aom: Add SVE2 impl of HBD dist_wtd_convolve_2d for 8-tap filters

From eefd5585a0c4c204fcf7d30065f8c2ca35c38a82 Mon Sep 17 00:00:00 2001
From: Salome Thirot <[EMAIL REDACTED]>
Date: Mon, 18 Mar 2024 15:03:01 +0000
Subject: [PATCH] Add SVE2 impl of HBD dist_wtd_convolve_2d for 8-tap filters

Add SVE2 implementation of av1_highbd_dist_wtd_convolve_2d for 8-tap
filters, as well as the corresponding tests. This gives up to 25% uplift
over the Neon implementation.

Change-Id: Ie7d8619e0a924b2da2fc04af16a707caaf8759c7
---
 aom_dsp/arm/mem_neon.h                        |  10 +
 .../arm/highbd_compound_convolve_sve2.c       | 356 ++++++++++++++++++
 av1/common/av1_rtcd_defs.pl                   |   2 +-
 test/av1_convolve_test.cc                     |   6 +
 4 files changed, 373 insertions(+), 1 deletion(-)

diff --git a/aom_dsp/arm/mem_neon.h b/aom_dsp/arm/mem_neon.h
index a4e9eb2e7..d749f1def 100644
--- a/aom_dsp/arm/mem_neon.h
+++ b/aom_dsp/arm/mem_neon.h
@@ -556,6 +556,16 @@ static INLINE void store_u16_8x2(uint16_t *s, ptrdiff_t dst_stride,
   vst1q_u16(s, s1);
 }
 
