Module 9: Performance & Optimization

Why is NumPy so fast? Most beginners use NumPy, but experts know how to optimize it for maximum performance. This module bridges that gap.

---

Lesson 19: Vectorization

Vectorization is the process of replacing explicit Python loops with array expressions.

Why standard loops are slow:

1. Type Checking: Python checks the type of every object in every iteration.
2. Dynamic Dispatch: Python looks up how to perform operations (like +) every single time.

Why Vectorization is fast:

1. Compiled Code: Operations happen in pre-compiled C/Fortran blocks.
2. CPU Optimization: NumPy uses SIMD (Single Instruction, Multiple Data) to process multiple numbers in one CPU cycle.

---

Lesson 20: Memory Management

Views vs. Copies (Review)

Creating copies (.copy()) consumes memory. Using views (via slicing or .ravel()) keeps your memory footprint small.

Memory Strides (Advanced)

A Stride is the number of bytes the CPU must skip in memory to reach the next element in an axis.

• Contiguous arrays (where data is side-by-side) are the fastest.
• You can check this with arr.strides.

Choosing the right dtype

Always use the smallest dtype possible.

• float64 uses 8 bytes per number.
• float32 uses 4 bytes.
• For huge datasets, this difference can mean being able to load the data into RAM vs. crashing.

---

Practice: Performance Comparison

Challenge: Using the time module in the editor above, compare the time it takes to find the square root (np.sqrt) of an array of 5 million items vs. using math.sqrt inside a for loop.

Key Takeaways

✅ Vectorization is mandatory for high-performance Python.
✅ Choose float32 over float64 to halve your memory usage if high precision isn't critical.
✅ Be aware of views vs copies to avoid wasting RAM.