Difference between revisions of "GPU/External Registers"
(Start cleaning up rotated screen clutter from PDC reg descriptions and add some mroe info) |
m (Fix a missing brace in calculation and make register more clear) |
||
Line 156: | Line 156: | ||
| The total width of a timing scanline. In other words, this is the horizontal refresh clock divider value. | | The total width of a timing scanline. In other words, this is the horizontal refresh clock divider value. | ||
− | HClock = PClock / ( | + | HClock = PClock / (HTotal + 1) |
|- | |- | ||
| 0x04 | | 0x04 | ||
Line 218: | Line 218: | ||
| Total height of the timing window. Can be interpreted as the vertical clock divider. | | Total height of the timing window. Can be interpreted as the vertical clock divider. | ||
− | VClock = | + | VClock = PClock / (HTotal + 1) / (VTotal + 1) |
− | Setting this to 494 lowers framerate to about 50.040660858 Hz (268111856 / 24 / (250 + 1) / (494 + 1)). | + | Setting this to 494 lowers framerate to about 50.040660858 Hz ((268111856 / 24) / (250 + 1) / (494 + 1)). |
|- | |- | ||
| 0x28 | | 0x28 |
Revision as of 21:36, 27 November 2021
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
Address mappings for the external registers. GSPGPU:WriteHWRegs takes these addresses relative to 0x1EB00000.
User VA | PA | Length | Name | Comments |
---|---|---|---|---|
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 | Bit31 = cmd-list busy, bit27 = PSC0 busy, bit26 = PSC1 busy. |
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
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
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.
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 | ? | Seems to determine the horizontal blanking interval.
Setting this to higher than Setting this to higher than |
0x08 | ? | must be >= REG#0x00 |
0x0C | ? | must be >= REG#0x08 |
0x10 | Window X start (LgyFb) | Offsets the viewing window on the display's physical X co-ordinate relative to its front porch end.
Outside of LgyFb changing this value only seems to cause weird pixel interpolation and blurryness. |
0x14 | Window X end (LgyFb) | Window X start + window width - 1 is written here to set the end display scan pixel offset.
Outside of LgyFb it seems to offset the screen to the left if this value is high enough, but can glitch out the syncing on the bottom screen. High enough values will make the screen skip too many "pixels". If this value is higher or equal to *some value* (aka. if less than one pixel per line is displayed on the screen) then the screen will lose synchronization. |
0x18 | Window Y start (LgyFb) | Offsets the viewing window on the display's physical Y co-ordinate relative to the first visible scanline. |
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:
This is configured by gsp in a way so that low u16 equals to HTotal, meaning the HSync interrupt will never fire. |
0x20 | Low: Window Y end (LgyFb)
High: ??? |
Low: Window Y start + window height - 1 is written here to set the last display scanline relative to the first visible scanline.
High: This is cleared to zero when displaying LgyFb. Outside of LgyFb this doesn't really seem to do anything useful.
|
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) / (250 + 1) / (494 + 1)). |
0x28 | ? | Seems to determine the vertical blanking interval.
Setting this to higher than Setting this to higher than |
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: framebuffer width
high u16: framebuffer height??? (seems to be unused) |
0x60 | ??? | low u16: timing data(?)
high u16: framebuffer total width (amount of pixels blitted regardless of framebuffer width) |
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 | Bit0-15: framebuffer format, bit16-31: unknown |
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
Bit | Description |
---|---|
2-0 | Color format |
3 | ? |
5-4 | Framebuffer scanline output mode (interlace config)
0 - A (output image as normal) 1 - AA (output a single line twice, aka framebuffer A is interlaced with itself) 2 - AB (interlace framebuffer A and framebuffer B) 3 - BA (same as above, but the line from framebuffer B is outputted first) 0 is used by bottom screen at all times. 1 is used by the top screen in 2D mode. 2 is used by top screen in 3D mode. 3 goes unused in userland. |
6 | Scan doubling enable?* (used by top screen) |
7 | ? |
9-8 | Value 1 = unknown: get rid of rainbow strip on top of screen, 3 = unknown: black screen. |
15-10 | Unused? |
- The weird thing about scan doubling, 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 column doubling). However on the top screen, it does scanline doubling instead. Considering that the bottom screen's table doesn't work on the top screen, this could give a hint as to how the top screen receives the pixel data from the PDC.
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 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 scan doubling are disabled, 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, 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 skipping lines (both framebuffers are treated as 240x800)
If only scan doubling is enabled, 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
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
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
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
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.
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.
Command List
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
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
- 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
- 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