+static INLINE void store_u16_8x3(uint16_t *s, ptrdiff_t dst_stride,
+                                 const uint16x8_t s0, const uint16x8_t s1,
+                                 const uint16x8_t s2) {
+  vst1q_u16(s, s0);
+  s += dst_stride;
+  vst1q_u16(s, s1);
+  s += dst_stride;
+  vst1q_u16(s, s2);
+}
+
 static INLINE void store_u16_8x4(uint16_t *s, ptrdiff_t dst_stride,
                                  const uint16x8_t s0, const uint16x8_t s1,
                                  const uint16x8_t s2, const uint16x8_t s3) {
diff --git a/av1/common/arm/highbd_compound_convolve_sve2.c b/av1/common/arm/highbd_compound_convolve_sve2.c
index d9ea83d55..dd55e346a 100644
--- a/av1/common/arm/highbd_compound_convolve_sve2.c
+++ b/av1/common/arm/highbd_compound_convolve_sve2.c
@@ -848,3 +848,359 @@ void av1_highbd_dist_wtd_convolve_y_sve2(
     }
   }
 }
+
+static INLINE uint16x8_t highbd_convolve8_8_2d_h(int16x8_t s0[8],
+                                                 int16x8_t filter,
+                                                 int64x2_t offset,
+                                                 int32x4_t shift) {
+  int64x2_t sum[8];
+
+  sum[0] = aom_sdotq_s16(offset, s0[0], filter);
+  sum[1] = aom_sdotq_s16(offset, s0[1], filter);
+  sum[2] = aom_sdotq_s16(offset, s0[2], filter);
+  sum[3] = aom_sdotq_s16(offset, s0[3], filter);
+  sum[4] = aom_sdotq_s16(offset, s0[4], filter);
+  sum[5] = aom_sdotq_s16(offset, s0[5], filter);
+  sum[6] = aom_sdotq_s16(offset, s0[6], filter);
+  sum[7] = aom_sdotq_s16(offset, s0[7], filter);
+
+  sum[0] = vpaddq_s64(sum[0], sum[1]);
+  sum[2] = vpaddq_s64(sum[2], sum[3]);
+  sum[4] = vpaddq_s64(sum[4], sum[5]);
+  sum[6] = vpaddq_s64(sum[6], sum[7]);
+
+  int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[2]));
+  int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum[4]), vmovn_s64(sum[6]));
+
+  sum0123 = vshlq_s32(sum0123, shift);
+  sum4567 = vshlq_s32(sum4567, shift);
+
+  return vcombine_u16(vqmovun_s32(sum0123), vqmovun_s32(sum4567));
+}
+
+static INLINE void highbd_dist_wtd_convolve_2d_horiz_8tap_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 int32x4_t shift = vdupq_n_s32(-conv_params->round_0);
+
+  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.
+
+  do {
+    const int16_t *s = (const int16_t *)src;
+    uint16_t *d = dst;
+    int w = width;
+
+    do {
+      int16x8_t s0[8], s1[8], s2[8], s3[8];
+      load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3],
+                   &s0[4], &s0[5], &s0[6], &s0[7]);
+      load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3],
+                   &s1[4], &s1[5], &s1[6], &s1[7]);
+      load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3],
+                   &s2[4], &s2[5], &s2[6], &s2[7]);
+      load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3],
+                   &s3[4], &s3[5], &s3[6], &s3[7]);
+
+      uint16x8_t d0 = highbd_convolve8_8_2d_h(s0, filter, offset_lo, shift);
+      uint16x8_t d1 = highbd_convolve8_8_2d_h(s1, filter, offset_lo, shift);
+      uint16x8_t d2 = highbd_convolve8_8_2d_h(s2, filter, offset_lo, shift);
+      uint16x8_t d3 = highbd_convolve8_8_2d_h(s3, filter, offset_lo, shift);
+
+      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[8], s1[8], s2[8];
+    load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3], &s0[4],
+                 &s0[5], &s0[6], &s0[7]);
+    load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3], &s1[4],
+                 &s1[5], &s1[6], &s1[7]);
+    load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3], &s2[4],
+                 &s2[5], &s2[6], &s2[7]);
+
+    uint16x8_t d0 = highbd_convolve8_8_2d_h(s0, filter, offset_lo, shift);
+    uint16x8_t d1 = highbd_convolve8_8_2d_h(s1, filter, offset_lo, shift);
+    uint16x8_t d2 = highbd_convolve8_8_2d_h(s2, filter, offset_lo, shift);
+
+    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,
+                                                 int64x2_t offset) {
+  int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0);
+  sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1);
+
+  int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0);
+  sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1);
+
+  int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23));
+
+  return vqrshrun_n_s32(sum0123, COMPOUND_ROUND1_BITS);
+}
+
+static INLINE uint16x8_t highbd_convolve8_8_2d_v(int16x8_t samples_lo[4],
+                                                 int16x8_t samples_hi[4],
+                                                 int16x8_t filter,
+                                                 int64x2_t offset) {
+  int64x2_t sum01 = aom_svdot_lane_s16(offset, samples_lo[0], filter, 0);
+  sum01 = aom_svdot_lane_s16(sum01, samples_hi[0], filter, 1);
