High bit-depth loop/arf/postproc filter functions

Adds high-bitdepth loopfilter, temporal filter and postproc functions

Change-Id: I81c8a9176890784686bc4f2af0d550d243b3b2d3

test/test.mk

@@ -128,6 +128,7 @@ LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct4x4_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct8x8_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += variance_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_subtract_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += lpf_8_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9)         += vp9_intrapred_test.cc
 
 ifeq ($(CONFIG_VP9_ENCODER),yes)

vp9/common/vp9_loopfilter_filters.c

@@ -17,6 +17,20 @@ static INLINE int8_t signed_char_clamp(int t) {
   return (int8_t)clamp(t, -128, 127);
 }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE int16_t signed_char_clamp_high(int t, int bd) {
+  switch (bd) {
+    case 10:
+      return (int16_t)clamp(t, -128*4, 128*4-1);
+    case 12:
+      return (int16_t)clamp(t, -128*16, 128*16-1);
+    case 8:
+    default:
+      return (int16_t)clamp(t, -128, 128-1);
+  }
+}
+#endif
+
 // should we apply any filter at all: 11111111 yes, 00000000 no
 static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit,
                                  uint8_t p3, uint8_t p2,
@@ -337,3 +351,390 @@ void vp9_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit,
                                 const uint8_t *limit, const uint8_t *thresh) {
   mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16);
 }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// Should we apply any filter at all: 11111111 yes, 00000000 no ?
+static INLINE int8_t high_filter_mask(uint8_t limit, uint8_t blimit,
+                                      uint16_t p3, uint16_t p2,
+                                      uint16_t p1, uint16_t p0,
+                                      uint16_t q0, uint16_t q1,
+                                      uint16_t q2, uint16_t q3, int bd) {
+  int8_t mask = 0;
+  int16_t limit16 = (uint16_t)limit << (bd - 8);
+  int16_t blimit16 = (uint16_t)blimit << (bd - 8);
+  mask |= (abs(p3 - p2) > limit16) * -1;
+  mask |= (abs(p2 - p1) > limit16) * -1;
+  mask |= (abs(p1 - p0) > limit16) * -1;
+  mask |= (abs(q1 - q0) > limit16) * -1;
+  mask |= (abs(q2 - q1) > limit16) * -1;
+  mask |= (abs(q3 - q2) > limit16) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t high_flat_mask4(uint8_t thresh,
+                                     uint16_t p3, uint16_t p2,
+                                     uint16_t p1, uint16_t p0,
+                                     uint16_t q0, uint16_t q1,
+                                     uint16_t q2, uint16_t q3, int bd) {
+  int8_t mask = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p1 - p0) > thresh16) * -1;
+  mask |= (abs(q1 - q0) > thresh16) * -1;
+  mask |= (abs(p2 - p0) > thresh16) * -1;
+  mask |= (abs(q2 - q0) > thresh16) * -1;
+  mask |= (abs(p3 - p0) > thresh16) * -1;
+  mask |= (abs(q3 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t high_flat_mask5(uint8_t thresh,
+                                     uint16_t p4, uint16_t p3,
+                                     uint16_t p2, uint16_t p1,
+                                     uint16_t p0, uint16_t q0,
+                                     uint16_t q1, uint16_t q2,
+                                     uint16_t q3, uint16_t q4, int bd) {
+  int8_t mask = ~high_flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p4 - p0) > thresh16) * -1;
+  mask |= (abs(q4 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+// Is there high edge variance internal edge:
+// 11111111_11111111 yes, 00000000_00000000 no ?
+static INLINE int16_t high_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0,
+                                    uint16_t q0, uint16_t q1, int bd) {
+  int16_t hev = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  hev |= (abs(p1 - p0) > thresh16) * -1;
+  hev |= (abs(q1 - q0) > thresh16) * -1;
+  return hev;
+}
+
+static INLINE void high_filter4(int8_t mask, uint8_t thresh, uint16_t *op1,
+                                uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
+                                int bd) {
+  int16_t filter1, filter2;
+  // ^0x80 equivalent to subtracting 0x80 from the values to turn them
+  // into -128 to +127 instead of 0 to 255.
+  int shift = bd - 8;
+  const int16_t ps1 = (int16_t)*op1 - (0x80 << shift);
+  const int16_t ps0 = (int16_t)*op0 - (0x80 << shift);
+  const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift);
+  const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift);
+  const uint16_t hev = high_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd);
+
+  // Add outer taps if we have high edge variance.
+  int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev;
+
+  // Inner taps.
+  filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask;
+
+  // Save bottom 3 bits so that we round one side +4 and the other +3
+  // if it equals 4 we'll set to adjust by -1 to account for the fact
+  // we'd round 3 the other way.
+  filter1 = signed_char_clamp_high(filter + 4, bd) >> 3;
+  filter2 = signed_char_clamp_high(filter + 3, bd) >> 3;
