summary here, will split to some issues(one issue for one items)
(1) clean up deprecated code for old dynamic-shape / fix-shape implementation, remove the stale code path when use-output-allocator=false.
(2)memory address space 0/1 correct use
(3)is compile-phase not touch constant reasonable?
(4)constant folding
(5)compile-phase optimization to reduce/eliminate runtime readback overhead, this topic maybe related to above (2)(3)(4)
(6)hip op definition review
(7)Pipeline design and code that does not conform to MLIR best practices review
(8)ORT session memory pool usage optimization
(9)share weights, re-design it as an EP formal feature
summary here, will split to some issues(one issue for one items)
(1) clean up deprecated code for old dynamic-shape / fix-shape implementation, remove the stale code path when use-output-allocator=false.
(2)memory address space 0/1 correct use
(3)is compile-phase not touch constant reasonable?
(4)constant folding
(5)compile-phase optimization to reduce/eliminate runtime readback overhead, this topic maybe related to above (2)(3)(4)
(6)hip op definition review
(7)Pipeline design and code that does not conform to MLIR best practices review
(8)ORT session memory pool usage optimization
(9)share weights, re-design it as an EP formal feature