Controller reading code: Difference between revisions

From NESdev Wiki
Jump to navigationJump to search
m (→‎Calculating Presses and Releases: a newline probably makes this more legible)
(→‎DPCM Safety: Clarifies that more than 1 bit may be deleted (Famicoms delete up to 3). Makes a less-sweeping claim about modern emulators supporting synced reads; for example, puNES currently deletes bits but has bad DMA timings that break sync.)
 
(7 intermediate revisions by 2 users not shown)
Line 139: Line 139:
</pre>
</pre>


=== DPCM Safety using Repeated Reads ===
=== DPCM Safety ===
-----


If your code is intended to be used with [[APU DMC]] playback, this code will need to be altered. The NES occasionally glitches the controller port twice in a row if sample playback is enabled, and games using samples need to work around this. The most common technique, as seen in ''Super Mario Bros. 3''<ref>[//forums.nesdev.org/viewtopic.php?p=151720#p151720 Super Mario Bros. 3 controller reread method]</ref> and other games, will read each controller twice, but in the event of a mismatch it will keep re-reading the controller until two results in a row are the same.
The NTSC CPU has a bug where the [[DMA]] that fetches DPCM sample data for the [[APU DMC|APU's DMC channel]] can cause extra reads of the joypads. This is fixed on PAL. For standard controllers, this bug results in at least one bit being deleted from the controller's serial report, causing the following buttons to be shifted over and usually resulting in the right direction appearing to be pressed. Software cannot directly detect that this bug has happened, so controller reading code in games using sampled audio has to be resilient against it. There are currently two approaches for dealing with this problem, each with significant pros and cons.
 
Note that both approaches share some limitations:
* They may become unsafe if interrupted by NMI or IRQ.
* Poor support across most emulators can make it hard to know if the code works correctly on hardware.
* Using third party controllers may make testing harder: official standard controllers shift in 1's on extra reads that cause obvious errant right presses, while third party controllers may shift in 0's.
* It is possible for joypad state to corrupt even if the controller reading code works as intended. Any read from $4000-401F provides an opportunity for DMC DMA to trigger a joypad read, and this can cause controllers with state that is slow or impossible to recover (e.g. [[Arkanoid controller]]) to lose that state before the controllers are intentionally read. Some sound engines cause such reads by doing indexed stores to sound registers. For controllers with normal strobe behavior where the full current state is immediately loaded on strobe, this is not a problem.
 
Neither approach is a one-size-fits-all solution; there are cases where one or the other solution does not work.
 
==== DPCM Safety using Repeated Reads ====
 
The most common technique, as seen in ''Super Mario Bros. 3''<ref>[//forums.nesdev.org/viewtopic.php?p=151720#p151720 Super Mario Bros. 3 controller reread method]</ref> and other games, will read each controller twice, but in the event of a mismatch it will keep re-reading the controller until two results in a row are the same.
<pre>
<pre>
readjoy2_safe:
readjoy2_safe:
Line 161: Line 174:
</pre>
</pre>


Note that the time between the start of one read and the end of the next read must be less than the length of the fastest DMC fetch period (432 cycles). For this reason, it is normal to read controllers one at a time with this method, rather than attempting both at once. (Note: <code>readjoy2</code> above takes too long to be suitable.) [[Game bugs|''Gimmick!'' has such a bug]] resulting from trying to read both at once.
Note that the time between the start of one read and the end of the next read must be less than the length of the fastest DMC fetch period minus 4 (428 cycles). This is because multiple bit deletions could cause both passes to result in the same incorrect data. For this reason, it is normal to read controllers one at a time with this method, rather than attempting both at once. (Note: <code>readjoy2</code> above takes too long to be suitable.) [[Game bugs|''Gimmick!'' has such a bug]] resulting from trying to read both at once.


