Function dma_copy_to_vram 

pub fn dma_copy_to_vram(
    _dma: &mut GpuDma,
    rect: VramRect,
    words: &[u32],
) -> bool

Upload words to rect over block-mode DMA (channel 2) instead of word-at-a-time FIFO writes, and wait for it. This mirrors PsyQ’s LoadImage: BS = words-per-row, BA = rows.

words holds the pixels as little-endian 32-bit words, two 16-bit halfwords each, row after row: (rect.w / 2) * rect.h words, and only that many are sent. Returns false, leaving the GPU untouched, when the transfer cannot be expressed this way: an odd or zero width, a zero height, a row wider than the GPU’s 16-word FIFO, or a slice shorter than the rectangle. Use upload_16bpp (the FIFO path) for those.

Opt-in: on real hardware the DMA controller can latch a channel busy-forever (CL2 probe, 2026-07-31). The completion wait is bounded, so a wedge costs a false and a partial upload rather than a hang.