List of TODOs in the code

[gussmith23]

No description provided.

[github-actions]

These TODOs were scraped from the code.

• (@gussmith23) Include some simple simplifying rewrites 🔗
• (@gussmith23) Include some simple simplifying rewrites 🔗
• (@gussmith23) This is starting to become a flaky test... 🔗
• (@gussmith23) add more checks. I should be using extraction to do this, 🔗
• (@gussmith23) Shouldn't always panic. 🔗
• (@gussmith23) Include some simple simplifying rewrites 🔗
• (@gussmith23) This is starting to become a flaky test... 🔗
• (@gussmith23) This is overflowing the stack 🔗
• (@gussmith23) This can be merged with the program itself. 🔗
• (@gussmith23) Include some simple simplifying rewrites 🔗
• (@gussmith23) I need a way to keep this in sync with the actual code 🔗
• (@gussmith23) This can be merged with the program itself. 🔗
• (@gussmith23) Rewrite to collapse move axis and transpose 🔗
• (@gussmith23) Change to just using transpose? 🔗
• (@gussmith23) I need a way to keep this in sync with the actual code 🔗
• (@gussmith23) This can be merged with the program itself. 🔗
• (@gussmith) Support batch dimension 🔗
• (@gussmith) Is this tolerance too big? 🔗
• (@gussmith23) This needs to be tested 🔗
• (@gussmith23) We shouldn't have to extract ANY computes! 🔗
• (@gussmith23) We shouldn't have to extract ANY computes! 🔗
• (@gussmith23) This needs to be tested 🔗
• (@gussmith23) Do this check in a more intelligent way. 🔗
• (@gussmith23) Do this check in a more intelligent way. 🔗
• (@gussmith23) potential bug 🔗
• (@gussmith23) the as f64 thing here is potentially a bug 🔗
• (@gussmith23) Terrible naming. Not related to extraction. 🔗
• (@gussmith23) This isn't always true and/or it doesn't matter 🔗
• (@gussmith23) This test doesn't have to depend on running CPLEX. 🔗
• (@gussmith23) Test incomplete 🔗
• (@gussmith23) This is a hack 🔗
• (@gussmith23) Turn on rustdoc lints 🔗
• (@gussmith23) hardcoded datatype 🔗
• (@gussmith23) Need to test w/ blocking 🔗
• (@gussmith23) Need to test w/ blocking 🔗
• (@gussmith23) hardcoded datatype 🔗
• (@gussmith23) Could add doctests 🔗
• (@gussmith23) Does not reason about ordering on hardware. 🔗
• (@gussmith23) Hardcoded to float32 🔗
• (@gussmith23) Assuming float32 output. 🔗
• (mike) : we probably could make codegen happen here 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• () : axis not used since 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• () : currently only supports maxpool2d shape in resnet 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) It would make our lives easier if we 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) These are just ripped from language.rs. 🔗
• (@gussmith23) how to assign unique names to each usage? 🔗
• (@gussmith23) Allocations should not be done ad-hoc 🔗
• (@gussmith23) Datatype assumption 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) Find a different way to name intermediates 🔗
• (@gussmith23) Needs test 🔗
• (@gussmith23) Do we want this RNG to use SEED? 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) Do we want this RNG to use SEED? 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) I think I should make this a conditional applier, and fold in 🔗
• (@gussmith23) make this cleaner 🔗
• (@gussmith23) should this be all or any or something else entirely? 🔗
• (@gussmith23) not sure all this should be public. 🔗
• (@gussmith23) left off here, I think I should actually do something 🔗
• (@gussmith23) I could write an even simpler rewrite rule that slices 🔗
• (@gussmith23) naming 🔗
• (@gussmith23) naming 🔗
• () 🔗
• (@gussmith23) naming 🔗
• (@gussmith23) I think we need more checks here, to make sure that the sizes 🔗
• (@gussmith23) this isn't the only way this could be done. 🔗
• (@gussmith23) would it make our pattern-matching life easier if (1) we 🔗
• (@gussmith23) check this 🔗
• (@gussmith23) Duplicated logic for cartprod shape calculation 🔗
• (@gussmith23) Test this 🔗
• (@gussmith23) This could be combined with the applier above. 🔗
• (@gussmith) Wouldn't need this if we had "access" op 🔗
• (@gussmith) Need to limit what gets sliced 🔗
• (@gussmith) Can also implement a collapse_nested_concatenate 🔗
• (@gussmith23) Test this 🔗
• (@gussmith23) Any constraints on these? 🔗
• (@gussmith23) Any constraints on these? 🔗
• (@gussmith23) This is a hack 🔗
• (@gussmith23) explain why... 🔗
• (gussmith23) on second thought, remove this; doesn't really do anything. 🔗
• (@gussmith23) Layout assumption 🔗
• (@gussmith23) Hardcoding to f32 🔗
• (@gussmith23) Hardcoding to f32 🔗
• (@gussmith23) Hardcoding to f32 🔗
• (@gussmith23) # random samples chosen arbitrarily 🔗
• (@gussmith23) Include some simple simplifying rewrites 🔗
• (@gussmith23) Do we want this RNG to use SEED? 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) Hardcoding test to f32 🔗
