aom: Add SVE2 impl of HBD dist_wtd_convolve_2d for 4-tap filters

From 8c0cdddac6987bf9fe600d380040bc768e8d6f91 Mon Sep 17 00:00:00 2001
From: Salome Thirot <[EMAIL REDACTED]>
Date: Mon, 18 Mar 2024 16:46:04 +0000
Subject: [PATCH] Add SVE2 impl of HBD dist_wtd_convolve_2d for 4-tap filters

Add SVE2 implementation of av1_highbd_dist_wtd_convolve_2d for 4-tap
filters. Only the horizontal pass makes use of SVE, the vertical pass
stays in Neon. This gives up to 30% uplift over the Neon implementation.

Change-Id: Ie9b8a2dd2ca7195184029b948a57f7468e71c11c
---
 aom_dsp/arm/mem_neon.h                        |  10 +
 .../arm/highbd_compound_convolve_sve2.c       | 305 +++++++++++++++++-
 2 files changed, 298 insertions(+), 17 deletions(-)

diff --git a/aom_dsp/arm/mem_neon.h b/aom_dsp/arm/mem_neon.h
index d749f1def..32a462a18 100644
--- a/aom_dsp/arm/mem_neon.h
+++ b/aom_dsp/arm/mem_neon.h
@@ -537,6 +537,16 @@ static INLINE void store_u16_8x8(uint16_t *s, ptrdiff_t dst_stride,
   vst1q_u16(s, s7);
 }
 
+static INLINE void store_u16_4x3(uint16_t *s, ptrdiff_t dst_stride,
+                                 const uint16x4_t s0, const uint16x4_t s1,
+                                 const uint16x4_t s2) {
+  vst1_u16(s, s0);
+  s += dst_stride;
+  vst1_u16(s, s1);
+  s += dst_stride;
+  vst1_u16(s, s2);
+}
+
 static INLINE void store_u16_4x4(uint16_t *s, ptrdiff_t dst_stride,
                                  const uint16x4_t s0, const uint16x4_t s1,
                                  const uint16x4_t s2, const uint16x4_t s3) {
diff --git a/av1/common/arm/highbd_compound_convolve_sve2.c b/av1/common/arm/highbd_compound_convolve_sve2.c
index dd55e346a..baffc0edb 100644
--- a/av1/common/arm/highbd_compound_convolve_sve2.c
+++ b/av1/common/arm/highbd_compound_convolve_sve2.c
@@ -887,10 +887,9 @@ static INLINE void highbd_dist_wtd_convolve_2d_horiz_8tap_sve2(
   const int64x2_t offset_lo = vcombine_s64(vcreate_s64(offset), vdup_n_s64(0));
   const int16x8_t filter = vld1q_s16(x_filter_ptr);
 
-  // We are only doing 8-tap vertical convolutions, therefore we know the
-  // intermediate height will be h + 7, so we can do the loop across the whole
-  // block 4 rows at a time and then process the last 3 rows separately. This
-  // will remain true when 4-tap specialisation is added.
+  // We are only doing 8-tap and 4-tap vertical convolutions, therefore we know
+  // that im_h % 4 = 3, so we can do the loop across the whole block 4 rows at
+  // a time and then process the last 3 rows separately.
 
   do {
     const int16_t *s = (const int16_t *)src;
@@ -946,6 +945,153 @@ static INLINE void highbd_dist_wtd_convolve_2d_horiz_8tap_sve2(
   } while (width != 0);
 }
 
