diff --git a/src/machine/machine_esp32.go b/src/machine/machine_esp32.go
index ed878677..9936f80d 100644
--- a/src/machine/machine_esp32.go
+++ b/src/machine/machine_esp32.go
@@ -4,7 +4,9 @@ package machine
 
 import (
 	"device/esp"
+	"errors"
 	"runtime/volatile"
+	"unsafe"
 )
 
 const peripheralClock = 80000000 // 80MHz
@@ -15,6 +17,10 @@ func CPUFrequency() uint32 {
 	return 160e6 // 160MHz
 }
 
+var (
+	ErrInvalidSPIBus = errors.New("machine: invalid SPI bus")
+)
+
 type PinMode uint8
 
 const (
@@ -26,6 +32,24 @@ const (
 
 // Configure this pin with the given configuration.
 func (p Pin) Configure(config PinConfig) {
+	// Output function 256 is a special value reserved for use as a regular GPIO
+	// pin. Peripherals (SPI etc) can set a custom output function by calling
+	// lowercase configure() instead with a signal name.
+	p.configure(config, 256)
+}
+
+// configure is the same as Configure, but allows for setting a specific input
+// or output signal.
+// Signals are always routed through the GPIO matrix for simplicity. Output
+// signals are configured in FUNCx_OUT_SEL_CFG which selects a particular signal
+// to output on a given pin. Input signals are configured in FUNCy_IN_SEL_CFG,
+// which sets the pin to use for a particular input signal.
+func (p Pin) configure(config PinConfig, signal uint32) {
+	if p == NoPin {
+		// This simplifies pin configuration in peripherals such as SPI.
+		return
+	}
+
 	var muxConfig uint32 // The mux configuration.
 
 	// Configure this pin as a GPIO pin.
@@ -56,6 +80,10 @@ func (p Pin) Configure(config PinConfig) {
 		} else {
 			esp.GPIO.ENABLE1_W1TS.Set(1 << (p - 32))
 		}
+		// Set the signal to read the output value from. It can be a peripheral
+		// output signal, or the special value 256 which indicates regular GPIO
+		// usage.
+		p.outFunc().Set(signal)
 	case PinInput, PinInputPullup, PinInputPulldown:
 		// Clear the 'output enable' bit.
 		if p < 32 {
@@ -63,9 +91,29 @@ func (p Pin) Configure(config PinConfig) {
 		} else {
 			esp.GPIO.ENABLE1_W1TC.Set(1 << (p - 32))
 		}
+		if signal != 256 {
+			// Signal is a peripheral function (not a simple GPIO). Connect this
+			// signal to the pin.
+			// Note that outFunc and inFunc work in the opposite direction.
+			// outFunc configures a pin to use a given output signal, while
+			// inFunc specifies a pin to use to read the signal from.
+			inFunc(signal).Set(esp.GPIO_FUNC_IN_SEL_CFG_SEL | uint32(p)<<esp.GPIO_FUNC_IN_SEL_CFG_IN_SEL_Pos)
+		}
 	}
 }
 
