EPR #419072 - [OpenCL2.0] Enable 16MB large on device queues
- Enable device queue creation up to 12MB. That should allow to run Intel SDK sample from the EPR that requires 6MB queue only.
- Currently a queue with >12.5MB size has a significant performance degradation. Thus the current max possible is 12MB. In general it's preferable to use the queue size more suitable for the task, rather than max possible.
Affected files ...
... //depot/stg/opencl/drivers/opencl/library/hsa/hsail/src/devenq/schedule.cl#10 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#115 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.hpp#38 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudefs.hpp#123 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#517 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpusched.hpp#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#372 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.hpp#131 edit
[ROCm/clr commit: 1386191b6c]
EPR #415638 - Improve APU performance
- Force remote allocation of local and persistent memory to Remote from RemoteUSWC:
- Use gpu copy for remote/pinned image/buffer.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#114 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.cpp#211 edit
[ROCm/clr commit: 637492a7dd]
ECR #304775 - Optimize oclBandwidthTest from nVidia SDK
- Cache pinned memory, since the benchmark sends the same transfer in a single batch. Thus we could avoid pin/unpin
- Swap SDMA engine allocation order. Blit manager allocates a queue on device, thus the first app queue was getting the paging second SDMA.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#112 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.hpp#37 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#339 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.hpp#121 edit
[ROCm/clr commit: dc8a3205ce]
ECR #304775 - Optimization for rectangular copies(Part2). Due to HW restriction of 14bits for src and dst pitch, its advantageous to choose optimal bpp. Higher the bpp the larger the byte pitch. This indirectly helps to reduce the number of packets for buffer copy(line by line vs a single sub_win raw packet)
ReviewBoardURL = http://ocltc.amd.com/reviews/r/5605/diff/
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#109 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.cpp#191 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.hpp#76 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp#64 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gslbe/src/rt/GSLContext.h#38 edit
[ROCm/clr commit: 5efe63df44]
ECR #304775 - Refactor code to do line by line copies for read\write Rect. This avoids taking the blit copy path which may be even slower.
ReviewBoardURL = http://ocltc.amd.com/reviews/r/5567/
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#108 edit
[ROCm/clr commit: a5e788c9f8]
ECR #304775 - Device enqueuing
- Use atomic fetch for enqueue flags
- Switch to a multithreaded scheduler
- Add a workaround for Linux host_multi_queue failures. Linux has only 2 queues, but the test allocates multiple host queues and the same HW ring can be used
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#106 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#449 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#127 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuschedcl.cpp#22 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#325 edit
[ROCm/clr commit: d2b905f18e]
ECR #304775 - Device enqueuing
- Switch to the single thread scheduler for now(the current version isn't friendly for single thread). Hopefully it's a temporary solution until synchronization issue with multithreaded scheduler will be identified.
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#104 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuschedcl.cpp#20 edit
[ROCm/clr commit: 6314b334ba]
foreman