sdpa vulkan flash attention with cooperative matrix optimization

[nihui]

z-image-ncnn 1024x1024

end2end	5060ti	9060xt	7900xtx
baseline	1m49s	1m27s	47.8s
+sdpa fa2	1m33s	1m8s	40.4s

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov-commenter]

Codecov Report

❌ Patch coverage is 95.91837% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 92.95%. Comparing base (b0be0c9) to head (b496a4a).

Files with missing lines	Patch %	Lines
src/layer/vulkan/sdpa_vulkan.cpp	95.91%	4 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##           master    #6528    +/-   ##
========================================
  Coverage   92.95%   92.95%            
========================================
  Files         809      809            
  Lines      256808   257186   +378     
========================================
+ Hits       238713   239079   +366     
- Misses      18095    18107    +12     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR introduces a cooperative-matrix–based Vulkan flash-attention kernel and integrates it into the SDPA_vulkan path to accelerate attention when cooperative matrices and BF16 are available.

Changes:

• Add a new sdpa_fa_cm compute shader implementing online softmax and QK/V matmuls using cooperative matrices and shared-memory tiling.
• Extend SDPA_vulkan to create/destroy a dedicated flash-attention pipeline and to route the forward path through the new kernel when conditions are met.
• Add helper printing functions to inspect Mat and VkMat data from the SDPA Vulkan implementation.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 9 comments.

File	Description
src/layer/vulkan/shader/sdpa_fa_cm.comp	New cooperative-matrix flash-attention shader (QK, online softmax, PV, and packed FP16/BF16 I/O).
src/layer/vulkan/sdpa_vulkan.h	Adds flash-attention pipeline pointer and unroll parameters to the SDPA_vulkan class.
src/layer/vulkan/sdpa_vulkan.cpp	Creates/destroys the flash-attention pipeline, wires it into forward, and adds CPU/Vulkan pretty-print debugging helpers.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.