Experimental high precision convolve for Wiener

Improves coding efficiency.

Change-Id: I7bb12190cdc4581097809a020355cdc8867fc1ad

aom_dsp/aom_dsp_rtcd_defs.pl

@@ -171,6 +171,9 @@ if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
   add_proto qw/void aom_convolve8_add_src/,       "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
   add_proto qw/void aom_convolve8_add_src_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
   add_proto qw/void aom_convolve8_add_src_vert/,  "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+  add_proto qw/void aom_convolve8_add_src_hip/,       "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+  add_proto qw/void aom_convolve8_add_src_horiz_hip/, "const uint8_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+  add_proto qw/void aom_convolve8_add_src_vert_hip/,  "const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
 
   specialize qw/aom_convolve8_add_src ssse3/;
   specialize qw/aom_convolve8_add_src_horiz ssse3/;
@@ -218,6 +221,9 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
     add_proto qw/void aom_highbd_convolve8_add_src/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
     add_proto qw/void aom_highbd_convolve8_add_src_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
     add_proto qw/void aom_highbd_convolve8_add_src_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+    add_proto qw/void aom_highbd_convolve8_add_src_hip/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+    add_proto qw/void aom_highbd_convolve8_add_src_horiz_hip/, "const uint8_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+    add_proto qw/void aom_highbd_convolve8_add_src_vert_hip/, "const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
 
     specialize qw/aom_highbd_convolve8_add_src/, "$sse2_x86_64";
     # The _horiz/_vert functions are currently unused, so we don't bother

av1/common/restoration.c

@@ -160,9 +160,16 @@ static void loop_wiener_filter_tile(uint8_t *data, int tile_idx, int width,
       int h = AOMMIN(MAX_SB_SIZE, (v_end - i + 15) & ~15);
       const uint8_t *data_p = data + i * stride + j;
       uint8_t *dst_p = dst + i * dst_stride + j;
+#if USE_WIENER_HIGH_INTERMEDIATE_PRECISION
+      aom_convolve8_add_src_hip(data_p, stride, dst_p, dst_stride,
+                                rst->rsi->wiener_info[tile_idx].hfilter, 16,
+                                rst->rsi->wiener_info[tile_idx].vfilter, 16, w,
+                                h);
+#else
       aom_convolve8_add_src(data_p, stride, dst_p, dst_stride,
                             rst->rsi->wiener_info[tile_idx].hfilter, 16,
                             rst->rsi->wiener_info[tile_idx].vfilter, 16, w, h);
+#endif  // USE_WIENER_HIGH_INTERMEDIATE_PRECISION
     }
 }
 
@@ -992,10 +999,17 @@ static void loop_wiener_filter_tile_highbd(uint16_t *data, int tile_idx,
       int h = AOMMIN(MAX_SB_SIZE, (v_end - i + 15) & ~15);
       const uint16_t *data_p = data + i * stride + j;
       uint16_t *dst_p = dst + i * dst_stride + j;
+#if USE_WIENER_HIGH_INTERMEDIATE_PRECISION
+      aom_highbd_convolve8_add_src_hip(
+          CONVERT_TO_BYTEPTR(data_p), stride, CONVERT_TO_BYTEPTR(dst_p),
+          dst_stride, rst->rsi->wiener_info[tile_idx].hfilter, 16,
+          rst->rsi->wiener_info[tile_idx].vfilter, 16, w, h, bit_depth);
+#else
       aom_highbd_convolve8_add_src(
           CONVERT_TO_BYTEPTR(data_p), stride, CONVERT_TO_BYTEPTR(dst_p),
           dst_stride, rst->rsi->wiener_info[tile_idx].hfilter, 16,
           rst->rsi->wiener_info[tile_idx].vfilter, 16, w, h, bit_depth);
+#endif  // USE_WIENER_HIGH_INTERMEDIATE_PRECISION
     }
 }
 

av1/common/restoration.h

@@ -81,6 +81,9 @@ extern "C" {
 #define WIENER_FILT_PREC_BITS 7
 #define WIENER_FILT_STEP (1 << WIENER_FILT_PREC_BITS)
 
+// Whether to use high intermediate precision filtering
+#define USE_WIENER_HIGH_INTERMEDIATE_PRECISION 0
+
 // Central values for the taps
 #define WIENER_FILT_TAP0_MIDV (3)
 #define WIENER_FILT_TAP1_MIDV (-7)