• (@gussmith23) ?axis should be 2 here, but we can't match an Int64 🔗
• (@gussmith23) I'm being lazy here 🔗
• (@gussmith23) Make sure TVM feature flag is getting tested in CI 🔗
• (@gussmith23) Give glenside-expression-creation helpers a new home 🔗
• (@gussmith23) Layout assumption 🔗
• (@gussmith23) Make grouped conv take advantage of new 🔗
• (@gussmith23) Layout assumption. 🔗
• (@gussmith23) layout assumption 🔗
• (@gussmith23) layout assumption 🔗
• (@gussmith23) Layout assumption 🔗
• (@gussmith23) Get rid of this hack eventually 🔗
• () , it might be that you have not 🔗
• (@gussmith23) Do we need to pass the recexpr into this closure? 🔗
• (@gussmith23) This is broken at the moment 🔗
• (@gussmith23) This is a hack 🔗
• (mike) : add operator support for these following 🔗
• (@gussmith23) How to actually constrain this? 🔗
• (mike) : support reducing on multiple axis? 🔗
• (@gussmith23) Layout assumption 🔗
• (@gussmith23) code is duplicated below. 🔗
• (@gussmith23) Hardcoding to f32 🔗
• (@gussmith23) Hardcoding to f32 🔗
• (@gussmith23) # random samples chosen arbitrarily 🔗
• (@gussmith23) Include some simple simplifying rewrites 🔗
• (@gussmith23) Do we want this RNG to use SEED? 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) output type assumption 🔗
• (@gussmith23) Hardcoding test to f32 🔗
• (@gussmith23) Relay/TVM doesn't seem to like nhwc w/o hwoi 🔗
• (@gussmith23) I'm being lazy here 🔗
• () : Uncomment out this test when reshape is implemented for the interpreter 🔗
• () : add tuple output support to test macro 🔗
• () : uncomment these tests once 🔗
• (@gussmith23) Interpreter stack overflows on large programs 🔗
• (@gussmith23) Trying out some new syntax... 🔗
• (@gussmith) What's going on here... 🔗
• (@gussmith23) More efficient way to do this? 🔗
• (@gussmith) Settle on vocab: "axis" or "dimension"? 🔗
• (@gussmith23) More efficient striding? 🔗
• (@gussmith23) Interpreter window striding is inefficient 🔗
• (@gussmith23) This seems dumb 🔗
• (@gussmith) Arbitrarily picked default access axis 🔗
• (@gussmith23) Does this already exist somewhere? 🔗
• (@gussmith23) Re-enable benchmarks using Criterion or the 🔗
• (@gussmith23) How to make the M argument "programmable"? 🔗
• (@gussmith23) do we need to specify rows and cols? You can infer these 🔗
• (@gussmith23) AccessWindows shouldn't be specific to filters. 🔗
• (@gussmith) Choose between ([Id; 1]) and (Id) and be consistent 🔗
• (@gussmith23) Implement access-slice-item 🔗
• (@gussmith23) Implement access-concatenate-item 🔗
• (@gussmith23) We need to just make a full-fledged Relay dialect. 🔗
• (@gussmith23) How to handle batch norms? 🔗
• (@gussmith23) We shouldn't always need this. Currently just used for 🔗
• (@gussmith) Multiple-arg compute feels clunky and ad-hoc. 🔗
• (@gussmith23) Add tests for flexasr-maxpool. 🔗
• (@gussmith23) Pick a better analysis name. 🔗
• (@gussmith23) Needed? 🔗
• (@gussmith23) Might want to replace this with full partial eval 🔗
• (@gussmith23) Pick a better analysis name. 🔗
• (@gussmith23) Make sure merge returns true infrequently 🔗
• (@gussmith23) Delete these 🔗
• () : add shape here 🔗
• (@gussmith23) This pattern appears a lot, and it's annoying. 🔗
• (@gussmith23) Assuming NCHW layout 🔗
• (@gussmith23) I'm just doing something arbitrary 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero_regions 🔗
• (@gussmith23) Implement zero_regions 🔗
• (@gussmith23) Remove this after figuring out padding issues 🔗
• (@gussmith23) Written in a rush. 🔗
• (@gussmith23) Remove this after figuring out padding issues 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) should really just have an "envelope" 🔗
• (@gussmith23) Implement zero_regions 🔗
• (@gussmith23) Implement zero_regions 🔗
• (@gussmith23) this should definitely not be commented out... 🔗
• (@gussmith23) DO NOT SUBMIT this is a hack b/c reshapes 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero_regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero_regions 🔗
• (@gussmith23) We only implement zero regions for 🔗
• (@gussmith23) How to access tensor literals? 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) have to figure out a way around this. 🔗
• (@gussmith23) Generalize AccessWindows to other accesses 🔗
• (@gussmith23) Implement zero regions 🔗
• (@gussmith23) Could probably clean this up with a for loop 🔗
• (@gussmith23) More tests of cart prod w/ padding 🔗
• (@gussmith) More access-insert-axis tests 🔗
• (@gussmith) More access-broadcast tests 🔗
• (@gussmith23) Add flags to control different rewrites 🔗
• (@gussmith23) temporary hack: prepend all the kernel definitions 🔗
• (@gussmith23) Add flags to control different rewrites 🔗
• (@gussmith23) Explain difference between extraction and hw gen 🔗