Project 2: Nicholas Liu

.gitignore

@@ -25,7 +25,8 @@ build
 .LSOverride
 
 # Icon must end with two \r
-Icon
+Icon
+
 
 # Thumbnails
 ._*
@@ -560,3 +561,4 @@ xcuserdata
 *.xccheckout
 *.moved-aside
 *.xcuserstate
+.vscode

README.md

@@ -3,12 +3,44 @@ CUDA Stream Compaction
 
 **University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 2**
 
-* (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)
+* Nicholas Liu
+  * [Linkedin](https://www.linkedin.com/in/liunicholas6/)
+* Tested on: Linux Mint 22 Wilma, AMD Ryzen 7 5800X @ 2.512GHz, 32GB RAM, GeForce GTX 1660 Ti
 
-### (TODO: Your README)
+# Project Description
 
-Include analysis, etc. (Remember, this is public, so don't put
-anything here that you don't want to share with the world.)
+This project implements scan (prefix-sum) in a variety of different ways. We test a CPU implementation, a GPU naive version, and a GPU work-efficient version against the implementation from the Thrust library.
 
+# Performance Analysis
+
+At very small array sizes, the CPU scan performs the best. As we increase the size of the array, the CPU begins to perform relatively worse, and is quickly eclipsed in speed by the Thrust implementation.
+
+Error bars shown are 1 sigma with 1000 runs of scan.
+
+![chart: cpu better](img/cpu-better.png)
+
+Once we reach a number of elements on the order of 2^23 or so, the work-efficient scan implemented also performs better than the CPU scan
+
+![chart: gpu better](img/gpu-better.png)
+
+| 256     | CPU       | 7.585e-05           | 1.684521478821987e-05  |   |
+|---------|-----------|---------------------|------------------------|---|
+| 256     | Naive     | 0.019890368         | 0.018129259603148012   |   |
+| 256     | Efficient | 0.041029536         | 0.03196243203484814    |   |
+| 256     | Thrust    | 0.022883072         | 0.07034156396333269    |   |
+| 4096    | CPU       | 0.00092551          | 0.00017840902353771144 |   |
+| 4096    | Naive     | 0.02674304          | 0.005283869080783783   |   |
+| 4096    | Efficient | 0.053911488         | 0.0037359112841419267  |   |
+| 4096    | Thrust    | 0.0177352           | 0.009973741521411908   |   |
+| 65536   | CPU       | 0.017869165         | 0.003946497566300506   |   |
+| 65536   | Naive     | 0.099199506         | 0.18741031849147752    |   |
+| 65536   | Efficient | 0.15358839999999999 | 0.34961337456361957    |   |
+| 65536   | Thrust    | 0.089322486         | 0.3254493673567683     |   |
+| 1048576 | CPU       | 0.27509278          | 0.048449150282649026   |   |
+| 1048576 | Naive     | 0.814048116         | 0.04939857649465461    |   |
+| 1048576 | Efficient | 0.520357476         | 0.11646960803214108    |   |
+| 1048576 | Thrust    | 0.199752738         | 0.07898137835466129    |   |
+| 8388608 | CPU       | 4.41064898          | 1.176601505235431      |   |
+| 8388608 | Naive     | 7.03792806          | 0.37055519290846245    |   |
+| 8388608 | Efficient | 3.77427417          | 0.2990962226148862     |   |
+| 8388608 | Thrust    | 0.519275972         | 0.20947768302315156    |   |

analysis/project/graph.py

@@ -0,0 +1,58 @@
+import matplotlib.pyplot as plt
+import csv
+import os
+import statistics
+from collections import defaultdict
+import numpy as np
+
+data_path = "../data"
+
+scan_types = ["CPU", "Naive", "Efficient", "Thrust"]
+array_sizes = []
+ys = {}
+errs = {}
+
+for scan_type in scan_types:
+    ys[scan_type] = []
+    errs[scan_type] = []
+
+filenames = [filename for filename in os.listdir(data_path)]
+filenames.sort(key=lambda name: int(name.split('.')[0]))
+# filenames = filenames[:-1]
+# filenames = [filenames[-1]]
+
+for filename in filenames:
+    array_sizes.append(1 << int(filename.split('.')[0]))
+    with open(data_path + "/" + filename) as file:
+        reader = csv.reader(file)
+        for (scan_type, line) in zip(scan_types, reader):
+            contents = [float(val) for val in line if val != '']
+            mean = statistics.mean(contents)
+            stdev = statistics.stdev(contents)
+            ys[scan_type].append(mean)
+            errs[scan_type].append(stdev)
+
+            print(array_sizes[-1], scan_type, mean, stdev)
+
+# xs = np.arange(len(array_sizes))
+# fig, ax = plt.subplots()
+# bar_width = 0.1
+
+# for (index, scan_type) in enumerate(scan_types):
+#     ax.bar(
+#         xs + (index - len(scan_types) / 2) * bar_width, ys[scan_type], bar_width, label = scan_type, yerr=errs[scan_type]
+#     )
+
+# ax.legend()
+
+# # ax.set_ylim(0, 1)
+# ax.set_ylabel("Milliseconds")
+
+# # ax.set_xlabel("Array size")
+# ax.set_xticks([])
+# # ax.set_xticks(xs)
+# # ax.set_xticklabels(array_sizes)
+
+# ax.set_title("Time to run scan on array of 2^23")
+
+# plt.savefig('chart.png')