aom: Add Armv8.0 Neon 6-tap filter path for convolve_horiz_scale

From a88627a130a7b455cc14201d23bd11a75e038b5f Mon Sep 17 00:00:00 2001
From: Gerda Zsejke More <[EMAIL REDACTED]>
Date: Thu, 30 May 2024 16:32:07 +0200
Subject: [PATCH] Add Armv8.0 Neon 6-tap filter path for convolve_horiz_scale

The filter values used in the scaling algorithm are specified in the
documentation of the inter-prediction process (chapter 7.11.3.4. [1]).
These filter values are defined in the av1_interp_filter_params_list
in filter.h. An important characteristic of these filters, except the
MULTITAP_SHARP filter, is that at indices 0 and 7 the values are 0.

Add an implementation for horizontal filtering that specialises on
6-tap filters. This way we avoid redundant work associated with 8-tap
filters. This approach is not applicable for the DotProd and I8MM
version of this horizontal filtering, given that the dot product
instructions accumulate the result of 4 multiplications.

[1]https://aomediacodec.github.io/av1-spec/av1-spec.pdf

Change-Id: I09fcc138289c3b9ee39099b6cdab740e7049260c
---
 av1/common/arm/av1_convolve_scale_neon.c | 184 +++++++++++++++++++++--
 1 file changed, 175 insertions(+), 9 deletions(-)

diff --git a/av1/common/arm/av1_convolve_scale_neon.c b/av1/common/arm/av1_convolve_scale_neon.c
index a972a19ef..88d126ead 100644
--- a/av1/common/arm/av1_convolve_scale_neon.c
+++ b/av1/common/arm/av1_convolve_scale_neon.c
@@ -68,12 +68,12 @@ static INLINE int16x8_t convolve8_8_h(const int16x8_t s0, const int16x8_t s1,
   return vshrq_n_s16(sum, ROUND0_BITS - 1);
 }
 
-static INLINE void convolve_horiz_scale_neon(const uint8_t *src, int src_stride,
-                                             int16_t *dst, int dst_stride,
-                                             int w, int h,
-                                             const int16_t *x_filter,
-                                             const int subpel_x_qn,
-                                             const int x_step_qn) {
+static INLINE void convolve_horiz_scale_8tap_neon(const uint8_t *src,
+                                                  int src_stride, int16_t *dst,
+                                                  int dst_stride, int w, int h,
+                                                  const int16_t *x_filter,
+                                                  const int subpel_x_qn,
+                                                  const int x_step_qn) {
   DECLARE_ALIGNED(16, int16_t, temp[8 * 8]);
   const int bd = 8;
 
@@ -191,6 +191,166 @@ static INLINE void convolve_horiz_scale_neon(const uint8_t *src, int src_stride,
   }
 }
 
