SIMD Optimization Skill

Parallelize tensor and array operations using SIMD.

When to Use

• Optimizing tensor operations
• Vectorizing element-wise computations
• Performance-critical loops (>1000 elements)
• Benchmark results show optimization potential

Quick Reference

from sys.info import simdwidthof

comptime width = simdwidthof[DType.float32]()

# SIMD vector add
for i in range(0, size, width):
    result.store(i, a.load[width](i) + b.load[width](i))

Workflow

1. Identify bottleneck - Profile code to find hot loops
2. Get SIMD width - Use simdwidthof[dtype]()
3. Vectorize loop - Process width elements per iteration
4. Handle remainder - Process leftover elements
5. Benchmark - Verify performance improvement (4x-8x expected)

Mojo-Specific Notes

• SIMD width varies by CPU and dtype (usually 8-16 for float32)
• Always handle remainder elements with scalar loop
• Prefer alias for compile-time SIMD width constants
• Test on target hardware - SIMD width is platform-specific

Error Handling

| Error | Cause | Solution | |-------|-------|----------| | Out of bounds | Remainder not handled | Add scalar remainder loop | | No speedup | Wrong SIMD width | Use simdwidthof[dtype]() | | Compilation fails | Type mismatch | Check load/store types match | | Segfault | Misaligned access | Ensure stride is correct |

References

• .claude/shared/mojo-guidelines.md - SIMD patterns section
• Mojo manual: SIMD documentation