fix the problem that v3's performance is worse than ck's

[minmengdie]

Motivation

fix the problem that v3's performance is worse than ck's

Technical Details

1. group mode uses a different grid dimension layout to launch kernel
2. update the fwd v3 kernel co file

Test Plan

MI300/MI308/MI355:
./benchmark_mha_fwd -mode=1 -b=32 -h=128 -h_k=8 -s=1024 -d=128 -iperm=0 -operm=0 -prec=bf16 -lse=1 -kname=1 -mask=2 -v=0
bash smoke_test_fwd_v3.sh

Test Result

perf result:
MI300:

MI308

MI355:

smoke test:
MI300:

MI308:

MI355:

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull Request Overview

This PR fixes a performance regression in the v3 FMHA (Fused Multi-Head Attention) implementation by introducing a new group-based kernel launch strategy. The changes replace the standard dispatcher with a group dispatcher that uses different grid/block dimensions to improve kernel launch efficiency.

Key changes:

• Added launch_kernel_group method with optimized grid dimension calculations
• Introduced fmha_fwd_v3_group_dispatcher function with modified tuning parameters
• Updated all kernel launch call sites to use the new group dispatcher

Reviewed Changes

Copilot reviewed 2 out of 18 changed files in this pull request and generated 3 comments.

File	Description
hsa/gfx950/fmha_v3_fwd/codegen.py	Adds group-based kernel launcher for gfx950 with tune_opt=5 and architecture-specific block dimensions
hsa/gfx942/fmha_v3_fwd/codegen.py	Adds group-based kernel launcher for gfx942 with tune_opt=4, then hardcoded to 0

---

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

[valarLip]

LGTM