ARTIQ/RTIO-performance-and-optimizations.md

@@ -7,9 +7,18 @@ This page briefly describes various observations, optimizations, and recommendat
 1. We did not observe a significant difference in performance when using _kernel function arguments_ over kernel invariants.
 1. Object attributes that are modified once or more during kernel execution are updated on the host once after the core device returns.
 1. We observed that the core device performance is very sensitive to changes in the program. Adding minor complexity (e.g. adding one local variable) might already measurable influence RTIO performance.
+1. Device initialization (e.g. for a 9912) can be very slow (i.e. > 1 ms) and such calls should be scheduled accordingly.
 
 ## Data types
 
+ARTIQ supports various data types including integers, floats, and arrays.
+
+1. 32-bit integer operations are preferred over 64-bit integer operations.
+1. Integer operations are preferred to expensive floating-point operations.
+1. It is possible to set the fast-math compile flag to allow aggressive optimizations on floating-point expressions at the cost of IEEE compliance.
+
+### Type inference and conversion
+
 ARTIQ implicitly infers most of the data types based on the program. Inferred data types are not always optimal.
 
 1. The Python `int` type often compiles to an `int64` type. We recommend to explicitly cast integer variables to the desired type using `np.int32` or `np.int64`.