Difference between revisions of "GPU/External Registers"

From 3dbrew
< GPU
Jump to navigation Jump to search
(Clarity bit6 alternative mode, clean up interlacing language, elaborate how fb interlacing works, and minor cleanup on the text below the fb format register table)
(Reword TextureCopy)
 
Line 555: Line 555:
 
=== TextureCopy ===
 
=== TextureCopy ===
  
When bit 3 of the control register is set, the hardware performs a TextureCopy-mode transfer. In this mode, all other bits of the control register (except for bit 2, which still needs to be set correctly) and the regular dimension registers are ignored, and no format conversions are done. Instead, it performs a raw data copy from the source to the destination, but with a configurable gap between lines. The total amount of bytes to copy is specified in the size register, and the hardware loops reading lines from the input and writing them to the output until this amount is copied. The "gap" specified in the input/output dimension register is the number of chunks to skip after each "width" chunks of the input/output, and is NOT counted towards the total size of the transfer.
+
When bit 3 of the control register is set, the hardware performs a TextureCopy-mode transfer: no format conversions are done, instead a raw data copy is performed from the source to the destination, with a configurable gap between lines. All bits of the control register are ignored, except for input/output dimensions, which are used for line width and gap, and bit 2, which must be set when gaps are used.
 +
 
 +
The total amount of bytes to copy is specified in the size register, the hardware loops reading lines from the input and writing them to the output until this amount is copied. The gap specifies the number of bytes to skip after each line read (a gap of 0 results in a contiguous read). Gaps do not count towards the total size of the transfer.
 +
 
 +
When setting line width and gap they must be divided by 2 (it can be thought as the calculation being done in bits, and the values being stripped of their lower 4 bits for the alignment). For example, if the left half of a 32x32 RGB8 texture is to be copied, the parameters will be:
 +
line width = (16 * 24) >> 4 = 24
 +
gap = line width
 +
size = (16 * 32 * 24) >> 4 = 768
  
 
By correctly calculating the input and output gap sizes it is possible to use this functionality to copy arbitrary sub-rectangles between differently-sized framebuffers or textures, which is one of its main uses over a regular no-conversion DisplayTransfer. When copying tiled textures/framebuffers it's important to remember that the contents of a tile are laid out sequentially in memory, and so this should be taken into account when calculating the transfer parameters.
 
By correctly calculating the input and output gap sizes it is possible to use this functionality to copy arbitrary sub-rectangles between differently-sized framebuffers or textures, which is one of its main uses over a regular no-conversion DisplayTransfer. When copying tiled textures/framebuffers it's important to remember that the contents of a tile are laid out sequentially in memory, and so this should be taken into account when calculating the transfer parameters.
  
Specifying invalid/junk values for the TextureCopy dimensions can result in the GPU hanging while attempting to process this TextureCopy.
+
Specifying invalid/junk values for the TextureCopy dimensions can result in the GPU hanging while attempting to process this TextureCopy. For instance, when in contiguous mode the size must be at least 16; when in gap mode, the size must be at least 192, and the line width must not be 0.
  
 
== Command List ==
 
== Command List ==

Latest revision as of 00:47, 17 December 2024

This page describes the address range accessible from the ARM11, used to configure the basic GPU functionality. For information about the internal registers used for 3D rendering, see GPU/Internal Registers.

Map[edit]

Address mappings for the external registers. GSPGPU:WriteHWRegs takes these addresses relative to 0x1EB00000.