Most often a controller will be read 2 times, and 3 or 4 in the case of a DPCM corruption, or the player pressing a button during the read. With the assistance of tools, a malicious controller input could change the buttons on every read, holding it in this reread loop indefinitely<ref>[https://arstechnica.com/gaming/2016/07/how-to-beat-super-mario-bros-3-in-less-than-a-second/ Ars Technica: How to beat Super Mario Bros. 3 in less than a second]</ref><ref>https://tasvideos.org/6466S Tool-Assisted Speedrun of Super Mario Bros. 3 which first demonstrated abusing controller reread routines.</ref>, but this is generally not an important edge case to account for.
Most often a controller will be read 2 times, and 3 or 4 in the case of a DPCM corruption, or the player pressing a button during the read. With the assistance of tools, a malicious controller input could change the buttons on every read, holding it in this reread loop indefinitely<ref>[https://arstechnica.com/gaming/2016/07/how-to-beat-super-mario-bros-3-in-less-than-a-second/ Ars Technica: How to beat Super Mario Bros. 3 in less than a second]</ref><ref>https://tasvideos.org/6466S Tool-Assisted Speedrun of Super Mario Bros. 3 which first demonstrated abusing controller reread routines.</ref>, but this is generally not an important edge case to account for.


=== DPCM Safety using OAM DMA ===
This approach to DPCM safety has these benefits:
* The code is easy to understand and modify.
* It is position-independent: branching across page boundaries does not inherently break it, and it can be done at any time in the frame.
* It works in all emulators, even if they don't implement the bug correctly or at all.
It also has these drawbacks:
* Repeated reads cannot be used for controllers that lose state on reads (e.g. [[SNES mouse]]), take a long time before data is available to read again (e.g. [[Arkanoid controller]]), or have too much data to read twice within the cycle budget (e.g. [[Famicom Network Controller]]).
* It can be hard to know if the code is safe:
** The budget limitations are not obvious because of obscure edge cases with the bug. All joypad reads for 2 complete passes must finish within one DMC period minus 4 (usually 428 cycles). Reading both controllers in a single pass takes too long; they must be handled individually.
** Hardware differences can make real hardware testing difficult: the bug behaves one way on the AV Famicom and NES, and another way on other Famicom models. The Famicom behavior may mitigate the bug when using repeated reads.
** Even most accurate emulators don't emulate all the ways repeated reads can fail. Mesen v2 does and is recommended.
* This approach does not prevent the bug, it merely works around it, so other negative side effects of the bug still happen:
** The sample byte read by the APU can be corrupted when the bug occurs. However, it is difficult to hear this corruption.
** The APU's frame IRQ may be lost when the bug occurs. However, this feature is very rarely used.
* It makes controller reads take at least twice as long and does not guarantee a maximum runtime. Normally, it will take up to 4 passes.


Because halts for DPCM fetches normally only occur on an put cycles, it is possible to get glitch-free controller reads by timing all '''$4016''' and '''$4017''' reads to fall on get cycles. This is made possible by the behavior of OAM DMA: the first cycle after an OAM DMA is normally guaranteed to be a get cycle.<ref>[http://forums.nesdev.org/viewtopic.php?p=171971 Forum post:] Rahsennor's OAM-synchronized controller read</ref> This is a relatively new technique and is not supported by some emulators.<ref>[http://forums.nesdev.org/viewtopic.php?f=2&t=14319&start=15#p172099 Forum post:] as of May 2016, Nintendulator and Nestopia do not accurately emulate OAM-synchronized controller reading.</ref> In the following example code, the controller1 and controller2 labels must be in zeropage for the timing to work.
==== DPCM Safety using OAM DMA ====


This solution might be ideal, using fewer cycles than doing repeated reads, and taking a consistent amount of time. It can also be adapted for DPCM-conflict free reading of controllers that require a longer report like [[four player adapters]], or the [[SNES mouse]] (see notes below).
Because halts for DPCM fetches have very specific timing, it is possible to get glitch-free controller reads by timing all '''$4016''' and '''$4017''' reads to fall on one cycle parity. This is because the CPU alternates between two kinds of cycles: 'get' and 'put'. DPCM fetches normally occur only on put cycles, and because the first cycle after OAM DMA is normally guaranteed to be a get cycle, timed code can use OAM DMA to ensure the joypad reads land on get cycles.
<ref>[http://forums.nesdev.org/viewtopic.php?p=171971 Forum post:] Rahsennor's OAM-synchronized controller read</ref> This is a relatively new technique and is not supported by several older emulators<ref>[http://forums.nesdev.org/viewtopic.php?f=2&t=14319&start=15#p172099 Forum post:] as of May 2016, Nintendulator and Nestopia do not accurately emulate OAM-synchronized controller reading.</ref>, but generally works in currently-supported emulators. In the following example code, the controller1 and controller2 labels must be in zeropage for the timing to work.