+
+  *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift);
+  *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift);
+
+  // Outer tap adjustments.
+  filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+  *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift);
+  *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift);
+}
+
+void vp9_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh, int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask = high_filter_mask(*limit, *blimit,
+                                         p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    high_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd);
+    ++s;
+  }
+}
+
+void vp9_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int p,
+                                        const uint8_t *blimit0,
+                                        const uint8_t *limit0,
+                                        const uint8_t *thresh0,
+                                        const uint8_t *blimit1,
+                                        const uint8_t *limit1,
+                                        const uint8_t *thresh1,
+                                        int bd) {
+  vp9_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
+}
+
+void vp9_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0],  q1 = s[1],  q2 = s[2],  q3 = s[3];
+    const int8_t mask = high_filter_mask(*limit, *blimit,
+                                         p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    high_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd);
+    s += pitch;
+  }
+}
+
+void vp9_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch,
+                                      const uint8_t *blimit0,
+                                      const uint8_t *limit0,
+                                      const uint8_t *thresh0,
+                                      const uint8_t *blimit1,
+                                      const uint8_t *limit1,
+                                      const uint8_t *thresh1,
+                                      int bd) {
+  vp9_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
+                              thresh1, 1, bd);
+}
+
+static INLINE void high_filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+                                uint16_t *op3, uint16_t *op2,
+                                uint16_t *op1, uint16_t *op0,
+                                uint16_t *oq0, uint16_t *oq1,
+                                uint16_t *oq2, uint16_t *oq3, int bd) {
+  if (flat && mask) {
+    const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+    // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+    *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+    *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+    *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+    *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+    *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+    *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+  } else {
+    high_filter4(mask, thresh, op1,  op0, oq0, oq1, bd);
+  }
+}
+
+void vp9_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit,
+                                   const uint8_t *limit, const uint8_t *thresh,
+                                   int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+    const int8_t mask = high_filter_mask(*limit, *blimit,
+                                         p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = high_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    high_filter8(mask, *thresh, flat,
+                 s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                 s, s + 1 * p, s + 2 * p, s + 3 * p, bd);
+    ++s;
+  }
+}
+
+void vp9_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int p,
+                                        const uint8_t *blimit0,
+                                        const uint8_t *limit0,
+                                        const uint8_t *thresh0,
+                                        const uint8_t *blimit1,
+                                        const uint8_t *limit1,
+                                        const uint8_t *thresh1,
+                                        int bd) {
+  vp9_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
+}
+
+void vp9_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int count, int bd) {
+  int i;
+
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask = high_filter_mask(*limit, *blimit,
+                                         p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = high_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    high_filter8(mask, *thresh, flat,
+                 s - 4, s - 3, s - 2, s - 1,
+                 s, s + 1, s + 2, s + 3,
+                 bd);
+    s += pitch;
+  }
+}
+
+void vp9_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch,
+                                      const uint8_t *blimit0,
+                                      const uint8_t *limit0,
+                                      const uint8_t *thresh0,
+                                      const uint8_t *blimit1,
