Quelle  highbd_variance_neon.c

/*
 *  Copyright (c) 2022 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <arm_neon.h>

#include "./vpx_dsp_rtcd.h"
#include "./vpx_config.h"

#include "vpx/vpx_integer.h"
#include "vpx_dsp/arm/mem_neon.h"
#include "vpx_dsp/arm/sum_neon.h"
#include "vpx_ports/mem.h"

// Process a block of width 4 two rows at a time.
static INLINE void highbd_variance_4xh_neon(const uint16_t *src_ptr,
                                            int src_stride,
                                            const uint16_t *ref_ptr,
                                            int ref_stride, int h,
                                            uint64_t *sse, int64_t *sum) {
  int16x8_t sum_s16 = vdupq_n_s16(0);
  int32x4_t sse_s32 = vdupq_n_s32(0);

  int i = h;
  do {
    const uint16x8_t s = load_unaligned_u16q(src_ptr, src_stride);
    const uint16x8_t r = load_unaligned_u16q(ref_ptr, ref_stride);

    int16x8_t diff = vreinterpretq_s16_u16(vsubq_u16(s, r));
    sum_s16 = vaddq_s16(sum_s16, diff);

    sse_s32 = vmlal_s16(sse_s32, vget_low_s16(diff), vget_low_s16(diff));
    sse_s32 = vmlal_s16(sse_s32, vget_high_s16(diff), vget_high_s16(diff));

    src_ptr += 2 * src_stride;
    ref_ptr += 2 * ref_stride;
    i -= 2;
  } while (i != 0);

  *sum = horizontal_add_int16x8(sum_s16);
  *sse = horizontal_add_int32x4(sse_s32);
}

// For 8-bit and 10-bit data, since we're using two int32x4 accumulators, all
// block sizes can be processed in 32-bit elements (1023*1023*64*16 = 1071645696
// for a 64x64 block).
static INLINE void highbd_variance_large_neon(const uint16_t *src_ptr,
                                              int src_stride,
                                              const uint16_t *ref_ptr,
                                              int ref_stride, int w, int h,
                                              uint64_t *sse, int64_t *sum) {
  int32x4_t sum_s32 = vdupq_n_s32(0);
  int32x4_t sse_s32[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };

  int i = h;
  do {
    int j = 0;
    do {
      const uint16x8_t s = vld1q_u16(src_ptr + j);
      const uint16x8_t r = vld1q_u16(ref_ptr + j);

      const int16x8_t diff = vreinterpretq_s16_u16(vsubq_u16(s, r));
      sum_s32 = vpadalq_s16(sum_s32, diff);

      sse_s32[0] =
          vmlal_s16(sse_s32[0], vget_low_s16(diff), vget_low_s16(diff));
      sse_s32[1] =
          vmlal_s16(sse_s32[1], vget_high_s16(diff), vget_high_s16(diff));

      j += 8;
    } while (j < w);

    src_ptr += src_stride;
    ref_ptr += ref_stride;
  } while (--i != 0);

  *sum = horizontal_add_int32x4(sum_s32);
  *sse = horizontal_long_add_uint32x4(vaddq_u32(
      vreinterpretq_u32_s32(sse_s32[0]), vreinterpretq_u32_s32(sse_s32[1])));
}

static INLINE void highbd_variance_8xh_neon(const uint16_t *src, int src_stride,
                                            const uint16_t *ref, int ref_stride,
                                            int h, uint64_t *sse,
                                            int64_t *sum) {
  highbd_variance_large_neon(src, src_stride, ref, ref_stride, 8, h, sse, sum);
}

static INLINE void highbd_variance_16xh_neon(const uint16_t *src,
                                             int src_stride,
                                             const uint16_t *ref,
                                             int ref_stride, int h,
                                             uint64_t *sse, int64_t *sum) {
  highbd_variance_large_neon(src, src_stride, ref, ref_stride, 16, h, sse, sum);
}

static INLINE void highbd_variance_32xh_neon(const uint16_t *src,
                                             int src_stride,
                                             const uint16_t *ref,
                                             int ref_stride, int h,
                                             uint64_t *sse, int64_t *sum) {
  highbd_variance_large_neon(src, src_stride, ref, ref_stride, 32, h, sse, sum);
}

static INLINE void highbd_variance_64xh_neon(const uint16_t *src,
                                             int src_stride,
                                             const uint16_t *ref,
                                             int ref_stride, int h,
                                             uint64_t *sse, int64_t *sum) {
  highbd_variance_large_neon(src, src_stride, ref, ref_stride, 64, h, sse, sum);
}

// For 12-bit data, we can only accumulate up to 128 elements in the sum of
// squares (4095*4095*128 = 2146435200), and because we're using two int32x4
// accumulators, we can only process up to 32 32-element rows (32*32/8 = 128)
// or 16 64-element rows before we have to accumulate into 64-bit elements.
// Therefore blocks of size 32x64, 64x32 and 64x64 are processed in a different
// helper function.