However, unlike the other solutions above, it can't be executed at any time. Instead it must be integrated into your NMI routine to coincide with your sprite OAM DMA once per frame.
This solution might be ideal, using fewer cycles than doing repeated reads, and taking a consistent amount of time. It can also be adapted for DPCM-conflict free reading of controllers that require a longer report, like [[four player adapters]] or the [[SNES mouse]] (see notes below).
 
However, unlike the repeated reads solution, it can't be executed at any time. Instead it must be integrated into your NMI routine to coincide with your sprite OAM DMA once per frame.


<pre>
<pre>
Line 185: Line 214:
     and #3            ; put get
     and #3            ; put get
     cmp #1            ; put get
     cmp #1            ; put get
     rol buttons+1, x  ; put get put get put get (X = 0; waste 1 cycle and 0 bytes for alignment)
     rol buttons+1, x  ; put get put get put get (X = 0; waste 1 cycle for alignment)
     lda $4016          ; put get put GET
     lda $4016          ; put get put GET
     and #3            ; put get
     and #3            ; put get
Line 195: Line 224:
Note that this example routine only reads two 8-bit controllers and does not take enough time to span more than one DPCM fetch. Routines longer than this must contend with two additional constraints:
Note that this example routine only reads two 8-bit controllers and does not take enough time to span more than one DPCM fetch. Routines longer than this must contend with two additional constraints:
* When DMC DMA is delayed by an odd number of cycles, it takes 3 cycles instead of 4, changing the cycle parity. If extending this function to read more bits, care must be taken so that ''all'' CPU write cycles are aligned. Instructions with a single write cycle must align the write to avoid conflict with the DPCM fetch, and double-write instructions like ROL need to align both writes so that the DPCM fetch falls on the first write.<ref>[//forums.nesdev.org/viewtopic.php?p=231604#p231604 Forum post:] demonstration of how ROL instruction affects alignment for OAM DMA synchronized controller reading.</ref> If an interrupt can occur during the routine, it must be aligned so the fetch can only fall on the second of the three automatic stack writes.
* When DMC DMA is delayed by an odd number of cycles, it takes 3 cycles instead of 4, changing the cycle parity. If extending this function to read more bits, care must be taken so that ''all'' CPU write cycles are aligned. Instructions with a single write cycle must align the write to avoid conflict with the DPCM fetch, and double-write instructions like ROL need to align both writes so that the DPCM fetch falls on the first write.<ref>[//forums.nesdev.org/viewtopic.php?p=231604#p231604 Forum post:] demonstration of how ROL instruction affects alignment for OAM DMA synchronized controller reading.</ref> If an interrupt can occur during the routine, it must be aligned so the fetch can only fall on the second of the three automatic stack writes.
* When DMC DMA occurs near the end of OAM DMA, it only steals 1 or 3 cycles, inverting the cycle parity. Every DMC period after that, a misaligned DPCM fetch will occur. Care must be taken to ensure this does not land on a joypad read.
* When DMC DMA occurs near the end of OAM DMA, it steals an odd number of cycles, inverting the cycle parity. Every DMC period after that, a misaligned DPCM fetch will occur. Care must be taken to ensure this does not land on a joypad read. For longer functions, these misaligned cases can be resynced by landing the DPCM fetch on a single write cycle.


