Prepare L21 flipped note

lectures/flipped/L21.md

@@ -25,13 +25,19 @@ taking a taxi or rideshare to and from the airport.
 
 ## Data parallelism (10 minutes)
 
+* The idea is that we evaluate a function, or kernel, at a set of points. The
+  set of points is also named *index space*. Each of the points corresponds to a
+  *work item*.
+
 * The unit of execution in a GPU is called a *warp*, which executes SIMT (Single
   Instruction Multiple Thread) instructions. There can be multiple units of
-  execution in a given GPU
+  execution in a given GPU.
 
 * Each data element is a work item. CUDA spawns a thread for each work item,
   with a unique thread ID; they are grouped into blocks (see [Grid of Thread
-  Blocks](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#thread-hierarchy-grid-of-thread-blocks))
+  Blocks](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#thread-hierarchy-grid-of-thread-blocks)).
+  Blocks can share memory locally, but each block has to be able to execute
+  independently
 
 * Discuss the memory types (A useful material may be
   [link](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#memory-hierarchy))
@@ -42,17 +48,20 @@ taking a taxi or rideshare to and from the airport.
   * constant memory (read-only, , optimized for different memory usages)
 
 * Branches. In practice, the hardware will execute all branches that any thread
-  in a warp executed (due to SIMT)
+  in a warp executed (due to SIMT).
+  [This](https://stackoverflow.com/questions/17223640/is-branch-divergence-really-so-bad)
+  may help.
 
-* Atomic functions
+* Atomic functions. For example, two work items somehow touch the same memory
+  location.
 
 ## Live coding: kernels (10 minutes)
 
-* show the simple sum kernel and run it (see `live-coding/L22/`)
+* show the simple sum kernel and run it (see `live-coding/L21/`)
 
 ## Activity
 
-TODO: find a better example, maybe the password cracking example from L24.
+TODO: find a better example
 
 Project the N-body problem code from L22. Each student is a work-item. Assign a
 grid number to each student on the 2D grid induced by the classroom layout. So,