Use reference counting on factories #2048
Conversation
github-actions
bot
added
loader
|
Currently as a draft as I want to run some performance tests to see if this affects performance at all. |
|
Compute Benchmarks level_zero run (with params: ): |
|
Compute Benchmarks level_zero run (): Summaryresult is better
Chartsapi_overhead_benchmark_sycl SubmitKernel out of order---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title api_overhead_benchmark_sycl SubmitKernel out of order
todayMarker off
dateFormat X
axisFormat %s
section SubmitKernel(api=sycl<br>Profiling=0<br>Ioq=0<br>DiscardEvents=0<br>NumKernels=10<br>KernelExecTime=1<br>MeasureCompletion=0)
This PR (48.64 μs) : crit, 0, 48
baseline (50.631 μs) : 0, 50
- : 0, 0
- : 0, 0
api_overhead_benchmark_sycl SubmitKernel in order---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title api_overhead_benchmark_sycl SubmitKernel in order
todayMarker off
dateFormat X
axisFormat %s
section SubmitKernel(api=sycl<br>Profiling=0<br>Ioq=1<br>DiscardEvents=0<br>NumKernels=10<br>KernelExecTime=1<br>MeasureCompletion=0)
This PR (46.76 μs) : crit, 0, 46
baseline (49.385 μs) : 0, 49
- : 0, 0
- : 0, 0
api_overhead_benchmark_ur SubmitKernel out of order---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title api_overhead_benchmark_ur SubmitKernel out of order
todayMarker off
dateFormat X
axisFormat %s
section SubmitKernel(api=ur<br>Profiling=0<br>Ioq=0<br>DiscardEvents=0<br>NumKernels=10<br>KernelExecTime=1<br>MeasureCompletion=0)
This PR (31.176 μs) : crit, 0, 31
baseline (31.93 μs) : 0, 31
- : 0, 0
- : 0, 0
api_overhead_benchmark_ur SubmitKernel in order---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title api_overhead_benchmark_ur SubmitKernel in order
todayMarker off
dateFormat X
axisFormat %s
section SubmitKernel(api=ur<br>Profiling=0<br>Ioq=1<br>DiscardEvents=0<br>NumKernels=10<br>KernelExecTime=1<br>MeasureCompletion=0)
This PR (25.163 μs) : crit, 0, 25
baseline (28.586 μs) : 0, 28
- : 0, 0
- : 0, 0
memory_benchmark_sycl QueueInOrderMemcpy from Device to Device, size 1024---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title memory_benchmark_sycl QueueInOrderMemcpy from Device to Device, size 1024
todayMarker off
dateFormat X
axisFormat %s
section QueueInOrderMemcpy(api=sycl<br>IsCopyOnly=0<br>sourcePlacement=Device<br>destinationPlacement=Device<br>size=1KB<br>count=100)
This PR (415.26 μs) : crit, 0, 415
baseline (423.457 μs) : 0, 423
- : 0, 0
- : 0, 0
memory_benchmark_sycl QueueInOrderMemcpy from Host to Device, size 1024---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title memory_benchmark_sycl QueueInOrderMemcpy from Host to Device, size 1024
todayMarker off
dateFormat X
axisFormat %s
section QueueInOrderMemcpy(api=sycl<br>IsCopyOnly=0<br>sourcePlacement=Host<br>destinationPlacement=Device<br>size=1KB<br>count=100)
This PR (206.845 μs) : crit, 0, 206
baseline (253.906 μs) : 0, 253
- : 0, 0
- : 0, 0
memory_benchmark_sycl QueueMemcpy from Device to Device, size 1024---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title memory_benchmark_sycl QueueMemcpy from Device to Device, size 1024
todayMarker off
dateFormat X
axisFormat %s
section QueueMemcpy(api=sycl<br>sourcePlacement=Device<br>destinationPlacement=Device<br>size=1KB)
This PR (9.95 μs) : crit, 0, 9
baseline (9.179 μs) : 0, 9
- : 0, 0
- : 0, 0
memory_benchmark_sycl StreamMemory, placement Device, type Triad, size 10240---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title memory_benchmark_sycl StreamMemory, placement Device, type Triad, size 10240
