; RP2040 PIO program for accelerating SCSI communication
; Run "pioasm scsi_accel.pio scsi_accel.pio.h" to regenerate the C header from this.
; GPIO mapping:
; - 0-7: DB0-DB7
; -   8: DBP
; Side set is REQ pin

.define REQ 17  ; was 9
.define ACK 26  ; was 10

; Delay from data setup to REQ assertion.
; deskew delay + cable skew delay = 55 ns minimum
; One clock cycle is 8 ns => delay 7 clocks
.define REQ_DLY 7

; Write to SCSI bus using asynchronous handshake.
; Data is written as 16-bit words that contain the 8 data bits + 1 parity bit.
; 7 bits in each word are discarded.
; Number of bytes to send must be multiple of 2.
.program scsi_accel_async_write
    .side_set 1

    pull ifempty block          side 1  ; Get data from TX FIFO
    out pins, 9                 side 1  ; Write data and parity bit
    out null, 7 [REQ_DLY-2]     side 1  ; Discard unused bits, wait for data preset time
    wait 1 gpio ACK             side 1  ; Wait for ACK to be inactive
    wait 0 gpio ACK             side 0  ; Assert REQ, wait for ACK low

; Read from SCSI bus using asynchronous handshake.
; Also works for synchronous mode down to 50 ns transfer period.
; Data is returned as 16-bit words that contain the 8 data bits + 1 parity bit.
; Number of bytes to receive minus 1 should be written to TX fifo.
; Number of bytes to receive must be divisible by 2.
.program scsi_accel_async_read
    .side_set 1

    pull block                  side 1  ; Get number of bytes to receive
    mov x, osr                  side 1  ; Store to counter X

start:
    wait 1 gpio ACK             side 1  ; Wait for ACK high
    wait 0 gpio ACK             side 0  ; Assert REQ, wait for ACK low
    in pins, 9                  side 1  ; Deassert REQ, read GPIO
    in null, 7                  side 1  ; Padding bits
    jmp x-- start               side 1  ; Decrement byte count and jump to start

; Data state machine for synchronous writes.
; Takes the lowest 9 bits of each 16 bit word and writes them to bus with REQ pulse.
; The delay times will be rewritten by C code to match the negotiated SCSI sync speed.
;
; Shifts one bit to ISR per every byte transmitted. This is used to control the transfer
; pace, the RX fifo acts as a counter to keep track of unacknowledged bytes. The C code
; can set the syncOffset by changing autopush threshold, e.g. threshold 3 = 12 bytes offset.
.program scsi_sync_write
    .side_set 1

    out pins, 9     [0]         side 1  ; Write data and parity bit, wait for deskew delay
    out null, 7     [0]         side 0  ; Assert REQ, wait for assert time
    in null, 1      [0]         side 1  ; Deassert REQ, wait for transfer period, wait for space in ACK buffer

; Data pacing state machine for synchronous writes.
; Takes one bit from ISR on every falling edge of ACK.
; The C code should set autopull threshold to match scsi_sync_write autopush threshold.
; System DMA will then move words from scsi_sync_write RX fifo to scsi_sync_write_pacer TX fifo.
.program scsi_sync_write_pacer
    wait 1 gpio ACK
    wait 0 gpio ACK   ; Wait for falling edge on ACK
    out null, 1       ; Let scsi_sync_write send one more byte