// Process a block of any size where the width is divisible by 8, with
// accumulation into 64-bit elements.
static INLINE void highbd_variance_xlarge_neon(
    const uint16_t *src_ptr, int src_stride, const uint16_t *ref_ptr,
    int ref_stride, int w, int h, int h_limit, uint64_t *sse, int64_t *sum) {
  int32x4_t sum_s32 = vdupq_n_s32(0);
  int64x2_t sse_s64 = vdupq_n_s64(0);

  // 'h_limit' is the number of 'w'-width rows we can process before our 32-bit
  // accumulator overflows. After hitting this limit we accumulate into 64-bit
  // elements.
  int h_tmp = h > h_limit ? h_limit : h;

  int i = 0;
  do {
    int32x4_t sse_s32[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
    do {
      int j = 0;
      do {
        const uint16x8_t s0 = vld1q_u16(src_ptr + j);
        const uint16x8_t r0 = vld1q_u16(ref_ptr + j);

        const int16x8_t diff = vreinterpretq_s16_u16(vsubq_u16(s0, r0));
        sum_s32 = vpadalq_s16(sum_s32, diff);

        sse_s32[0] =
            vmlal_s16(sse_s32[0], vget_low_s16(diff), vget_low_s16(diff));
        sse_s32[1] =
            vmlal_s16(sse_s32[1], vget_high_s16(diff), vget_high_s16(diff));

        j += 8;
      } while (j < w);

      src_ptr += src_stride;
      ref_ptr += ref_stride;
      i++;
    } while (i < h_tmp);

    sse_s64 = vpadalq_s32(sse_s64, sse_s32[0]);
    sse_s64 = vpadalq_s32(sse_s64, sse_s32[1]);
    h_tmp += h_limit;
  } while (i < h);

  *sum = horizontal_add_int32x4(sum_s32);
  *sse = (uint64_t)horizontal_add_int64x2(sse_s64);
}

static INLINE void highbd_variance_32xh_xlarge_neon(
    const uint16_t *src, int src_stride, const uint16_t *ref, int ref_stride,
    int h, uint64_t *sse, int64_t *sum) {
  highbd_variance_xlarge_neon(src, src_stride, ref, ref_stride, 32, h, 32, sse,
                              sum);
}

static INLINE void highbd_variance_64xh_xlarge_neon(
    const uint16_t *src, int src_stride, const uint16_t *ref, int ref_stride,
    int h, uint64_t *sse, int64_t *sum) {
  highbd_variance_xlarge_neon(src, src_stride, ref, ref_stride, 64, h, 16, sse,
                              sum);
}

#define HBD_VARIANCE_WXH_8_NEON(w, h)                                 \
  uint32_t vpx_highbd_8_variance##w##x##h##_neon(                     \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \
      int ref_stride, uint32_t *sse) {                                \
    int sum;                                                          \
    uint64_t sse_long = 0;                                            \
    int64_t sum_long = 0;                                             \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                     \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                     \
    highbd_variance_##w##xh_neon(src, src_stride, ref, ref_stride, h, \
                                 &sse_long, &sum_long);               \
    *sse = (uint32_t)sse_long;                                        \
    sum = (int)sum_long;                                              \
    return *sse - (uint32_t)(((int64_t)sum * sum) / (w * h));         \
  }

#define HBD_VARIANCE_WXH_10_NEON(w, h)                                \
  uint32_t vpx_highbd_10_variance##w##x##h##_neon(                    \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \
      int ref_stride, uint32_t *sse) {                                \
    int sum;                                                          \
    int64_t var;                                                      \
    uint64_t sse_long = 0;                                            \
    int64_t sum_long = 0;                                             \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                     \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                     \
    highbd_variance_##w##xh_neon(src, src_stride, ref, ref_stride, h, \
                                 &sse_long, &sum_long);               \
    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);                 \
    sum = (int)ROUND_POWER_OF_TWO(sum_long, 2);                       \
    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (w * h));         \
    return (var >= 0) ? (uint32_t)var : 0;                            \
  }