+
+  int64x2_t sum23 = aom_svdot_lane_s16(offset, samples_lo[1], filter, 0);
+  sum23 = aom_svdot_lane_s16(sum23, samples_hi[1], filter, 1);
+
+  int64x2_t sum45 = aom_svdot_lane_s16(offset, samples_lo[2], filter, 0);
+  sum45 = aom_svdot_lane_s16(sum45, samples_hi[2], filter, 1);
+
+  int64x2_t sum67 = aom_svdot_lane_s16(offset, samples_lo[3], filter, 0);
+  sum67 = aom_svdot_lane_s16(sum67, samples_hi[3], filter, 1);
+
+  int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23));
+  int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67));
+
+  return vcombine_u16(vqrshrun_n_s32(sum0123, COMPOUND_ROUND1_BITS),
+                      vqrshrun_n_s32(sum4567, COMPOUND_ROUND1_BITS));
+}
+
+static INLINE void highbd_dist_wtd_convolve_2d_vert_8tap_sve2(
+    const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride,
+    int width, int height, const int16_t *y_filter_ptr, int offset) {
+  const int16x8_t y_filter = vld1q_s16(y_filter_ptr);
+  const int64x2_t offset_s64 = vdupq_n_s64(offset);
+
+  uint16x8x3_t merge_block_tbl = vld1q_u16_x3(kDotProdMergeBlockTbl);
+  // Scale indices by size of the true vector length to avoid reading from an
+  // 'undefined' portion of a vector on a system with SVE vectors > 128-bit.
+  uint16x8_t correction0 =
+      vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000000000000ULL));
+  merge_block_tbl.val[0] = vaddq_u16(merge_block_tbl.val[0], correction0);
+
+  uint16x8_t correction1 =
+      vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100000000ULL));
+  merge_block_tbl.val[1] = vaddq_u16(merge_block_tbl.val[1], correction1);
+
+  uint16x8_t correction2 =
+      vreinterpretq_u16_u64(vdupq_n_u64(svcnth() * 0x0001000100010000ULL));
+  merge_block_tbl.val[2] = vaddq_u16(merge_block_tbl.val[2], correction2);
+
+  if (width == 4) {
+    int16_t *s = (int16_t *)src;
+    int16x4_t s0, s1, s2, s3, s4, s5, s6;
+    load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
+    s += 7 * src_stride;
+
+    // This operation combines a conventional transpose and the sample permute
+    // required before computing the dot product.
+    int16x8_t s0123[2], s1234[2], s2345[2], s3456[2];
+    transpose_concat_4x4(s0, s1, s2, s3, s0123);
+    transpose_concat_4x4(s1, s2, s3, s4, s1234);
+    transpose_concat_4x4(s2, s3, s4, s5, s2345);
+    transpose_concat_4x4(s3, s4, s5, s6, s3456);
+
+    do {
+      int16x4_t s7, s8, s9, s10;
+      load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10);
+
+      int16x8_t s4567[2], s5678[2], s6789[2], s789A[2];
+      // Transpose and shuffle the 4 lines that were loaded.
+      transpose_concat_4x4(s7, s8, s9, s10, s789A);
+
+      // Merge new data into block from previous iteration.
+      aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[0], s4567);
+      aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[1], s5678);
+      aom_tbl2x2_s16(s3456, s789A, merge_block_tbl.val[2], s6789);
+
+      uint16x4_t d0 =
+          highbd_convolve8_4_2d_v(s0123, s4567, y_filter, offset_s64);
+      uint16x4_t d1 =
+          highbd_convolve8_4_2d_v(s1234, s5678, y_filter, offset_s64);
+      uint16x4_t d2 =
+          highbd_convolve8_4_2d_v(s2345, s6789, y_filter, offset_s64);
+      uint16x4_t d3 =
+          highbd_convolve8_4_2d_v(s3456, s789A, y_filter, offset_s64);
+
+      store_u16_4x4(dst, dst_stride, d0, d1, d2, d3);
+
+      // Prepare block for next iteration - re-using as much as possible.
+      // Shuffle everything up four rows.
+      s0123[0] = s4567[0];
+      s0123[1] = s4567[1];
+      s1234[0] = s5678[0];
+      s1234[1] = s5678[1];
+      s2345[0] = s6789[0];
+      s2345[1] = s6789[1];
+      s3456[0] = s789A[0];
+      s3456[1] = s789A[1];
+
+      s += 4 * src_stride;
+      dst += 4 * dst_stride;
+      height -= 4;
+    } while (height != 0);
+  } else {
+    do {
+      int h = height;
+      int16_t *s = (int16_t *)src;
+      uint16_t *d = dst;
+
+      int16x8_t s0, s1, s2, s3, s4, s5, s6;
+      load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
+      s += 7 * src_stride;
+
+      // This operation combines a conventional transpose and the sample permute
+      // required before computing the dot product.
+      int16x8_t s0123[4], s1234[4], s2345[4], s3456[4];
+      transpose_concat_8x4(s0, s1, s2, s3, s0123);
+      transpose_concat_8x4(s1, s2, s3, s4, s1234);
+      transpose_concat_8x4(s2, s3, s4, s5, s2345);
+      transpose_concat_8x4(s3, s4, s5, s6, s3456);
+
+      do {
+        int16x8_t s7, s8, s9, s10;
+        load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10);
+        int16x8_t s4567[4], s5678[4], s6789[4], s789A[4];
+
+        // Transpose and shuffle the 4 lines that were loaded.
+        transpose_concat_8x4(s7, s8, s9, s10, s789A);
+
