Project 1: Yichao Wang

README.md

@@ -1,11 +1,32 @@
 **University of Pennsylvania, CIS 565: GPU Programming and Architecture,
 Project 1 - Flocking**
 
-* (TODO) YOUR NAME HERE
-  * (TODO) [LinkedIn](), [personal website](), [twitter](), etc.
-* Tested on: (TODO) Windows 22, i7-2222 @ 2.22GHz 22GB, GTX 222 222MB (Moore 2222 Lab)
+* Yichao Wang
+  * [LinkedIn](https://www.linkedin.com/in/wangyic/) 
+* Tested on: Windows 10 Home 64-bit (10.0, Build 18363)
+  * Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz (8 CPUs)
+  * GeForce GTX 1060	6.1
 
-### (TODO: Your README)
+**Boids Animation**
 
-Include screenshots, analysis, etc. (Remember, this is public, so don't put
-anything here that you don't want to share with the world.)
+![](images/screenshot.PNG)
+
+![](images/demo.gif)
+
+**Performance Analysis**
+
+![Figure1](images/withoutV.png)
+
+![](images/withV.png)
+
+![](images/blockSize.png)
+
+From above plots, we can see:
+
+1. For each implementation, as the # of boids increases, the performance descreases. This is because we need to process more data as the # of boids increases.
+
+2. For each implementation, changing of block size does not affect performance. This is because changing of block size does not change the amount of threads we used to process data.
+
+3. The performance improves with the more coherent uniform grid. Because we reduce the random access by eliminating ```dev_particleArrayIndices``` and use contiguous memory by ordering boid's position and velocity in the coherent kernel.
+
+4. Surprisingly, checking 27 neighboring has better performance than checking 8 neighboring. This is because we have many if statements when checking 8 neighboring. The if statements make the branch in warp diverge which results in more run time.