kws_cortex_m

KWS-style Cortex-M example

A host runner that exercises the new 2D layers (Conv2D, DepthwiseConv2D, PointwiseConv2D, MaxPool2D, GlobalAvgPool2D) plus the benchmark harness, arranged as a keyword-spotting-style pipeline.

Pipeline

input [20 x 20 x 1]
  -> Conv2D 3x3, 8 filters           -> [18 x 18 x 8]
  -> MaxPool2D 2x2                   -> [9  x 9  x 8]
  -> DepthwiseConv2D 3x3             -> [7  x 7  x 8]
  -> PointwiseConv2D 8 -> 16         -> [7  x 7  x 16]
  -> GlobalAvgPool2D                 -> [16]
  -> PointwiseConv2D (dense) 16 -> 10

Everything is statically allocated. No heap, no RTTI, no exceptions.

Build and run

make release
make run

Sample output (CSV + summary):

name,weight_bytes,activation_bytes,cycles
conv2d_3x3_8,640,10368,...
maxpool2d_2x2,5184,2592,...
...

weight_bytes includes both weights and gradients (Conv2D stores them together for on-device training). For inference-only MCU builds, a future templated variant can drop the gradient array and roughly halve the weight footprint.

Porting to a Cortex-M target

1. Cycle counter. Compile with -DTINYMIND_BENCH_CORTEX_M so bench::readCycleCounter() reads DWT->CYCCNT instead of the host nanosecond fallback. Call bench::enableCycleCounter() once at startup.
2. Output sink. Replace std::cout with a UART wrapper. A minimal UartSink is provided in port_stub.hpp; wire the three functions to your HAL.
3. Input front-end. Replace the synthetic random input with your MFCC extractor. The FFT at cpp/fft1d.hpp plus the sin/cos tables provide the signal-processing pieces.
4. Trained weights. Load via setFilterWeight / setChannelWeight. See examples/pytorch/ for a PyTorch export pattern.

Why these dimensions?

The 20x20 input keeps the example compile-fast and readable. Real KWS models typically use 40x49 MFCC tiles. All dimensions are compile-time template parameters — swap them in the using aliases at the top of kws_cortex_m.cpp and the rest of the pipeline follows.