support 4d mat in more operators

[nihui]

Add 4d Mat handling across generic and optimized backends, extend operator tests, and document supported input and output Mat dimensions.

[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 99.36190% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 96.01%. Comparing base (866b73d) to head (6d81c8a).
⚠️ Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layer/loongarch/shufflechannel_loongarch.cpp	87.50%	2 Missing ⚠️
src/layer/arm/dequantize_arm.cpp	90.00%	1 Missing ⚠️
src/layer/arm/dequantize_arm_asimdhp.cpp	85.71%	1 Missing ⚠️
src/layer/loongarch/quantize_loongarch.cpp	93.75%	1 Missing ⚠️
src/layer/mips/quantize_mips.cpp	93.75%	1 Missing ⚠️
src/layer/mvn.cpp	80.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6737      +/-   ##
==========================================
+ Coverage   95.98%   96.01%   +0.03%     
==========================================
  Files         965      965              
  Lines      404957   405109     +152     
==========================================
+ Hits       388681   388968     +287     
+ Misses      16276    16141     -135     

☔ 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.

[nihui]

@codex review

[Copilot]

Pull request overview

This PR extends NCNN’s operator implementations and the pnnx conversion/test suite to support 4D Mat (w, h, d, c) across more layers and backends (generic CPU + arch-specific + Vulkan), with expanded coverage via both C++ layer tests and pnnx Torch operator tests.

Changes:

• Add/extend 4D tensor handling in multiple layers (math ops, normalization/norm layers, shuffle/split, quantize/dequantize/requantize) across generic, x86/ARM/RISC-V/MIPS/LoongArch, and Vulkan paths.
• Update Vulkan pipelines/shaders to treat 4D as flattened height×depth where appropriate and adjust indexing/grouping logic for norm-related kernels.
• Expand C++ layer tests and pnnx Torch tests to include 4D-shaped inputs; add new tests and a new pnnx NCNN pass for nn.InstanceNorm3d.

