flexible coopmat mnk and unified elempack for vulkan convolution 1x1s1d1

[nihui]

No description provided.

[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

Attention: Patch coverage is 43.46290% with 160 lines in your changes missing coverage. Please review.

Project coverage is 95.77%. Comparing base (a1f5d5b) to head (b705bdc).

Files with missing lines	Patch %	Lines
src/layer/vulkan/convolution_vulkan.cpp	50.61%	120 Missing ⚠️
src/gpu.cpp	0.00%	40 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6154      +/-   ##
==========================================
- Coverage   95.78%   95.77%   -0.01%     
==========================================
  Files         835      835              
  Lines      264943   265103     +160     
==========================================
+ Hits       253783   253910     +127     
- Misses      11160    11193      +33     

☔ 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.

[github-actions]

The binary size change of libncnn.so (bytes)

architecture	base size	pr size	difference
x86_64	15602512	15606928	+4416 ⚠️
armhf	6611348	6615712	+4364 ⚠️
aarch64	9921904	9987680	+65776 ⚠️

[nihui]

7900xtx	baseline	pr6514
squeezenet	0.5	0.42	-16.00%
mobilenet	0.46	0.4	-13.04%
mobilenet_v2	0.76	0.63	-17.11%
mobilenet_v3	0.88	0.78	-11.36%
shufflenet	0.51	0.51	0.00%
shufflenet_v2	0.73	0.68	-6.85%
mnasnet	0.81	0.66	-18.52%
proxylessnasnet	0.83	0.67	-19.28%
efficientnet_b0	1.44	1.25	-13.19%
efficientnetv2_b0	2.88	2.57	-10.76%
regnety_400m	1.1	1	-9.09%
blazeface	0.4	0.38	-5.00%
googlenet	1.73	1.46	-15.61%
resnet18	0.73	0.73	0.00%
alexnet	0.53	0.53	0.00%
vgg16	1.33	1.33	0.00%
resnet50	1.73	1.49	-13.87%
squeezenet_ssd	1.46	1.36	-6.85%
mobilenet_ssd	1.17	1.03	-11.97%
mobilenet_yolo	0.84	0.7	-16.67%
mobilenetv2_yolov3	2.43	2.44	0.41%
yolov4-tiny	3.21	3.11	-3.12%
nanodet_m	1.45	1.37	-5.52%
yolo-fastest-1.1	0.77	0.7	-9.09%
yolo-fastestv2	0.8	0.75	-6.25%
vision_transformer	7.47	7.48	0.13%
FastestDet	0.77	0.7	-9.09%
			-8.80%

[Copilot]

Pull Request Overview

This PR integrates flexible cooperative matrix (coopmat) support into the Vulkan 1x1 convolution path by selecting optimal M/N/K tile sizes at runtime and unifying specialized shaders into a single generic shader that handles different element and output packs.

• Added get_optimal_cooperative_matrix_mnk utility in GpuInfo to pick best coopmat dimensions.
• Replaced two fixed-size pack4 shaders with one parameterized convolution_1x1s1d1_cm.comp.
• Extended Convolution_vulkan to configure, pack weight and dispatch the unified shader using dynamic M/N/K and elempack settings.

Reviewed Changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/gpu.cpp	Implemented runtime coopmat MNK selection and fallback to fp32
src/gpu.h	Declared new get_optimal_cooperative_matrix_mnk utility
src/layer/vulkan/shader/convolution_1x1s1d1_cm.comp	Added a single unified 1x1 coopmat shader with dynamic tile sizes
src/layer/vulkan/convolution_vulkan.h	Added coopmat config members (use_cooperative_matrix, coopmat_*, UNROLL_*)
src/layer/vulkan/convolution_vulkan.cpp	Hooked up coopmat parameters, weight packing, pipeline creation and dispatch logic

Comments suppressed due to low confidence (2)

src/layer/vulkan/convolution_vulkan.h:62

• [nitpick] Member names like UNROLL_SG_M use uppercase and underscores, which is inconsistent with the project's camelCase field naming. Consider renaming to unrollSgM or similar for consistency.

    int UNROLL_SG_M;

src/gpu.h:385

• [nitpick] Consider adding documentation comments explaining each parameter and the return behavior of get_optimal_cooperative_matrix_mnk, as its signature is complex and understanding M/N/K selection logic will aid future maintainers.

    void get_optimal_cooperative_matrix_mnk(int M, int N, int K, VkComponentTypeKHR type, VkComponentTypeKHR acctype, VkScopeKHR scope, int& coopmat_M, int& coopmat_N, int& coopmat_K) const;