# HG changeset patch # User darkshikari # Date 1280885888 0 # Node ID 18117b5bb7dced812a9ebe0b43cec36f6eff28ba # Parent 565ccbf260151b48b6458a062ca3d6d696ed2897 VP8: simplify decode_block_coeffs to avoid having to track nonzero coeffs Slightly faster. diff -r 565ccbf26015 -r 18117b5bb7dc vp8.c --- a/vp8.c Wed Aug 04 00:08:16 2010 +0000 +++ b/vp8.c Wed Aug 04 01:38:08 2010 +0000 @@ -826,17 +826,20 @@ int i, int zero_nhood, int16_t qmul[2]) { uint8_t *token_prob = probs[i][zero_nhood]; - int nonzero = 0; int coeff; + if (!vp56_rac_get_prob_branchy(c, token_prob[0])) + return 0; + goto skip_eob; + do { if (!vp56_rac_get_prob_branchy(c, token_prob[0])) // DCT_EOB - return nonzero; + return i; skip_eob: if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0 if (++i == 16) - return nonzero; // invalid input; blocks should end with EOB + return i; // invalid input; blocks should end with EOB token_prob = probs[i][0]; goto skip_eob; } @@ -870,13 +873,10 @@ } token_prob = probs[i+1][2]; } + block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i]; + } while (++i < 16); - // todo: full [16] qmat? load into register? - block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i]; - nonzero = ++i; - } while (i < 16); - - return nonzero; + return i; } static av_always_inline @@ -910,9 +910,9 @@ // luma blocks for (y = 0; y < 4; y++) for (x = 0; x < 4; x++) { - nnz_pred = l_nnz[y] + t_nnz[x]; + nnz_pred = l_nnz[y] + t_nnz[x];START_TIMER; nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start, - nnz_pred, s->qmat[segment].luma_qmul); + nnz_pred, s->qmat[segment].luma_qmul);STOP_TIMER("test"); // nnz+block_dc may be one more than the actual last index, but we don't care s->non_zero_count_cache[y][x] = nnz + block_dc; t_nnz[x] = l_nnz[y] = !!nnz;