Alif E8 board --> Run MV2/MV3

[psiddh]

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

[psiddh]

• [cortex_m] Fix linear weight layout: transpose in AOT pass, align meta/ref impl #16782
• Cortex-M backend: Fix quantizer edge-cases #16870

[psiddh]

• Summary: MV2 CortexM PassManager changes for Alif E8 #17009 (MV2 CortexM PassManager changes for Alif E8)

[psiddh]

Very early and raw data

profiling with native MLEK profiler + ARM DWT_CYCCNT cycle counter (not yet with ETDump)

MobileNetV2 INT8 Inference Profiling - Alif E8 DevKit

Setup:

• Model: MobileNetV2 INT8 quantized (fully lowered to CMSIS-NN)
• Hardware: Alif Ensemble E8 DevKit (Cortex-M55 @ 400MHz + Helium MVE)
• Backend: ExecutorTorch + CMSIS-NN kernels

Time Breakdown @ 400MHz:

Operator	Calls	Time (ms)	%
CONV2D	35	409.3ms	53.1%
MEAN	1	14.1ms	1.8%
ADD	10	9.2ms	1.2%
LINEAR	1	5.6ms	0.7%
QUANTIZE	2	3.3ms	0.4%
DEQUANTIZE	2	0.5ms	0.07%
Framework overhead	-	329.3ms	42.7%
TOTAL	-	771.2ms	100%

Key Observations:

1. CONV2D dominates at 53% - expected for MobileNetV2
2. MEAN is surprisingly expensive (14ms) - uses portable code, not CMSIS-NN optimized
3. ReLU6/HARDTANH not shown - fused into CMSIS-NN convolutions via activation clipping
4. ~43% is ExecutorTorch framework overhead ?? (dispatch, tensor metadata, memory ops) (* Need to understand this better )

Next Steps:

• Investigate CONV2D layer performance - see if we can squeeze more out of CMSIS-NN
• Look into framework overhead reduction (??)

[psiddh]

Adding tests for dialect testing for MV2: #17300

[psiddh]

Added Program fix that was decomposing the linear ops to addmm, complicating the cortex m lowering flow : https://github.com/pytorch/executorch/pull/17493/changes

[psiddh]

#17075 --> cleaned up and ready for review