+static INLINE uint16x4_t highbd_convolve4_4_2d_h(int16x8_t s0, int16x8_t filter,
+                                                 int64x2_t offset,
+                                                 int32x4_t shift,
+                                                 uint16x8x2_t permute_tbl) {
+  int16x8_t permuted_samples0 = aom_tbl_s16(s0, permute_tbl.val[0]);
+  int16x8_t permuted_samples1 = aom_tbl_s16(s0, permute_tbl.val[1]);
+
+  int64x2_t sum01 = aom_svdot_lane_s16(offset, permuted_samples0, filter, 0);
+  int64x2_t sum23 = aom_svdot_lane_s16(offset, permuted_samples1, filter, 0);
+
+  int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23));
+  sum0123 = vshlq_s32(sum0123, shift);
+
+  return vqmovun_s32(sum0123);
+}
+
+static INLINE uint16x8_t highbd_convolve4_8_2d_h(int16x8_t s0[4],
+                                                 int16x8_t filter,
+                                                 int64x2_t offset,
+                                                 int32x4_t shift,
+                                                 uint16x8_t tbl) {
+  int64x2_t sum04 = aom_svdot_lane_s16(offset, s0[0], filter, 0);
+  int64x2_t sum15 = aom_svdot_lane_s16(offset, s0[1], filter, 0);
+  int64x2_t sum26 = aom_svdot_lane_s16(offset, s0[2], filter, 0);
+  int64x2_t sum37 = aom_svdot_lane_s16(offset, s0[3], filter, 0);
+
+  int32x4_t sum0415 = vcombine_s32(vmovn_s64(sum04), vmovn_s64(sum15));
+  sum0415 = vshlq_s32(sum0415, shift);
+
+  int32x4_t sum2637 = vcombine_s32(vmovn_s64(sum26), vmovn_s64(sum37));
+  sum2637 = vshlq_s32(sum2637, shift);
+
+  uint16x8_t res = vcombine_u16(vqmovun_s32(sum0415), vqmovun_s32(sum2637));
+  return aom_tbl_u16(res, tbl);
+}
+
+static INLINE void highbd_dist_wtd_convolve_2d_horiz_4tap_sve2(
+    const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride,
+    int width, int height, const int16_t *x_filter_ptr,
+    ConvolveParams *conv_params, const int offset) {
+  const int64x2_t offset_vec = vdupq_n_s64(offset);
+  const int32x4_t shift = vdupq_n_s32(-conv_params->round_0);
+  const int16x4_t x_filter = vld1_s16(x_filter_ptr + 2);
+  const int16x8_t filter = vcombine_s16(x_filter, vdup_n_s16(0));
+
+  // We are only doing 8-tap and 4-tap vertical convolutions, therefore we know
+  // that im_h % 4 = 3, so we can do the loop across the whole block 4 rows at
+  // a time and then process the last 3 rows separately.
+
+  if (width == 4) {
+    uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl);
+
+    const int16_t *s = (const int16_t *)(src);
+
+    do {
+      int16x8_t s0, s1, s2, s3;
+      load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3);
+
+      uint16x4_t d0 =
+          highbd_convolve4_4_2d_h(s0, filter, offset_vec, shift, permute_tbl);
+      uint16x4_t d1 =
+          highbd_convolve4_4_2d_h(s1, filter, offset_vec, shift, permute_tbl);
+      uint16x4_t d2 =
+          highbd_convolve4_4_2d_h(s2, filter, offset_vec, shift, permute_tbl);
