aom_dsp/x86/avg_intrin_sse2.c

/*
 * Copyright (c) 2016, 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 <immintrin.h>

#include "config/aom_dsp_rtcd.h"
#include "aom/aom_integer.h"
#include "aom_dsp/x86/bitdepth_conversion_sse2.h"
#include "aom_ports/mem.h"

static void hadamard_col8_sse2(__m128i *in, int iter) {
  __m128i a0 = in[0];
  __m128i a1 = in[1];
  __m128i a2 = in[2];
  __m128i a3 = in[3];
  __m128i a4 = in[4];
  __m128i a5 = in[5];
  __m128i a6 = in[6];
  __m128i a7 = in[7];

  __m128i b0 = _mm_add_epi16(a0, a1);
  __m128i b1 = _mm_sub_epi16(a0, a1);
  __m128i b2 = _mm_add_epi16(a2, a3);
  __m128i b3 = _mm_sub_epi16(a2, a3);
  __m128i b4 = _mm_add_epi16(a4, a5);
  __m128i b5 = _mm_sub_epi16(a4, a5);
  __m128i b6 = _mm_add_epi16(a6, a7);
  __m128i b7 = _mm_sub_epi16(a6, a7);

  a0 = _mm_add_epi16(b0, b2);
  a1 = _mm_add_epi16(b1, b3);
  a2 = _mm_sub_epi16(b0, b2);
  a3 = _mm_sub_epi16(b1, b3);
  a4 = _mm_add_epi16(b4, b6);
  a5 = _mm_add_epi16(b5, b7);
  a6 = _mm_sub_epi16(b4, b6);
  a7 = _mm_sub_epi16(b5, b7);

  if (iter == 0) {
    b0 = _mm_add_epi16(a0, a4);
    b7 = _mm_add_epi16(a1, a5);
    b3 = _mm_add_epi16(a2, a6);
    b4 = _mm_add_epi16(a3, a7);
    b2 = _mm_sub_epi16(a0, a4);
    b6 = _mm_sub_epi16(a1, a5);
    b1 = _mm_sub_epi16(a2, a6);
    b5 = _mm_sub_epi16(a3, a7);

    a0 = _mm_unpacklo_epi16(b0, b1);
    a1 = _mm_unpacklo_epi16(b2, b3);
    a2 = _mm_unpackhi_epi16(b0, b1);
    a3 = _mm_unpackhi_epi16(b2, b3);
    a4 = _mm_unpacklo_epi16(b4, b5);
    a5 = _mm_unpacklo_epi16(b6, b7);
    a6 = _mm_unpackhi_epi16(b4, b5);
    a7 = _mm_unpackhi_epi16(b6, b7);

    b0 = _mm_unpacklo_epi32(a0, a1);
    b1 = _mm_unpacklo_epi32(a4, a5);
    b2 = _mm_unpackhi_epi32(a0, a1);
    b3 = _mm_unpackhi_epi32(a4, a5);
    b4 = _mm_unpacklo_epi32(a2, a3);
    b5 = _mm_unpacklo_epi32(a6, a7);
    b6 = _mm_unpackhi_epi32(a2, a3);
    b7 = _mm_unpackhi_epi32(a6, a7);

    in[0] = _mm_unpacklo_epi64(b0, b1);
    in[1] = _mm_unpackhi_epi64(b0, b1);
    in[2] = _mm_unpacklo_epi64(b2, b3);
    in[3] = _mm_unpackhi_epi64(b2, b3);
    in[4] = _mm_unpacklo_epi64(b4, b5);
    in[5] = _mm_unpackhi_epi64(b4, b5);
    in[6] = _mm_unpacklo_epi64(b6, b7);
    in[7] = _mm_unpackhi_epi64(b6, b7);
  } else {
    in[0] = _mm_add_epi16(a0, a4);
    in[7] = _mm_add_epi16(a1, a5);
    in[3] = _mm_add_epi16(a2, a6);
    in[4] = _mm_add_epi16(a3, a7);
    in[2] = _mm_sub_epi16(a0, a4);
    in[6] = _mm_sub_epi16(a1, a5);
    in[1] = _mm_sub_epi16(a2, a6);
    in[5] = _mm_sub_epi16(a3, a7);
  }
}