Reviewed changes

Copilot reviewed 155 out of 155 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tools/pnnx/tests/ncnn/test_torch_pow.py	Extend pow test to include 4D input/output cases.
tools/pnnx/tests/ncnn/test_torch_erf.py	New Torch erf test covering 1D/3D/4D inputs.
tools/pnnx/tests/ncnn/test_torch_cumsum.py	Extend cumsum test to cover 4D and multiple axes.
tools/pnnx/tests/ncnn/test_nn_Softplus.py	New nn.Softplus module test including 4D input.
tools/pnnx/tests/ncnn/test_nn_RMSNorm.py	Extend RMSNorm module test to include (batch+4D) shapes.
tools/pnnx/tests/ncnn/test_nn_PReLU.py	Extend PReLU module test to include (batch+4D) shapes.
tools/pnnx/tests/ncnn/test_nn_LayerNorm.py	Extend LayerNorm module test to include (batch+4D) shapes.
tools/pnnx/tests/ncnn/test_nn_InstanceNorm3d.py	New nn.InstanceNorm3d module test (batch+4D -> 4D Mat).
tools/pnnx/tests/ncnn/test_nn_GLU.py	Extend GLU module test to include 4D input and more dims.
tools/pnnx/tests/ncnn/test_nn_ChannelShuffle.py	Extend ChannelShuffle module test to include 4D input.
tools/pnnx/tests/ncnn/test_F_softplus.py	New functional softplus test including 4D input.
tools/pnnx/tests/ncnn/test_F_rms_norm.py	Extend functional rms_norm to include (batch+4D) shapes.
tools/pnnx/tests/ncnn/test_F_prelu.py	Extend functional prelu to include (batch+4D) shapes.
tools/pnnx/tests/ncnn/test_F_pad.py	Extend functional pad test with 4D padding modes.
tools/pnnx/tests/ncnn/test_F_normalize.py	Extend functional normalize test to include 4D input.
tools/pnnx/tests/ncnn/test_F_layer_norm.py	Extend functional layer_norm to include (batch+4D) shapes.
tools/pnnx/tests/ncnn/test_F_instance_norm.py	New functional instance_norm test covering 3D+4D mats.
tools/pnnx/tests/ncnn/test_F_glu.py	Extend functional GLU test to include 4D input and axes.
tools/pnnx/tests/ncnn/CMakeLists.txt	Register new/extended pnnx-ncnn tests (erf, softplus, instance_norm, InstanceNorm3d).
tools/pnnx/src/pass_ncnn/nn_InstanceNorm3d.cpp	New pass mapping nn.InstanceNorm3d to NCNN InstanceNorm.
tools/pnnx/src/pass_ncnn/F_normalize.cpp	Extend normalize pass to accept rank-4 tensors (after batch removal).
tools/pnnx/src/CMakeLists.txt	Build-system wiring for the new InstanceNorm3d pass.
tests/test_statisticspooling.cpp	New layer test including 4D inputs for StatisticsPooling.
tests/test_split.cpp	New layer test covering 4D Split output counts.
tests/test_softplus.cpp	Add 4D RandomMat coverage and improved debug print.
tests/test_shufflechannel.cpp	Refactor helper to accept Mat; add 4D ShuffleChannel coverage.
tests/test_scale.cpp	Add dims==4 scale sizing + new 4D test cases.
tests/test_rmsnorm.cpp	Extend RMSNorm tests to include 4D mats + improved debug print.
tests/test_requantize.cpp	Add 4D requantize tests and adjust pack8 forcing for riscv.
tests/test_quantize.cpp	Add 4D quantize randomization + new 4D test cases.
tests/test_prelu.cpp	Add 4D PReLU test cases + improved debug print.
tests/test_power.cpp	Add 4D Power test cases + improved debug print.
tests/test_padding.cpp	Add additional padding test parameter coverage (incl. 3D padding params).
tests/test_normalize.cpp	Add 4D Normalize tests + improved debug print.
tests/test_noop.cpp	Add explicit 4D Noop tests + improved debug print.
tests/test_mvn.cpp	New MVN layer test with 4D inputs.
tests/test_memorydata.cpp	Add 4D MemoryData coverage (incl. param 11=d) + helper factoring.
tests/test_log.cpp	Add 4D Log test cases.
tests/test_layernorm.cpp	Add 4D LayerNorm tests + improved debug print.
tests/test_instancenorm.cpp	Add 4D InstanceNorm tests + improved debug print.
tests/test_input.cpp	New Input layer test including 4D inputs (param 11=d).
tests/test_hardswish.cpp	Add 4D HardSwish tests + improved debug print.
tests/test_hardsigmoid.cpp	Fix HardSigmoid beta ParamDict index; add 4D tests + improved debug print.
tests/test_glu.cpp	Add explicit 4D GLU axis coverage + improved debug print.
tests/test_expanddims.cpp	Add additional axis coverage (incl. negative axes) relevant to higher-rank behavior.
tests/test_exp.cpp	Add 4D Exp test cases.
tests/test_erf.cpp	Add 4D Erf test cases + improved debug print.
tests/test_dropout.cpp	Add 4D Dropout test cases + improved debug print.
tests/test_dequantize.cpp	Add 4D dequantize tests and adjust pack8 forcing for riscv.