+// outFunc returns the FUNCx_OUT_SEL_CFG register used for configuring the
+// output function selection.
+func (p Pin) outFunc() *volatile.Register32 {
+	return (*volatile.Register32)(unsafe.Pointer((uintptr(unsafe.Pointer(&esp.GPIO.FUNC0_OUT_SEL_CFG)) + uintptr(p)*4)))
+}
+
+// inFunc returns the FUNCy_IN_SEL_CFG register used for configuring the input
+// function selection.
+func inFunc(signal uint32) *volatile.Register32 {
+	return (*volatile.Register32)(unsafe.Pointer((uintptr(unsafe.Pointer(&esp.GPIO.FUNC0_IN_SEL_CFG)) + uintptr(signal)*4)))
+}
+
 // Set the pin to high or low.
 // Warning: only use this on an output pin!
 func (p Pin) Set(value bool) {
@@ -232,3 +280,224 @@ func (uart UART) WriteByte(b byte) error {
 	uart.Bus.TX_FIFO.Set(b)
 	return nil
 }
+
+// Serial Peripheral Interface on the ESP32.
+type SPI struct {
+	Bus *esp.SPI_Type
+}
+
+var (
+	// SPI0 and SPI1 are reserved for use by the caching system etc.
+	SPI2 = SPI{esp.SPI2}
+	SPI3 = SPI{esp.SPI3}
+)
+
+// SPIConfig configures a SPI peripheral on the ESP32. Make sure to set at least
+// SCK, SDO and SDI (possibly to NoPin if not in use). The default for LSBFirst
+// (false) and Mode (0) are good for most applications. The frequency defaults
+// to 1MHz if not set but can be configured up to 40MHz. Possible values are
+// 40MHz and integer divisions from 40MHz such as 20MHz, 13.3MHz, 10MHz, 8MHz,
+// etc.
+type SPIConfig struct {
+	Frequency uint32
+	SCK       Pin
+	SDO       Pin
+	SDI       Pin
+	LSBFirst  bool
+	Mode      uint8
+}
+
+// Configure and make the SPI peripheral ready to use.
+func (spi SPI) Configure(config SPIConfig) error {
+	if config.Frequency == 0 {
+		config.Frequency = 1e6 // default to 1MHz
+	}
+
+	// Configure the SPI clock. This assumes a peripheral clock of 80MHz.
+	var clockReg uint32
+	if config.Frequency >= 40e6 {
+		// Don't use a prescaler, but directly connect to the APB clock. This
+		// results in a SPI clock frequency of 40MHz.
+		clockReg |= esp.SPI_CLOCK_CLK_EQU_SYSCLK
+	} else {
+		// Use a prescaler for frequencies below 40MHz. They will get rounded
+		// down to the next possible frequency (20MHz, 13.3MHz, 10MHz, 8MHz,
+		// 6.7MHz, 5.7MHz, 5MHz, etc).
+		// This code is much simpler than how ESP-IDF configures the frequency,
+		// but should be just as accurate. The only exception is for frequencies
+		// below 4883Hz, which will need special support.
+		if config.Frequency < 4883 {
+			// The current lower limit is 4883Hz.
+			// The hardware supports lower frequencies by setting the h and n
+			// variables, but that's not yet implemented.
+			config.Frequency = 4883
+		}
+		// The prescaler value is 40e6 / config.Frequency, but rounded up so
+		// that the actual frequency is never higher than the frequency
+		// requested in config.Frequency.
+		var (
+			pre uint32 = (40e6 + config.Frequency - 1) / config.Frequency
+			n   uint32 = 2 // this value seems to equal the number of ticks per SPI clock tick
+			h   uint32 = 1 // must be half of n according to the formula in the reference manual
+			l   uint32 = n // must equal n according to the reference manual
+		)
+		clockReg |= (pre - 1) << esp.SPI_CLOCK_CLKDIV_PRE_Pos
+		clockReg |= (n - 1) << esp.SPI_CLOCK_CLKCNT_N_Pos
+		clockReg |= (h - 1) << esp.SPI_CLOCK_CLKCNT_H_Pos
+		clockReg |= (l - 1) << esp.SPI_CLOCK_CLKCNT_L_Pos
+	}
+	spi.Bus.CLOCK.Set(clockReg)
+
+	// SPI_CTRL_REG controls bit order.
+	var ctrlReg uint32
+	if config.LSBFirst {
+		ctrlReg |= esp.SPI_CTRL_WR_BIT_ORDER
+		ctrlReg |= esp.SPI_CTRL_RD_BIT_ORDER
+	}
+	spi.Bus.CTRL.Set(ctrlReg)
+
+	// SPI_CTRL2_REG, SPI_USER_REG and SPI_PIN_REG control SPI clock polarity
+	// (mode), among others.
+	var ctrl2Reg, userReg, pinReg uint32
+	// For mode configuration, see table 29 in the reference manual (page 128).
+	var delayMode uint32
+	switch config.Mode {
+	case 0:
+		delayMode = 2
+	case 1:
+		delayMode = 1
+		userReg |= esp.SPI_USER_CK_OUT_EDGE
+	case 2:
+		delayMode = 1
+		userReg |= esp.SPI_USER_CK_OUT_EDGE
+		pinReg |= esp.SPI_PIN_CK_IDLE_EDGE
+	case 3:
+		delayMode = 2
+		pinReg |= esp.SPI_PIN_CK_IDLE_EDGE
+	}
+	// Extra configuration necessary for correct data input at high frequencies.
+	// This is only necessary when MISO goes through the GPIO matrix (which it
+	// currently does).
+	if config.Frequency >= 40e6 {
+		// Delay mode must be set to 0 and SPI_USR_DUMMY_CYCLELEN should be set
+		// to 0 (the default).
+		userReg |= esp.SPI_USER_USR_DUMMY
+	} else if config.Frequency >= 20e6 {
+		// Nothing to do here, delay mode should be set to 0 according to the
+		// datasheet.
+	} else {
+		// Follow the delay mode as given in table 29 on page 128 of the
+		// reference manual.
+		// Note that this is only specified for SPI frequency of 10MHz and
+		// below (≤Fapb/8), so 13.3MHz appears to be left unspecified.
+		ctrl2Reg |= delayMode << esp.SPI_CTRL2_MOSI_DELAY_MODE_Pos
+	}
+	// Enable full-duplex communication.
+	userReg |= esp.SPI_USER_DOUTDIN
+	userReg |= esp.SPI_USER_USR_MOSI
+	// Write values to registers.
+	spi.Bus.CTRL2.Set(ctrl2Reg)
+	spi.Bus.USER.Set(userReg)
+	spi.Bus.PIN.Set(pinReg)
+
+	// Configure pins.
+	// TODO: use direct output if possible, if the configured pins match the
+	// possible direct configurations (e.g. for SPI2, when SCK is pin 14 etc).
+	if spi.Bus == esp.SPI2 {
+		config.SCK.configure(PinConfig{Mode: PinOutput}, 8)  // HSPICLK
+		config.SDI.configure(PinConfig{Mode: PinInput}, 9)   // HSPIQ
+		config.SDO.configure(PinConfig{Mode: PinOutput}, 10) // HSPID
+	} else if spi.Bus == esp.SPI3 {
+		config.SCK.configure(PinConfig{Mode: PinOutput}, 63) // VSPICLK
+		config.SDI.configure(PinConfig{Mode: PinInput}, 64)  // VSPIQ
+		config.SDO.configure(PinConfig{Mode: PinOutput}, 65) // VSPID
+	} else {
+		// Don't know how to configure this bus.
+		return ErrInvalidSPIBus
+	}
+
+	return nil
+}
+
+// Transfer writes/reads a single byte using the SPI interface. If you need to
+// transfer larger amounts of data, Tx will be faster.
+func (spi SPI) Transfer(w byte) (byte, error) {
+	spi.Bus.MISO_DLEN.Set(7 << esp.SPI_MISO_DLEN_USR_MISO_DBITLEN_Pos)
+	spi.Bus.MOSI_DLEN.Set(7 << esp.SPI_MOSI_DLEN_USR_MOSI_DBITLEN_Pos)
+
+	spi.Bus.W0.Set(uint32(w))
+
+	// Send/receive byte.
+	spi.Bus.CMD.Set(esp.SPI_CMD_USR)
+	for spi.Bus.CMD.Get() != 0 {
+	}
+
+	// The received byte is stored in W0.
+	return byte(spi.Bus.W0.Get()), nil
+}
+
+// Tx handles read/write operation for SPI interface. Since SPI is a syncronous write/read
+// interface, there must always be the same number of bytes written as bytes read.
+// This is accomplished by sending zero bits if r is bigger than w or discarding
+// the incoming data if w is bigger than r.
+//
+func (spi SPI) Tx(w, r []byte) error {
+	toTransfer := len(w)
+	if len(r) > toTransfer {
+		toTransfer = len(r)
+	}
+
+	for toTransfer != 0 {
+		// Do only 64 bytes at a time.
+		chunkSize := toTransfer
+		if chunkSize > 64 {
+			chunkSize = 64
+		}
+
+		// Fill tx buffer.
+		transferWords := (*[16]volatile.Register32)(unsafe.Pointer(uintptr(unsafe.Pointer(&spi.Bus.W0))))
+		var outBuf [16]uint32
+		txSize := 64
+		if txSize > len(w) {
+			txSize = len(w)
+		}
+		for i := 0; i < txSize; i++ {
+			outBuf[i/4] = outBuf[i/4] | uint32(w[i])<<((i%4)*8)
+		}
+		for i, word := range outBuf {
+			transferWords[i].Set(word)
+		}
+
+		// Do the transfer.
+		spi.Bus.MISO_DLEN.Set((uint32(chunkSize)*8 - 1) << esp.SPI_MISO_DLEN_USR_MISO_DBITLEN_Pos)
+		spi.Bus.MOSI_DLEN.Set((uint32(chunkSize)*8 - 1) << esp.SPI_MOSI_DLEN_USR_MOSI_DBITLEN_Pos)
+		spi.Bus.CMD.Set(esp.SPI_CMD_USR)
+		for spi.Bus.CMD.Get() != 0 {
+		}
+
+		// Read rx buffer.
+		rxSize := 64
+		if rxSize > len(r) {
+			rxSize = len(r)
+		}
+		for i := 0; i < rxSize; i++ {
+			r[i] = byte(transferWords[i/4].Get() >> ((i % 4) * 8))
+		}
+
+		// Cut off some part of the output buffer so the next iteration we will
+		// only send the remaining bytes.
+		if len(w) < chunkSize {
+			w = nil
+		} else {
+			w = w[chunkSize:]
+		}
+		if len(r) < chunkSize {
+			r = nil
+		} else {
+			r = r[chunkSize:]
+		}
+		toTransfer -= chunkSize
+	}
+
+	return nil
+}