advect¶

Differentiable Physics Kernel using NVIDIA Warp. Supports Inverse Design via gradients.

Classes¶

AdvectEngine

Functions¶

traffic_step_kernel(positions, velocities, ...)

Module Contents¶

traffic_step_kernel(positions, velocities, traffic_light_pos, traffic_light_timings, time, dt, sharpness, bounds_min, bounds_max)¶

class AdvectEngine(num_agents=100000)¶

num_agents = 100000¶

device = 'cpu'¶

bounds_min = (0.0, 0.0)¶

bounds_max = (5000.0, 5000.0)¶

current_time = 0.0¶

pos_buffer¶

vel_buffer¶

traffic_lights_pos¶

traffic_lights_timings¶

tape = None¶

step(time, dt=0.1)¶

Run differentiable step with Tape.

tick(dt=1.0 / 60.0)¶

Compatiblity method for Refinery/Sentinel.

backward_pass(loss_tensor)¶

Triggers backward propagation to compute gradients for traffic lights.