tests/test_deepcopy.cpp	Add 4D DeepCopy test cases + improved debug print.
tests/test_cumulativesum.cpp	Add explicit 4D CumulativeSum axis coverage + improved debug print.
tests/test_cast.cpp	Skip GPU fp16p cast test if fp16 packed not supported.
tests/test_bnll.cpp	Add 4D BNLL test cases + improved debug print.
tests/test_absval.cpp	Add 4D AbsVal test cases + improved debug print.
tests/CMakeLists.txt	Register new tests (Input/MVN/Split/StatisticsPooling).
src/layer/x86/shufflechannel_x86.cpp	Extend ShuffleChannel x86 to treat 4D as whd per channel and preserve shape.
src/layer/x86/rmsnorm_x86.cpp	Add 4D RMSNorm handling paths (fp32 + bf16).
src/layer/x86/requantize_x86.cpp	Extend requantize x86 to support dims==4 allocation and processing size whd.
src/layer/x86/quantize_x86.cpp	Extend quantize x86 to support dims==4 allocation and processing size whd.
src/layer/x86/quantize_bf16s.h	Extend bf16 quantize helper to handle dims==4.
src/layer/x86/prelu_x86.cpp	Extend PReLU x86 to apply over dims==4 (size whd).
src/layer/x86/layernorm_x86.cpp	Add 4D LayerNorm handling (fp32 + bf16).
src/layer/x86/instancenorm_x86.cpp	Extend InstanceNorm x86 to compute stats over whd.
src/layer/x86/dropout_x86.cpp	Extend Dropout x86 to process dims==4 (size whd).
src/layer/x86/dequantize_x86.cpp	Extend dequantize x86 to process dims==4 (size whd).
src/layer/x86/dequantize_bf16s.h	Extend bf16 dequantize helper to allocate/process dims==4.
src/layer/vulkan/shufflechannel_vulkan.cpp	Treat 4D as flattened h*d in pipeline constants/specialization; preserve shape on output.
src/layer/vulkan/scale_vulkan.cpp	Treat 4D as flattened h*d for pipeline sizing/constants.
src/layer/vulkan/rmsnorm_vulkan.cpp	Extend grouping logic for dims==4 and flattened h*d layout.
src/layer/vulkan/prelu_vulkan.cpp	Treat 4D as flattened h*d in pipeline sizing/constants.
src/layer/vulkan/normalize_vulkan.cpp	Treat 4D as flattened h*d for reductions and constants.
src/layer/vulkan/layernorm_vulkan.cpp	Extend grouping logic for dims==4 and flattened h*d layout.
src/layer/vulkan/innerproduct_vulkan.cpp	Flatten includes depth dimension when preparing shape for GEMM path.
src/layer/vulkan/deepcopy_vulkan.cpp	Treat 4D as flattened h*d in pipeline sizing/constants.
src/layer/vulkan/shader/rmsnorm_norm.comp	Generalize coeff indexing for flattened h*d layout.
src/layer/vulkan/shader/rmsnorm_norm_pack4.comp	Same as above for pack4 path.
src/layer/vulkan/shader/padding_3d.comp	Implement replicate/reflect for 3D padding shader.
src/layer/vulkan/shader/padding_3d_pack4.comp	Same as above for pack4 path.
src/layer/vulkan/shader/layernorm_sub_mean_square.comp	Generalize group indexing for flattened h*d layout.
src/layer/vulkan/shader/layernorm_sub_mean_square_pack4.comp	Same as above for pack4 path.
src/layer/vulkan/shader/layernorm_norm.comp	Generalize group/inner indexing for flattened h*d layout.
src/layer/vulkan/shader/layernorm_norm_pack4.comp	Same as above for pack4 path.
src/layer/vulkan/shader/instancenorm_reduce_sum4_fp32.comp	Fix tail handling for reduce-sum4 over width.
src/layer/statisticspooling.cpp	Include depth in pooling size and use faster variance accumulation.
src/layer/shufflechannel.cpp	Extend generic ShuffleChannel to treat 4D as whd per channel.
src/layer/scale.cpp	Extend Scale to support dims==4 (size whd).
src/layer/rmsnorm.cpp	Add generic 4D RMSNorm implementation.
src/layer/requantize.cpp	Extend generic requantize to allocate/process dims==4 (size whd).
src/layer/quantize.cpp	Extend generic quantize to allocate/process dims==4 (size whd).
src/layer/prelu.cpp	Extend generic PReLU to support dims==4 (size whd).
src/layer/power.cpp	Extend Power to process dims==4 (size whd).
src/layer/normalize.cpp	Extend Normalize to process dims==4 (size whd).
src/layer/mvn.cpp	Extend MVN output shape and internal size to include depth.
src/layer/log.cpp	Extend Log to process dims==4 (size whd).
src/layer/layernorm.cpp	Add generic 4D LayerNorm implementation.
src/layer/instancenorm.cpp	Extend InstanceNorm to compute stats over whd.
src/layer/innerproduct.cpp	Include depth in flattened size computation (whd).
src/layer/hardswish.cpp	Extend HardSwish to process dims==4 (size whd).
src/layer/hardsigmoid.cpp	Extend HardSigmoid to process dims==4 (size whd).
src/layer/glu.cpp	Add explicit dims==4 GLU implementations for all axes.
