Horizontal gathers
IMAGE_GATHER4H and IMAGE_GATHER4H_PCK gather one channel across a horizontal group of texels. The measured direct 1D and 2D subset is covered by 1,242 GPU dispatches.
Repository notes through October 10, 2026
What the instruction returns
A vertical gather reads four texels in Y. The H form reads them in X. GATHER4H writes the selected channel of those texels. GATHER4H_PCK writes the packed raw texel stream. The destination width follows DMASK, and registers the instruction does not own stay unchanged.
Edges use the sampler's addressing rules. A texel that falls outside the image is handled rather than read from a neighbouring allocation. The October 9 probe ran 1,242 GPU dispatches over the measured direct 1D and 2D subset.
What stays out
A16, D16, R128, indirect resource tables, array, cube and MSAA views, and compressed formats are not part of the measured subset. Those descriptor and control modes remain explicit limitations.
The gather work shares the translator and the Vulkan image path with the multi-texel loads. It does not change detile addressing, and it does not claim a frame-rate result.
What is implemented
- IMAGE_GATHER4H and IMAGE_GATHER4H_PCK for the measured direct 1D and 2D cases.
- DMASK destination widths and preservation of unrelated registers.
- Edge handling according to the sampler, checked with 1,242 GPU dispatches.
What is still missing
- Array, cube, MSAA, compressed and several descriptor modes are not in the measured subset.
- A16, D16, R128 and indirect tables remain limitations.
- No title frame rate is inferred from the probe.
Related mechanisms
What these pages are