University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 0
- Shubham Sharma
- Tested on: Windows 10, i7-9750H @ 2.26GHz, 16GB, GTX 1660ti 6GB (Personal Computer). *GPU Compute Capability: 7.5
Several techniques were implemented in this research to simulate the movements of Boid particles. We started with a rudimentary implementation and then tweaked it to increase parallelism and reduce memory requests.
Simulation Begins: Early on in execution the boids are scattered randomly throughout the space
Simulation Continues: Boids begin to travel in flocks with their neighbors forming groups
Simulation Begins
Boids Flocking
Simulation Begins
Boids Flocking
Boids move according to three rules. These were implemented across threads to allow for a level of parallelism across GPU cores.
-Boids try to fly towards the centre of mass of neighboring boids. -Boids try to keep a small distance away from other objects (including other boids). -Boids try to match velocity with near boids.
In the first iteration of the implementation, each boid checked all other boids naively to see how the boid in question should modify its velocity.
We don't need to examine all other boids every time step because boids are only influenced by their neighbors. Instead, we can figure out how to see which boids are in close proximity.
We finish this by setting up a uniform network of cells over the recreation space, breaking it up into cubic locales. These locales can be conceptualised as Frameworks. Presently for each boid we must decide which network cell contains the boid. As a last preprocessing step, we are able walk through the recently sorted framework cell esteem cluster to compute beginning and finishing positions for each cell that's valuable when performing the boid computations.
##2.3 Grids with Memory Coherency In the uniform grid section, for each boid we put away the record of where we can discover its position and speed. Presently, we make an improvement which permits us to utilize a similar list inside the cell exhibit. This lessens the quantity of redirections and arbitrary gets to that we need to perform.
FPS is utilized as the essential measurement of execution investigation. Edge are uncapped (v-sync off), which implies the reproduction is running at its maximum value.
Below Graphs Represent recoreded FPS with respect to the number of boids. Below data is averaged out over 10 seconds of execution.
- For Each Implementation increasing the number of Boids increase the number of checks for a boid to calculate its resultant properties. In case of Naive approach each boid is compared with every other boid leading to the worst time complexity of O(N^2). Therefore accross all the implementations we can clearly see a decrease in performance.
- Changing the block count does not result in change of Performance. Since the kernels get divided up across the SMs on the GPU and each SM can run one warp at a time, there is neither an increase or decrease in performance.
- With the coherent uniform grid Implementation the performance increases greatly. This behaviour can be attributed toward faster access time from the memory since the data is stored in contiguos memore locations and the intermediary step of memory redicrection was removed.
- In our current Scenario decreasing the cell weidth resulted in an increase in performance. This can be attributed to decrease in the bouding volume of a grid cube. In the scenario where boids are more dispersed we end up checking lesser boids as compared to before but in a scenario where boids areaccumulated closer to each other this behaviour may change.
I have deleted old Videos and added Gifs as Github can showcase them, also updated the images and changed a negative sign for one of the rules in my kernel.cu code. Please dont count this commit towards late submission as i have already consulted Prof. Shehzan Mohammed regarding this. Thank you!