Rule. Avoid allocating and freeing dynamic heap memory (such as via malloc(), calloc(), aom_malloc(), or aom_memalign()) inside per-block, prediction, search, or inner-loop DSP/SIMD routines. For temporary buffers with bounded dimensions (e.g., maximum superblock size 128x128), use fixed-size stack-allocated arrays guarded by dimension assertions (e.g., assert(bw <= 128 && bh <= 128)). For large, aligned, or reusable scratch buffers (such as SIMD restoration unit stats buffers), hoist dynamic memory allocations to frame-, tile-, or search-level context structures (such as pick_lr_ctxt) and pass buffer pointers into processing kernels.
Why. Performing dynamic memory allocation and deallocation inside per-block or inner-loop functions introduces substantial heap allocator overhead, system calls, cache penalties, and error-handling failure paths in performance-critical code paths. Fixed-size stack arrays provide fast, zero-allocation local storage for bounded block buffers, while hoisting allocations to frame or context setup amortizes allocation overhead across all blocks and eliminates per-block failure checks.
Bad:
// Bad: Dynamic heap allocation for a bounded prediction block uint8_t *pred_block = (uint8_t *)malloc(bw * bh * sizeof(uint8_t)); if (!pred_block) return; generate_prediction(pred_block, bw, bh); free(pred_block); // Bad: Dynamic scratch memory allocation inside per-unit SIMD routine int16_t *dgd_avg = (int16_t *)aom_memalign( 32, sizeof(*dgd_avg) * 6 * RESTORATION_UNITSIZE_MAX * RESTORATION_UNITSIZE_MAX); // ... compute stats ... aom_free(dgd_avg);
Good:
// Good: Fixed-size stack array bounded by maximum superblock dimensions assert(bw <= 128 && bh <= 128); uint8_t pred_block[128 * 128]; const int pred_block_stride = 128; generate_prediction(pred_block, bw, bh, pred_block_stride); // Good: Hoist scratch allocations to frame or context setup structures CHECK_MEM_ERROR(cm, cpi->pick_lr_ctxt.dgd_avg, (int16_t *)aom_memalign(32, buf_size)); rsc.dgd_avg = cpi->pick_lr_ctxt.dgd_avg; void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd, const uint8_t *src, int16_t *dgd_avg, int16_t *src_avg, ...) { // Use passed-in scratch buffers directly without allocating }