shithub: libvpx

Info  •  Files  •  Log  •  Branches

Use 4-tap interp filter in speed 1 sub-pel motion search

Added the 4-tap interp filter, and used it for speed 1 sub-pel motion
search. Speed 2 motion search still used bilinear filter as before.

Speed 1 borg test showed good bit savings.
        avg_psnr:  ovr_psnr:    ssim:
lowres:  -1.125    -1.179      -1.021
midres:  -0.717    -0.710      -0.543
hdres:   -0.357    -0.370      -0.342
Speed test at speed 1 showed ~10% encoder time increase, which was
partially because of no SIMD version of 4-tap filter.

Change-Id: Ic9b48cdc6a964538c20144108526682d64348301

--- a/test/convolve_test.cc

+++ b/test/convolve_test.cc

@@ -789,7 +789,7 @@

   }

 }

-const int kNumFilterBanks = 4;

+const int kNumFilterBanks = 5;

 const int kNumFilters = 16;

 TEST(ConvolveTest, FiltersWontSaturateWhenAddedPairwise) {

--- a/vp9/common/vp9_filter.c

+++ b/vp9/common/vp9_filter.c

@@ -63,6 +63,20 @@

   { 0, -3, 2, 41, 63, 29, -2, -2 },   { 0, -3, 1, 38, 64, 32, -1, -3 }

 };

-const InterpKernel *vp9_filter_kernels[4] = {

-  sub_pel_filters_8, sub_pel_filters_8lp, sub_pel_filters_8s, bilinear_filters

+// 4-tap filter

+DECLARE_ALIGNED(256, static const InterpKernel,

+                sub_pel_filters_4[SUBPEL_SHIFTS]) = {

+  { 0, 0, 0, 128, 0, 0, 0, 0 },     { 0, 0, -3, 125, 8, -2, 0, 0 },

+  { 0, 0, -6, 120, 18, -4, 0, 0 },  { 0, 0, -8, 115, 27, -6, 0, 0 },

+  { 0, 0, -10, 108, 37, -7, 0, 0 }, { 0, 0, -11, 101, 47, -9, 0, 0 },

+  { 0, 0, -11, 93, 56, -10, 0, 0 }, { 0, 0, -12, 85, 66, -11, 0, 0 },

+  { 0, 0, -11, 75, 75, -11, 0, 0 }, { 0, 0, -11, 66, 85, -12, 0, 0 },

+  { 0, 0, -10, 56, 93, -11, 0, 0 }, { 0, 0, -9, 47, 101, -11, 0, 0 },

+  { 0, 0, -7, 37, 108, -10, 0, 0 }, { 0, 0, -6, 27, 115, -8, 0, 0 },

+  { 0, 0, -4, 18, 120, -6, 0, 0 },  { 0, 0, -2, 8, 125, -3, 0, 0 }

+};

+

+const InterpKernel *vp9_filter_kernels[5] = {

+  sub_pel_filters_8, sub_pel_filters_8lp, sub_pel_filters_8s, bilinear_filters,

+  sub_pel_filters_4

 };

--- a/vp9/common/vp9_filter.h

+++ b/vp9/common/vp9_filter.h

@@ -25,6 +25,7 @@

 #define EIGHTTAP_SHARP 2

 #define SWITCHABLE_FILTERS 3 /* Number of switchable filters */

 #define BILINEAR 3

+#define FOURTAP 4

 // The codec can operate in four possible inter prediction filter mode:

 // 8-tap, 8-tap-smooth, 8-tap-sharp, and switching between the three.

 #define SWITCHABLE_FILTER_CONTEXTS (SWITCHABLE_FILTERS + 1)

@@ -32,7 +33,7 @@

 typedef uint8_t INTERP_FILTER;

-extern const InterpKernel *vp9_filter_kernels[4];

+extern const InterpKernel *vp9_filter_kernels[5];

 #ifdef __cplusplus

 }  // extern "C"

--- a/vp9/encoder/vp9_mcomp.c

+++ b/vp9/encoder/vp9_mcomp.c

@@ -760,7 +760,14 @@

   unsigned int cost_array[5];

   int kr, kc;

   MvLimits subpel_mv_limits;

-  const InterpKernel *kernel = vp9_filter_kernels[EIGHTTAP];

+

+  // TODO(yunqing): need to add 4-tap filter optimization to speed up the

+  // encoder.

+  const InterpKernel *kernel = (use_accurate_subpel_search > 0)

+                                   ? ((use_accurate_subpel_search == USE_4_TAPS)

+                                          ? vp9_filter_kernels[FOURTAP]

+                                          : vp9_filter_kernels[EIGHTTAP])

+                                   : vp9_filter_kernels[BILINEAR];

   vp9_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, ref_mv);

   minc = subpel_mv_limits.col_min;

--- a/vp9/encoder/vp9_speed_features.c

+++ b/vp9/encoder/vp9_speed_features.c

@@ -288,7 +288,7 @@

     sf->exhaustive_searches_thresh =

         (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ? (1 << 23)

                                                                 : INT_MAX;

-    sf->use_accurate_subpel_search = USE_2_TAPS;

+    sf->use_accurate_subpel_search = USE_4_TAPS;

   }

   if (speed >= 2) {

@@ -330,6 +330,8 @@

             good_quality_mesh_patterns[mesh_density_level][i].interval;

       }

     }

+

+    sf->use_accurate_subpel_search = USE_2_TAPS;

   }

   if (speed >= 3) {