User VA PA Length Name Comments
0x1EF00000 0x10400000 4 Hardware ID Bit2: new model
0x1EF00004 0x10400004 4 ?
0x1EF00010 0x10400010 16 Memory Fill1 "PSC0" GX command 2
0x1EF00020 0x10400020 16 Memory Fill2 "PSC1" GX command 2
0x1EF00030 0x10400030 4 VRAM bank control Bits 8-11 = bank[i] disabled; other bits are unused.
0x1EF00034 0x10400034 4 GPU Busy Bit26 = PSC0, bit27 = PSC1, Bit30 = PPF, Bit31 = P3D
0x1EF00050 0x10400050 4 ? Writes 0x22221200 on GPU init.
0x1EF00054 0x10400054 4 ? Writes 0xFF2 on GPU init.
0x1EF000C0 0x104000C0 4 Backlight control Writes 0x0 to allow backlights to turn off, 0x20000000 to force them always on.
0x1EF00400 0x10400400 0x100 Framebuffer Setup "PDC0" (top screen)
0x1EF00500 0x10400500 0x100 Framebuffer Setup "PDC1" (bottom)
0x1EF00C00 0x10400C00 ? Transfer Engine "DMA"
0x1EF01000/0x10401000 - 0x1EF01C00/0x10401C00 maps to GPU internal registers. These registers are usually not read/written directly here, but are written using the command list interface below (corresponding to the GPUREG_CMDBUF_* internal registers)
0x1EF01000 0x10401000 0x4 ? Writes 0 on GPU init and before the Command List is used
0x1EF01080 0x10401080 0x4 ? Writes 0x12345678 on GPU init.
0x1EF010C0 0x104010C0 0x4 ? Writes 0xFFFFFFF0 on GPU init.
0x1EF010D0 0x104010D0 0x4 ? Writes 1 on GPU init.
0x1EF014?? 0x104014?? 0x14 "PPF" ?
0x1EF018E0 0x104018E0 0x14 Command List "P3D"

Memory Fill[edit]

User VA Description
0x1EF000X0 Buffer start physaddr >> 3
0x1EF000X4 Buffer end physaddr >> 3
0x1EF000X8 Fill value
0x1EF000XC Control. bit0: start/busy, bit1: finished, bit8-9: fill-width (0=16bit, 1=3=24bit, 2=32bit)

Memory fills are used to initialize buffers in memory with a given value, similar to memset. A memory fill is triggered by setting bit0 in the control register. Doing so aborts any running memory fills on that filling unit. Upon completion, the hardware unsets bit0 and sets bit1 and fires interrupt PSC0.

These registers are used by GX SetMemoryFill.

LCD Source Framebuffer Setup[edit]

All of these registers must be accessed with 32bit operations regardless of the registers' actual bit size.

The naming of these parameters reflects the physical characteristics of the displays, and not the way the 3DS is normally held.

To make sense of these values, the 3DS must be held in a way, so that the bottom screen is in the left hand, and the top screen is in the right hand, and that way the first pixel will be in the top-left corner, as it should be. If the 3DS is held normally, the first pixel is in the bottom-left corner.

All pixel and scanline timing values are 12bits, unless noted. This also applies to those fields where two u16 are combined into one register. Each u16 field is only 12bits in size. timin

The horizontal timing parameter order is as follows (values may overflow through HTotal register value):

0x10 < 0x14 <= 0x60.LO <= 0x04 <= 0x60.HI <= 0x08 <= 0x0C <= 0x10
0x18 <= 0x60.LO

Timing starts from HCount == 0, then each absolute value in the beforementioned register chain triggers when HCount == register, latching the primitive display controller into a new mode. There is an inherent latch order, where if two simultenaous events occur, one event wins over another.

Known latched modes (in order):
- HSync (triggers a line to the LCD to move to the next line)
- Back porch (area between HSync and border being displayed, no pixels pushed, min 16 pixel clocks, otherwise the screen gets glitchy)
- Left border start (no image data is being displayed, just a configurable solid color)
- Image start (pixel data is being DMA'd from video memory or main RAM)
- Right border start/Image end (border color is being displayed after the main image)
- Unknown synchronization (supposed to be probably right border end, but this mode seems to be broken or not do anything)
- Front porch (no pixels pushed, 68 clock min, otherwise the screen doesn't sync properly, and really glitches out)
Offset Name Comments
0x00 HTotal The total width of a timing scanline. In other words, this is the horizontal refresh clock divider value.

HClock = PClock / (HTotal + 1)

0x04 HStart Determines when the image is going to be displayed in the visible region (register 0x60).
0x08 HBR Right border start(?). Does nothing.

While this register seems to have no impact on the image whatsoever, it still has to be set to a valid value.

0x0C HPF Front porch. The image is blanked during this period, and no pixels are pushed to the LCD.

Unknown why, but a single dot of red is displayed before entering this mode.

0x10 HSync Triggers a HSync pulse.

Based on behavior, this needs to last at least a pixel clock for the LCD to register the sync.

0x14 HPB Back porch? Has to be at least one bigger than HSync, otherwise HSync never triggers.

The display is blank, and the LCD displays nothing in this period (doesn't push pixels).

0x18 HBL Left border trigger treshold. Enables pushing pixels to the display.

If this value is smaller than the back porch, then the back porch period will be zero, and the border will be immediately displayed upon entering the back porch period.

Can be lower than HSync, as the back porch is what takes the controller out of HSync.

Must be <= HDisp start (reg 0x60 low u16), otherwise no pixels will be pushed due to a glitched state.

0x1C H Interrupt timing Made up from two u16 values, PDC interrupt line is asserted when HCount == low u16, and most likely deasserted when HCount == high u16.

There seems to be some limitations though:

  • low u16 must be smaller than high u16
  • if low u16 is less than HTotal then high u16 must also be smaller than HTotal
  • setting low u16 to >= HTotal disables the interrupt ever firing

This is configured by gsp in a way so that low u16 equals to HTotal, meaning the HSync interrupt will never fire.

0x20 low u16: ???

high u16: ???

???
0x24 VTotal Total height of the timing window. Can be interpreted as the vertical clock divider.

VClock = PClock / (HTotal + 1) / (VTotal + 1)

Setting this to 494 lowers framerate to about 50.040660858 Hz ((268111856 / 24) / (450 + 1) / (494 + 1)).

0x28 ? Seems to determine the vertical blanking interval.


Setting this to lower than VTotal - VDisp will cut off the top VTotal - VDisp - thisvalue lines.

Setting this to higher than VTotal - VDisp will make the image be pushed downwards with the overscan color visible.

Setting this to higher than HTotal will make the GPU skip vertical pixel data synchronization (hence filling the screen with the rest of the pixel data past the given screen framebuffer size). Also will skip thisvalue + somevalue - HTotal lines into the "global" pixel buffer.

0x30 ? Total amount of vertical scanlines in the pixel buffer, must be bigger than *an unknown blanking-like value*. If this value is less than VDisp then the last two scanlines will be repeated interlaced until VDisp is reached.
0x34 VDisp(?) Total amonut of vertical scanlines displayed (only for top screen it seems like). If this value is less than VTotal then the rest of the scanlines will not be updated on the screen, so those will slowly fade out. Must be bigger than *an unknown blanking-like value*, otherwise an underflow will happen.
0x38 Vertical data offset(?) ??? Seems to offset the screen upwards if this value is high enough. If this value is higher or equal to *some value* (aka. if less than one scanline is displayed on the screen) then the screen will lose synchronization.
0x40 V Interrupt timing Similar to H Interrupt timing (0x1C), except the comparison is done against VCount, the limitations are emposed on VTotal, and the interrupt that fires is VSync.

One important note is that it seems like the VSync interrupt always fires at HCount == 0, and there doesn't seem to be a register to control this behavior.

0x44 ??? similar functionality to 0x10
0x48 ??? bit0 seems to disable HSync, bit8 seems to disable VSync, rest of the bits aren't writable.
0x4C Overscan filler color 24bits(? top 8bits ignored)

When the visible region is being drawn, but the timing parameters are set up in a way that the framebuffer is smaller than the visible region, it will be filled by this color.

0x50 HCount Horizontal "beam position" counter. Note that this value does not equal to the current pixel being drawn.
0x54 VCount Vertical "beam position" counter. Note that the scanline being drawn isn't equal to this value.
0x5C ??? low u16: Image width (including some offset?)

high u16: Image height??? (seems to be unused)

0x60 HDisp low u16: Image start (border --> pixel data)

high u16: Image end (pixel data --> border)

0x64 ??? low u16: unknown

high u16: framebuffer total height (amount of scanlines blitted regardless of framebuffer height)

0x68 Framebuffer A first address For top screen, this is the left eye 3D framebuffer.
0x6C Framebuffer A second address For top screen, this is the left eye 3D framebuffer.
0x70 Framebuffer format and other settings See framebuffer format
0x74 PDC control Bit 0: Enable display controller.