+      uint16x4_t d3 =
+          highbd_convolve4_4_2d_h(s3, filter, offset_vec, shift, permute_tbl);
+
+      store_u16_4x4(dst, dst_stride, d0, d1, d2, d3);
+
+      s += 4 * src_stride;
+      dst += 4 * dst_stride;
+      height -= 4;
+    } while (height > 4);
+
+    // Process final 3 rows.
+    int16x8_t s0, s1, s2;
+    load_s16_8x3(s, src_stride, &s0, &s1, &s2);
+
+    uint16x4_t d0 =
+        highbd_convolve4_4_2d_h(s0, filter, offset_vec, shift, permute_tbl);
+    uint16x4_t d1 =
+        highbd_convolve4_4_2d_h(s1, filter, offset_vec, shift, permute_tbl);
+    uint16x4_t d2 =
+        highbd_convolve4_4_2d_h(s2, filter, offset_vec, shift, permute_tbl);
+
+    store_u16_4x3(dst, dst_stride, d0, d1, d2);
+
+  } else {
+    uint16x8_t idx = vld1q_u16(kDeinterleaveTbl);
+
+    do {
+      const int16_t *s = (const int16_t *)(src);
+      uint16_t *d = dst;
+      int w = width;
+
+      do {
+        int16x8_t s0[4], s1[4], s2[4], s3[4];
+        load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]);
+        load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]);
+        load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]);
+        load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]);
+
+        uint16x8_t d0 =
+            highbd_convolve4_8_2d_h(s0, filter, offset_vec, shift, idx);
+        uint16x8_t d1 =
+            highbd_convolve4_8_2d_h(s1, filter, offset_vec, shift, idx);
+        uint16x8_t d2 =
+            highbd_convolve4_8_2d_h(s2, filter, offset_vec, shift, idx);
+        uint16x8_t d3 =
+            highbd_convolve4_8_2d_h(s3, filter, offset_vec, shift, idx);
+
+        store_u16_8x4(d, dst_stride, d0, d1, d2, d3);
+
+        s += 8;
+        d += 8;
+        w -= 8;
+      } while (w != 0);
+      src += 4 * src_stride;
+      dst += 4 * dst_stride;
+      height -= 4;
+    } while (height > 4);
+
+    // Process final 3 rows.
+    const int16_t *s = (const int16_t *)(src);
+
+    do {
+      int16x8_t s0[4], s1[4], s2[4];
+      load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]);
+      load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]);
+      load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]);
+
+      uint16x8_t d0 =
+          highbd_convolve4_8_2d_h(s0, filter, offset_vec, shift, idx);
+      uint16x8_t d1 =
+          highbd_convolve4_8_2d_h(s1, filter, offset_vec, shift, idx);
+      uint16x8_t d2 =
+          highbd_convolve4_8_2d_h(s2, filter, offset_vec, shift, idx);
+
+      store_u16_8x3(dst, dst_stride, d0, d1, d2);
+
+      s += 8;
+      dst += 8;
+      width -= 8;
+    } while (width != 0);
+  }
+}
+
 static INLINE uint16x4_t highbd_convolve8_4_2d_v(int16x8_t samples_lo[2],
                                                  int16x8_t samples_hi[2],
                                                  int16x8_t filter,
@@ -1130,6 +1276,111 @@ static INLINE void highbd_dist_wtd_convolve_2d_vert_8tap_sve2(
   }
 }
 