todayMarker off
dateFormat X
axisFormat %s
section StreamMemory(api=sycl<br>type=Triad<br>size=10KB<br>useEvents=0<br>contents=Zeros<br>memoryPlacement=Device)
This PR (1.903 μs) : crit, 0, 1
baseline (1.854 μs) : 0, 1
- : 0, 0
- : 0, 0
api_overhead_benchmark_sycl ExecImmediateCopyQueue out of order from Device to Device, size 1024---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title api_overhead_benchmark_sycl ExecImmediateCopyQueue out of order from Device to Device, size 1024
todayMarker off
dateFormat X
axisFormat %s
section ExecImmediateCopyQueue(api=sycl<br>IsCopyOnly=1<br>MeasureCompletionTime=0<br>src=Device<br>dst=Device<br>size=1KB<br>ioq=0)
This PR (4.598 μs) : crit, 0, 4
baseline (4.506 μs) : 0, 4
- : 0, 0
- : 0, 0
api_overhead_benchmark_sycl ExecImmediateCopyQueue in order from Device to Host, size 1024---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title api_overhead_benchmark_sycl ExecImmediateCopyQueue in order from Device to Host, size 1024
todayMarker off
dateFormat X
axisFormat %s
section ExecImmediateCopyQueue(api=sycl<br>IsCopyOnly=1<br>MeasureCompletionTime=0<br>src=Host<br>dst=Host<br>size=1KB<br>ioq=1)
This PR (3.575 μs) : crit, 0, 3
baseline (3.613 μs) : 0, 3
- : 0, 0
- : 0, 0
miscellaneous_benchmark_sycl VectorSum---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title miscellaneous_benchmark_sycl VectorSum
todayMarker off
dateFormat X
axisFormat %s
section VectorSum(api=sycl<br>numberOfElementsX=512<br>numberOfElementsY=256<br>numberOfElementsZ=256)
This PR (862.055 μs) : crit, 0, 862
baseline (863.651 μs) : 0, 863
- : 0, 0
- : 0, 0
Velocity-Bench Hashtable---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title Velocity-Bench Hashtable
todayMarker off
dateFormat X
axisFormat %s
section hashtable
This PR (243.842459 M keys/sec) : crit, 0, 243
baseline (178.291413 M keys/sec) : 0, 178
- : 0, 0
- : 0, 0
Velocity-Bench Bitcracker---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title Velocity-Bench Bitcracker
todayMarker off
dateFormat X
axisFormat %s
section bitcracker
This PR (35.7031 s) : crit, 0, 35
baseline (35.8407 s) : 0, 35
- : 0, 0
- : 0, 0
Velocity-Bench CudaSift---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title Velocity-Bench CudaSift
todayMarker off
dateFormat X
axisFormat %s
section cudaSift
This PR (280.397 ms) : crit, 0, 280
baseline (283.294 ms) : 0, 283
- : 0, 0
- : 0, 0
Velocity-Bench Easywave---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title Velocity-Bench Easywave
todayMarker off
dateFormat X
axisFormat %s
section easywave
This PR (457 ms) : crit, 0, 457
baseline (457.0 ms) : 0, 457
- : 0, 0
- : 0, 0
Velocity-Bench QuickSilver---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title Velocity-Bench QuickSilver
todayMarker off
dateFormat X
axisFormat %s
section QuickSilver
This PR (116.11 MMS/CTT) : crit, 0, 116
baseline (115.63 MMS/CTT) : 0, 115
- : 0, 0
- : 0, 0
Velocity-Bench Sobel Filter---
config:
gantt:
rightPadding: 10
leftPadding: 120
sectionFontSize: 10
numberSectionStyles: 2
---
gantt
title Velocity-Bench Sobel Filter
todayMarker off
dateFormat X
axisFormat %s
section sobel_filter
This PR (798.662 ms) : crit, 0, 798
baseline (934.963 ms) : 0, 934
- : 0, 0
- : 0, 0
