2023-11-15 09:09:13 PM

topics/metal-splats.md

@@ -11,6 +11,32 @@
 - [WIP PR implementing ML metal compute kernels in HF Candle](https://github.com/huggingface/candle/pull/1230/files)
 - [Good slides on bitonic sorting](https://wiki.rice.edu/confluence/download/attachments/4435861/comp322-s12-lec28-slides-JMC.pdf?version=1&modificationDate=1333163955158) (linked from [here](https://people.cs.rutgers.edu/~venugopa/parallel_summer2012/bitonic_overview.html))
 
+## 2023-11-15
+
+* Think I have a handle on the bitonic sorting algorithm! Will start playing around with a CPU version before going for the GPU version.
+* Implemented a working CPU version after a couple of bugs. Did you know C++ vectors don't do bounds checks by default? I didn't and nor did my professional C++ programmer friend.
+
+Bitonic sort is ~3x slower than the other algorithms at this data scale:
+
+```
+Generating 1048576 random integers
+Generated 1048576 random integers
+std::sort() execution time: 29 ms
+sort_radix() execution time: 34 ms
+sort_bitonic() execution time: 105 ms
+```
+
+Extending to about 10x slower at higher data scale:
+
+```
+Generating 16777216 random integers
+Generated 16777216 random integers
+std::sort() execution time: 296 ms
+sort_radix() execution time: 523 ms
+sort_bitonic() execution time: 2297 ms
+```
+
+
 ## 2023-11-12
 
 ### Algorithms / Reference