See [[DMA]] for detailed information on DMA timing.
See [[DMA]] for detailed information on DMA timing. However, game developers writing synced code may find it easier to consider just the specific case of DMC DMA used for reloading the sample buffer. These timings do not apply to the initial DMC DMA load caused by writing to $4015, and they are not sufficient for emulating DMA timings; emulator authors should refer to [[DMA]] for comprehensive information. A DMC DMA reload takes these cycle counts in these cases:
* 4 cycles if it falls on a CPU read cycle
* 3 cycles if it falls on an odd number of CPU write cycles
* 4 cycles if it falls on an even number of CPU write cycles
* An even number of cycles if it occurs during OAM DMA
* An odd number of cycles if it occurs at the end of OAM DMA (changing cycle parity for the following code)
 
This approach to DPCM safety has these benefits:
* It can support any controller because it only does one pass and has no cycle budget.
* It is nearly as fast as standard reads.
* It avoids the bug and thus does not corrupt the sample byte or interfere with the APU's frame IRQ.
It also has these drawbacks:
* It uses timed code and relies on complicated DMA behavior, so it is harder to understand and modify.
* Moving the code can break it by causing branches to cross a page boundary, changing the timing.
* It must directly follow OAM DMA. Unusual features that skip OAM DMA on gameplay frames are not compatible with this approach. (On lag frames, normally both OAM DMA and controller reads are skipped. This is not a problem for synced reads.)
* It does not work in emulators that emulate the bug but have incorrect DMA timing (e.g. Nestopia). Many modern emulators either don't emulate the bug or have correct-enough DMA timing.
* Synced code that lasts longer than one DMC period minus 4 (usually 428 cycles) is difficult to write correctly. This is because write cycles must also be synced and specific desync cases occur if DMC DMA lands at the end of OAM DMA. However, this is the only approach that supports code of this length.


== Other Useful Operations ==
== Other Useful Operations ==
=== Button Flags ===
It is helpful to define the buttons as a series of bit flags:
<pre>
BUTTON_A      = 1 << 7
BUTTON_B      = 1 << 6
BUTTON_SELECT = 1 << 5
BUTTON_START  = 1 << 4
BUTTON_UP    = 1 << 3
BUTTON_DOWN  = 1 << 2
BUTTON_LEFT  = 1 << 1
BUTTON_RIGHT  = 1 << 0
</pre>
And then buttons can be checked as follows:
<pre>
    lda buttons
    and #BUTTON_A | BUTTON_B
    beq notPressingAorB
    ; Handle presses.
notPressingAorB:
</pre>
=== Calculating Presses and Releases ===
To calculate newly pressed and newly released buttons by comparing against the last frame's buttons:
<pre>
    ; newly pressed buttons: not held last frame, and held now
    lda last_frame_buttons, x
    eor #%11111111
    and buttons, x
    sta pressed_buttons, x
    ; newly released buttons: not held now, and held last frame
    lda buttons, x
    eor #%11111111
    and last_frame_buttons, x
    sta released_buttons, x
</pre>


=== Directional Safety ===
=== Directional Safety ===


To reject opposing presses (Up+Down and Left+Right), which are possible on a worn Control Pad:
Opposing directions (Up+Down and Left+Right) are possible on non-standard controllers, or even on a worn out standard one. Both directions can be tested for at the same time in an efficient bitwise operation:
<pre>
<pre>
     lda buttons, x
     lda buttons, x
Line 209: Line 294:
     lsr a
     lsr a
     and buttons, x    ; to Down and Right
     and buttons, x    ; to Down and Right
     beq not_updown
     beq not_opposing
         ; Use previous frame's directions
         ; Use previous frame's directions
         lda buttons, x
         lda buttons, x
Line 216: Line 301:
         eor last_frame_buttons, x
         eor last_frame_buttons, x
         sta buttons, x
         sta buttons, x
     not_updown:
     not_opposing:
</pre>
</pre>


To instead reject all diagonal presses, simulating a 4-way joystick:
Diagonal directions can also be detected at the same time as opposing ones, by testing if more than 1 directional bit is set. This could be used as part of a 4-way joystick style control scheme:
<pre>
<pre>
     lda buttons, x
     lda buttons, x
Line 237: Line 322:
</pre>
</pre>


=== Calculating Presses and Releases ===
As an example, the code above rejects the detected cases by using the direction bits from the previous frame, but depending on the situation another response might be more appropriate (e.g. cancel just the opposing bits, or clear all directions).
 
