[Performance]: Static ViT model significantly slower than dynamic with same input shape

mertalev commented 1 month ago

OpenVINO Version

tag 2024.1.0

Operating System

Debian Bookworm (with latest intel-opencl-icd: 24.22.29735.21-1)

Device used for inference

iGPU

OpenVINO installation

PyPi

Programming Language

Python

Hardware Architecture

x86 (64 bits)

Model used

https://huggingface.co/immich-app/ViT-B-16-plus-240__laion400m_e32/resolve/6e7e12096396bc9280b6db412ac30eef0a9cbe3e/visual/vit-plus-dyn-sim.onnx

Model quantization

No

Target Platform

Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 46 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 22 On-line CPU(s) list: 0-21 Vendor ID: GenuineIntel Model name: Intel(R) Core(TM) Ultra 7 155H CPU family: 6 Model: 170 Thread(s) per core: 2 Core(s) per socket: 16 Socket(s): 1 Stepping: 4 CPU(s) scaling MHz: 22% CPU max MHz: 4800.0000 CPU min MHz: 400.0000 BogoMIPS: 5990.40 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 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_dea dline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb intel_ppin ssbd ibrs ibpb stibp ibrs_enhanced tp r_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel _pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect user_shstk avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_w indow hwp_epp hwp_pkg_req hfi vnmi umip pku ospke waitpkg gfni vaes vpclmulqdq rdpid bus_lock_detect movdiri movdir64b fsrm md_c lear serialize arch_lbr ibt flush_l1d arch_capabilities Virtualization features:
Virtualization: VT-x NUMA:
NUMA node(s): 1 NUMA node0 CPU(s): 0-21 Vulnerabilities:
Gather data sampling: Not affected Itlb multihit: Not affected L1tf: Not affected Mds: Not affected Meltdown: Not affected Mmio stale data: Not affected Reg file data sampling: Not affected Retbleed: Not affected Spec rstack overflow: Not affected Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS Not affected; BHI BHI_DIS_S Srbds: Not affected Tsx async abort: Not affected

Performance issue description

The MatMul ops in the static version of this model are extremely slow compared to the dynamic version. The linked model is dynamic for easy conversion to static dimensions.

I converted this model into OV format with this command: ovc <model_file> --compress_to_fp16 false. I then renamed the model and ran the same command, but with the --input [1,3,240,240] flag. Running benchmark_app for both OV models, I get very different results:

Static

Command line parameters
-d;GPU.0
-m;vit-plus-dyn-1.xml
-pc;
-report_type;detailed_counters
-data_shape;[1,3,240,240]

Configuration setup
topology;Model1
target device;GPU.0
API;async
inference_only;True
precision;UNSPECIFIED
batch size;1
number of iterations;None
number of parallel infer requests;4
duration (ms);60000

Execution results
read model time (ms);34.44
compile model time (ms);1027.40
first inference time (ms);109.96
total execution time (ms);60384.59
total number of iterations;692
latency (ms);348.69
avg latency;348.07
min latency;177.24
max latency;371.51
throughput;11.46

Dynamic

Command line parameters
-d;GPU.0
-m;vit-plus-dyn.xml
-pc;
-report_type;detailed_counters
-data_shape;[1,3,240,240]

Configuration setup
topology;Model1
target device;GPU.0
API;async
inference_only;False
precision;UNSPECIFIED
batch size;1
number of iterations;None
number of parallel infer requests;4
duration (ms);60000

Execution results
read model time (ms);46.83
compile model time (ms);4733.11
first inference time (ms);344.01
total execution time (ms);60084.50
total number of iterations;2616
latency (ms);89.58
avg latency;91.69
min latency;51.45
max latency;167.14
throughput;43.54

The dynamic model has nearly 4x lower latency and nearly 4x higher throughput. The higher throughput could make sense if it had batched inputs, but the inputs were [1,3,240,240] in both cases. The latency should also be higher, not lower, in the dynamic model.

Looking at the individual kernel times, these ops stand out:

Static

MatMul_1021;Status.EXECUTED;FullyConnected;fully_connected_gpu_bf_tiled__f16;3.961;0.002
MatMul_1033;Status.EXECUTED;FullyConnected;fully_connected_gpu_bf_tiled__f16;1.694;0.001
MatMul_1062;Status.EXECUTED;FullyConnected;fully_connected_gpu_bf_tiled__f16;1.092;0.002

Dynamic

MatMul_1021;Status.EXECUTED;FullyConnected;fully_connected_gpu_bf_tiled__f16;0.668;0.014
MatMul_1033;Status.EXECUTED;FullyConnected;fully_connected_gpu_bf_tiled__f16;0.499;0.016
MatMul_1062;Status.EXECUTED;FullyConnected;fully_connected_gpu_bf_tiled__f16;0.349;0.017

I wondered if the 4x difference could be related to the line number of parallel infer requests;4, but adding -nireq 1 and re-running didn't change the relative gap between the two:

Static

[ INFO ] Execution Devices:['GPU.0']
[ INFO ] Count:            610 iterations
[ INFO ] Duration:         60156.68 ms
[ INFO ] Latency:
[ INFO ]    Median:        98.04 ms
[ INFO ]    Average:       98.43 ms
[ INFO ]    Min:           95.16 ms
[ INFO ]    Max:           109.08 ms
[ INFO ] Throughput:   10.14 FPS

Dynamic

[ INFO ] Execution Devices:['GPU.0']
[ INFO ] Count:            2307 iterations
[ INFO ] Duration:         60036.19 ms
[ INFO ] Latency:
[ INFO ]    Median:        25.56 ms
[ INFO ]    Average:       25.83 ms
[ INFO ]    Min:           23.67 ms
[ INFO ]    Max:           45.64 ms
[ INFO ] Throughput:   38.43 FPS

Step-by-step reproduction

Download the provided ONNX model
Convert to OV format with ovc <model_file> --compress_to_fp16 false
Rename ONNX model to avoid overwriting the OV model
Convert again to OV format with ovc <model_file> --compress_to_fp16 false --input [1,3,240,240]
Run benchmark_app -d GPU.0 -m <model_file.xml> -pc -report_type detailed_counters -data_shape [1,3,240,240] for both models and compare results

Issue submission checklist

[X] I'm reporting a performance issue. It's not a question.
[X] I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
[X] There is reproducer code and related data files such as images, videos, models, etc.

wenjiew commented 1 month ago

@riverlijunjie Is this similar to the issue you are observing? Thanks!

riverlijunjie commented 1 month ago

@riverlijunjie Is this similar to the issue you are observing? Thanks!

Not the same issue, what I observed is that FC kernel was in poor performance.

wenjiew commented 1 month ago

@geunhwan to take a look and provide feedback. Thanks!

openvinotoolkit / openvino