Bit 8: HBlank IRQ mask (0 = enabled). Bit 9: VBlank IRQ mask (0 = enabled). Bit 10: Error IRQ mask? (0 = enabled). Bit 16: Output enable?

0x78 Framebuffer select and status Bit 0: Next framebuffer to display (after VBlank).

Bit 4: Currently displaying framebuffer? Bit 8: Reset FIFO? Bit 16: HBlank IRQ status/ack. Write 1 to aknowledge. Bit 17: VBlank IRQ status/ack. Bit 18: Error IRQ status/ack?

0x80 Color lookup table index select 8bits, write-only
0x84 Color lookup table indexed element Contains the value of the color lookup table indexed by the above register, 24bits, RGB8 (0x00BBGGRR)

Accessing this register will increase the index register by one

0x90 Framebuffer stride 32bits (bottom 3bits ignored?)

Distance in bytes between the start of two framebuffer rows (must be a multiple of 8).

In other words, this can be interpreted as the amount to add to the framebuffer pointer after displaying a scanline.

Setting this to zero will cause only the first line of the image to be displayed repeated on the entire display. With the HSync interrupt it's possible to "race the beam" to (ab)use this feature.

Because of this simplicity, writing a negative value here VFlips the image, although that requires the framebuffer pointer register to be set to the start of the last scanline, instead of at the start of the framebuffer.

0x94 Framebuffer B first address For top screen, this is the right eye 3D framebuffer. Unused for bottom screen in userland.
0x98 Framebuffer B second address For top screen, this is the right eye 3D framebuffer. Unused for bottom screen in userland.

Framebuffer format[edit]

Bit Description
2-0 Color format
5-4 Framebuffer interlacing mode
0 - A  (no interlacing)
1 - AA (scanline doubling)
2 - AB (interlace enable)
3 - BA (same as above, but the fields are inverted)

In AB and BA interlace modes, a scanline from each framebuffer is output in an alternating manner. In AB mode, Framebuffer A is output on the frist display scanline. Similarly, in BA mode, Framebuffer B gets output to the first display scanline.

The way AB and BA modes work, is that a scanline is output, the framebuffer stride value is added to the internal scanline pointer value, and the other framebuffer is selected. And this alternates until the end of the draw region.

AA interlacing works like AB interlacing, except both internal framebuffer pointers are set to the Framebuffer A pointer value.

In A mode (no interlacing), it doesn't switch to the other framebuffer at the end of outpuitting a scanline to the display.

Bottom screen has this set to 0 (A mode, no interlacing) at all times. Top screen uses AB interlacing in 3D mode (with 3D slider enabled), and A mode (no interlacing) in 2D mode.

6 Alternative pixel output mode*
7 ?
9-8 DMA size
0 -  4 FCRAM words (32 bytes)
1 -  8 FCRAM words (64 bytes)
2 - 16 FCRAM words (128 bytes)
3 - ???

FCRAM doesn't support DMA size 3, as it can only burst up to 16 words (128 bytes), and will show a black screen instead.

31-16 Unknown


* The weird thing about bit6, is that it works different between the bottom and top LCD. On the bottom LCD, it doubles the number of outputted pixels (so the same pixel is outputted twice, effectively doing pixel/column doubling). However on the top screen, it does scanline doubling instead. Most likely the top screen receives two pixels at once per clock unit, outputting two scanlines simultaneously.

On a 2DS, it seems to have no effect on the top part of the display, and on the bottom screen it just shifts the framebuffer to the right two pixels.

GSP module only allows the LCD stereoscopy (3D) to be enabled when bit5=1 and bit6=0 here. When GSP module updates this register, GSP module will automatically disable the stereoscopy if those bits are not set for enabling stereoscopy.


