From b30f17d4b8b7070ee7b3f455bdee626a9b58fe99 Mon Sep 17 00:00:00 2001
From: Ari Lemmetti <ari.lemmetti@gmail.com>
Date: Mon, 10 Aug 2015 17:49:31 +0300
Subject: [PATCH] Add fractional pixel sampling for chroma

---
 src/inter.c                           | 23 +++++-------------
 src/strategies/generic/ipol-generic.c | 35 +++++++++++++++++++++++++++
 src/strategies/generic/ipol-generic.h |  3 +++
 3 files changed, 44 insertions(+), 17 deletions(-)

diff --git a/src/inter.c b/src/inter.c
index 2898c620..9a65526f 100644
--- a/src/inter.c
+++ b/src/inter.c
@@ -139,28 +139,17 @@ void inter_recon_lcu(const encoder_state_t * const state, const kvz_picture * co
       extend_borders(xpos>>1, ypos>>1, (mv[0]>>2)>>1, (mv[1]>>2)>>1, state->tile->lcu_offset_x * (LCU_WIDTH>>1), state->tile->lcu_offset_y * (LCU_WIDTH>>1),
           ref->u, ref->width>>1, ref->height>>1, FILTER_SIZE_C, width_c, width_c, octpel_src_u);
 
-      filter_inter_octpel_chroma(state->encoder_control, octpel_src_off_u, width_c+FILTER_SIZE_C, width_c,
-                                 width_c, octpel_dst_u, width_c*8, c_off_x, c_off_y);
+      sample_octpel_chroma_generic(state->encoder_control, octpel_src_off_u, width_c + FILTER_SIZE_C, width_c,
+        width_c, lcu->rec.u + ((ypos >> 1) % LCU_WIDTH_C)*LCU_WIDTH_C + ((xpos >> 1) % LCU_WIDTH_C), LCU_WIDTH_C, c_off_x, c_off_y, mv);
 
       //Fractional chroma V
-      extend_borders(xpos>>1, ypos>>1, (mv[0]>>2)>>1, (mv[1]>>2)>>1, state->tile->lcu_offset_x * (LCU_WIDTH>>1), state->tile->lcu_offset_y * (LCU_WIDTH>>1),
-          ref->v, ref->width>>1, ref->height>>1, FILTER_SIZE_C, width_c, width_c, octpel_src_v);
+      extend_borders(xpos >> 1, ypos >> 1, (mv[0] >> 2) >> 1, (mv[1] >> 2) >> 1, state->tile->lcu_offset_x * (LCU_WIDTH_C >> 1), state->tile->lcu_offset_y * (LCU_WIDTH_C >> 1),
+        ref->v, ref->width >> 1, ref->height >> 1, FILTER_SIZE_C, width_c, width_c, octpel_src_v);
 
-      filter_inter_octpel_chroma(state->encoder_control, octpel_src_off_v, width_c+FILTER_SIZE_C, width_c,
-                   width_c, octpel_dst_v, width_c*8, c_off_x, c_off_y);
+      sample_octpel_chroma_generic(state->encoder_control, octpel_src_off_v, width_c + FILTER_SIZE_C, width_c,
+        width_c, lcu->rec.v + ((ypos >> 1) % LCU_WIDTH_C)*LCU_WIDTH_C + ((xpos >> 1) % LCU_WIDTH_C), LCU_WIDTH_C, c_off_x, c_off_y, mv);
 
 
-      //Sample fractional pixels for chroma
-      for(y = 0; y < width_c; ++y) {
-        int y_in_lcu = ((y+(ypos>>1)) & ((LCU_WIDTH>>1)-1));
-        int qpel_y = y*8+c_off_y;
-        for(x = 0; x < width_c; ++x) {
-          int x_in_lcu = ((x+(xpos>>1)) & ((LCU_WIDTH>>1)-1));
-          int qpel_x = x*8+c_off_x;
-          lcu->rec.u[y_in_lcu * dst_width_c + x_in_lcu] = (kvz_pixel)octpel_dst_u[qpel_y*(width_c*8)+qpel_x];
-          lcu->rec.v[y_in_lcu * dst_width_c + x_in_lcu] = (kvz_pixel)octpel_dst_v[qpel_y*(width_c*8)+qpel_x];
-        }
-      }
     }
 