#define HBD_VARIANCE_WXH_12_NEON(w, h)                                \
  uint32_t vpx_highbd_12_variance##w##x##h##_neon(                    \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \
      int ref_stride, uint32_t *sse) {                                \
    int sum;                                                          \
    int64_t var;                                                      \
    uint64_t sse_long = 0;                                            \
    int64_t sum_long = 0;                                             \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                     \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                     \
    highbd_variance_##w##xh_neon(src, src_stride, ref, ref_stride, h, \
                                 &sse_long, &sum_long);               \
    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);                 \
    sum = (int)ROUND_POWER_OF_TWO(sum_long, 4);                       \
    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (w * h));         \
    return (var >= 0) ? (uint32_t)var : 0;                            \
  }

#define HBD_VARIANCE_WXH_12_XLARGE_NEON(w, h)                                \
  uint32_t vpx_highbd_12_variance##w##x##h##_neon(                           \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,        \
      int ref_stride, uint32_t *sse) {                                       \
    int sum;                                                                 \
    int64_t var;                                                             \
    uint64_t sse_long = 0;                                                   \
    int64_t sum_long = 0;                                                    \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                            \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                            \
    highbd_variance_##w##xh_xlarge_neon(src, src_stride, ref, ref_stride, h, \
                                        &sse_long, &sum_long);               \
    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);                        \
    sum = (int)ROUND_POWER_OF_TWO(sum_long, 4);                              \
    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (w * h));                \
    return (var >= 0) ? (uint32_t)var : 0;                                   \
  }

// 8-bit
HBD_VARIANCE_WXH_8_NEON(4, 4)
HBD_VARIANCE_WXH_8_NEON(4, 8)

HBD_VARIANCE_WXH_8_NEON(8, 4)
HBD_VARIANCE_WXH_8_NEON(8, 8)
HBD_VARIANCE_WXH_8_NEON(8, 16)

HBD_VARIANCE_WXH_8_NEON(16, 8)
HBD_VARIANCE_WXH_8_NEON(16, 16)
HBD_VARIANCE_WXH_8_NEON(16, 32)

HBD_VARIANCE_WXH_8_NEON(32, 16)
HBD_VARIANCE_WXH_8_NEON(32, 32)
HBD_VARIANCE_WXH_8_NEON(32, 64)

HBD_VARIANCE_WXH_8_NEON(64, 32)
HBD_VARIANCE_WXH_8_NEON(64, 64)

// 10-bit
HBD_VARIANCE_WXH_10_NEON(4, 4)
HBD_VARIANCE_WXH_10_NEON(4, 8)

HBD_VARIANCE_WXH_10_NEON(8, 4)
HBD_VARIANCE_WXH_10_NEON(8, 8)
HBD_VARIANCE_WXH_10_NEON(8, 16)

HBD_VARIANCE_WXH_10_NEON(16, 8)
HBD_VARIANCE_WXH_10_NEON(16, 16)
HBD_VARIANCE_WXH_10_NEON(16, 32)

HBD_VARIANCE_WXH_10_NEON(32, 16)
HBD_VARIANCE_WXH_10_NEON(32, 32)
HBD_VARIANCE_WXH_10_NEON(32, 64)

HBD_VARIANCE_WXH_10_NEON(64, 32)
HBD_VARIANCE_WXH_10_NEON(64, 64)

// 12-bit
HBD_VARIANCE_WXH_12_NEON(4, 4)
HBD_VARIANCE_WXH_12_NEON(4, 8)

HBD_VARIANCE_WXH_12_NEON(8, 4)
HBD_VARIANCE_WXH_12_NEON(8, 8)
HBD_VARIANCE_WXH_12_NEON(8, 16)

HBD_VARIANCE_WXH_12_NEON(16, 8)
HBD_VARIANCE_WXH_12_NEON(16, 16)
HBD_VARIANCE_WXH_12_NEON(16, 32)

HBD_VARIANCE_WXH_12_NEON(32, 16)
HBD_VARIANCE_WXH_12_NEON(32, 32)
HBD_VARIANCE_WXH_12_XLARGE_NEON(32, 64)

HBD_VARIANCE_WXH_12_XLARGE_NEON(64, 32)
HBD_VARIANCE_WXH_12_XLARGE_NEON(64, 64)

