SWDEV-219121 - [Navi][ROCm] Add performance counter support
This is initial implementation for Navi 10 performance counter in OpenCL runtime.
Tests:
1. http://ocltc.amd.com:8111/viewModification.html?modId=130609&personal=true&tab=vcsModificationBuilds
2. ./ocltst -m oclruntime.so -t OCLPerfCounters
Before this code change, the segmenation fault happens inside OpenCL runtime. After this code change, the error happened inside HSA. Error message is generated in function hsa_ven_amd_aqlprofile_start. Inside this function, the C++ try block create exception. The exception error in HSA is "GFXIP is not supported(gfx1010)". We need HSA to add support for Navi.
ReviewBoards: http://ocltc.amd.com/reviews/r/18463/
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/roccounters.cpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/roccounters.hpp#4 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#94 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/runtime/OCLPerfCounters.cpp#50 edit
[ROCm/clr commit: 68107b160f]
SWDEV-216708 - [hipclang-vdi-rocm][FBA-77]hipGetDeviceCount() query should not trigger any queue creation.
- KFD queues get created when we call hsa_executable_load_agent_code_object when creating a blitProgram at init. Delay blit creation until when needed
- Queue also gets created when we init a xferQueue. Delay that too until when needed.
If we want to have a lite init, there are multiple changes needed in both OpenCL and HIP. Thats rather a feature to work on later.
ReviewBoardURL = http://ocltc.amd.com/reviews/r/18416/diff/
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.cpp#34 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#148 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#47 edit
[ROCm/clr commit: 1384e82e60]
SWDEV-204782 - introduce hostcall
Hostcall is a service that allows a kernel to submit requests to the
host using shared buffers, and block until a response is received. This
will eventually replace the shared buffer currently used for printf, and
repurposes the same hidden kernel argument.
When the runtime launches a kernel that requires the hostcall service it
performs the following actions:
- Launch a hostcall listener thread if it is not already running.
- Locate the hostcall buffer for the corresponding hardware queue, or
create a new one.
- Register the new hostcall buffer with the listener thread.
- Set the hostcall buffer pointer as an implicit argument to the kernel.
Affected files ...
... //depot/stg/opencl/drivers/opencl/make/hip.git/tests/Makefile#21 edit
... //depot/stg/opencl/drivers/opencl/make/hip.git/tests/build/Makefile.hip_tests#31 edit
... //depot/stg/opencl/drivers/opencl/make/hip.git/tests/scripts/hip_hostcall_tests.txt#1 add
... //depot/stg/opencl/drivers/opencl/make/hip.git/tests/scripts/run_all_tests.sh#22 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.cpp#30 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.hpp#19 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#143 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#45 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rochostcall.cpp#1 add
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rochostcall.hpp#1 add
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#92 edit
[ROCm/clr commit: 416f273e94]
SWDEV-198863 - Options for hip-clang-vdi path to provide the chicken bits, or functional equivalents to HCC_DB (phase 1)
1. The log macros is turned off for release build. So log functions has zero impact to release build.
2. The log macros have level, mask, condition control. So we can have more control to avoid log flooding.
I also adjusted some existing log to use new log functions.
1. To excercise and test the new log functions.
2. To improve performance slightly.
3. The change is mainly for HIP-ROCM, we can move more in next phases for PAL or ORCA.
4. I make these log feature unavailable for release build. We can revert to old log functions for release build in a case by case method.
Tests:
1. http://ocltc.amd.com:8111/viewModification.html?modId=128289&personal=true&tab=vcsModificationBuildshttp://ocltc.amd.com:8111/viewModification.html?modId=128358&personal=true&tab=vcsModificationBuilds
2. release build, run hip program, there is no log
3. fastdebug build, run hip program,
export LOG_LEVEL=3
export GPU_LOG_MASK=4294967295
There was a lot of logs.
4. fastdebug build, run hip program,
export LOG_LEVEL=2
export GPU_LOG_MASK=4294967295
There was no logs.
5. fastdebug build, run hip program,
export LOG_LEVEL=3
export GPU_LOG_MASK=4294967294
There was much less logs.
6. fastdebug build, run hip program,
export LOG_LEVEL=3
export GPU_LOG_MASK=47102
There was even much less logs. The logs was expected according to the mask.