+static INLINE int16x4_t convolve6_4_h(const int16x4_t s0, const int16x4_t s1,
+                                      const int16x4_t s2, const int16x4_t s3,
+                                      const int16x4_t s4, const int16x4_t s5,
+                                      const int16x8_t filter,
+                                      const int32x4_t horiz_const) {
+  int16x4_t filter_lo = vget_low_s16(filter);
+  int16x4_t filter_hi = vget_high_s16(filter);
+
+  int32x4_t sum = horiz_const;
+  // Filter values at indices 0 and 7 are 0.
+  sum = vmlal_lane_s16(sum, s0, filter_lo, 1);
+  sum = vmlal_lane_s16(sum, s1, filter_lo, 2);
+  sum = vmlal_lane_s16(sum, s2, filter_lo, 3);
+  sum = vmlal_lane_s16(sum, s3, filter_hi, 0);
+  sum = vmlal_lane_s16(sum, s4, filter_hi, 1);
+  sum = vmlal_lane_s16(sum, s5, filter_hi, 2);
+
+  return vshrn_n_s32(sum, ROUND0_BITS);
+}
+
+static INLINE int16x8_t convolve6_8_h(const int16x8_t s0, const int16x8_t s1,
+                                      const int16x8_t s2, const int16x8_t s3,
+                                      const int16x8_t s4, const int16x8_t s5,
+                                      const int16x8_t filter,
+                                      const int16x8_t horiz_const) {
+  int16x4_t filter_lo = vget_low_s16(filter);
+  int16x4_t filter_hi = vget_high_s16(filter);
+
+  int16x8_t sum = horiz_const;
+  // Filter values at indices 0 and 7 are 0.
+  sum = vmlaq_lane_s16(sum, s0, filter_lo, 1);
+  sum = vmlaq_lane_s16(sum, s1, filter_lo, 2);
+  sum = vmlaq_lane_s16(sum, s2, filter_lo, 3);
+  sum = vmlaq_lane_s16(sum, s3, filter_hi, 0);
+  sum = vmlaq_lane_s16(sum, s4, filter_hi, 1);
+  sum = vmlaq_lane_s16(sum, s5, filter_hi, 2);
+
+  // We halved the filter values so -1 from right shift.
+  return vshrq_n_s16(sum, ROUND0_BITS - 1);
+}
+
+static INLINE void convolve_horiz_scale_6tap_neon(const uint8_t *src,
+                                                  int src_stride, int16_t *dst,
+                                                  int dst_stride, int w, int h,
+                                                  const int16_t *x_filter,
+                                                  const int subpel_x_qn,
+                                                  const int x_step_qn) {
+  DECLARE_ALIGNED(16, int16_t, temp[8 * 8]);
+  const int bd = 8;
+
+  if (w == 4) {
+    // The shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding shifts.
+    const int32x4_t horiz_offset =
+        vdupq_n_s32((1 << (bd + FILTER_BITS - 1)) + (1 << (ROUND0_BITS - 1)));
+
+    do {
+      int x_qn = subpel_x_qn;
+
+      // Process a 4x4 tile.
+      for (int r = 0; r < 4; ++r) {
+        const uint8_t *const s = &src[x_qn >> SCALE_SUBPEL_BITS];
+
+        const ptrdiff_t filter_offset =
+            SUBPEL_TAPS * ((x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS);
+        const int16x8_t filter = vld1q_s16(x_filter + filter_offset);
+
+        uint8x8_t t0, t1, t2, t3;
+        load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
+
+        transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
+
+        int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+        int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+        int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+        int16x4_t s3 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+        int16x4_t s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+        int16x4_t s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+
+        int16x4_t d0 =
+            convolve6_4_h(s0, s1, s2, s3, s4, s5, filter, horiz_offset);
+
+        vst1_s16(&temp[r * 4], d0);
+        x_qn += x_step_qn;
+      }
+
+      // Transpose the 4x4 result tile and store.
+      int16x4_t d0, d1, d2, d3;
+      load_s16_4x4(temp, 4, &d0, &d1, &d2, &d3);
+
+      transpose_elems_inplace_s16_4x4(&d0, &d1, &d2, &d3);
+
+      store_s16_4x4(dst, dst_stride, d0, d1, d2, d3);
+
+      dst += 4 * dst_stride;
+      src += 4 * src_stride;
+      h -= 4;
+    } while (h > 0);
+  } else {
+    // The shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding shifts.
+    // The additional -1 is needed because we are halving the filter values.
+    const int16x8_t horiz_offset =
+        vdupq_n_s16((1 << (bd + FILTER_BITS - 2)) + (1 << (ROUND0_BITS - 2)));
+
+    do {
+      int x_qn = subpel_x_qn;
+      int16_t *d = dst;
+      int width = w;
+
+      do {
+        // Process an 8x8 tile.
+        for (int r = 0; r < 8; ++r) {
+          const uint8_t *const s = &src[(x_qn >> SCALE_SUBPEL_BITS)];
+
+          const ptrdiff_t filter_offset =
+              SUBPEL_TAPS * ((x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS);
+          int16x8_t filter = vld1q_s16(x_filter + filter_offset);
+          // Filter values are all even so halve them to allow convolution
+          // kernel computations to stay in 16-bit element types.
+          filter = vshrq_n_s16(filter, 1);
+
+          uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7;
+          load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+
+          transpose_elems_u8_8x8(t0, t1, t2, t3, t4, t5, t6, t7, &t0, &t1, &t2,
+                                 &t3, &t4, &t5, &t6, &t7);
+
+          int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t1));
+          int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t2));
+          int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t3));
+          int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t4));
+          int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t5));
+          int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+          int16x8_t d0 =
+              convolve6_8_h(s0, s1, s2, s3, s4, s5, filter, horiz_offset);
+
+          vst1q_s16(&temp[r * 8], d0);
+
+          x_qn += x_step_qn;
+        }
+
+        // Transpose the 8x8 result tile and store.
+        int16x8_t d0, d1, d2, d3, d4, d5, d6, d7;
+        load_s16_8x8(temp, 8, &d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
+
+        transpose_elems_inplace_s16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
+
+        store_s16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
+
+        d += 8;
+        width -= 8;
+      } while (width != 0);
+
+      dst += 8 * dst_stride;
+      src += 8 * src_stride;
+      h -= 8;
+    } while (h > 0);
+  }
+}
+
 void av1_convolve_2d_scale_neon(const uint8_t *src, int src_stride,
                                 uint8_t *dst, int dst_stride, int w, int h,
                                 const InterpFilterParams *filter_params_x,
@@ -222,9 +382,15 @@ void av1_convolve_2d_scale_neon(const uint8_t *src, int src_stride,
   const ptrdiff_t vert_offset = (filter_params_y->taps / 2 - 1) * src_stride;
 
   // Horizontal filter
-  convolve_horiz_scale_neon(
-      src - horiz_offset - vert_offset, src_stride, im_block, im_stride, w,
-      im_h, filter_params_x->filter_ptr, subpel_x_qn, x_step_qn);
+  if (filter_params_x->interp_filter == MULTITAP_SHARP) {
+    convolve_horiz_scale_8tap_neon(
+        src - horiz_offset - vert_offset, src_stride, im_block, im_stride, w,
+        im_h, filter_params_x->filter_ptr, subpel_x_qn, x_step_qn);
+  } else {
+    convolve_horiz_scale_6tap_neon(
+        src - horiz_offset - vert_offset, src_stride, im_block, im_stride, w,
+        im_h, filter_params_x->filter_ptr, subpel_x_qn, x_step_qn);
+  }
 
   // Vertical filter
   if (filter_params_y->interp_filter == MULTITAP_SHARP) {