SWDEV-2 - Change OpenCL version number from 2909 to 2910.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/utils/versions.hpp#2657 edit
[ROCm/clr commit: 39d84b3142]
* Fix hipcc for hip-clang.
If there is -g, do not add -O3 by default.
If HIP_VDI_HOME is not set, set HIP_VDI_HOME based on hipcc directory for HIP/VDI runtime.
For HIP/VDI runtime, set HIP_CLANG_PATH and DEVICE_LIB_PATH based on HIP_VDI_HOME only if they exist.
This allows using HIP/VDI runtime with hip-clang installed at /opt/rocm/llvm and device lib installed
at /opt/rocm/lib.
* Fix HIP_VDI_HOME for hipcc called from /opt/rocm/bin
[ROCm/clr commit: 34b1c6c5b2]
SWDEV-145570 - Support loading fat binary generated through --genco by hipModuleLoad.
hip-clang --genco generates fat binary instead of code object. To support that
we need to extract code object from fat binary in hipModuleLoadData. This is
needed for hipRTC since multiple GPU archs may be passed.
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#27 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_platform.cpp#31 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.cpp#42 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#308 edit
[ROCm/clr commit: bea21af6ad]
SWDEV-145570 - Support loading fat binary generated through --genco by hipModuleLoad.
hip-clang --genco generates fat binary instead of code object. To support that
we need to extract code object from fat binary in hipModuleLoadData. This is
needed for hipRTC since multiple GPU archs may be passed.
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#27 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_platform.cpp#31 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.cpp#42 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#308 edit
[ROCm/clr commit: c166e2513c]
SWDEV-2 - Change OpenCL version number from 2908 to 2909.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/utils/versions.hpp#2656 edit
[ROCm/clr commit: 09621adaca]
SWDEV-2 - Change OpenCL version number from 2907 to 2908.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/utils/versions.hpp#2655 edit
[ROCm/clr commit: 3ffa7150da]
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.
[ROCm/clr commit: de89102528]
SWDEV-2 - Change OpenCL version number from 2906 to 2907.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/utils/versions.hpp#2654 edit
[ROCm/clr commit: d6882ece45]
SWDEV-2 - Change OpenCL version number from 2905 to 2906.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/utils/versions.hpp#2653 edit
[ROCm/clr commit: 064f5d2c01]
SWDEV-189990, SWDEV-190337 - Update OpenCL/PAL gfx10 counter blocks again to match GPUPerfAPI.
SWDEV-190337:
- Update the GFX10 block list (gfx10BlockIsPal) to match GPUPerfAPI, which expects the OpenCL block lists to match OpenGL (UGL). Note: The list now matches the expectations of GPUPerfAPI. For TA/TD/TCP, nearly all GFX10 ASICs only require 10 or 12 instances (Arden would require 14, Mero would require 8, but not sure if those are supported by OCL), but we are using 16 instances to match UGL.
- Make sure the blockIdToIndexSelect array contains all the blocks supported by PAL (add a static_assert to ensure this)
- Refactor the PCIndexSelect enum. This enum is used to determine how to sum up counters across multiple block instances. The following types are now supported:
Instance -- no autosumming; instances have a one-to-one correlation with PAL
ShaderEngine -- the block is instanced per shader engine, and OpenCL will autosum counters across all PAL instances, providing a single value for all of PAL's instances
ShaderArray -- the block is instanced per shader array, and OpenCL will autosum counters across shader arrays, providing a single value for each instance within a shader array. For example, if a block has four instances per shader array, PAL would expose 16 instances total on Navi10 (2 SEs, 2 SAs per SE), but OpenCL will expose four instances
ComputeUnit -- the block is instanced per compute unit, and OpenCL will autosum counters across shader arrays, providing a single value for each compute-unit-per-shader-array. For example, if a block is instanced per compute unit, then PAL would expose 40 instances on a 40CU Navi10. OpenCL would support 10 instances (2 CUs-per-WGP, 5 WGPs-per-SA), autosummed across shader arrays.
SWDEV-189990:
- Revert GFX9 and GFX10 tests back to using the MCVML2 counter it was using previously (prior to CL 1766829). This is counter index 2, which the test calls "BigK bank 0 hits". In the aforementioned change list, I updated the counter index from 2 to 14, since index 14 is the actual counter that represents "BigK bank 0 hits". Counter index 2 is the number of hits, not "bigK" hits. This previous change caused a test regression reported in SWDEV-189990. By reverting the code to use counter 2, the expected value in the test should be correct. Perhaps a better update would be to change the description in the source from "BigK bank 0 hits" to "bank 0 hits", but for now, I'm just going to go back to what the test was doing before.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palcounters.cpp#21 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palcounters.hpp#11 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/runtime/OCLPerfCounters.cpp#47 edit
[ROCm/clr commit: fde741c2b4]
Zuhaib Khan
foreman
Konstantin Pyzhov
Maneesh Gupta
Jenkins
Yaxun (Sam) Liu
Evgeny Mankov
Siu Chi Chan
Aaron Enye Shi
Laurent Morichetti
Brian Sumner