(nit) Do you have plans to usesrc_desc(), ... queries elsewhere?
If not, why not callquery_md() directly in methods that create memory objects (like here)?

Hi, we usesrc_desc() asprimitive_desc class to make the code consistent with oneDNN.

(readability) maybe use a better name thanm? (e.g. mem_arg_cache? )

Also, if you decide to hoist make_args calls out of the execute function, you might need to make this an unordered_map.

mem_arg_cache is better thenm. I'll take it.
We adoptedunordered_map at first. But when we tested on a BMG client with a low-end CPU,unordered_map would affect the performance. So, we chose the simpler array.

what are the slots 0 and 1 reserved for?

Since we just provide int4 gemm this time. The slots 0 and 1 are reserved forscale andzp for int4 gemm.

For that test to work, there is an assumption that memory::desc is constant across multiple calls to execute. Just two notes on that:

• this will no more be true if you start using oneDNN runtime_dimension feature (in that case, you will have to call make_args with proper shape at every execute call).
• if you don't use oneDNN runtime_dimension feature, you actually can hoist make_args calls out of this execute call (e.g. in primitive_ext constructor), and only useset_data_handle here.

why not take transpose and types as parameter, and hash them as part of the key?
This would allow to have just a single cache instead of separate cache for each combination (and it would be harder to control total cache size).
Also, do you plan to have separate cache for each primitive?

Yes, cache only rely on information of shapes but nothing else. Other factors can be processed by switch cases. Which will greatly shrink hash conflict and speed up lookup.

using unordered_map inprimitive_ext would help hiding this implementation detai.

Please do not change the test logic of int4 gemm ut.

1. GPTQ XPU is not supported at AO now. The ut is only intended for RTN recipe.
2. Usingdq weight to construct ref is not recommended. We formerly found that it would hide bugs in weight prepack.
3. The UT follows the usage of pytorch designed cases, we want keep a same logic.