Strange Timing in SPI

[RudolfAtHome]

I found a strange timing in SPI.cpp and analyzed it:
The "beginTransaction (...)" function takes 67 µs! The following 16-bit SPI transfer then takes 1.09 µs. Why is SPI slowed down so much?
I analyzed the function calls and probably identified the cause:

//---- in \rp2040\hardware\rp2040\1.9.4\libraries\SPI ----_
_void SPIClassRP2040::beginTransaction(SPISettings settings)
utBegin2 = micros(); // My time measure start
spi_init(spi, spis.getClockFreq()); // -> the blocking function
utBegin2 = micros() - utBegin2; // My time measure end -> 67 µs
//---- in \rp2040\hardware\rp2040\1.9.4\pico-sdk\src\rp2_common\hardware_spi\spi.c ----
uint spi_init(spi_inst_t *spi, uint baudrate)
static inline void spi_unreset(spi_inst_t *spi)
unreset_block_wait(spi == spi0 ? RESETS_RESET_SPI0_BITS : RESETS_RESET_SPI1_BITS);
static inline void unreset_block_wait(uint32_t bits) // -> this is the cause (?)

[earlephilhower]

Interesting observation! I imagine the clock startup time could be pretty long.

Maybe we can only call spi_init() when the SPI clock frequency changes, and never disable the SPI block until a SPI.end()! It feels like a simple tweak, do you think you could give it a try? Basically, cache the last clock and options and if they're the same as before we might be able to skip the init. OTW, call the spi deinit/spi reinit and cache these new settings...

[earlephilhower]

See #394 which should fix this and speed up transactions a lot.