DMA beta
New, hand-written section - see the note on the cheat sheet.
Direct Memory Access (DMA) moves blocks of data between main RAM and the PS2's other components without the EE copying them by hand. Since the GS, VU1, and the IOP (over SIF) are all separate destinations outside the EE's own memory, DMA is how data actually gets to any of them in a real program - not an optimization you bolt on later.
Ten channels, one job each
This SDK's real, vendored channel list (ee/dma/include/dma.h):
| Channel | Value | Direction |
|---|---|---|
DMA_CHANNEL_VIF0 | 0 | EE → VU0 |
DMA_CHANNEL_VIF1 | 1 | EE → VU1 |
DMA_CHANNEL_GIF | 2 | EE (or VU1) → GS |
DMA_CHANNEL_fromIPU / DMA_CHANNEL_toIPU | 3 / 4 | The Image Processing Unit - the IPU is what actually does the MPEG2 decode work core/mpeg wraps (see the cheat sheet) |
DMA_CHANNEL_fromSIF0 / DMA_CHANNEL_toSIF1 | 5 / 6 | EE ↔ IOP (see EE, IOP, and SIF) |
DMA_CHANNEL_SIF2 | 7 | A second, separate EE↔IOP path (used less often than SIF0/1) |
DMA_CHANNEL_fromSPR / DMA_CHANNEL_toSPR | 8 / 9 | Scratchpad RAM (a small, fast EE-local memory) |
The channel you use follows directly from where the data is going - sending vertex/texture data to the GS means the GIF channel, sending an RPC request to the IOP means a SIF channel, and so on. There's no generic "DMA to anywhere" call.
Why dmakit exists
Queuing a DMA transfer for real means building a DMA tag (a small header describing the transfer - source, size, and what to do when it's done or when to chain into the next tag) and writing it to the right channel's registers. dmakit (world tier) wraps this: building tag chains, waiting on completion, and the common patterns (send this buffer, then this one, then interrupt me) instead of hand-rolling tag bytes for every call site.
Why this matters for how you structure code
Because DMA transfers run concurrently with the CPU, a real program built around it looks different from one that just calls a blocking "draw now" function: you queue work, do something else (build the next frame's data, run game logic), and only synchronize when you actually need the result or need to reuse a buffer the DMA engine hasn't finished reading from yet. Reusing a source buffer before its DMA transfer has actually completed is a classic PS2 bug - the data the GS/VU ends up seeing is whatever was in memory at the moment the DMA engine got to it, not at the moment you called the function.
See also
- Graphics Synthesizer - the GIF channel's destination
- EE, IOP, and SIF - the SIF channels' destination
- Vector Units - the VIF0/VIF1 channels' destination