Experimental high precision convolve for Wiener

Improves coding efficiency.

Change-Id: I7bb12190cdc4581097809a020355cdc8867fc1ad

aom_dsp/aom_convolve.c

@@ -446,6 +446,125 @@ void aom_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride,
   convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                    x_step_q4, filters_y, y0_q4, y_step_q4, w, h);
 }
+
+#define EXTRAPREC_BITS 2
+#define EXTRAPREC_CLAMP_LIMIT (512 << EXTRAPREC_BITS)
+static void convolve_add_src_horiz_hip(const uint8_t *src, ptrdiff_t src_stride,
+                                       uint16_t *dst, ptrdiff_t dst_stride,
+                                       const InterpKernel *x_filters, int x0_q4,
+                                       int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS);
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] =
+          (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS),
+                          0, EXTRAPREC_CLAMP_LIMIT - 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_add_src_vert_hip(const uint16_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const InterpKernel *y_filters, int y0_q4,
+                                      int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k,
+          sum = ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS);
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] =
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS));
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_add_src_hip(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const InterpKernel *const x_filters, int x0_q4,
+                                 int x_step_q4,
+                                 const InterpKernel *const y_filters, int y0_q4,
+                                 int y_step_q4, int w, int h) {
+  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_add_src_horiz_hip(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                             src_stride, temp, MAX_SB_SIZE, x_filters, x0_q4,
+                             x_step_q4, w, intermediate_height);
+  convolve_add_src_vert_hip(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+                            MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4,
+                            y_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_horiz_hip_c(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) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_add_src_horiz_hip(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                             x_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_vert_hip_c(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) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_add_src_vert_hip(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                            y_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_hip_c(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) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve_add_src_hip(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                       x_step_q4, filters_y, y0_q4, y_step_q4, w, h);
+}
 #endif  // CONFIG_LOOP_RESTORATION
 
 #if CONFIG_HIGHBITDEPTH
@@ -850,5 +969,133 @@ void aom_highbd_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride,
   highbd_convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                           x_step_q4, filters_y, y0_q4, y_step_q4, w, h, bd);
 }
+
+static void highbd_convolve_add_src_horiz_hip(
+    const uint8_t *src8, ptrdiff_t src_stride, uint16_t *dst,
+    ptrdiff_t dst_stride, const InterpKernel *x_filters, int x0_q4,
+    int x_step_q4, int w, int h, int bd) {
+  const int extraprec_clamp_limit = (EXTRAPREC_CLAMP_LIMIT << (bd - 8));
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS);
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] =
+          (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS),
+                          0, extraprec_clamp_limit - 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_add_src_vert_hip(
+    const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst8,
+    ptrdiff_t dst_stride, const InterpKernel *y_filters, int y0_q4,
+    int y_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k,
+          sum = ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS);
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel_highbd(
+          ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS), bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_add_src_hip(
+    const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+    ptrdiff_t dst_stride, const InterpKernel *const x_filters, int x0_q4,
+    int x_step_q4, const InterpKernel *const y_filters, int y0_q4,
+    int y_step_q4, int w, int h, int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  highbd_convolve_add_src_horiz_hip(
+      src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, MAX_SB_SIZE,
+      x_filters, x0_q4, x_step_q4, w, intermediate_height, bd);
+  highbd_convolve_add_src_vert_hip(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+                                   MAX_SB_SIZE, dst, dst_stride, y_filters,
+                                   y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_horiz_hip_c(
+    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 bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_add_src_horiz_hip(src, src_stride, dst, dst_stride, filters_x,
+                                    x0_q4, x_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_vert_hip_c(
+    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 bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_add_src_vert_hip(src, src_stride, dst, dst_stride, filters_y,
+                                   y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_hip_c(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 bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  highbd_convolve_add_src_hip(src, src_stride, dst, dst_stride, filters_x,
+                              x0_q4, x_step_q4, filters_y, y0_q4, y_step_q4, w,
+                              h, bd);
+}
+
 #endif  // CONFIG_LOOP_RESTORATION
 #endif  // CONFIG_HIGHBITDEPTH

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)