#define HIGHBD_GET_VAR(S)                                             \
  void vpx_highbd_8_get##S##x##S##var_neon(                           \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \
      int ref_stride, uint32_t *sse, int *sum) {                      \
    uint64_t sse_long = 0;                                            \
    int64_t sum_long = 0;                                             \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                     \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                     \
    highbd_variance_##S##xh_neon(src, src_stride, ref, ref_stride, S, \
                                 &sse_long, &sum_long);               \
    *sse = (uint32_t)sse_long;                                        \
    *sum = (int)sum_long;                                             \
  }                                                                   \
                                                                      \
  void vpx_highbd_10_get##S##x##S##var_neon(                          \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \
      int ref_stride, uint32_t *sse, int *sum) {                      \
    uint64_t sse_long = 0;                                            \
    int64_t sum_long = 0;                                             \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                     \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                     \
    highbd_variance_##S##xh_neon(src, src_stride, ref, ref_stride, S, \
                                 &sse_long, &sum_long);               \
    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);                 \
    *sum = (int)ROUND_POWER_OF_TWO(sum_long, 2);                      \
  }                                                                   \
                                                                      \
  void vpx_highbd_12_get##S##x##S##var_neon(                          \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \
      int ref_stride, uint32_t *sse, int *sum) {                      \
    uint64_t sse_long = 0;                                            \
    int64_t sum_long = 0;                                             \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                     \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                     \
    highbd_variance_##S##xh_neon(src, src_stride, ref, ref_stride, S, \
                                 &sse_long, &sum_long);               \
    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);                 \
    *sum = (int)ROUND_POWER_OF_TWO(sum_long, 4);                      \
  }

HIGHBD_GET_VAR(8)
HIGHBD_GET_VAR(16)

static INLINE uint32_t highbd_mse_wxh_neon(const uint16_t *src_ptr,
                                           int src_stride,
                                           const uint16_t *ref_ptr,
                                           int ref_stride, int w, int h) {
  uint32x4_t sse_u32[2] = { vdupq_n_u32(0), vdupq_n_u32(0) };

  int i = h;
  do {
    int j = 0;
    do {
      uint16x8_t s = vld1q_u16(src_ptr + j);
      uint16x8_t r = vld1q_u16(ref_ptr + j);

      uint16x8_t diff = vabdq_u16(s, r);

      sse_u32[0] =
          vmlal_u16(sse_u32[0], vget_low_u16(diff), vget_low_u16(diff));
      sse_u32[1] =
          vmlal_u16(sse_u32[1], vget_high_u16(diff), vget_high_u16(diff));

      j += 8;
    } while (j < w);

    src_ptr += src_stride;
    ref_ptr += ref_stride;
  } while (--i != 0);

  return horizontal_add_uint32x4(vaddq_u32(sse_u32[0], sse_u32[1]));
}

static INLINE uint32_t highbd_mse8_8xh_neon(const uint16_t *src_ptr,
                                            int src_stride,
                                            const uint16_t *ref_ptr,
                                            int ref_stride, int h) {
  return highbd_mse_wxh_neon(src_ptr, src_stride, ref_ptr, ref_stride, 8, h);
}

static INLINE uint32_t highbd_mse8_16xh_neon(const uint16_t *src_ptr,
                                             int src_stride,
                                             const uint16_t *ref_ptr,
                                             int ref_stride, int h) {
  return highbd_mse_wxh_neon(src_ptr, src_stride, ref_ptr, ref_stride, 16, h);
}

#define HIGHBD_MSE_WXH_NEON(w, h)                                         \
  uint32_t vpx_highbd_8_mse##w##x##h##_neon(                              \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,     \
      int ref_stride, uint32_t *sse) {                                    \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                         \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                         \
    *sse = highbd_mse8_##w##xh_neon(src, src_stride, ref, ref_stride, h); \
    return *sse;                                                          \
  }                                                                       \
                                                                          \
  uint32_t vpx_highbd_10_mse##w##x##h##_neon(                             \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,     \
      int ref_stride, uint32_t *sse) {                                    \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                         \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                         \
    *sse = highbd_mse_wxh_neon(src, src_stride, ref, ref_stride, w, h);   \
    *sse = ROUND_POWER_OF_TWO(*sse, 4);                                   \
    return *sse;                                                          \
  }                                                                       \
                                                                          \
  uint32_t vpx_highbd_12_mse##w##x##h##_neon(                             \
      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,     \
      int ref_stride, uint32_t *sse) {                                    \
    uint16_t *src = CONVERT_TO_SHORTPTR(src_ptr);                         \
    uint16_t *ref = CONVERT_TO_SHORTPTR(ref_ptr);                         \
    *sse = highbd_mse_wxh_neon(src, src_stride, ref, ref_stride, w, h);   \
    *sse = ROUND_POWER_OF_TWO(*sse, 8);                                   \
    return *sse;                                                          \
  }

HIGHBD_MSE_WXH_NEON(16, 16)
HIGHBD_MSE_WXH_NEON(16, 8)
HIGHBD_MSE_WXH_NEON(8, 16)
HIGHBD_MSE_WXH_NEON(8, 8)

#undef HIGHBD_MSE_WXH_NEON