DetailsSubmitKernel(api=sycl Profiling=0 Ioq=0 DiscardEvents=0 NumKernels=10 KernelExecTime=1 MeasureCompletion=0)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/api_overhead_benchmark_sycl --test=SubmitKernel --csv --noHeaders --Ioq=0 --DiscardEvents=0 --MeasureCompletion=0 --iterations=100000 --Profiling=0 --NumKernels=10 --KernelExecTime=1 Output:TestCase,Mean,Median,StdDev,Min,Max,Type SubmitKernel(api=sycl Profiling=0 Ioq=1 DiscardEvents=0 NumKernels=10 KernelExecTime=1 MeasureCompletion=0)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/api_overhead_benchmark_sycl --test=SubmitKernel --csv --noHeaders --Ioq=1 --DiscardEvents=0 --MeasureCompletion=0 --iterations=100000 --Profiling=0 --NumKernels=10 --KernelExecTime=1 Output:TestCase,Mean,Median,StdDev,Min,Max,Type SubmitKernel(api=ur Profiling=0 Ioq=0 DiscardEvents=0 NumKernels=10 KernelExecTime=1 MeasureCompletion=0)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/api_overhead_benchmark_ur --test=SubmitKernel --csv --noHeaders --Ioq=0 --DiscardEvents=0 --MeasureCompletion=0 --iterations=100000 --Profiling=0 --NumKernels=10 --KernelExecTime=1 Output:TestCase,Mean,Median,StdDev,Min,Max,Type SubmitKernel(api=ur Profiling=0 Ioq=1 DiscardEvents=0 NumKernels=10 KernelExecTime=1 MeasureCompletion=0)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/api_overhead_benchmark_ur --test=SubmitKernel --csv --noHeaders --Ioq=1 --DiscardEvents=0 --MeasureCompletion=0 --iterations=100000 --Profiling=0 --NumKernels=10 --KernelExecTime=1 Output:TestCase,Mean,Median,StdDev,Min,Max,Type QueueInOrderMemcpy(api=sycl IsCopyOnly=0 sourcePlacement=Device destinationPlacement=Device size=1KB count=100)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/memory_benchmark_sycl --test=QueueInOrderMemcpy --csv --noHeaders --iterations=10000 --IsCopyOnly=0 --sourcePlacement=Device --destinationPlacement=Device --size=1024 --count=100 Output:TestCase,Mean,Median,StdDev,Min,Max,Type QueueInOrderMemcpy(api=sycl IsCopyOnly=0 sourcePlacement=Host destinationPlacement=Device size=1KB count=100)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/memory_benchmark_sycl --test=QueueInOrderMemcpy --csv --noHeaders --iterations=10000 --IsCopyOnly=0 --sourcePlacement=Host --destinationPlacement=Device --size=1024 --count=100 Output:TestCase,Mean,Median,StdDev,Min,Max,Type QueueMemcpy(api=sycl sourcePlacement=Device destinationPlacement=Device size=1KB)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/memory_benchmark_sycl --test=QueueMemcpy --csv --noHeaders --iterations=10000 --sourcePlacement=Device --destinationPlacement=Device --size=1024 Output:TestCase,Mean,Median,StdDev,Min,Max,Type StreamMemory(api=sycl type=Triad size=10KB useEvents=0 contents=Zeros memoryPlacement=Device)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/memory_benchmark_sycl --test=StreamMemory --csv --noHeaders --iterations=10000 --type=Triad --size=10240 --memoryPlacement=Device --useEvents=0 --contents=Zeros Output:TestCase,Mean,Median,StdDev,Min,Max,Type ExecImmediateCopyQueue(api=sycl IsCopyOnly=1 MeasureCompletionTime=0 src=Device dst=Device size=1KB ioq=0)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/api_overhead_benchmark_sycl --test=ExecImmediateCopyQueue --csv --noHeaders --iterations=100000 --ioq=0 --IsCopyOnly=1 --MeasureCompletionTime=0 --src=Device --dst=Device --size=1024 Output:TestCase,Mean,Median,StdDev,Min,Max,Type ExecImmediateCopyQueue(api=sycl IsCopyOnly=1 MeasureCompletionTime=0 src=Host dst=Host size=1KB ioq=1)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/api_overhead_benchmark_sycl --test=ExecImmediateCopyQueue --csv --noHeaders --iterations=100000 --ioq=1 --IsCopyOnly=1 --MeasureCompletionTime=0 --src=Host --dst=Host --size=1024 Output:TestCase,Mean,Median,StdDev,Min,Max,Type VectorSum(api=sycl numberOfElementsX=512 numberOfElementsY=256 numberOfElementsZ=256)Environment Variables:Command:/home/test-user/bench_workdir/compute-benchmarks-build/bin/miscellaneous_benchmark_sycl --test=VectorSum --csv --noHeaders --iterations=1000 --numberOfElementsX=512 --numberOfElementsY=256 --numberOfElementsZ=256 Output:TestCase,Mean,Median,StdDev,Min,Max,Type hashtableEnvironment Variables:Command:/home/test-user/bench_workdir/hashtable/hashtable_sycl --no-verify Output:hashtable - total time for whole calculation: 0.550428 s bitcrackerEnvironment Variables:Command:/home/test-user/bench_workdir/bitcracker/bitcracker -f /home/test-user/bench_workdir/velocity-bench-repo/bitcracker/hash_pass/img_win8_user_hash.txt -d /home/test-user/bench_workdir/velocity-bench-repo/bitcracker/hash_pass/user_passwords_60000.txt -b 60000 Output:---------> BitCracker: BitLocker password cracking tool <--------- ==================================
|
RossBrunton
force-pushed
the
ross/refc
branch
from
cc693d5
to
6898550
Compare
|
@pbalcer thanks for kicking off the benchmark build. Looks like at the very least there are no major regressions. I also did a few local rough tests on a few synthetic cases and didn't find significant slowdown there. I'll undraft this and create an associated llvm MR. |
Pre-commit MR for oneapi-src/unified-runtime#2048
|
e2e tests show many regressions. edit: the fact that all UR tests pass, while a lot of SYCL tests are failing, indiactes to me that we have a pretty big gap in tests. Maybe add one after you root cause the failures? |
Pre-commit MR for oneapi-src/unified-runtime#2048
RossBrunton
force-pushed
the
ross/refc
branch
from
6898550
to
e00cb3e
Compare
Pre-commit MR for oneapi-src/unified-runtime#2048
RossBrunton
force-pushed
the
ross/refc
branch
from
e00cb3e
to
2875a5e
Compare
Pre-commit MR for oneapi-src/unified-runtime#2048
Pre-commit MR for oneapi-src/unified-runtime#2048
RossBrunton
force-pushed
the
ross/refc
branch
from
2875a5e
to
3a6ed49
Compare
Pre-commit MR for oneapi-src/unified-runtime#2048
RossBrunton
force-pushed
the
ross/refc
branch
from
3a6ed49
to
6095fcd
Compare
github-actions
bot
added
ci/cd
RossBrunton
force-pushed
the
ross/refc
branch
2 times, most recently
from
16eea0a
to
972e943
Compare
Previously the factories used by ur_ldrddi (used when there are multiple backends) would add newly created objects to a map, but never release them. This patch adds reference counting semantics to the allocation, retention and release methods. A lot of changes were also made to fix use-after-free issues, specifically: * The `release` functions now no longer use the handle after freeing it. * `urDeviceTest` no longer frees devices that it dosen't own. * Some tests for reference counting now explicitly retain an extra copy before releasing them. No tests were added; this should be covered by tests for urThingRetain. Closes: oneapi-src#1784 .
RossBrunton
force-pushed
the
ross/refc
branch
from
972e943
to
ccab45f
Compare
Pre-commit MR for oneapi-src/unified-runtime#2048
Previously the factories used by ur_ldrddi (used when there are multiple backends) would add newly created objects to a map, but never release them. This patch adds reference counting semantics to the allocation, retention and release methods.
No tests were added; this should be covered by tests for urThingRetain.
Closes: #1784 .