static INLINE void hadamard_8x8_sse2(const int16_t *src_diff,
                                     ptrdiff_t src_stride, tran_low_t *coeff,
                                     int is_final) {
  __m128i src[8];
  src[0] = _mm_load_si128((const __m128i *)src_diff);
  src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
  src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
  src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
  src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
  src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
  src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
  src[7] = _mm_load_si128((const __m128i *)(src_diff += src_stride));

  hadamard_col8_sse2(src, 0);
  hadamard_col8_sse2(src, 1);

  if (is_final) {
    store_tran_low(src[0], coeff);
    coeff += 8;
    store_tran_low(src[1], coeff);
    coeff += 8;
    store_tran_low(src[2], coeff);
    coeff += 8;
    store_tran_low(src[3], coeff);
    coeff += 8;
    store_tran_low(src[4], coeff);
    coeff += 8;
    store_tran_low(src[5], coeff);
    coeff += 8;
    store_tran_low(src[6], coeff);
    coeff += 8;
    store_tran_low(src[7], coeff);
  } else {
    int16_t *coeff16 = (int16_t *)coeff;
    _mm_store_si128((__m128i *)coeff16, src[0]);
    coeff16 += 8;
    _mm_store_si128((__m128i *)coeff16, src[1]);
    coeff16 += 8;
    _mm_store_si128((__m128i *)coeff16, src[2]);
    coeff16 += 8;
    _mm_store_si128((__m128i *)coeff16, src[3]);
    coeff16 += 8;
    _mm_store_si128((__m128i *)coeff16, src[4]);
    coeff16 += 8;
    _mm_store_si128((__m128i *)coeff16, src[5]);
    coeff16 += 8;
    _mm_store_si128((__m128i *)coeff16, src[6]);
    coeff16 += 8;
    _mm_store_si128((__m128i *)coeff16, src[7]);
  }
}

void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
                           tran_low_t *coeff) {
  hadamard_8x8_sse2(src_diff, src_stride, coeff, 1);
}

static INLINE void hadamard_16x16_sse2(const int16_t *src_diff,
                                       ptrdiff_t src_stride, tran_low_t *coeff,
                                       int is_final) {
  // For high bitdepths, it is unnecessary to store_tran_low
  // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
  // next stage.  Output to an intermediate buffer first, then store_tran_low()
  // in the final stage.
  DECLARE_ALIGNED(32, int16_t, temp_coeff[16 * 16]);
  int16_t *t_coeff = temp_coeff;
  int16_t *coeff16 = (int16_t *)coeff;
  int idx;
  for (idx = 0; idx < 4; ++idx) {
    const int16_t *src_ptr =
        src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
    hadamard_8x8_sse2(src_ptr, src_stride, (tran_low_t *)(t_coeff + idx * 64),
                      0);
  }

  for (idx = 0; idx < 64; idx += 8) {
    __m128i coeff0 = _mm_load_si128((const __m128i *)t_coeff);
    __m128i coeff1 = _mm_load_si128((const __m128i *)(t_coeff + 64));
    __m128i coeff2 = _mm_load_si128((const __m128i *)(t_coeff + 128));
    __m128i coeff3 = _mm_load_si128((const __m128i *)(t_coeff + 192));