src/layer/exp.cpp	Extend Exp to process dims==4 (size whd).
src/layer/erf.cpp	Extend Erf to process dims==4 (size whd).
src/layer/dropout.cpp	Extend Dropout to process dims==4 (size whd).
src/layer/dequantize.cpp	Extend Dequantize to process dims==4 (size whd).
src/layer/cumulativesum.cpp	Add explicit dims==4 cumulative-sum implementations for all axes.
src/layer/bnll.cpp	Extend BNLL to process dims==4 (size whd).
src/layer/absval.cpp	Extend AbsVal to process dims==4 (size whd).
src/layer/riscv/requantize_riscv.cpp	Extend requantize riscv to allocate/process dims==4 (size whd).
src/layer/riscv/quantize_riscv.cpp	Extend quantize riscv to allocate/process dims==4 (size whd).
src/layer/riscv/quantize_riscv_zfh.cpp	Extend fp16 quantize riscv (zfh) to allocate/process dims==4.
src/layer/riscv/prelu_riscv.cpp	Extend PReLU riscv to support dims==4 (size whd).
src/layer/riscv/prelu_riscv_zfh.cpp	Extend fp16 PReLU riscv (zfh) to support dims==4.
src/layer/riscv/layernorm_riscv.cpp	Add 4D LayerNorm handling on riscv.
src/layer/riscv/layernorm_riscv_zfh.cpp	Add 4D fp16 LayerNorm handling on riscv (zfh).
src/layer/riscv/instancenorm_riscv.cpp	Extend InstanceNorm riscv to compute stats over whd.
src/layer/riscv/instancenorm_riscv_zfh.cpp	Extend fp16 InstanceNorm riscv (zfh) to compute stats over whd.
src/layer/riscv/dequantize_riscv.cpp	Extend dequantize riscv to process dims==4 (size whd).
src/layer/riscv/dequantize_riscv_zfh.cpp	Extend fp16 dequantize riscv (zfh) to allocate/process dims==4.
src/layer/mips/shufflechannel_mips.cpp	Extend ShuffleChannel mips to treat 4D as whd and preserve shape.
src/layer/mips/scale_mips.cpp	Extend Scale mips to support dims==4 (size whd).
src/layer/mips/rmsnorm_mips.cpp	Add 4D RMSNorm handling on mips (fp32 + bf16).
src/layer/mips/requantize_mips.cpp	Extend requantize mips to allocate/process dims==4 (size whd).
src/layer/mips/prelu_mips.cpp	Extend PReLU mips to support dims==4 (size whd).
src/layer/mips/layernorm_mips.cpp	Add 4D LayerNorm handling on mips (fp32 + bf16).
src/layer/mips/instancenorm_mips.cpp	Extend InstanceNorm mips to compute stats over whd.
src/layer/mips/dequantize_mips.cpp	Extend dequantize mips to allocate/process dims==4 and compute size whd.
src/layer/loongarch/scale_loongarch.cpp	Extend Scale loongarch to support dims==4 (size whd).
src/layer/loongarch/rmsnorm_loongarch.cpp	Add 4D RMSNorm handling on loongarch (fp32 + bf16).
src/layer/loongarch/requantize_loongarch.cpp	Extend requantize loongarch to allocate/process dims==4 (size whd).
src/layer/loongarch/prelu_loongarch.cpp	Extend PReLU loongarch to support dims==4 (size whd).
src/layer/loongarch/layernorm_loongarch.cpp	Add 4D LayerNorm handling on loongarch (fp32 + bf16).
src/layer/loongarch/instancenorm_loongarch.cpp	Extend InstanceNorm loongarch to compute stats over whd.
src/layer/loongarch/dequantize_loongarch.cpp	Extend dequantize loongarch to allocate/process dims==4 and compute size whd.
src/layer/arm/shufflechannel_arm.cpp	Extend ShuffleChannel arm to treat 4D as whd and preserve shape.
src/layer/arm/scale_arm.cpp	Extend Scale arm to support dims==4; update fallback condition.
src/layer/arm/rmsnorm_arm.cpp	Add 4D RMSNorm handling on arm (fp32 + bf16).
src/layer/arm/rmsnorm_arm_asimdhp.cpp	Add 4D fp16 RMSNorm handling on arm.
src/layer/arm/requantize_arm.cpp	Extend requantize arm to allocate/process dims==4 (size whd).
src/layer/arm/quantize_arm_asimdhp.cpp	Extend fp16 quantize arm to allocate/process dims==4.
src/layer/arm/prelu_arm.cpp	Extend PReLU arm to support dims==4 (size whd).
src/layer/arm/prelu_arm_asimdhp.cpp	Extend fp16 PReLU arm to support dims==4.
src/layer/arm/layernorm_arm.cpp	Add 4D LayerNorm handling on arm (fp32 + bf16).
src/layer/arm/layernorm_arm_asimdhp.cpp	Add 4D fp16 LayerNorm handling on arm.
src/layer/arm/instancenorm_arm.cpp	Extend InstanceNorm arm to compute stats over whd.
src/layer/arm/instancenorm_arm_asimdhp.cpp	Extend fp16 InstanceNorm arm to compute stats over whd.
src/layer/arm/dequantize_arm.cpp	Extend dequantize arm to allocate/process dims==4 (incl. bf16).
src/layer/arm/dequantize_arm_asimdhp.cpp	Extend fp16 dequantize arm to allocate/process dims==4.

---

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

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep them coming!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".