Rule. In SIMD or DSP accumulation kernels (such as variance, SAD, SSE, or sum of squares), check theoretical maximum sums against accumulator bit-depth limits across discrete block sizes. If the maximum sum cannot overflow 32-bit vector registers for any supported size, eliminate chunking loops and accumulate directly in 32-bit registers with a single post-loop 64-bit reduction. When accumulator overflow is only possible for the largest discrete block size (e.g., 128x128), branch directly on that size and use fixed chunk constants instead of dynamic pixel multiplication and division.
Why. Intermediate chunking sub-loops, horizontal reductions, and dynamic dimension calculations (w * h, limit / w) introduce unnecessary arithmetic, branching, and register shuffling in hot DSP loops. Determining overflow boundaries statically against known discrete block sizes maximizes vector pipeline throughput and avoids runtime division.
Bad:
// Bad: Unnecessary intermediate chunking and reduction when max sum fits in 32-bit do { __m128i v_acc_q = _mm_setzero_si128(); const int step = AOMMIN(height - r, 32); for (int j = 0; j < step; j += 4) { ... } v_acc_q = _mm_add_epi32(v_acc_q, _mm_srli_epi64(v_acc_q, 32)); sum += (uint64_t)_mm_cvtsi128_si32(v_acc_q); } while (r < height); // Bad: Dynamic pixel multiplication and division when only 128x128 can overflow } else if (w * h <= max_pel_allowed_per_ovf) { hbd_obmc_variance_w8n(pre, stride, wsrc, mask, &sse, &sum, w, h); } else { int h_per_ovf = max_pel_allowed_per_ovf / w; hbd_obmc_variance_w8n(pre, stride, wsrc, mask, &sse, &sum, w, h_per_ovf); }
Good:
// Good: Single continuous accumulation loop with post-loop 64-bit widening do { const __m128i v_acc_d = sum_squares_i16_4x4_sse2(src, stride); v_acc_q = _mm_add_epi32(v_acc_q, v_acc_d); src += stride << 2; r += 4; } while (r < height); return (uint64_t)xx_cvtsi128_si64(reduce_to_64(v_acc_q)); // Good: Direct dimension check with fixed constant chunk height } else if (w < 128 || h < 128) { hbd_obmc_variance_w8n(pre, stride, wsrc, mask, &sse, &sum, w, h); } else { assert(w == 128 && h == 128); hbd_obmc_variance_w8n(pre, stride, wsrc, mask, &sse, &sum, w, 64); }