To calculate newly pressed and newly released buttons:
<pre>
    ; newly pressed buttons: not held last frame, and held now
    lda last_frame_buttons, x
    eor #%11111111
    and buttons, x
    sta pressed_buttons, x
 
    ; newly released buttons: not held now, and held last frame
    lda buttons, x
    eor #%11111111
    and last_frame_buttons, x
    sta released_buttons, x
</pre>
 
=== Button Flags ===
 
It is helpful to define the buttons as a series of bit flags:
<pre>
BUTTON_A      = 1 << 7
BUTTON_B      = 1 << 6
BUTTON_SELECT = 1 << 5
BUTTON_START  = 1 << 4
BUTTON_UP    = 1 << 3
BUTTON_DOWN  = 1 << 2
BUTTON_LEFT  = 1 << 1
BUTTON_RIGHT  = 1 << 0
</pre>
 
And then buttons can be checked as follows:
<pre>
    lda buttons
    and #BUTTON_A | BUTTON_B
    beq notPressingAorB
    ; Handle presses.
notPressingAorB:
</pre>


== External Examples ==
== External Examples ==

Latest revision as of 07:16, 23 August 2024

This page contains example code for reading the NES controller.

See also: Controller reading

Controller Reading Code

Basic Example

This is a tutorial example of the bare minimum needed to read the controller. It will explain the basic principles in detail, but once understood, you may wish to continue to the Standard Read example that follows, as a more complete and ready-to-use code example.

This code describes an efficient method of reading the standard controller using ca65 syntax.

The result byte buttons should be placed in zero page to save a cycle each time through the loop.

; we reserve one byte for storing the data that is read from controller
.zeropage
buttons .res 1

When reading from JOYPAD* what is read might be different from $01/$00 for various reasons. (See Controller reading.) In this code the only concern is bit 0 read from JOYPAD*..

JOYPAD1 = $4016
JOYPAD2 = $4017

This is the end result that will be stored in buttons. 1 if the button was pressed, 0 otherwise.

bit     7         6         5         4         3         2         1         0    
button A B Select Start Up Down Left Right

This subroutine takes 132 cycles to execute but ignores the Famicom expansion controller. Many controller reading subroutines use the X or Y register to count 8 times through the loop. But this one uses a more clever ring counter technique: $01 is loaded into the result first, and once eight bits are shifted in, the 1 bit will be shifted out, terminating the loop.

; At the same time that we strobe bit 0, we initialize the ring counter
; so we're hitting two birds with one stone here
readjoy:
    lda #$01
    ; While the strobe bit is set, buttons will be continuously reloaded.
    ; This means that reading from JOYPAD1 will only return the state of the
    ; first button: button A.
    sta JOYPAD1
    sta buttons
    lsr a        ; now A is 0
    ; By storing 0 into JOYPAD1, the strobe bit is cleared and the reloading stops.
    ; This allows all 8 buttons (newly reloaded) to be read from JOYPAD1.
    sta JOYPAD1
loop:
    lda JOYPAD1
    lsr a        ; bit 0 -> Carry
    rol buttons  ; Carry -> bit 0; bit 7 -> Carry
    bcc loop
    rts

Continue to the next example for a more complete read routine that handles both controllers and the standard Famicom expansion controllers.

Standard Read for 2 Controllers and Famicom

Adding support for controllers on the Famicom's DA15 expansion port and for player 2's controller is straightforward. Something similar to the following routine is used in most Famicom games. Even though the expansion port is unused on the NES, the unconnected bit will read as 0, so this solution works safely with both Famicom and NES hardware.

.zeropage
buttons: .res 2     ; space for 2 reads

.code
readjoyx2:
    ldx #$00
    jsr readjoyx    ; X=0: read controller 1
    inx
    ; fall through to readjoyx below, X=1: read controller 2

readjoyx:           ; X register = 0 for controller 1, 1 for controller 2
    lda #$01
    sta JOYPAD1
    sta buttons, x
    lsr a
    sta JOYPAD1
loop:
    lda JOYPAD1, x
    and #%00000011  ; ignore bits other than controller
    cmp #$01        ; Set carry if and only if nonzero
    rol buttons, x  ; Carry -> bit 0; but 7 -> Carry
    bcc loop
    rts

If playing DPCM samples, there is an additional reread step to prevent errors (see below).

Note that the and to ignore bits is not optional, as the upper bits of JOYPAD1 read are not guaranteed to be 1 or 0[1].

