Memory API

#include "aicraft/memory.h"

Arena Allocator

Aicraft uses a custom arena allocator for all tensor memory. This provides:

• O(1) allocation — simple bump pointer
• O(1) deallocation — restore to checkpoint
• Zero fragmentation — contiguous memory block
• Constant memory during training loops

Initialisation

ac_init();                     // Default arena (16 MB)
ac_init_with_arena(64 * 1024); // Custom arena size (64 KB)

Checkpoint / Restore

AcMemCheckpoint cp = ac_mem_checkpoint();

// Allocate tensors, run forward/backward...
AcTensor *x = ac_tensor_rand((int[]){1, 784}, 2);
AcTensor *y = ac_forward_seq(net, 2, x);

ac_mem_restore(cp);  // Free everything since checkpoint

Memory Query

size_t used  = ac_mem_used();      // Bytes currently allocated
size_t total = ac_mem_capacity();  // Total arena size
size_t free  = ac_mem_free();      // Remaining capacity

Why Not malloc?

Feature	Arena	malloc/free
Allocation speed	O(1) bump	O(n) search
Deallocation	O(1) restore	O(1) per free
Fragmentation	None	Grows over time
Bulk free	Checkpoint/restore	Manual tracking
Overhead	0 bytes/alloc	16-32 bytes/alloc

Training Loop Pattern

for (int epoch = 0; epoch < N; epoch++) {
    for (int batch = 0; batch < batches; batch++) {
        AcMemCheckpoint cp = ac_mem_checkpoint();

        // All intermediate tensors allocated here
        AcTensor *pred = ac_forward_seq(net, 2, x);
        AcTensor *loss = ac_cross_entropy(pred, target);
        ac_backward(loss);
        ac_optimizer_step(opt);

        ac_mem_restore(cp);  // Memory usage stays constant
    }
}