Quelle  av1_txfm1d_sse4.h   Sprache: C

/*
 * Copyright (c) 2018, 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.
 */

#ifndef AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_
#define AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_

#include <smmintrin.h>
#include "av1/common/av1_txfm.h"
#include "av1/common/x86/av1_txfm_sse4.h"

#ifdef __cplusplus
extern "C" {
#endif

void av1_fdct32_sse4_1(__m128i *input, __m128i *output, int cos_bit,
                       const int stride);
void av1_fdct64_sse4_1(__m128i *input, __m128i *output, int8_t cos_bit,
                       const int instride, const int outstride);

void av1_idtx32_sse4_1(__m128i *input, __m128i *output, int cos_bit,
                       const int col_num);

static inline void transpose_32_4x4(int stride, const __m128i *input,
                                    __m128i *output) {
  __m128i temp0 = _mm_unpacklo_epi32(input[0 * stride], input[2 * stride]);
  __m128i temp1 = _mm_unpackhi_epi32(input[0 * stride], input[2 * stride]);
  __m128i temp2 = _mm_unpacklo_epi32(input[1 * stride], input[3 * stride]);
  __m128i temp3 = _mm_unpackhi_epi32(input[1 * stride], input[3 * stride]);

  output[0 * stride] = _mm_unpacklo_epi32(temp0, temp2);
  output[1 * stride] = _mm_unpackhi_epi32(temp0, temp2);
  output[2 * stride] = _mm_unpacklo_epi32(temp1, temp3);
  output[3 * stride] = _mm_unpackhi_epi32(temp1, temp3);
}

// the entire input block can be represent by a grid of 4x4 blocks
// each 4x4 blocks can be represent by 4 vertical __m128i
// we first transpose each 4x4 block internally
// then transpose the grid
static inline void transpose_32(int txfm_size, const __m128i *input,
                                __m128i *output) {
  const int num_per_128 = 4;
  const int row_size = txfm_size;
  const int col_size = txfm_size / num_per_128;
  int r, c;

  // transpose each 4x4 block internally
  for (r = 0; r < row_size; r += 4) {
    for (c = 0; c < col_size; c++) {
      transpose_32_4x4(col_size, &input[r * col_size + c],
                       &output[c * 4 * col_size + r / 4]);
    }
  }
}

// out0 = in0*w0 + in1*w1
// out1 = -in1*w0 + in0*w1
#define btf_32_sse4_1_type0(w0, w1, in0, in1, out0, out1, bit) \
  do {                                                         \
    const __m128i ww0 = _mm_set1_epi32(w0);                    \
    const __m128i ww1 = _mm_set1_epi32(w1);                    \
    const __m128i in0_w0 = _mm_mullo_epi32(in0, ww0);          \
    const __m128i in1_w1 = _mm_mullo_epi32(in1, ww1);          \
    out0 = _mm_add_epi32(in0_w0, in1_w1);                      \
    out0 = av1_round_shift_32_sse4_1(out0, bit);               \
    const __m128i in0_w1 = _mm_mullo_epi32(in0, ww1);          \
    const __m128i in1_w0 = _mm_mullo_epi32(in1, ww0);          \
    out1 = _mm_sub_epi32(in0_w1, in1_w0);                      \
    out1 = av1_round_shift_32_sse4_1(out1, bit);               \
  } while (0)

// out0 = in0*w0 + in1*w1
// out1 = in1*w0 - in0*w1
#define btf_32_sse4_1_type1(w0, w1, in0, in1, out0, out1, bit) \
  do {                                                         \
    btf_32_sse4_1_type0(w1, w0, in1, in0, out0, out1, bit);    \
  } while (0)

// out0 = in0*w0 + in1*w1
// out1 = -in1*w0 + in0*w1
#define btf_32_type0_sse4_1_new(ww0, ww1, in0, in1, out0, out1, r, bit) \
  do {                                                                  \
    const __m128i in0_w0 = _mm_mullo_epi32(in0, ww0);                   \
    const __m128i in1_w1 = _mm_mullo_epi32(in1, ww1);                   \
    out0 = _mm_add_epi32(in0_w0, in1_w1);                               \
    out0 = _mm_add_epi32(out0, r);                                      \
    out0 = _mm_srai_epi32(out0, bit);                                   \
    const __m128i in0_w1 = _mm_mullo_epi32(in0, ww1);                   \
    const __m128i in1_w0 = _mm_mullo_epi32(in1, ww0);                   \
    out1 = _mm_sub_epi32(in0_w1, in1_w0);                               \
    out1 = _mm_add_epi32(out1, r);                                      \
    out1 = _mm_srai_epi32(out1, bit);                                   \
  } while (0)

// out0 = in0*w0 + in1*w1
// out1 = in1*w0 - in0*w1
#define btf_32_type1_sse4_1_new(ww0, ww1, in0, in1, out0, out1, r, bit) \
  do {                                                                  \
    btf_32_type0_sse4_1_new(ww1, ww0, in1, in0, out0, out1, r, bit);    \
  } while (0)

#ifdef __cplusplus
}
#endif

#endif  // AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_