Alternative 2 Controllers Read

Alternatively, we could combine both controller reads into 1 loop with a single strobe, though this routine is not safe to use with DPCM samples playing (see below).

.zeropage
buttons: .res 2

.code
readjoy2:
    lda #$01
    sta JOYPAD1
    sta buttons+1 ; player 2's buttons double as a ring counter
    lsr a
    sta JOYPAD1
loop:
    lda JOYPAD1
    and #%00000011
    cmp #$01
    rol buttons+0
    lda JOYPAD2
    and #%00000011
    cmp #$01
    rol buttons+1
    bcc loop
    rts

DPCM Safety


The NTSC CPU has a bug where the DMA that fetches DPCM sample data for the APU's DMC channel can cause extra reads of the joypads. This is fixed on PAL. For standard controllers, this bug results in at least one bit being deleted from the controller's serial report, causing the following buttons to be shifted over and usually resulting in the right direction appearing to be pressed. Software cannot directly detect that this bug has happened, so controller reading code in games using sampled audio has to be resilient against it. There are currently two approaches for dealing with this problem, each with significant pros and cons.

Note that both approaches share some limitations:

  • They may become unsafe if interrupted by NMI or IRQ.
  • Poor support across most emulators can make it hard to know if the code works correctly on hardware.
  • Using third party controllers may make testing harder: official standard controllers shift in 1's on extra reads that cause obvious errant right presses, while third party controllers may shift in 0's.
  • It is possible for joypad state to corrupt even if the controller reading code works as intended. Any read from $4000-401F provides an opportunity for DMC DMA to trigger a joypad read, and this can cause controllers with state that is slow or impossible to recover (e.g. Arkanoid controller) to lose that state before the controllers are intentionally read. Some sound engines cause such reads by doing indexed stores to sound registers. For controllers with normal strobe behavior where the full current state is immediately loaded on strobe, this is not a problem.

Neither approach is a one-size-fits-all solution; there are cases where one or the other solution does not work.

DPCM Safety using Repeated Reads

The most common technique, as seen in Super Mario Bros. 3[2] and other games, will read each controller twice, but in the event of a mismatch it will keep re-reading the controller until two results in a row are the same.

readjoy2_safe:
    ldx #$00
    jsr readjoyx_safe  ; X=0: safe read controller 1
    inx
    ; fall through to readjoyx_safe, X=1: safe read controller 2

readjoyx_safe:
    jsr readjoyx
reread:
    lda buttons, x
    pha
    jsr readjoyx
    pla
    cmp buttons, x
    bne reread
    rts

Note that the time between the start of one read and the end of the next read must be less than the length of the fastest DMC fetch period minus 4 (428 cycles). This is because multiple bit deletions could cause both passes to result in the same incorrect data. For this reason, it is normal to read controllers one at a time with this method, rather than attempting both at once. (Note: readjoy2 above takes too long to be suitable.) Gimmick! has such a bug resulting from trying to read both at once.

Most often a controller will be read 2 times, and 3 or 4 in the case of a DPCM corruption, or the player pressing a button during the read. With the assistance of tools, a malicious controller input could change the buttons on every read, holding it in this reread loop indefinitely[3][4], but this is generally not an important edge case to account for.

This approach to DPCM safety has these benefits:

  • The code is easy to understand and modify.
  • It is position-independent: branching across page boundaries does not inherently break it, and it can be done at any time in the frame.
  • It works in all emulators, even if they don't implement the bug correctly or at all.

It also has these drawbacks:

  • Repeated reads cannot be used for controllers that lose state on reads (e.g. SNES mouse), take a long time before data is available to read again (e.g. Arkanoid controller), or have too much data to read twice within the cycle budget (e.g. Famicom Network Controller).
  • It can be hard to know if the code is safe:
    • The budget limitations are not obvious because of obscure edge cases with the bug. All joypad reads for 2 complete passes must finish within one DMC period minus 4 (usually 428 cycles). Reading both controllers in a single pass takes too long; they must be handled individually.
    • Hardware differences can make real hardware testing difficult: the bug behaves one way on the AV Famicom and NES, and another way on other Famicom models. The Famicom behavior may mitigate the bug when using repeated reads.
    • Even most accurate emulators don't emulate all the ways repeated reads can fail. Mesen v2 does and is recommended.
  • This approach does not prevent the bug, it merely works around it, so other negative side effects of the bug still happen:
    • The sample byte read by the APU can be corrupted when the bug occurs. However, it is difficult to hear this corruption.
    • The APU's frame IRQ may be lost when the bug occurs. However, this feature is very rarely used.
  • It makes controller reads take at least twice as long and does not guarantee a maximum runtime. Normally, it will take up to 4 passes.