When both interlacing and alternative mode is disabled (bit6=0), the full resolution of the top screen (240x800) can be utilized if the PDC registers are updated to accomodate this higher resolution. GSP contains tables for this mode (gsp mode == 1). GSP automatically applies this mode if both bit5 and bit6 are cleared. This is also the default, and the only valid mode for the bottom screen in userland.

If only AB interlacing is enabled (bit5=1, bit6=0), gsp detects this as a request to switch to 3D mode (gsp mode == 2), and enables the parallax barrier. It's unknown how to control this, but some other PDC registers control if interlacing should be done by true interleaving (both framebuffers are treated as 240x400), or by skipping lines (both framebuffers are treated as 240x800).

If only alternative mode is enabled (bit5=0, bit6=1), gsp detects it as a request to switch back to 2D mode for the top screen (gsp mode == 0). This is also the default mode for the top screen.

Both interlacing and scan doubling can't be enabled in usermode, but it works as expected in baremetal.

Framebuffer color formats[edit]

Value Description
0 GL_RGBA8_OES
1 GL_RGB8_OES
2 GL_RGB565_OES
3 GL_RGB5_A1_OES
4 GL_RGBA4_OES

Color components are laid out in reverse byte order, with the most significant bits used first (i.e. non-24-bit pixels are stored as a little-endian values). For instance, a raw data stream of two GL_RGB565_OES pixels looks like GGGBBBBB RRRRRGGG GGGBBBBB RRRRRGGG.

Color formats 5, 6, and 7 are blocked by gsp, but they behave as pixel-doubled RGBA8 (not line doubling, but instead the same pixel is output twice) if used outside of userland.

Transfer Engine[edit]

Register address Description
0x1EF00C00 Input physical address >> 3
0x1EF00C04 Output physical address >> 3
0x1EF00C08 DisplayTransfer output width (bits 0-15) and height (bits 16-31).
0x1EF00C0C DisplayTransfer input width and height.
0x1EF00C10 Transfer flags. (See below)
0x1EF00C14 GSP module writes value 0 here prior to writing to 0x1EF00C18, for cmd3.
0x1EF00C18 Setting bit0 starts the transfer. Upon completion, bit0 is unset and bit8 is set.
0x1EF00C1C ?
0x1EF00C20 TextureCopy total amount of data to copy, in bytes.
0x1EF00C24 TextureCopy input line width (bits 0-15) and gap (bits 16-31), in 16 byte units.
0x1EF00C28 TextureCopy output line width and gap.

These registers are used by GX command 3 and 4. For cmd4, *0x1EF00C18 |= 1 is used instead of just writing value 1. The DisplayTransfer registers are only used if bit 3 of the flags is unset and ignored otherwise. The TextureCopy registers are likewise only used if bit 3 is set, and ignored otherwise.

Flags Register - 0x1EF00C10[edit]

Bit Description
0 When set, the framebuffer data is flipped vertically.
1 When set, the input framebuffer is treated as linear and converted to tiled in the output, converts tiled->linear when unset.
2 This bit is required when the output width is less than the input width for the hardware to properly crop the lines, otherwise the output will be mis-aligned.
3 Uses a TextureCopy mode transfer. See below for details.
4 Not writable
5 Don't perform tiled-linear conversion. Incompatible with bit 1, so only tiled-tiled transfers can be done, not linear-linear.
7-6 Not writable
10-8 Input framebuffer color format, value0 and value1 are the same as the LCD Source Framebuffer Formats (usually zero)
11 Not writable
14-12 Output framebuffer color format
15 Not writable
16 Use 32x32 block tiling mode, instead of the usual 8x8 one. Output dimensions must be multiples of 32, even if cropping with bit 2 set above.
17-23 Not writable
24-25 Scale down the input image using a box filter. 0 = No downscale, 1 = 2x1 downscale. 2 = 2x2 downscale, 3 = invalid
31-26 Not writable

TextureCopy[edit]

When bit 3 of the control register is set, the hardware performs a TextureCopy-mode transfer: no format conversions are done, instead a raw data copy is performed from the source to the destination, with a configurable gap between lines. All bits of the control register are ignored, except for input/output dimensions, which are used for line width and gap, and bit 2, which must be set when gaps are used.

