Fix SIMD scatter kernel: move simd_shuffle outside divergent control flow

[konard]

Summary

• Fix SIMD uniform control flow violation in radix_scatter_simd kernel
• Move simd_shuffle operations outside the if (valid) block to ensure all 32 threads in the SIMD group participate
• Update case study documentation with proper root cause analysis

Problem

The SIMD scatter kernel was failing verification because simd_shuffle was called inside a divergent if (valid) block. On Apple Silicon GPUs, all threads in a SIMD group (32 threads) must execute SIMD group operations together for correct behavior.

Root Cause

When array sizes don't evenly divide by 32 (the SIMD group size), some threads become "invalid" (they don't have data to process). The old code skipped the entire shuffle computation for invalid threads:

if (valid) {
    // Only valid threads execute simd_shuffle - WRONG!
    for (uint lane = 0; lane < simd_lane; lane++) {
        uint other_digit = simd_shuffle(digit, lane);
        // ...
    }
}

This caused undefined behavior according to Apple Developer Forums:

"For correct behaviour all threads in SIMD group should execute these instructions."

Solution

Move simd_shuffle loops outside the conditional block so all threads participate, while only counting matches for valid threads:

// All threads execute simd_shuffle together (uniform control flow)
for (uint lane = 0; lane < simd_lane; lane++) {
    uint other_digit = simd_shuffle(digit, lane);
    if (valid && other_digit == digit) {  // Only valid threads count
        simd_rank++;
    }
}

Invalid threads have digit = RADIX_SIZE (256) which won't match any valid digit (0-255), so they don't affect the counting.

Test Plan

• CI passes (compiles on Linux without Metal)
• Manual test on macOS with Apple Silicon: cargo run --release -- 2684354
• Verify SIMD radix sort produces correct sorted output
• Compare SIMD and basic kernel results match exactly

References

• Fixes ERROR: GPU radix sort (SIMD) failed verification! #17
• Previous fix attempt (incomplete): PR Fix SIMD scatter kernel rank computation bug #18
• Apple Developer Forums - simdgroup issues
• Apple G13 GPU Architecture Reference

🤖 Generated with Claude Code

[konard]

🤖 Solution Draft Log

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost estimation:

• Public pricing estimate: $3.753484 USD
• Calculated by Anthropic: $2.961328 USD
• Difference: $-0.792156 (-21.10%)
  📎 Log file uploaded as GitHub Gist (573KB)
  🔗 View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.