DPCM Safety using OAM DMA

Because halts for DPCM fetches have very specific timing, it is possible to get glitch-free controller reads by timing all $4016 and $4017 reads to fall on one cycle parity. This is because the CPU alternates between two kinds of cycles: 'get' and 'put'. DPCM fetches normally occur only on put cycles, and because the first cycle after OAM DMA is normally guaranteed to be a get cycle, timed code can use OAM DMA to ensure the joypad reads land on get cycles. [5] This is a relatively new technique and is not supported by several older emulators[6], but generally works in currently-supported emulators. In the following example code, the controller1 and controller2 labels must be in zeropage for the timing to work.

This solution might be ideal, using fewer cycles than doing repeated reads, and taking a consistent amount of time. It can also be adapted for DPCM-conflict free reading of controllers that require a longer report, like four player adapters or the SNES mouse (see notes below).

However, unlike the repeated reads solution, it can't be executed at any time. Instead it must be integrated into your NMI routine to coincide with your sprite OAM DMA once per frame.

    lda #OAM
    sta $4014          ; ------ OAM DMA ------
    ldx #1             ; get put          <- strobe code must take an odd number of cycles total
    stx buttons+0      ; get put get      <- buttons must be in the zeropage
    stx $4016          ; put get put get
    dex                ; put get
    stx $4016          ; put get put get
loop:
    lda $4017          ; put get put GET  <- loop code must take an even number of cycles total
    and #3             ; put get
    cmp #1             ; put get
    rol buttons+1, x   ; put get put get put get (X = 0; waste 1 cycle for alignment)
    lda $4016          ; put get put GET
    and #3             ; put get
    cmp #1             ; put get
    rol buttons+0      ; put get put get put
    bcc loop           ; get put [get]    <- this branch must not be allowed to cross a page

Note that this example routine only reads two 8-bit controllers and does not take enough time to span more than one DPCM fetch. Routines longer than this must contend with two additional constraints:

  • When DMC DMA is delayed by an odd number of cycles, it takes 3 cycles instead of 4, changing the cycle parity. If extending this function to read more bits, care must be taken so that all CPU write cycles are aligned. Instructions with a single write cycle must align the write to avoid conflict with the DPCM fetch, and double-write instructions like ROL need to align both writes so that the DPCM fetch falls on the first write.[7] If an interrupt can occur during the routine, it must be aligned so the fetch can only fall on the second of the three automatic stack writes.
  • When DMC DMA occurs near the end of OAM DMA, it steals an odd number of cycles, inverting the cycle parity. Every DMC period after that, a misaligned DPCM fetch will occur. Care must be taken to ensure this does not land on a joypad read. For longer functions, these misaligned cases can be resynced by landing the DPCM fetch on a single write cycle.

See DMA for detailed information on DMA timing. However, game developers writing synced code may find it easier to consider just the specific case of DMC DMA used for reloading the sample buffer. These timings do not apply to the initial DMC DMA load caused by writing to $4015, and they are not sufficient for emulating DMA timings; emulator authors should refer to DMA for comprehensive information. A DMC DMA reload takes these cycle counts in these cases:

  • 4 cycles if it falls on a CPU read cycle
  • 3 cycles if it falls on an odd number of CPU write cycles
  • 4 cycles if it falls on an even number of CPU write cycles
  • An even number of cycles if it occurs during OAM DMA
  • An odd number of cycles if it occurs at the end of OAM DMA (changing cycle parity for the following code)

This approach to DPCM safety has these benefits:

  • It can support any controller because it only does one pass and has no cycle budget.
  • It is nearly as fast as standard reads.
  • It avoids the bug and thus does not corrupt the sample byte or interfere with the APU's frame IRQ.

