shithub: libvpx

Info  •  Files  •  Log  •  Branches

Tune tokenization/detokenization flow for speed-up

This commit optimizes the tokenization and detokenization operational
flow for speed-up. It makes the coding process about 0.3% faster at
speed 0.

Change-Id: I28008df7482874e4b5f237f2d418ff82a249dd56

--- a/vp9/decoder/vp9_detokenize.c

+++ b/vp9/decoder/vp9_detokenize.c

@@ -118,22 +118,22 @@

     }

     case TX_8X8: {

       scan = get_scan_8x8(get_tx_type_8x8(type, xd));

-      above_ec = (A[0] + A[1]) != 0;

-      left_ec = (L[0] + L[1]) != 0;

+      above_ec = !!*(uint16_t *)A;

+      left_ec  = !!*(uint16_t *)L;

       band_translate = vp9_coefband_trans_8x8plus;

       break;

     }

     case TX_16X16: {

       scan = get_scan_16x16(get_tx_type_16x16(type, xd));

-      above_ec = (A[0] + A[1] + A[2] + A[3]) != 0;

-      left_ec = (L[0] + L[1] + L[2] + L[3]) != 0;

+      above_ec = !!*(uint32_t *)A;

+      left_ec  = !!*(uint32_t *)L;

       band_translate = vp9_coefband_trans_8x8plus;

       break;

     }

     case TX_32X32:

       scan = vp9_default_scan_32x32;

-      above_ec = (A[0] + A[1] + A[2] + A[3] + A[4] + A[5] + A[6] + A[7]) != 0;

-      left_ec = (L[0] + L[1] + L[2] + L[3] + L[4] + L[5] + L[6] + L[7]) != 0;

+      above_ec = !!*(uint64_t *)A;

+      left_ec  = !!*(uint64_t *)L;

       band_translate = vp9_coefband_trans_8x8plus;

       break;

   }

--- a/vp9/encoder/vp9_tokenize.c

+++ b/vp9/encoder/vp9_tokenize.c

@@ -136,9 +136,7 @@

   const int eob = xd->plane[plane].eobs[block];

   const PLANE_TYPE type = xd->plane[plane].plane_type;

   const int16_t *qcoeff_ptr = BLOCK_OFFSET(xd->plane[plane].qcoeff, block, 16);

-  const BLOCK_SIZE_TYPE sb_type = (mbmi->sb_type < BLOCK_SIZE_SB8X8) ?

-                                   BLOCK_SIZE_SB8X8 : mbmi->sb_type;

-  const int bwl = b_width_log2(sb_type);

+  const int bwl = b_width_log2(bsize);

   const int off = block >> (2 * tx_size);

   const int mod = bwl - tx_size - xd->plane[plane].subsampling_x;

   const int aoff = (off & ((1 << mod) - 1)) << tx_size;

@@ -168,22 +166,22 @@

       band_translate = vp9_coefband_trans_4x4;

       break;

     case TX_8X8:

-      above_ec = (A[0] + A[1]) != 0;

-      left_ec = (L[0] + L[1]) != 0;

+      above_ec = !!*(uint16_t *)A;

+      left_ec  = !!*(uint16_t *)L;

       seg_eob = 64;

       scan = get_scan_8x8(get_tx_type_8x8(type, xd));

       band_translate = vp9_coefband_trans_8x8plus;

       break;

     case TX_16X16:

-      above_ec = (A[0] + A[1] + A[2] + A[3]) != 0;

-      left_ec = (L[0] + L[1] + L[2] + L[3]) != 0;

+      above_ec = !!*(uint32_t *)A;

+      left_ec  = !!*(uint32_t *)L;

       seg_eob = 256;

       scan = get_scan_16x16(get_tx_type_16x16(type, xd));

       band_translate = vp9_coefband_trans_8x8plus;

       break;

     case TX_32X32:

-      above_ec = (A[0] + A[1] + A[2] + A[3] + A[4] + A[5] + A[6] + A[7]) != 0;

-      left_ec = (L[0] + L[1] + L[2] + L[3] + L[4] + L[5] + L[6] + L[7]) != 0;

+      above_ec = !!*(uint64_t *)A;

+      left_ec  = !!*(uint64_t *)L;

       seg_eob = 1024;

       scan = vp9_default_scan_32x32;

       band_translate = vp9_coefband_trans_8x8plus;

@@ -225,7 +223,7 @@

       if (!t->skip_eob_node)

         ++cpi->common.counts.eob_branch[tx_size][type][ref][band][pt];

     }

-    token_cache[scan[c]] = vp9_pt_energy_class[token];

+    token_cache[rc] = vp9_pt_energy_class[token];

     ++t;

   } while (c < eob && ++c < seg_eob);

@@ -233,9 +231,8 @@

   if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0) {

     set_contexts_on_border(xd, bsize, plane, tx_size, c, aoff, loff, A, L);

   } else {

-    for (pt = 0; pt < (1 << tx_size); pt++) {

-      A[pt] = L[pt] = c > 0;

-    }

+    vpx_memset(A, c > 0, sizeof(ENTROPY_CONTEXT) * (1 << tx_size));

+    vpx_memset(L, c > 0, sizeof(ENTROPY_CONTEXT) * (1 << tx_size));

   }

 }