VRAM: Difference between revisions

From NeoGeo Development Wiki
Jump to navigation Jump to search
m (minor edits)
m (Minor details, some rephrasing)
Line 11: Line 11:
* {{Reg|REG_VRAMMOD ($3C0004)}} is the signed value added to the VRAM address after a write
* {{Reg|REG_VRAMMOD ($3C0004)}} is the signed value added to the VRAM address after a write


The original SNK documentation specifies that the address register should be set directly instead of using {{Reg|REG_VRAMMOD}} when crossing VRAM zones ($0000~$7FFF to/from $8000~$FFFF). This is probably because the GPU's internal adder is 15 bit only.
The original [[Official development manual|SNK documentation]] specifies that the address register should be set directly instead of counting on incrementation when crossing VRAM zones ($0000~$7FFF to/from $8000~$FFFF) i.e. when bit 15 has to change. This is a side effect of simplifications in the [[GPU]].


== Memory map ==
== Memory map ==
Line 31: Line 31:
|$7500
|$7500
|$7FFF
|$7FFF
|[[Fix bankswitching|Extension]]
|[[Fix bankswitching|Fix map extension]]
|-
|-
|$8000
|$8000
Line 54: Line 54:
|}
|}


Even if VRAM can be modified at any time (even during active display), since it is not dual-ported, some timing restrictions have to be met in order to avoid skipping writes or reading data too early after setting the address. This is because the LSPC allocates a number of "access slots" per video line for the CPU between rendering steps.
Even if VRAM can be accessed at any time (even during active display), it is not dual-ported so some timing restrictions have to be met in order to avoid skipping writes or reading data too early after setting the address. This is because LSPC has to interleave VRAM access between the rendering process and the CPU.


== Timing considerations ==
== Timing considerations ==
The following timings are given to be 100% sure that no read or write to VRAM will fail. It is possible to sometimes shorten them with extreme guessing methods on the CPU side, but it's practically useless.
The following timings are given to be 100% sure that no read or write to VRAM will fail. It might be possible to sometimes shorten them with extreme guessing methods on the CPU's side, but it's practically useless.


After a write:
After a write:
Line 65: Line 65:
After an address change, reads should be made after at least 16 CPU cycles.
After an address change, reads should be made after at least 16 CPU cycles.


Note that these timings only affect the VRAM access, and not the other LSPC registers.
Note that these restrictions only concern VRAM access, and not the internal LSPC registers.


[[Category:Video system]]
[[Category:Video system]]

Revision as of 02:36, 31 March 2016

VRAM stands for Video RAM.

File:Aes cxk5814.jpg
One of two CXK5814 2KiB RAM chips used for the upper zone of the VRAM ($8000~$87FF) on a AES system

The NeoGeo has 68KiB (physically 64KiB + 4KiB, respectively called lower and upper, or slow and fast) of VRAM accessible as 16 bits words, which is used to store sprite attributes, the fix layer tile map and sprite lists for video rendering. Contrary to other systems, the VRAM here does not contain actual graphics.

Access to VRAM is done through 3 Memory mapped registers handled by the GPU, it isn't mapped in the 68k address space. Every VRAM address points to a word, not a byte.

The original SNK documentation specifies that the address register should be set directly instead of counting on incrementation when crossing VRAM zones ($0000~$7FFF to/from $8000~$FFFF) i.e. when bit 15 has to change. This is a side effect of simplifications in the GPU.

Memory map

Start End Description
$0000 $6FFF SCB1
$7000 $74FF Fix map
$7500 $7FFF Fix map extension
$8000 $81FF SCB2
$8200 $83FF SCB3
$8400 $85FF SCB4
$8600 $867F Sprite list for even scanlines
$8680 $86FF Sprite list for odd scanlines

Even if VRAM can be accessed at any time (even during active display), it is not dual-ported so some timing restrictions have to be met in order to avoid skipping writes or reading data too early after setting the address. This is because LSPC has to interleave VRAM access between the rendering process and the CPU.

Timing considerations

The following timings are given to be 100% sure that no read or write to VRAM will fail. It might be possible to sometimes shorten them with extreme guessing methods on the CPU's side, but it's practically useless.

After a write:

  • Another write should be made after at least 12 cycles.
  • An address change should be made after at least 16 cycles.

After an address change, reads should be made after at least 16 CPU cycles.

Note that these restrictions only concern VRAM access, and not the internal LSPC registers.