git submodule update --init --recursive
export PICO_SDK_PATH=/path/to/your/pico-sdk
mkdir build
cd build
cmake ..
make
More complete info can be extracted from GitHub action script.
A simple app demonstrating the use of PIO and DMA for efficient data transfers to a 480x320 TFT display. The graphics library creates a monochromatic image in memory which is expanded to 16-bit color codes in PIO block and transferred via 8-bit parallel interface.
A simple app controlled by one button, controlling a WS2812 strip/matrix. Can be used to test strips/matrices.
| Button press/sequence | Action |
|---|---|
| Hold | Adjust brightness |
| Tap and hold | Adjust color |
| One tap | Toggle 3-minute timer |
| Two taps | Toggle between 'Off' and 'Max Red' |
| Three taps | Cycle through presets |
Reading audio from a INMP441 MEMS microphone using PIO+DMA, computing fixed point FFT (CMSIS-DSP) and displaying an ASCII spectrogram on serial.
rpi_inmp441_fft_demo.mp4
rpi_inmp441_fft_demo_2.mp4
TFLM Micro Speech model is ported to Raspberry Pi Pico (RP2040, Cortex-M0+). Sound is from a MEMS I2S microphone (INMP441). The CPU is clocked at 250MHz. Data from the microphone is transferred using PIO+DMA. With this configuration real-time speech analysis is possible (stride is 20ms at 16kHz sample rate and single inference takes ~19ms).
rpi_tflm_micro_speech_demo.mp4
TinyML keyword spotting demo. The model is Shallow RNN
architecture with a FastGRNN cell converted
to ~200 lines of C code.
At 280MHz Core0 is setting up DMA transfers from I2S microphone
and doing feature extraction (dc blocking -> preemphasis -> FFT -> power -> log -> Mel filterbank)
in frames, nine frames a second.
The core utilization is ~70%. Core1 every 105ms does NN inference which takes about 85ms (77% core utilization).
The microphone (INMP441) connections are as follows:
tinygl_1b test app (dithered glxgears)
glxgears_1b.mov
tinygl_1b test app - model exported from Blender.
