av1/encoder/arm/wedge_utils_neon.c

/*
 * Copyright (c) 2022, Alliance for Open Media. All rights reserved.
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <arm_neon.h>
#include <assert.h>

#include "aom_dsp/arm/sum_neon.h"
#include "av1/common/reconinter.h"

#define MAX_MASK_VALUE (1 << WEDGE_WEIGHT_BITS)

/**
 * See av1_wedge_sse_from_residuals_c for details of the parameters and
 * computation.
 */
uint64_t av1_wedge_sse_from_residuals_neon(const int16_t *r1, const int16_t *d,
                                           const uint8_t *m, int N) {
  assert(N % 64 == 0);

  uint64x2_t v_csse[2] = { vdupq_n_u64(0), vdupq_n_u64(0) };

  int i = 0;
  do {
    int32x4_t sum[4];
    int32x4_t sse[2];
    int16x4_t sum_s16[4];

    const int16x8_t r1_l = vld1q_s16(r1 + i);
    const int16x8_t r1_h = vld1q_s16(r1 + i + 8);
    const int16x8_t d_l = vld1q_s16(d + i);
    const int16x8_t d_h = vld1q_s16(d + i + 8);
    // The following three lines are a bit inelegant compared to using a pair
    // of vmovl_u8()... but it forces the compiler to emit a ZIP1, ZIP2 pair -
    // which can be executed in parallel with the subsequent SSHL instructions.
    // (SSHL can only be executed on half of the Neon pipes in modern Arm
    // cores, whereas ZIP1/2 can be executed on all of them.)
    const uint8x16x2_t m_u16 = vzipq_u8(vld1q_u8(m + i), vdupq_n_u8(0));
    const int16x8_t m_l = vreinterpretq_s16_u8(m_u16.val[0]);
    const int16x8_t m_h = vreinterpretq_s16_u8(m_u16.val[1]);

    sum[0] = vshll_n_s16(vget_low_s16(r1_l), WEDGE_WEIGHT_BITS);
    sum[1] = vshll_n_s16(vget_high_s16(r1_l), WEDGE_WEIGHT_BITS);
    sum[2] = vshll_n_s16(vget_low_s16(r1_h), WEDGE_WEIGHT_BITS);
    sum[3] = vshll_n_s16(vget_high_s16(r1_h), WEDGE_WEIGHT_BITS);

    sum[0] = vmlal_s16(sum[0], vget_low_s16(m_l), vget_low_s16(d_l));
    sum[1] = vmlal_s16(sum[1], vget_high_s16(m_l), vget_high_s16(d_l));
    sum[2] = vmlal_s16(sum[2], vget_low_s16(m_h), vget_low_s16(d_h));
    sum[3] = vmlal_s16(sum[3], vget_high_s16(m_h), vget_high_s16(d_h));

    sum_s16[0] = vqmovn_s32(sum[0]);
    sum_s16[1] = vqmovn_s32(sum[1]);
    sum_s16[2] = vqmovn_s32(sum[2]);
    sum_s16[3] = vqmovn_s32(sum[3]);

    sse[0] = vmull_s16(sum_s16[0], sum_s16[0]);
    sse[1] = vmull_s16(sum_s16[2], sum_s16[2]);
    sse[0] = vmlal_s16(sse[0], sum_s16[1], sum_s16[1]);
    sse[1] = vmlal_s16(sse[1], sum_s16[3], sum_s16[3]);

    v_csse[0] = vpadalq_u32(v_csse[0], vreinterpretq_u32_s32(sse[0]));
    v_csse[1] = vpadalq_u32(v_csse[1], vreinterpretq_u32_s32(sse[1]));

    i += 16;
  } while (i < N);

  uint64_t csse = horizontal_add_u64x2(vaddq_u64(v_csse[0], v_csse[1]));
  return ROUND_POWER_OF_TWO(csse, 2 * WEDGE_WEIGHT_BITS);
}

int8_t av1_wedge_sign_from_residuals_neon(const int16_t *ds, const uint8_t *m,
                                          int N, int64_t limit) {
  int32x4_t acc[4] = { vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0),
                       vdupq_n_s32(0) };

  do {
    int16x8_t ds_l = vld1q_s16(ds);
    int16x8_t ds_h = vld1q_s16(ds + 8);

    int8x16_t m_s8 = vreinterpretq_s8_u8(vld1q_u8(m));
    int16x8_t m_l = vmovl_s8(vget_low_s8(m_s8));
    int16x8_t m_h = vmovl_s8(vget_high_s8(m_s8));

    acc[0] = vmlal_s16(acc[0], vget_low_s16(ds_l), vget_low_s16(m_l));
    acc[1] = vmlal_s16(acc[1], vget_high_s16(ds_l), vget_high_s16(m_l));
    acc[2] = vmlal_s16(acc[2], vget_low_s16(ds_h), vget_low_s16(m_h));
    acc[3] = vmlal_s16(acc[3], vget_high_s16(ds_h), vget_high_s16(m_h));

    ds += 16;
    m += 16;
    N -= 16;
  } while (N != 0);

  int64x2_t sum = vpaddlq_s32(acc[0]);
  sum = vpadalq_s32(sum, acc[1]);
  sum = vpadalq_s32(sum, acc[2]);
  sum = vpadalq_s32(sum, acc[3]);

  return (horizontal_add_s64x2(sum) > limit);
}

void av1_wedge_compute_delta_squares_neon(int16_t *d_ptr, const int16_t *a_ptr,
                                          const int16_t *b_ptr, int N) {
  do {
    int16x8_t a = vld1q_s16(a_ptr);
    int16x8_t b = vld1q_s16(b_ptr);

    int32x4_t sq_lo = vmull_s16(vget_low_s16(a), vget_low_s16(a));
    int32x4_t sq_hi = vmull_s16(vget_high_s16(a), vget_high_s16(a));

    sq_lo = vmlsl_s16(sq_lo, vget_low_s16(b), vget_low_s16(b));
    sq_hi = vmlsl_s16(sq_hi, vget_high_s16(b), vget_high_s16(b));

    int16x8_t res = vcombine_s16(vqmovn_s32(sq_lo), vqmovn_s32(sq_hi));

    vst1q_s16(d_ptr, res);

    d_ptr += 8;
    a_ptr += 8;
    b_ptr += 8;
    N -= 8;
  } while (N != 0);
}