vulkan int8 packing quantize dequantize requantize

[nihui]

No description provided.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 94.41118% with 28 lines in your changes missing coverage. Please review.

Project coverage is 95.81%. Comparing base (7557f5c) to head (2110e77).
Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
src/gpu.cpp	83.78%	12 Missing ⚠️
src/allocator.cpp	9.09%	10 Missing ⚠️
src/layer/vulkan/dequantize_vulkan.cpp	98.44%	2 Missing ⚠️
src/layer/vulkan/quantize_vulkan.cpp	98.36%	2 Missing ⚠️
src/layer/vulkan/requantize_vulkan.cpp	98.54%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3731      +/-   ##
==========================================
- Coverage   95.81%   95.81%   -0.01%     
==========================================
  Files         831      834       +3     
  Lines      264948   265433     +485     
==========================================
+ Hits       253851   254315     +464     
- Misses      11097    11118      +21     

☔ 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	15593968	15661624	+67656 ⚠️
armhf	6602932	6657988	+55056 ⚠️
aarch64	9983872	9989392	+5520 ⚠️

[Copilot]

Pull Request Overview

This pull request adds comprehensive int8 support to Vulkan operations across ncnn including quantization, dequantization, requantization, and packing. Key changes include new/int8 pipelines and shader modules, modifications in GPU utility functions to handle int8, and corresponding test updates for quantize/dequantize and packing operations.

Reviewed Changes

Copilot reviewed 34 out of 34 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/test_quantize.cpp & oom variant	Replace existing scale_data creation for 1D tests (use constant 1)
tests/test_packing.cpp	Renaming and restructuring of packing functions to separate FP32 and int8 tests
src/net.cpp	Added handling for int8 arithmetic adjustments
Multiple shader files (e.g., requantize_.comp, quantize_.comp)	New shader code supporting int8 operations
src/gpu.cpp, src/allocator.cpp	Updates to utility operators, image format selection, and int8 support
src/command.cpp	Adjustments to record int8 transfers
Other Vulkan layer files (requantize_vulkan., quantize_vulkan., dequantize_vulkan.*, packing_vulkan.cpp)	Pipeline creation and operator updates to incorporate int8 support

Comments suppressed due to low confidence (3)

tests/test_quantize.cpp:27

• The change from using a.w to a constant '1' for creating scale_data in 1D cases is significant; please add an inline comment clarifying the reasoning for this modification to aid future maintainers.

        if (a.dims == 1) scale_data.create(1);

src/allocator.cpp:900

• For elempack == 8, using VK_FORMAT_R8G8B8A8_SINT may not correctly represent 8 individual int8 channels since this format typically supports 4 channels; please verify that the chosen image format meets the int8 packing requirements or update the format selection accordingly.

        if (elempack == 8) format = VK_FORMAT_R8G8B8A8_SINT;

tests/test_packing.cpp:217

• [nitpick] The renaming from 'test_packing_gpu_buffer' to 'test_packing_gpu_fp32' and the addition of an int8 variant are nontrivial changes; please ensure that all test calls and related documentation are updated to reflect this new naming and functionality split.

}