+static INLINE uint16x4_t highbd_convolve4_4_2d_v(
+    const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
+    const int16x4_t s3, const int16x4_t filter, const int32x4_t offset) {
+  int32x4_t sum = vmlal_lane_s16(offset, s0, filter, 0);
+  sum = vmlal_lane_s16(sum, s1, filter, 1);
+  sum = vmlal_lane_s16(sum, s2, filter, 2);
+  sum = vmlal_lane_s16(sum, s3, filter, 3);
+
+  return vqrshrun_n_s32(sum, COMPOUND_ROUND1_BITS);
+}
+
+static INLINE uint16x8_t highbd_convolve4_8_2d_v(
+    const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
+    const int16x8_t s3, const int16x4_t filter, const int32x4_t offset) {
+  int32x4_t sum0 = vmlal_lane_s16(offset, vget_low_s16(s0), filter, 0);
+  sum0 = vmlal_lane_s16(sum0, vget_low_s16(s1), filter, 1);
+  sum0 = vmlal_lane_s16(sum0, vget_low_s16(s2), filter, 2);
+  sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), filter, 3);
+
+  int32x4_t sum1 = vmlal_lane_s16(offset, vget_high_s16(s0), filter, 0);
+  sum1 = vmlal_lane_s16(sum1, vget_high_s16(s1), filter, 1);
+  sum1 = vmlal_lane_s16(sum1, vget_high_s16(s2), filter, 2);
+  sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), filter, 3);
+
+  return vcombine_u16(vqrshrun_n_s32(sum0, COMPOUND_ROUND1_BITS),
+                      vqrshrun_n_s32(sum1, COMPOUND_ROUND1_BITS));
+}
+
+static INLINE void highbd_dist_wtd_convolve_2d_vert_4tap_neon(
+    const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride,
+    int w, int h, const int16_t *y_filter_ptr, const int offset) {
+  const int16x4_t y_filter = vld1_s16(y_filter_ptr + 2);
+  const int32x4_t offset_vec = vdupq_n_s32(offset);
+
+  if (w == 4) {
+    const int16_t *s = (const int16_t *)src_ptr;
+    uint16_t *d = dst_ptr;
+
+    int16x4_t s0, s1, s2;
+    load_s16_4x3(s, src_stride, &s0, &s1, &s2);
+    s += 3 * src_stride;
+
+    do {
+      int16x4_t s3, s4, s5, s6;
+      load_s16_4x4(s, src_stride, &s3, &s4, &s5, &s6);
+
+      uint16x4_t d0 =
+          highbd_convolve4_4_2d_v(s0, s1, s2, s3, y_filter, offset_vec);
+      uint16x4_t d1 =
+          highbd_convolve4_4_2d_v(s1, s2, s3, s4, y_filter, offset_vec);
+      uint16x4_t d2 =
+          highbd_convolve4_4_2d_v(s2, s3, s4, s5, y_filter, offset_vec);
+      uint16x4_t d3 =
+          highbd_convolve4_4_2d_v(s3, s4, s5, s6, y_filter, offset_vec);
+
+      store_u16_4x4(d, dst_stride, d0, d1, d2, d3);
+
+      s0 = s4;
+      s1 = s5;
+      s2 = s6;
+
+      s += 4 * src_stride;
+      d += 4 * dst_stride;
+      h -= 4;
+    } while (h != 0);
+  } else {
+    do {
+      int height = h;
+      const int16_t *s = (const int16_t *)src_ptr;
+      uint16_t *d = dst_ptr;
+
+      int16x8_t s0, s1, s2;
+      load_s16_8x3(s, src_stride, &s0, &s1, &s2);
+      s += 3 * src_stride;
+
+      do {
+        int16x8_t s3, s4, s5, s6;
+        load_s16_8x4(s, src_stride, &s3, &s4, &s5, &s6);
+
+        uint16x8_t d0 =
+            highbd_convolve4_8_2d_v(s0, s1, s2, s3, y_filter, offset_vec);
+        uint16x8_t d1 =
+            highbd_convolve4_8_2d_v(s1, s2, s3, s4, y_filter, offset_vec);
+        uint16x8_t d2 =
+            highbd_convolve4_8_2d_v(s2, s3, s4, s5, y_filter, offset_vec);
+        uint16x8_t d3 =
+            highbd_convolve4_8_2d_v(s3, s4, s5, s6, y_filter, offset_vec);
+
+        store_u16_8x4(d, dst_stride, d0, d1, d2, d3);
+
+        s0 = s4;
+        s1 = s5;
+        s2 = s6;
+
+        s += 4 * src_stride;
+        d += 4 * dst_stride;
+        height -= 4;
+      } while (height != 0);
+      src_ptr += 8;
+      dst_ptr += 8;
+      w -= 8;
+    } while (w != 0);
+  }
+}
+
 void av1_highbd_dist_wtd_convolve_2d_sve2(
     const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
     int h, const InterpFilterParams *filter_params_x,
@@ -1143,20 +1394,22 @@ void av1_highbd_dist_wtd_convolve_2d_sve2(
   CONV_BUF_TYPE *dst16 = conv_params->dst;
   int dst16_stride = conv_params->dst_stride;
   const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn);
+  const int clamped_x_taps = x_filter_taps < 4 ? 4 : x_filter_taps;
 