    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);

    b0 = _mm_srai_epi16(b0, 1);
    b1 = _mm_srai_epi16(b1, 1);
    b2 = _mm_srai_epi16(b2, 1);
    b3 = _mm_srai_epi16(b3, 1);

    coeff0 = _mm_add_epi16(b0, b2);
    coeff1 = _mm_add_epi16(b1, b3);
    coeff2 = _mm_sub_epi16(b0, b2);
    coeff3 = _mm_sub_epi16(b1, b3);

    if (is_final) {
      store_tran_low(coeff0, coeff);
      store_tran_low(coeff1, coeff + 64);
      store_tran_low(coeff2, coeff + 128);
      store_tran_low(coeff3, coeff + 192);
      coeff += 8;
    } else {
      _mm_store_si128((__m128i *)coeff16, coeff0);
      _mm_store_si128((__m128i *)(coeff16 + 64), coeff1);
      _mm_store_si128((__m128i *)(coeff16 + 128), coeff2);
      _mm_store_si128((__m128i *)(coeff16 + 192), coeff3);
      coeff16 += 8;
    }

    t_coeff += 8;
  }
}

void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
                             tran_low_t *coeff) {
  hadamard_16x16_sse2(src_diff, src_stride, coeff, 1);
}

void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
                             tran_low_t *coeff) {
  // For high bitdepths, it is unnecessary to store_tran_low
  // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
  // next stage.  Output to an intermediate buffer first, then store_tran_low()
  // in the final stage.
  DECLARE_ALIGNED(32, int16_t, temp_coeff[32 * 32]);
  int16_t *t_coeff = temp_coeff;
  int idx;
  for (idx = 0; idx < 4; ++idx) {
    const int16_t *src_ptr =
        src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
    hadamard_16x16_sse2(src_ptr, src_stride,
                        (tran_low_t *)(t_coeff + idx * 256), 0);
  }

  for (idx = 0; idx < 256; idx += 8) {
    __m128i coeff0 = _mm_load_si128((const __m128i *)t_coeff);
    __m128i coeff1 = _mm_load_si128((const __m128i *)(t_coeff + 256));
    __m128i coeff2 = _mm_load_si128((const __m128i *)(t_coeff + 512));
    __m128i coeff3 = _mm_load_si128((const __m128i *)(t_coeff + 768));

    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);

    b0 = _mm_srai_epi16(b0, 2);
    b1 = _mm_srai_epi16(b1, 2);
    b2 = _mm_srai_epi16(b2, 2);
    b3 = _mm_srai_epi16(b3, 2);

    coeff0 = _mm_add_epi16(b0, b2);
    coeff1 = _mm_add_epi16(b1, b3);
    store_tran_low(coeff0, coeff);
    store_tran_low(coeff1, coeff + 256);

    coeff2 = _mm_sub_epi16(b0, b2);
    coeff3 = _mm_sub_epi16(b1, b3);
    store_tran_low(coeff2, coeff + 512);
    store_tran_low(coeff3, coeff + 768);

    coeff += 8;
    t_coeff += 8;
  }
}

int aom_satd_sse2(const tran_low_t *coeff, int length) {
  int i;
  const __m128i zero = _mm_setzero_si128();
  __m128i accum = zero;

  for (i = 0; i < length; i += 8) {
    const __m128i src_line = load_tran_low(coeff);
    const __m128i inv = _mm_sub_epi16(zero, src_line);
    const __m128i abs = _mm_max_epi16(src_line, inv);  // abs(src_line)
    const __m128i abs_lo = _mm_unpacklo_epi16(abs, zero);
    const __m128i abs_hi = _mm_unpackhi_epi16(abs, zero);
    const __m128i sum = _mm_add_epi32(abs_lo, abs_hi);
    accum = _mm_add_epi32(accum, sum);
    coeff += 8;
  }

  {  // cascading summation of accum
    __m128i hi = _mm_srli_si128(accum, 8);
    accum = _mm_add_epi32(accum, hi);
    hi = _mm_srli_epi64(accum, 32);
    accum = _mm_add_epi32(accum, hi);
  }

  return _mm_cvtsi128_si32(accum);
}