apple-bottom is a numerical computing library. Security considerations include:

• Memory safety: Buffer overflows, use-after-free, double-free
• Integer overflows: In matrix dimension calculations
• Resource exhaustion: GPU memory allocation failures
• Data integrity: Numerical precision guarantees

This library does NOT handle:

• Network communication
• User authentication
• File system access, except where a diagnostic variable asks for it: AB_TRACE and AB_PROFILE_JSON write files at the paths they name, and AB_DISPATCH_TABLE reads one. The Metal shaders are compiled from source built into the library.
• Cryptographic operations

If you discover a security vulnerability, please report it responsibly:

1. DO NOT open a public issue
2. Email: grantdheileman@gmail.com
3. Include:
   • Description of the vulnerability
   • Steps to reproduce
   • Potential impact
   • Suggested fix (if any)

• Acknowledgment: Within 48 hours
• Initial assessment: Within 1 week
• Fix timeline: Depends on severity
  • Critical: 24-48 hours
  • High: 1 week
  • Medium: 2 weeks
  • Low: Next release

Always validate matrix dimensions before calling apple-bottom:

// Good
if (ab_matrix_check_size(rows, cols) == AB_OK) {  // rows * cols within limits
    ABMatrix m = ab_matrix_create(rows, cols);
}

// Bad - no validation
ABMatrix m = ab_matrix_create(user_input_rows, user_input_cols);

Check available GPU memory before large allocations:

// Each NxN matrix uses 8*N*N bytes (DD format)
// 8192x8192 = 512MB per matrix
size_t required = 8 * rows * cols;
// Ensure this fits in available unified memory

Always check return values:

ABStatus status = ab_dgemm(A, B, C);
if (status != AB_OK) {
    // Handle error appropriately
    fprintf(stderr, "DGEMM failed: %s\n", ab_status_string(status));
}

1. Threads: ab_init() and ab_shutdown() are safe to call from any thread, and calls on different matrices may come from different threads; one matrix, pool or session must not be used by two threads at once
2. One command queue: all GPU work shares one Metal command queue
3. Unified memory: Large matrices compete with system RAM