Faster x264_exp2fix8

22->13 cycles on Core 2 with mfpmath=sse

diff --git a/common/common.h b/common/common.h
index f78a35c1d5d73f4d577d87d49d777bed20772fce..a4dc3cde80074d6db4c89b04821d1a16029c9673 100644 (file)
--- a/common/common.h
+++ b/common/common.h
@@ -161,15 +161,14 @@ static const uint8_t exp2_lut[64] = {
     177, 182, 186, 191, 196, 201, 206, 211, 216, 221, 226, 232, 237, 242, 248, 253,
 };
 
+/* Not a general-purpose function; multiplies input by -1/6 to convert
+ * qp to qscale. */
 static ALWAYS_INLINE int x264_exp2fix8( float x )
 {
-    int i, f;
-    x += 8;
-    if( x <= 0 ) return 0;
-    if( x >= 16 ) return 0xffff;
-    i = x;
-    f = (x-i)*64;
-    return (exp2_lut[f]+256) << i >> 8;
+    if( x >= 512.f/6.f ) return 0;
+    if( x <= -512.f/6.f ) return 0xffff;
+    int i = x*(-64.f/6.f) + 512;
+    return (exp2_lut[i&63]+256) << (i>>6) >> 8;
 }
 
 static const float log2_lut[128] = {

diff --git a/encoder/ratecontrol.c b/encoder/ratecontrol.c
index a6f864f3750759db437846e1c1fc3b47858242c0..d33b00b6b8b24095dee1ed9acee44517df63f576 100644 (file)
--- a/encoder/ratecontrol.c
+++ b/encoder/ratecontrol.c
@@ -251,7 +251,7 @@ void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
             frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] =
             frame->f_qp_offset_aq[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
             if( h->frames.b_have_lowres )
-                frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj*(-1.f/6.f));
+                frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj);
         }
 }
 

diff --git a/encoder/slicetype.c b/encoder/slicetype.c
index 0fa5b1f411d370fb38b4876995bb2fc9592eac90..75dc020701c75fd256959a3231f8e474b8d23e37 100644 (file)
--- a/encoder/slicetype.c
+++ b/encoder/slicetype.c
@@ -389,7 +389,7 @@ static int x264_slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **fram
             int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
             int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy];
             float qp_adj = frames[b]->f_qp_offset[i_mb_xy];
-            i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj*(-1.f/6.f)) + 128) >> 8;
+            i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
             row_satd[ h->mb.i_mb_y ] += i_mb_cost;
             if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->sps->i_mb_height - 1 &&
                  h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->sps->i_mb_width - 1) ||
@@ -410,6 +410,9 @@ static void x264_macroblock_tree_finish( x264_t *h, x264_frame_t *frame, int b_b
         memcpy( frame->f_qp_offset, frame->f_qp_offset_aq, sizeof( frame->f_qp_offset ) );
     else
     {
+        /* Allow the strength to be adjusted via qcompress, since the two
+         * concepts are very similar. */
+        float strength = 5.0f * (1.0f - h->param.rc.f_qcompress);
         for( mb_index = 0; mb_index < h->mb.i_mb_count; mb_index++ )
         {
             int intra_cost = (frame->i_intra_cost[mb_index] * frame->i_inv_qscale_factor[mb_index]+128)>>8;
@@ -417,9 +420,7 @@ static void x264_macroblock_tree_finish( x264_t *h, x264_frame_t *frame, int b_b
             {
                 int propagate_cost = frame->i_propagate_cost[mb_index];
                 float log2_ratio = x264_log2(intra_cost + propagate_cost) - x264_log2(intra_cost);
-                /* Allow the constant to be adjusted via qcompress, since the two
-                 * concepts are very similar. */
-                frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - 5.0 * (1.0 - h->param.rc.f_qcompress) * log2_ratio;
+                frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - strength * log2_ratio;
             }
         }
     }