   mv[0] >>= 2;
diff --git a/src/strategies/generic/ipol-generic.c b/src/strategies/generic/ipol-generic.c
index 804723d8..a6706223 100644
--- a/src/strategies/generic/ipol-generic.c
+++ b/src/strategies/generic/ipol-generic.c
@@ -383,6 +383,41 @@ void filter_inter_octpel_chroma_generic(const encoder_control_t * const encoder,
   }
 }
 
+void sample_octpel_chroma_generic(const encoder_control_t * const encoder, kvz_pixel *src, int16_t src_stride, int width, int height,kvz_pixel *dst, int16_t dst_stride, int8_t hor_flag, int8_t ver_flag, const int16_t mv[2])
+{
+  //TODO: horizontal and vertical only filtering
+  int32_t x, y;
+  int16_t shift1 = KVZ_BIT_DEPTH - 8;
+  int32_t shift2 = 6;
+  int32_t shift3 = 14 - KVZ_BIT_DEPTH;
+  int32_t offset23 = 1 << (shift2 + shift3 - 1);
+
+  int8_t *hor_filter = g_chroma_filter[mv[0] & 7];
+  int8_t *ver_filter = g_chroma_filter[mv[1] & 7];
+
+#define FILTER_SIZE_C (FILTER_SIZE / 2)
+#define FILTER_OFFSET_C (FILTER_OFFSET / 2)
+  int16_t flipped_hor_filtered[(LCU_WIDTH_C + 1) + FILTER_SIZE_C][(LCU_WIDTH_C + 1) + FILTER_SIZE_C];
+
+  // Filter horizontally and flip x and y
+  for (x = 0; x < width; ++x) {
+    for (y = 0; y < height + FILTER_SIZE_C; ++y) {
+      int ypos = y - FILTER_OFFSET_C;
+      int xpos = x - FILTER_OFFSET_C;
+      flipped_hor_filtered[x][y] = four_tap_filter_hor_generic(hor_filter, &src[src_stride*ypos + xpos]) >> shift1;
+    }
+  }
+
+  // Filter vertically and flip x and y
+  for (x = 0; x < width; ++x) {
+    for (y = 0; y < height; ++y) {
+      int ypos = y;
+      int xpos = x;
+      dst[y*dst_stride + x] = fast_clip_32bit_to_pixel(((four_tap_filter_hor_16bit_generic(ver_filter, &flipped_hor_filtered[xpos][ypos]) + offset23) >> shift2) >> shift3);
+    }
+  }
+}
+
 void extend_borders_generic(int xpos, int ypos, int mv_x, int mv_y, int off_x, int off_y, kvz_pixel *ref, int ref_width, int ref_height,
   int filterSize, int width, int height, kvz_pixel *dst) {
 
diff --git a/src/strategies/generic/ipol-generic.h b/src/strategies/generic/ipol-generic.h
index 64607ee2..f1a6264f 100644
--- a/src/strategies/generic/ipol-generic.h
+++ b/src/strategies/generic/ipol-generic.h
@@ -22,6 +22,9 @@
 #include "encoder.h"
 
 int strategy_register_ipol_generic(void* opaque);
+
+//TODO: create strategies from sample functions
 void sample_quarterpel_luma_generic(const encoder_control_t * const encoder, kvz_pixel *src, int16_t src_stride, int width, int height, kvz_pixel *dst, int16_t dst_stride, int8_t hor_flag, int8_t ver_flag, const int16_t mv[2]);
+void sample_octpel_chroma_generic(const encoder_control_t * const encoder, kvz_pixel *src, int16_t src_stride, int width, int height, kvz_pixel *dst, int16_t dst_stride, int8_t hor_flag, int8_t ver_flag, const int16_t mv[2]);
 
 #endif //STRATEGIES_IPOL_GENERIC_H_