7. Tested step 2 to 6 similarily in Windows and Linux
ReviewBoard: http://ocltc.amd.com/reviews/r/18215
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_internal.hpp#46 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_memory.cpp#82 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_stream.cpp#26 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hiprtc_internal.hpp#2 edit
... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_svm.cpp#29 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/comgrctx.cpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.cpp#29 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.cpp#68 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#137 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#91 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#100 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/commandqueue.cpp#32 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/runtime.cpp#40 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/debug.hpp#10 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#323 edit
[ROCm/clr commit: 3f6e18bf6b]
SWDEV-208424 - ROCr language runtime should not free code object until executable destroy
- Keep the code object reader alive until the program destruction. Update HSAIL path only, since LC path already handles it correctly.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.cpp#107 edit
[ROCm/clr commit: a1c86d10b5]
SWDEV-184710 - Support hipLaunchCooperativeKernelMultiDevice()
- Add support for multi grid launch in hip
- Detect the new hidden argument and pass the required information for the kernel launch
- Memory for synchronization is allocated as a single object and then the offset for each GPU is found
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#44 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#343 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.cpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.hpp#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#82 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#136 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#42 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#90 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#30 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#99 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.hpp#97 edit
[ROCm/clr commit: 6e7e97987f]
SWDEV-198862 - Options for hip-clang-vdi path to provide the chicken bits, or functional equivalents to HCC_OPT_FLUSH
Add HCC_OPT_FLUSH flag to use fence scope agent when possible for HIP VDI. The flag is defaulted to turn on, similiar to HIP HCC.
Add AMD_OCL_OPT_FLUSH to use fence scope agent when possible for OpenCL. This was tested in Windows and PAL. Default is off.
This flag can be used for future OpenCL test.
Tests:
1. http://ocltc.amd.com:8111/viewModification.html?modId=127189&personal=true&tab=vcsModificationBuilds
The teamcity test includes HIP - VDI - Rocm tests.
2. VEGA10 , Windows, HIP, 110 hiptests PASS.
3. VEGA10 , Linux AMDGPU PRO, HIP - PAL, 110 hiptests PASS.
Newer:
http://ocltc.amd.com:8111/viewModification.html?modId=127193&personal=true&tab=vcsModificationBuilds
Reviewboard: http://ocltc.amd.com/reviews/r/18092/
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/device.cpp#247 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#342 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#89 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#29 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#321 edit
[ROCm/clr commit: d43f2b6372]
SWDEV-204782 - store extra information per HSA queue
The new struct QueueInfo is used to store metadata about each HSA
queue. For hostcall, this structure will eventually contain a pointer to
the hostcall buffer allocated to each HSA queue.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#135 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#41 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#88 edit
[ROCm/clr commit: 1820fe21cf]
SWDEV-189650 - [HIP-CLANG][HIP/VDI/PAL] Hangs on test hip_threadfence_system
1. In HIP + VDI + ROCm, allow SVM atomic in VEGA10 and later ASIC. GFX8 (Tonga) was enabled before.
2. In HIP + VDI + PAL Linux driver, allow SVM atomic in VEGA10 and later ASIC.
Tests:
1. In HIP + VDI + ROCm, hip_threadfence_system test passed.
2. In HIP + VDI + PAL + Linux , hip_threadfence_system test passed.
3. OpenCL + PAL, clinfo and ocltest runtime test pass.
4. OpenCL + ROCM, clinfo and ocltest runtime test pass.
5. Windows 10, VEGA 10, clinfo and and ocltest runtime test pass. hip_threadfence_system test passed by skipping the test.
Teamcity presubmission test:
http://ocltc.amd.com:8111/viewModification.html?modId=127083&personal=true&tab=vcsModificationBuildshttp://ocltc.amd.com:8111/viewModification.html?modId=127076&personal=true&tab=vcsModificationBuilds
ReviewBoard: http://ocltc.amd.com/reviews/r/18077/
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_memory.cpp#73 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#171 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#80 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#31 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#134 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocmemory.cpp#44 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#320 edit
[ROCm/clr commit: d3b6a9731c]
SWDEV-201128 - [HIP] test_snli_cuda failure
Default to sync packet
Make sure GPU_NUM_MEM_DEPENDENCY is 0 for HIP
No sync packet is only used when there are mem dependency check
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_context.cpp#22 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#86 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#28 edit
[ROCm/clr commit: 78cf6e473e]
SWDEV-79445 - OCL generic changes and code clean-up
- Remove the atomic flag from the pool of kernel arguments
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#85 edit
[ROCm/clr commit: dbcb93693f]
SWDEV-193423 - HIP/VDI - Support for lazy hsa queue creation
- Add queue pool support for HSA HW queues. GPU_MAX_HW_QUEUES controls the pool size. The current default value is 4 (the number of active pipes on GPU).
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#132 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#38 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#81 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#24 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#314 edit
[ROCm/clr commit: 950dccb283]
SWDEV-174198 - Put back the fix of segfault running integer_ops.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#80 edit
[ROCm/clr commit: 44401e2509]
SWDEV-174198 - Back out changelist 1968648.
Since it causes soft-hang in multiple_device_context test.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#79 edit
[ROCm/clr commit: 5781701a9c]
foreman