[kurquhar]

Overview

Extend scalar DMA broadcasting support on hexagon backend

Additional information

Hexagon does not currently support this op condition, but it is needed by stable-diffusion.cpp. There will be a future companion PR for the corresponding backend ops test support.

Specifically, FastWan has the following geometry for ADD:
src1 = [1,1,1,C] as bias WEIGHTS
src0 = [W,H,T,C]

Before this change:

• the existing scalar DMA kernel was not selected during graph processing (ne13!=1)
• the scalar kernel was selected instead (ne10==1)
• execute_op_binary() returns HTP_STATUS_NO_SUPPORT at runtime when it encounters an extended DMA tensor coupled with the scalar kernel

After this change:

• the extended scalar DMA kernel now also supports non-unity higher dimensions e.g. ne13==ne03
• the extended scalar DMA kernel is selected via an additional gate (ne10==1 && dma64 && WEIGHTS && contig && fits)

Verification:

• Targeted weighted scalar ADD/MUL HTP regression passes.
• Broad HTP ADD,SUB,MUL,DIV sweep passed: 287/287 supported cases.

Requirements

• I have read and agree with the contributing guidelines yes
• AI usage disclosure: Yes: to reason, code the kernel extension, and perform A/B testing.

Assisted-by: OpenCode

[max-krasnyansky]

Closing for a sec to undraft another PR

[max-krasnyansky]

@lhez for another review and ack

[max-krasnyansky]

@ggerganov @CISC we added a couple more tests here for binary ops that were not covered and cause issues in ggml-hexagon.
The CI runs look clean for all backends. I'm going to merge. Let me know if you want a follow up on the tests.