shithub: libvpx

Info  •  Files  •  Log  •  Branches

Reduce bmi buffer length from 16 to 4

This commit removes the use of bmi_ in the first-pass encoding by
forcing encode_intra4x4block_ to use DC_PRED, followed by DCT_DCT
only, as John suggested. This makes the need for bmi buffer only
up to 4 entries, instead of 16.

Change-Id: I3410007dfae789ee46a09ae20c39d3ce3c7954aa

--- a/vp9/common/vp9_blockd.h

+++ b/vp9/common/vp9_blockd.h

@@ -254,7 +254,7 @@

 typedef struct {

   MB_MODE_INFO mbmi;

-  union b_mode_info bmi[16];

+  union b_mode_info bmi[4];

 } MODE_INFO;

 struct scale_factors {

--- a/vp9/encoder/vp9_encodeintra.c

+++ b/vp9/encoder/vp9_encodeintra.c

@@ -31,19 +31,17 @@

   } else {

     int i;

-    for (i = 0; i < 16; i++) {

-      x->e_mbd.mode_info_context->bmi[i].as_mode.first = B_DC_PRED;

+    for (i = 0; i < 16; i++)

       encode_intra4x4block(x, i, BLOCK_SIZE_MB16X16);

-    }

   }

   return vp9_get_mb_ss(x->plane[0].src_diff);

 }

+// This function is used only by the firstpass encoding.

 static void encode_intra4x4block(MACROBLOCK *x, int ib,

                                  BLOCK_SIZE_TYPE bsize) {

   MACROBLOCKD * const xd = &x->e_mbd;

-  TX_TYPE tx_type;

   uint8_t* const src =

       raster_block_offset_uint8(xd, bsize, 0, ib,

                                 x->plane[0].src.buf, x->plane[0].src.stride);

@@ -58,26 +56,17 @@

   assert(ib < (1 << (bwl + bhl)));

-  vp9_intra4x4_predict(&x->e_mbd, ib, bsize,

-                       xd->mode_info_context->bmi[ib].as_mode.first,

+  vp9_intra4x4_predict(&x->e_mbd, ib, bsize, DC_PRED,

                        dst, xd->plane[0].dst.stride);

   vp9_subtract_block(4, 4, src_diff, 4 << bwl,

                      src, x->plane[0].src.stride,

                      dst, xd->plane[0].dst.stride);

-  tx_type = get_tx_type_4x4(&x->e_mbd, ib);

-  if (tx_type != DCT_DCT) {

-    vp9_short_fht4x4(src_diff, coeff, 4 << bwl, tx_type);

-    x->quantize_b_4x4(x, ib, tx_type, 16);

-    vp9_short_iht4x4_add(BLOCK_OFFSET(xd->plane[0].dqcoeff, ib, 16), dst,

-                         xd->plane[0].dst.stride, tx_type);

-  } else {

-    x->fwd_txm4x4(src_diff, coeff, 8 << bwl);

-    x->quantize_b_4x4(x, ib, tx_type, 16);

-    vp9_inverse_transform_b_4x4_add(&x->e_mbd, xd->plane[0].eobs[ib],

-                                BLOCK_OFFSET(xd->plane[0].dqcoeff, ib, 16),

-                                dst, xd->plane[0].dst.stride);

-  }

+  x->fwd_txm4x4(src_diff, coeff, 8 << bwl);

+  x->quantize_b_4x4(x, ib, DCT_DCT, 16);

+  vp9_inverse_transform_b_4x4_add(&x->e_mbd, xd->plane[0].eobs[ib],

+                              BLOCK_OFFSET(xd->plane[0].dqcoeff, ib, 16),

+                              dst, xd->plane[0].dst.stride);

 }

 void vp9_encode_intra16x16mby(VP9_COMMON *const cm, MACROBLOCK *x) {

-- 

⑨