[Performance] Low performance of vectorized kernels

[cfgfung]

🐛 Describe the bug

It is found that the vectorized kernels are not performing well. For instance, copy_() is just utilizing only 40-50% of the theoretical bandwidth.

Versions

Collecting environment information... PyTorch version: 2.6.0a0+git95f869c Is debug build: False CUDA used to build PyTorch: None ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64) GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 Clang version: Could not collect CMake version: version 3.22.1 Libc version: glibc-2.35

Python version: 3.10.15 | packaged by conda-forge | (main, Sep 30 2024, 17:51:04) [GCC 13.3.0] (64-bit runtime) Python platform: Linux-5.15.0-73-generic-x86_64-with-glibc2.35 Is CUDA available: False CUDA runtime version: No CUDA CUDA_MODULE_LOADING set to: N/A GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 52 bits physical, 57 bits virtual Byte Order: Little Endian CPU(s): 96 On-line CPU(s) list: 0-95 Vendor ID: GenuineIntel Model name: Intel(R) Xeon(R) Gold 6336Y CPU @ 2.40GHz CPU family: 6 Model: 106 Thread(s) per core: 2 Core(s) per socket: 24 Socket(s): 2 Stepping: 6 CPU max MHz: 3600.0000 CPU min MHz: 800.0000 BogoMIPS: 4800.00 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 pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 invpcid_single ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 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 split_lock_detect wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 2.3 MiB (48 instances) L1i cache: 1.5 MiB (48 instances) L2 cache: 60 MiB (48 instances) L3 cache: 72 MiB (2 instances) NUMA node(s): 2 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 Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable Vulnerability Retbleed: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected

Versions of relevant libraries: [pip3] numpy==2.1.2 [pip3] optree==0.13.0 [pip3] torch==2.6.0a0+git95f869c [conda] numpy 2.1.2 pypi_0 pypi [conda] optree 0.13.0 pypi_0 pypi [conda] torch 2.6.0a0+git95f869c pypi_0 pypi ~$ xpu-smi discovery +-----------+--------------------------------------------------------------------------------------+ | Device ID | Device Information | +-----------+--------------------------------------------------------------------------------------+ | 0 | Device Name: Intel(R) Data Center GPU Flex 170 | | | Vendor Name: Intel(R) Corporation | | | SOC UUID: 00000000-0000-0000-947c-ef6e26c509ae | | | PCI BDF Address: 0000:4d:00.0 | | | DRM Device: /dev/dri/card1 | | | Function Type: physical | +-----------+--------------------------------------------------------------------------------------+

[xytintel]

@cfgfung Note that it's reasonable not to hit the peak bandwidth for broadcast and dynamic cast cases, because the input/output layout can't be accessed in a coalescing manner.

[cfgfung]

Found the root cause. The root cause is that when we get an API to get the vector_size / SIMD lane from the LevelZero runtime, the runtime returns just half of the vector width. e.g.: For float32, it returns width 2 instead of width 4. I have created an internal JIRA ticket for the related teams to solve the problem

@xytintel Shall we add a hotfix on our side? We can add a simple heuristic to correct the width 2 to width 4. I have tried that on copy_() operator and the performance shows 70% improvement. This will help all other vectorized kernels.