+        // Merge new data into block from previous iteration.
+        aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[0], s4567);
+        aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[1], s5678);
+        aom_tbl2x4_s16(s3456, s789A, merge_block_tbl.val[2], s6789);
+
+        uint16x8_t d0 =
+            highbd_convolve8_8_2d_v(s0123, s4567, y_filter, offset_s64);
+        uint16x8_t d1 =
+            highbd_convolve8_8_2d_v(s1234, s5678, y_filter, offset_s64);
+        uint16x8_t d2 =
+            highbd_convolve8_8_2d_v(s2345, s6789, y_filter, offset_s64);
+        uint16x8_t d3 =
+            highbd_convolve8_8_2d_v(s3456, s789A, y_filter, offset_s64);
+
+        store_u16_8x4(d, dst_stride, d0, d1, d2, d3);
+
+        // Prepare block for next iteration - re-using as much as possible.
+        // Shuffle everything up four rows.
+        s0123[0] = s4567[0];
+        s0123[1] = s4567[1];
+        s0123[2] = s4567[2];
+        s0123[3] = s4567[3];
+        s1234[0] = s5678[0];
+        s1234[1] = s5678[1];
+        s1234[2] = s5678[2];
+        s1234[3] = s5678[3];
+        s2345[0] = s6789[0];
+        s2345[1] = s6789[1];
+        s2345[2] = s6789[2];
+        s2345[3] = s6789[3];
+        s3456[0] = s789A[0];
+        s3456[1] = s789A[1];
+        s3456[2] = s789A[2];
+        s3456[3] = s789A[3];
+
+        s += 4 * src_stride;
+        d += 4 * dst_stride;
+        h -= 4;
+      } while (h != 0);
+      src += 8;
+      dst += 8;
+      width -= 8;
+    } while (width != 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,
+    const InterpFilterParams *filter_params_y, const int subpel_x_qn,
+    const int subpel_y_qn, ConvolveParams *conv_params, int bd) {
+  DECLARE_ALIGNED(16, uint16_t,
+                  im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint16_t,
+                  im_block2[(MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]);
+
+  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 y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn);
+
+  if (x_filter_taps != 8 || y_filter_taps != 8) {
+    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_stride = MAX_SB_SIZE;
+  const int vert_offset = y_filter_taps / 2 - 1;
+  const int horiz_offset = x_filter_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 =
+      (1 << (bd + FILTER_BITS - 1)) + (1 << (conv_params->round_0 - 1));
+  const int y_offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int round_offset_conv_y = (1 << y_offset_bits);
+
+  const uint16_t *src_ptr = src - vert_offset * src_stride - horiz_offset;
+
+  const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel(
+      filter_params_x, subpel_x_qn & SUBPEL_MASK);
+  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 (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 (conv_params->use_dist_wtd_comp_avg) {
+      if (bd == 12) {
+        highbd_12_dist_wtd_comp_avg_neon(im_block2, im_stride, dst, dst_stride,
+                                         w, h, conv_params);
+
+      } else {
+        highbd_dist_wtd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w,
+                                      h, conv_params, bd);
+      }
+    } else {
+      if (bd == 12) {
+        highbd_12_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, h,
+                                conv_params);
+
+      } else {
+        highbd_comp_avg_neon(im_block2, im_stride, dst, dst_stride, w, h,
+                             conv_params, bd);
+      }
+    }
+  } 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);
+  }
+}
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index 9113e4457..c0831330d 100644
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -606,7 +606,7 @@ ()
   specialize qw/av1_dist_wtd_convolve_x sse2 avx2 neon neon_dotprod neon_i8mm/;
   specialize qw/av1_dist_wtd_convolve_y sse2 avx2 neon/;
   if(aom_config("CONFIG_AV1_HIGHBITDEPTH") eq "yes") {
-    specialize qw/av1_highbd_dist_wtd_convolve_2d sse4_1 avx2 neon/;
+    specialize qw/av1_highbd_dist_wtd_convolve_2d sse4_1 avx2 neon sve2/;
     specialize qw/av1_highbd_dist_wtd_convolve_x sse4_1 avx2 neon sve2/;
     specialize qw/av1_highbd_dist_wtd_convolve_y sse4_1 avx2 neon sve2/;
     specialize qw/av1_highbd_dist_wtd_convolve_2d_copy sse4_1 avx2 neon/;
diff --git a/test/av1_convolve_test.cc b/test/av1_convolve_test.cc
index 40a71667c..b2392276c 100644
--- a/test/av1_convolve_test.cc
+++ b/test/av1_convolve_test.cc
@@ -2464,6 +2464,12 @@ INSTANTIATE_TEST_SUITE_P(
     BuildHighbdLumaParams(av1_highbd_dist_wtd_convolve_2d_neon));
 #endif
 
+#if HAVE_SVE2
+INSTANTIATE_TEST_SUITE_P(
+    SVE2, AV1Convolve2DHighbdCompoundTest,
+    BuildHighbdLumaParams(av1_highbd_dist_wtd_convolve_2d_sve2));
+#endif
+
 #endif  // CONFIG_AV1_HIGHBITDEPTH
 
 }  // namespace