mojo-simd-optimize
Apply SIMD optimizations to Mojo code for parallel computation. Use when optimizing performance-critical tensor and array operations.
When & Why to Use This Skill
This Claude skill automates the application of SIMD (Single Instruction, Multiple Data) optimizations to Mojo code, specifically targeting performance-critical tensor and array operations. It streamlines the process of vectorizing loops and managing hardware-specific SIMD widths to achieve significant computational speedups, typically ranging from 4x to 8x.
Use Cases
- Accelerating Tensor Math: Automatically transform standard loops into SIMD-optimized versions for faster machine learning inference and mathematical computations in Mojo.
- Performance Tuning: Refactor element-wise array computations to leverage hardware-level vectorization in performance-critical applications.
- High-Volume Data Processing: Efficiently parallelize loops containing over 1000 elements by utilizing platform-specific SIMD widths and 'simdwidthof' constants.
- Robust Code Refactoring: Implement safe SIMD patterns that include mandatory scalar fallback loops to handle remainder elements and prevent out-of-bounds errors.
| name | mojo-simd-optimize |
|---|---|
| description | "Apply SIMD optimizations to Mojo code for parallel computation. Use when optimizing performance-critical tensor and array operations." |
| mcp_fallback | none |
| category | mojo |
| agent | test-engineer |
| user-invocable | false |
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
- Identify bottleneck - Profile code to find hot loops
- Get SIMD width - Use
simdwidthof[dtype]() - Vectorize loop - Process
widthelements per iteration - Handle remainder - Process leftover elements
- 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
aliasfor 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