add time event support

[Chao1Han]

Refer pytorch/pytorch@ab557421a47, Implemented the event cache, added the start event, and enabled support for event timing (as event timing no longer causes hangs after version 2021.17). Its usage scenario lies in getDurationFromEvent, which is part of the FR feature triggered by the watchdog thread; therefore, no test case is currently available.

Usage

    pg = dist.distributed_c10d._get_default_group()
    pg._enable_collectives_timing()
    x = torch.ones([2, 2]).to(device)
    num_repeats = 10
    for _ in range(num_repeats):
        dist.all_reduce(x)
    time.sleep(1)
    t = pickle.loads(torch._C._distributed_c10d._dump_xccl_trace())
    for seq in range(num_repeats):
        duration = t["entries"][seq]["duration_ms"]
        print(duration)

[Copilot]

Pull Request Overview

This PR adds time event support to the XCCL (XPU Collective Communication Library) system by introducing event caching and timing capabilities. The changes enable performance measurement and event management for XPU operations through a caching mechanism.

Key changes:

• Introduces XPUEventCache class for efficient event object reuse and timing support
• Adds timing functionality to WorkXCCL with start/end events and duration calculation
• Updates point-to-point communication operations to support timing and preprocessing/postprocessing hooks

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
src/xccl/XPUEventCache.hpp	Defines the XPUEventCache class interface for managing cached XPU events
src/xccl/XPUEventCache.cpp	Implements event caching logic with timing support and thread-local device mapping
src/xccl/ProcessGroupXCCL.hpp	Adds timing support fields and template method overloads for point-to-point operations
src/xccl/ProcessGroupXCCL.cpp	Integrates event caching, timing functionality, and refactors point-to-point operations

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[guangyey]

I’m curious about how much performance improvement we could get by using the event cache. I guess nothing...
Overall LGTM.
I need some more time to go through the code.