We can reuse allocations of tensors in case of same shape product && data type (e.g. tensor reshape)
This requires extending shape inference for all supported operations and writing custom arena allocator.
All those alloc reusage should be planned load-time and integrated into an execution plan
We can reuse allocations of tensors in case of same shape product && data type (e.g. tensor reshape)
This requires extending shape inference for all supported operations and writing custom arena allocator.
All those alloc reusage should be planned load-time and integrated into an execution plan