+                                      const uint8_t *limit1,
+                                      const uint8_t *thresh1,
+                                      int bd) {
+  vp9_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
+                              thresh1, 1, bd);
+}
+
+static INLINE void high_filter16(int8_t mask, uint8_t thresh,
+                                 uint8_t flat, uint8_t flat2,
+                                 uint16_t *op7, uint16_t *op6,
+                                 uint16_t *op5, uint16_t *op4,
+                                 uint16_t *op3, uint16_t *op2,
+                                 uint16_t *op1, uint16_t *op0,
+                                 uint16_t *oq0, uint16_t *oq1,
+                                 uint16_t *oq2, uint16_t *oq3,
+                                 uint16_t *oq4, uint16_t *oq5,
+                                 uint16_t *oq6, uint16_t *oq7, int bd) {
+  if (flat2 && flat && mask) {
+    const uint16_t p7 = *op7;
+    const uint16_t p6 = *op6;
+    const uint16_t p5 = *op5;
+    const uint16_t p4 = *op4;
+    const uint16_t p3 = *op3;
+    const uint16_t p2 = *op2;
+    const uint16_t p1 = *op1;
+    const uint16_t p0 = *op0;
+    const uint16_t q0 = *oq0;
+    const uint16_t q1 = *oq1;
+    const uint16_t q2 = *oq2;
+    const uint16_t q3 = *oq3;
+    const uint16_t q4 = *oq4;
+    const uint16_t q5 = *oq5;
+    const uint16_t q6 = *oq6;
+    const uint16_t q7 = *oq7;
+
+    // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+    *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0, 4);
+    *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1, 4);
+    *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2, 4);
+    *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3, 4);
+    *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4, 4);
+    *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5, 4);
+    *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
+    *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
+    *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
+    *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
+    *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+    *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+    *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+    *oq6 = ROUND_POWER_OF_TWO(p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+  } else {
+    high_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3,
+                 bd);
+  }
+}
+
+void vp9_highbd_lpf_horizontal_16_c(uint16_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit, const uint8_t *thresh,
+                                    int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask = high_filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = high_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat2 = high_flat_mask5(
+        1, s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+        q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p], bd);
+
+    high_filter16(mask, *thresh, flat, flat2,
+                  s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
+                  s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                  s, s + 1 * p, s + 2 * p, s + 3 * p,
+                  s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p,
+                  bd);
+    ++s;
+  }
+}
+
+static void high_mb_lpf_vertical_edge_w(uint16_t *s, int p,
+                                        const uint8_t *blimit,
+                                        const uint8_t *limit,
+                                        const uint8_t *thresh,
+                                        int count, int bd) {
+  int i;
+
+  for (i = 0; i < count; ++i) {
+    const uint16_t p3 = s[-4];
+    const uint16_t p2 = s[-3];
+    const uint16_t p1 = s[-2];
+    const uint16_t p0 = s[-1];
+    const uint16_t q0 = s[0];
+    const uint16_t q1 = s[1];
+    const uint16_t q2 = s[2];
+    const uint16_t q3 = s[3];
+    const int8_t mask = high_filter_mask(*limit, *blimit,
+                                         p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = high_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat2 = high_flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+                                         q0, s[4], s[5], s[6], s[7], bd);
+
+    high_filter16(mask, *thresh, flat, flat2,
+                  s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+                  s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7,
+                  bd);
+    s += p;
+  }
+}
+
+void vp9_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit,
+                                  const uint8_t *limit, const uint8_t *thresh,
+                                  int bd) {
+  high_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd);
+}
+
+void vp9_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p,
+                                       const uint8_t *blimit,
+                                       const uint8_t *limit,
+                                       const uint8_t *thresh,
+                                       int bd) {
+  high_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH

vp9/common/vp9_postproc.c

@@ -19,6 +19,9 @@
 #include "vpx_scale/vpx_scale.h"
 #include "vpx_scale/yv12config.h"
 
+#if CONFIG_VP9_HIGHBITDEPTH
+#include "vp9/common/vp9_common.h"
+#endif
 #include "vp9/common/vp9_onyxc_int.h"
 #include "vp9/common/vp9_postproc.h"
 #include "vp9/common/vp9_systemdependent.h"
@@ -152,6 +155,84 @@ void vp9_post_proc_down_and_across_c(const uint8_t *src_ptr,
   }
 }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_post_proc_down_and_across_c(const uint16_t *src_ptr,
+                                            uint16_t *dst_ptr,
+                                            int src_pixels_per_line,
+                                            int dst_pixels_per_line,
+                                            int rows,
+                                            int cols,
+                                            int flimit) {
+  uint16_t const *p_src;
+  uint16_t *p_dst;
+  int row;
+  int col;
+  int i;
+  int v;
+  int pitch = src_pixels_per_line;
+  uint16_t d[8];
+
+  for (row = 0; row < rows; row++) {
+    // post_proc_down for one row.
+    p_src = src_ptr;
+    p_dst = dst_ptr;
+
+    for (col = 0; col < cols; col++) {
+      int kernel = 4;
+      int v = p_src[col];
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i * pitch]) > flimit)
+          goto down_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i * pitch];
+      }
+
+      v = (kernel >> 3);
+
+    down_skip_convolve:
+      p_dst[col] = v;
+    }
+
+    /* now post_proc_across */
+    p_src = dst_ptr;
+    p_dst = dst_ptr;
+
+    for (i = 0; i < 8; i++)
+      d[i] = p_src[i];
+
+    for (col = 0; col < cols; col++) {
+      int kernel = 4;
+      v = p_src[col];
+
+      d[col & 7] = v;
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i]) > flimit)
+          goto across_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i];
+      }
+
+      d[col & 7] = (kernel >> 3);
+
+    across_skip_convolve:
+      if (col >= 2)
+        p_dst[col - 2] = d[(col - 2) & 7];
+    }
+
+    /* handle the last two pixels */
+    p_dst[col - 2] = d[(col - 2) & 7];
+    p_dst[col - 1] = d[(col - 1) & 7];
+
+
+    /* next row */
+    src_ptr += pitch;
+    dst_ptr += dst_pixels_per_line;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
 static int q2mbl(int x) {
   if (x < 20) x = 20;
 
@@ -162,10 +243,46 @@ static int q2mbl(int x) {
 void vp9_mbpost_proc_across_ip_c(uint8_t *src, int pitch,
                                  int rows, int cols, int flimit) {
   int r, c, i;
-
   uint8_t *s = src;
   uint8_t d[16];
 
+  for (r = 0; r < rows; r++) {
+    int sumsq = 0;
+    int sum = 0;
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i] * s[i];
+      sum += s[i];
+      d[i + 8] = 0;
+    }
+
+    for (c = 0; c < cols + 8; c++) {
+      int x = s[c + 7] - s[c - 8];
+      int y = s[c + 7] + s[c - 8];
+
+      sum += x;
+      sumsq += x * y;
+
+      d[c & 15] = s[c];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[c & 15] = (8 + sum + s[c]) >> 4;
+      }
+
+      s[c - 8] = d[(c - 8) & 15];
+    }
+    s += pitch;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_mbpost_proc_across_ip_c(uint16_t *src, int pitch,
+                                        int rows, int cols, int flimit) {
+  int r, c, i;
+
+  uint16_t *s = src;
+  uint16_t d[16];
+
 
   for (r = 0; r < rows; r++) {
     int sumsq = 0;
@@ -196,6 +313,7 @@ void vp9_mbpost_proc_across_ip_c(uint8_t *src, int pitch,
     s += pitch;
   }
 }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
 
 void vp9_mbpost_proc_down_c(uint8_t *dst, int pitch,
                             int rows, int cols, int flimit) {
@@ -229,6 +347,40 @@ void vp9_mbpost_proc_down_c(uint8_t *dst, int pitch,
   }
 }