-
Notifications
You must be signed in to change notification settings - Fork 2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Performance]:Why is there a reorder op after variadicsplit op ? #24412
Comments
Hi @sitabulaixizawaluduo, did you check the latest 2024.1 release if it has the same issue? @dmitry-gorokhov, @mg-intel, please pay attention to performance degradation in 2024.0 release Best regards, |
Thanks for reply ! I have not try 2024.1. Do you mean that 2024.0 does have a performance drop? |
@sitabulaixizawaluduo Reorder op is reponsible for memory copy in this context. In 2023 Split operation does nothing and real memory copy (from Split input to model outputs) is performed by Reorder ops. |
Thanks! You can use the code above to get the onnx file, and then build the IR file with the ovc command. |
Which code? Maybe I missing smt. |
The code is shared in #24288 . Sharing the code here. `import numpy as np index = [1, 1, 1, 1, 1, 1, 1, 1, 10, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 30, 30, 30, 1, 1, 1, 1, 1, 1, 1, 1, 30, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 30, 1, 30, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] initializers= [] input_1 = helper.make_tensor_value_info('input_1', TensorProto.FLOAT, [256,279,81]) outputs_list = [] node_def = onnx.helper.make_node( |
I found that this reorder op doesn't appear all the time, if one of the branches after split continues after the split operation, then the reorder op doesn't appear. |
OpenVINO Version
2024.0.0
Operating System
Ubuntu 22.04 (LTS)
Device used for inference
CPU
OpenVINO installation
Build from source
Programming Language
Python
Hardware Architecture
x86 (64 bits)
Model used
recommend
Model quantization
No
Target Platform
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 52 bits physical, 57 bits virtual
CPU(s): 96
On-line CPU(s) list: 0-95
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 143
Model name: Intel(R) Xeon(R) Gold 6442Y
Stepping: 8
Frequency boost: enabled
CPU MHz: 2601.000
CPU max MHz: 2601.0000
CPU min MHz: 800.0000
BogoMIPS: 5200.00
Virtualization: VT-x
L1d cache: 2.3 MiB
L1i cache: 1.5 MiB
L2 cache: 96 MiB
L3 cache: 120 MiB
NUMA node0 CPU(s): 0-23,48-71
NUMA node1 CPU(s): 24-47,72-95
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Vulnerable: eIBRS with unprivileged eBPF
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2
ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology
nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 ds_cpl vmx smx est tm2 ssse3 sdbg fma cx1
6 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_l
m abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ib
pb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi
2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha
_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local spl
it_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx
512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid bus_lock_detect cldemo
te movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr avx512_fp16 amx_tile flush_l
1d arch_capabilities
Performance issue description
I found that the VariadicSplit op is slower in version 2024.0.0 than in 2023.0.0. I processed the same onnx file with both versions of mo, 2023 and 2024, and at this point the structure of the two is the same when I look at it through Netron,
2024
2023
but when I processed it through benchmark_app,
taskset -c 0-23 benchmark_app -m split.xml -report_type detailed_counters -nstreams 24 -nthreads 24 -hint none -exec_graph_path benchmark_new.xml
and looked at the benchmark.xml file, I found that the 2023 version has an extra reorder node after the split op.
2024
2023
detailed.csv
2024
2023
1、What is the operation that causes the difference between the two version?
2、Why does the varadicsplit operation show not run in version 2023 but executed in version 2024?
Step-by-step reproduction
No response
Issue submission checklist
The text was updated successfully, but these errors were encountered: