[OPT] x86: optimize PixelShuffle with SIMD block transpose

[crafcat7]

Summary

Adds an x86-specific implementation of PixelShuffle (src/layer/x86/pixelshuffle_x86.{h,cpp}) that replaces the original scalar stride-store loop with an SSE/AVX block-transpose, yielding a ~3× speedup on pure-PixelShuffle workloads for both single- and multi-thread paths, with no regressions on any other layer or model.

Motivation

The generic PixelShuffle::forward in src/layer/pixelshuffle.cpp uses a stride store
inside the inner-most loop:

for (int j = 0; j < w; j++)
{
    outptr[0] = sptr[0];
    sptr++;
    outptr += upscale_factor; // stride store
}

Each write jumps upscale_factor floats, which defeats auto-vectorization and causes
repeated dirty cache-line traffic. For common super-resolution networks (ESPCN,
RealSR, waifu2x, etc.) the upscale heads are r=2 or r=4, so an SSE/AVX block
transpose followed by contiguous stores is a natural fit.

Algorithm

For each (output_channel p, sub-row sh) we gather r source channel pointers and
walk them in lockstep. Each tile of vec_width × r is transposed and written
contiguously:

• r == 2: AVX unpacklo_ps / unpackhi_ps + permute2f128 → 8 columns per step;
  SSE unpacklo_ps / unpackhi_ps → 4 columns per step.
• r == 4: SSE _MM_TRANSPOSE4_PS → 4 columns per step (writes 16 contiguous floats).
• r == 3 or r > 4: fall back to the base scalar implementation (correct but slow;
  rarely used in practice).
• Both mode=0 (CRD) and mode=1 (DCR) are supported; only the source channel
  index formula differs.

The new layer keeps support_packing = false (inherited from the base), so ncnn's
packing machinery automatically unpacks to pack1 before forward runs; the fast
path stays simple and only handles pack1 + fp32.

Correctness

ctest -R test_pixelshuffle --output-on-failure
Passed

Covers the existing test cases (r=1/2/3/4, mode=0/1, assorted channel counts).

Performance

Environment: Linux / WSL2, AMD Ryzen 7 9800X3D (Zen5, full AVX-512 family), g++,
-O3 -DNDEBUG, AVX/AVX2/FMA/AVX-512 all enabled. Measurements from commit
900ca77 using benchncnn with loop=100 and taskset -c 0-7; reported as
the min metric (most stable at sub-millisecond workloads).

Pure-PixelShuffle workloads

Network	Shape	Threads	Baseline (ms)	Optimized (ms)	Speedup
pixelshuffle_demo	256×256×64 → 1024×1024×4 (r=4)	1	0.85	0.29	2.93×
pixelshuffle_demo	same	8	0.28	0.09	3.11×
pixelshuffle_stacked_demo	two stacked r=4	1	0.97	0.32	3.03×
pixelshuffle_stacked_demo	same	8	0.34	0.10	3.40×

End-to-end super-resolution network (ESPCN topology)

Synthetic ESPCN benchmark/espcn_demo.param:
Input 256×256×3 → Conv5×5 (64) → ReLU → Conv3×3 (32) → ReLU → Conv3×3 (48) → PixelShuffle r=4 → 1024×1024×3.

Threads	Baseline (ms)	Optimized (ms)	Speedup
1	18.66	18.77	≈1× (within noise)
8	10.96	10.12	1.08×

Conv work dominates this topology (≈4.8 GMACs vs. a pure-data-movement
PixelShuffle). The layer-level speedup is real but its absolute magnitude
(~0.7 ms) is small compared to the conv total. The optimization matters most
in networks where PixelShuffle is a larger fraction of the runtime
(upsampling heads, lightweight SR / style transfer, embedded / single-thread
deployments).

No regressions

All 35+ built-in benchmark models (squeezenet, mobilenet v1/v2/v3,
shufflenet v1/v2, googlenet, resnet18/50, vgg16, yolo v3/v4-tiny, nanodet,
efficientnet b0 / v2-b0, vision_transformer, etc.) were measured before/after
at both threads=1 and threads=8. All timings are within noise (±3%). None
of these networks use PixelShuffle.

Files

• src/layer/x86/pixelshuffle_x86.h — new header
• src/layer/x86/pixelshuffle_x86.cpp — new implementation

[codecov-commenter]

Codecov Report

❌ Patch coverage is 98.73418% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 93.96%. Comparing base (71b1a61) to head (266825e).
⚠️ Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layer/x86/pixelshuffle_x86.cpp	98.73%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##           master    #6690    +/-   ##
========================================
  Coverage   93.96%   93.96%            
========================================
  Files         932      933     +1     
  Lines      299059   299668   +609     
========================================
+ Hits       280998   281597   +599     
- Misses      18061    18071    +10     

☔ 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]

Thanks for your contribution !