Remove padding agents using interface entities

[aaravpandya]

This is a big PR since remove the padding agents impacts almost everything.

We introduce new archetypes AgentInterface and RoadInterface with the components that we want to export. To fill in the observation, we simply use new structs AgentInterfaceEntity and RoadInterfaceEntity that store a a reference to these interface entities. While collecting observations, we can simply get the corresponding arrays and fill the values while ignoring the padded values.

While we make as many entities from these new archetypes as specified by consts, they do not have physics components and are not registered with the BVH. Also, these entities are not iterated through in the task graph. The effect is that the taskgraph is only run for the actual number of agents in the Sim and we simply reference the "export components" by id and fill in the observations.

[aaravpandya]

I am getting incorrect values in this branch while executing on CUDA. The values are exactly correct in every index for CPU exec mode. I am unsure of how to debug this.

What this PR tries to do - Before this change, we simply had Agent and PhysicsEntity. We had to create padding agents and physicsentities to make up for irregular number of entities in the dataset per world. To remove the padding agents, I introduced AgentInterface and RoadInterface archetypes. Now we only create as many agents as are needed per world, but have consts::kMaxAgentCount number of AgentInterface entities. Each agent stores a reference to an AgentInterface using a AgentInterfaceEntity struct. We fill in the observations in the AgentInterface which are then exported, while keeping zeros/ones in the padded AgentInterface entities depending on semantics. This is designed in a similar way as gpu_hideseek but in the reverse manner. The Agent stores the reference to AgentInterface.
The relevant code for entity creation is in level_gen.cpp::343 and the Archetypes are defined in types.hpp::241.
After the Agents are created, we make sure we have consts::kMaxAgentCount number of AgentInterface here - level_gen.cpp::299

How to reproduce the error - Simply run headless on CPU and then CUDA using ./headless CPU 1 for 1 step. Print any tensor (I am printing Done tensor in here as it is very easy to infer as it is a scalar value). The values ouput by the CPU exec mode are correct. The values output by CUDA are not correct.

Also, ctest can be run by changing exec modes. The tests pass on CPU but fail on CUDA.

I am not sure how to debug this or where the error could even reside.

@eugenevinitsky

[aaravpandya]

@shacklettbp Hi Brennan, would you mind taking a look ? I am not sure how to proceed with this issue.

For reference, this is how the output looks like and I think the main issue is that CUDA exec mode is returning different values than CPU. The CPU values are correct.

(madrona) (base) aarav@emerge2-desktop:~/gpudrive/build$ ./headless CPU 1
Done
[ 0 0 1 1 1 0 0 0 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ]
[ 1 0 0 1 1 1 1 1 0 1 1 1 1 0 0 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ]
FPS 3377.298340
Agent-Normalized FPS 178996.812500
(madrona) (base) aarav@emerge2-desktop:~/gpudrive/build$ ./headless CUDA 1
Initialization finished
Done
[ 0 0 0 0 1 0 1 0 1 0 1 0 0 0 0 0 1 0 1 0 1 0 1 0 1 0 0 0 0 0 1 0 1 1 1 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ]
[ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ]
[ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ]
FPS 1700.869263
Agent-Normalized FPS 0.000000

[shacklettbp]

You need to sort the new interface entities when using the GPU backend. It doesn't look like you're doing that currently. I assume this could be done once during the initialization task graph.

[eugenevinitsky]

Well that was quick, thanks!