skipped.
Note that earlier host memory is always mapped to device 0 and HIP
always performs peer checking for all kinds of hipMemcpy. Since the
peer checking process requires context locking, hipMemcpy from/to host
memory always grabs device 0's context lock. Therefore, if there is
another thread holding the context lock of device 0 (e.g.,
hipDeviceSynchronize on device 0), hipMemcpy will have to wait for the
lock until it can actually perform memcpy. This can significantly
deteriorate execution performance.
Signed-off-by: Sarunya Pumma <sarunya.pumma@amd.com>
Maneesh Gupta
Evgeny Mankov
sunway513
Jenkins
Rahul Garg
Evgeny Mankov
Yaxun Sam Liu
foreman
Wen-Heng (Jack) Chung
Aaron Enye Shi
lthakur
Sarunya Pumma