The total amount of bytes to copy is specified in the size register, the hardware loops reading lines from the input and writing them to the output until this amount is copied. The gap specifies the number of bytes to skip after each line read (a gap of 0 results in a contiguous read). Gaps do not count towards the total size of the transfer.

When setting line width and gap they must be divided by 2 (it can be thought as the calculation being done in bits, and the values being stripped of their lower 4 bits for the alignment). For example, if the left half of a 32x32 RGB8 texture is to be copied, the parameters will be:

line width = (16 * 24) >> 4 = 24
gap = line width
size = (16 * 32 * 24) >> 4 = 768

By correctly calculating the input and output gap sizes it is possible to use this functionality to copy arbitrary sub-rectangles between differently-sized framebuffers or textures, which is one of its main uses over a regular no-conversion DisplayTransfer. When copying tiled textures/framebuffers it's important to remember that the contents of a tile are laid out sequentially in memory, and so this should be taken into account when calculating the transfer parameters.

Specifying invalid/junk values for the TextureCopy dimensions can result in the GPU hanging while attempting to process this TextureCopy. For instance, when in contiguous mode the size must be at least 16; when in gap mode, the size must be at least 192, and the line width must not be 0.

Command List[edit]

Register address Description
0x1EF018E0 Buffer size in bytes >> 3
0x1EF018E8 Buffer physical address >> 3
0x1EF018F0 Setting bit0 to 1 enables processing GPU command execution. Upon completion, bit0 seems to be reset to 0.

These 3 registers are used by GX command 1. This is used for GPU commands.

Framebuffers[edit]

These LCD framebuffers normally contain the last rendered frames from the GPU. The framebuffers are drawn from left-to-right, instead of top-to-bottom.(Thus the beginning of the framebuffer is drawn starting at the left side of the screen)

Both of the 3D screen left/right framebuffers are displayed regardless of the 3D slider's state, however when the 3D slider is set to "off" the 3D effect is disabled. Normally when the 3D slider's state is set to "off" the left/right framebuffer addresses are set to the same physical address. When the 3D effect is disabled and the left/right framebuffers are set to separate addresses, the LCD seems to alternate between displaying the left/right framebuffer each frame.

Init Values from nngxInitialize for Top Screen[edit]

  • 0x1EF00400 = 0x1C2
  • 0x1EF00404 = 0xD1
  • 0x1EF00408 = 0x1C1
  • 0x1EF0040C = 0x1C1
  • 0x1EF00410 = 0
  • 0x1EF00414 = 0xCF
  • 0x1EF00418 = 0xD1
  • 0x1EF0041C = 0x1C501C1
  • 0x1EF00420 = 0x10000
  • 0x1EF00424 = 0x19D
  • 0x1EF00428 = 2
  • 0x1EF0042C = 0x1C2
  • 0x1EF00430 = 0x1C2
  • 0x1EF00434 = 0x1C2
  • 0x1EF00438 = 1
  • 0x1EF0043C = 2
  • 0x1EF00440 = 0x1960192
  • 0x1EF00444 = 0
  • 0x1EF00448 = 0
  • 0x1EF0045C = 0x19000F0
  • 0x1EF00460 = 0x1c100d1
  • 0x1EF00464 = 0x1920002
  • 0x1EF00470 = 0x80340
  • 0x1EF0049C = 0

More Init Values from nngxInitialize for Top Screen[edit]

  • 0x1EF00468 = 0x18300000, later changed by GSP module when updating state, framebuffer
  • 0x1EF0046C = 0x18300000, later changed by GSP module when updating state, framebuffer
  • 0x1EF00494 = 0x18300000
  • 0x1EF00498 = 0x18300000
  • 0x1EF00478 = 1, doesn't stay 1, read as 0
  • 0x1EF00474 = 0x10501