It also has these drawbacks:

  • It uses timed code and relies on complicated DMA behavior, so it is harder to understand and modify.
  • Moving the code can break it by causing branches to cross a page boundary, changing the timing.
  • It must directly follow OAM DMA. Unusual features that skip OAM DMA on gameplay frames are not compatible with this approach. (On lag frames, normally both OAM DMA and controller reads are skipped. This is not a problem for synced reads.)
  • It does not work in emulators that emulate the bug but have incorrect DMA timing (e.g. Nestopia). Many modern emulators either don't emulate the bug or have correct-enough DMA timing.
  • Synced code that lasts longer than one DMC period minus 4 (usually 428 cycles) is difficult to write correctly. This is because write cycles must also be synced and specific desync cases occur if DMC DMA lands at the end of OAM DMA. However, this is the only approach that supports code of this length.

Other Useful Operations

Button Flags

It is helpful to define the buttons as a series of bit flags:

BUTTON_A      = 1 << 7
BUTTON_B      = 1 << 6
BUTTON_SELECT = 1 << 5
BUTTON_START  = 1 << 4
BUTTON_UP     = 1 << 3
BUTTON_DOWN   = 1 << 2
BUTTON_LEFT   = 1 << 1
BUTTON_RIGHT  = 1 << 0

And then buttons can be checked as follows:

    lda buttons
    and #BUTTON_A | BUTTON_B
    beq notPressingAorB
    ; Handle presses.
notPressingAorB:

Calculating Presses and Releases

To calculate newly pressed and newly released buttons by comparing against the last frame's buttons:

    ; newly pressed buttons: not held last frame, and held now
    lda last_frame_buttons, x
    eor #%11111111
    and buttons, x
    sta pressed_buttons, x

    ; newly released buttons: not held now, and held last frame
    lda buttons, x
    eor #%11111111
    and last_frame_buttons, x
    sta released_buttons, x

Directional Safety

Opposing directions (Up+Down and Left+Right) are possible on non-standard controllers, or even on a worn out standard one. Both directions can be tested for at the same time in an efficient bitwise operation:

    lda buttons, x
    and #%00001010    ; Compare Up and Left...
    lsr a
    and buttons, x    ; to Down and Right
    beq not_opposing
        ; Use previous frame's directions
        lda buttons, x
        eor last_frame_buttons, x
        and #%11110000
        eor last_frame_buttons, x
        sta buttons, x
    not_opposing:

Diagonal directions can also be detected at the same time as opposing ones, by testing if more than 1 directional bit is set. This could be used as part of a 4-way joystick style control scheme:

    lda buttons, x
    and #%00001111    ; If A & (A - 1) is nonzero, A has more than one bit set
    beq not_diagonal
    sec
    sbc #1
    and buttons, x
    beq not_diagonal
        ; Use previous frame's directions
        lda buttons, x
        eor last_frame_buttons, x
        and #%11110000
        eor last_frame_buttons, x
        sta buttons, x
    not_diagonal:

As an example, the code above rejects the detected cases by using the direction bits from the previous frame, but depending on the situation another response might be more appropriate (e.g. cancel just the opposing bits, or clear all directions).

External Examples

  • Forum post: Blargg's DMC-fortified controller read routine
  • Forum post: Rahsennor's OAM-synchronized controller read
  • Forum post: Drag's bitwise DMC-safe controller reading
  • pads.s: pinobatch's NROM-template controller read routine

References

  1. Controller reading: Unconnected data lines and open bus
  2. Super Mario Bros. 3 controller reread method
  3. Ars Technica: How to beat Super Mario Bros. 3 in less than a second
  4. https://tasvideos.org/6466S Tool-Assisted Speedrun of Super Mario Bros. 3 which first demonstrated abusing controller reread routines.
  5. Forum post: Rahsennor's OAM-synchronized controller read
  6. Forum post: as of May 2016, Nintendulator and Nestopia do not accurately emulate OAM-synchronized controller reading.
  7. Forum post: demonstration of how ROL instruction affects alignment for OAM DMA synchronized controller reading.