   const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn);
+  const int clamped_y_taps = y_filter_taps < 4 ? 4 : y_filter_taps;
 
-  if (x_filter_taps != 8 || y_filter_taps != 8) {
+  if (x_filter_taps == 6 || y_filter_taps == 6) {
     av1_highbd_dist_wtd_convolve_2d_neon(
         src, src_stride, dst, dst_stride, w, h, filter_params_x,
         filter_params_y, subpel_x_qn, subpel_y_qn, conv_params, bd);
     return;
   }
 
-  const int im_h = h + y_filter_taps - 1;
+  const int im_h = h + clamped_y_taps - 1;
   const int im_stride = MAX_SB_SIZE;
-  const int vert_offset = y_filter_taps / 2 - 1;
-  const int horiz_offset = x_filter_taps / 2 - 1;
+  const int vert_offset = clamped_y_taps / 2 - 1;
+  const int horiz_offset = clamped_x_taps / 2 - 1;
   // The extra shim of (1 << (conv_params->round_0 - 1)) allows us to use a
   // faster non-rounding non-saturating left shift.
   const int round_offset_conv_x =
@@ -1171,14 +1424,26 @@ void av1_highbd_dist_wtd_convolve_2d_sve2(
   const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel(
       filter_params_y, subpel_y_qn & SUBPEL_MASK);
 
-  highbd_dist_wtd_convolve_2d_horiz_8tap_sve2(src_ptr, src_stride, im_block,
-                                              im_stride, w, im_h, x_filter_ptr,
-                                              conv_params, round_offset_conv_x);
+  if (x_filter_taps <= 4) {
+    highbd_dist_wtd_convolve_2d_horiz_4tap_sve2(
+        src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr,
+        conv_params, round_offset_conv_x);
+  } else {
+    highbd_dist_wtd_convolve_2d_horiz_8tap_sve2(
+        src_ptr, src_stride, im_block, im_stride, w, im_h, x_filter_ptr,
+        conv_params, round_offset_conv_x);
+  }
 
   if (conv_params->do_average) {
-    highbd_dist_wtd_convolve_2d_vert_8tap_sve2(im_block, im_stride, im_block2,
-                                               im_stride, w, h, y_filter_ptr,
-                                               round_offset_conv_y);
+    if (y_filter_taps <= 4) {
+      highbd_dist_wtd_convolve_2d_vert_4tap_neon(im_block, im_stride, im_block2,
+                                                 im_stride, w, h, y_filter_ptr,
+                                                 round_offset_conv_y);
+    } else {
+      highbd_dist_wtd_convolve_2d_vert_8tap_sve2(im_block, im_stride, im_block2,
+                                                 im_stride, w, h, y_filter_ptr,
+                                                 round_offset_conv_y);
+    }
     if (conv_params->use_dist_wtd_comp_avg) {
       if (bd == 12) {
         highbd_12_dist_wtd_comp_avg_neon(im_block2, im_stride, dst, dst_stride,
@@ -1199,8 +1464,14 @@ void av1_highbd_dist_wtd_convolve_2d_sve2(
       }
     }
   } else {
-    highbd_dist_wtd_convolve_2d_vert_8tap_sve2(im_block, im_stride, dst16,
-                                               dst16_stride, w, h, y_filter_ptr,
-                                               round_offset_conv_y);
+    if (y_filter_taps <= 4) {
+      highbd_dist_wtd_convolve_2d_vert_4tap_neon(
+          im_block, im_stride, dst16, dst16_stride, w, h, y_filter_ptr,
+          round_offset_conv_y);
+    } else {
+      highbd_dist_wtd_convolve_2d_vert_8tap_sve2(
+          im_block, im_stride, dst16, dst16_stride, w, h, y_filter_ptr,
+          round_offset_conv_y);
+    }
   }
 }