Source code
Revision control
Copy as Markdown
Other Tools
/*
* Copyright (c) 2023, 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 "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "mem_neon.h"
#include "sum_neon.h"
static inline void obmc_variance_8x1_s16_neon(int16x8_t pre_s16,
const int32_t *wsrc,
const int32_t *mask,
int32x4_t *ssev,
int32x4_t *sumv) {
// For 4xh and 8xh we observe it is faster to avoid the double-widening of
// pre. Instead we do a single widening step and narrow the mask to 16-bits
// to allow us to perform a widening multiply. Widening multiply
// instructions have better throughput on some micro-architectures but for
// the larger block sizes this benefit is outweighed by the additional
// instruction needed to first narrow the mask vectors.
int32x4_t wsrc_s32_lo = vld1q_s32(&wsrc[0]);
int32x4_t wsrc_s32_hi = vld1q_s32(&wsrc[4]);
int16x8_t mask_s16 = vuzpq_s16(vreinterpretq_s16_s32(vld1q_s32(&mask[0])),
vreinterpretq_s16_s32(vld1q_s32(&mask[4])))
.val[0];
int32x4_t diff_s32_lo =
vmlsl_s16(wsrc_s32_lo, vget_low_s16(pre_s16), vget_low_s16(mask_s16));
int32x4_t diff_s32_hi =
vmlsl_s16(wsrc_s32_hi, vget_high_s16(pre_s16), vget_high_s16(mask_s16));
// ROUND_POWER_OF_TWO_SIGNED(value, 12) rounds to nearest with ties away
// from zero, however vrshrq_n_s32 rounds to nearest with ties rounded up.
// This difference only affects the bit patterns at the rounding breakpoints
// exactly, so we can add -1 to all negative numbers to move the breakpoint
// one value across and into the correct rounding region.
diff_s32_lo = vsraq_n_s32(diff_s32_lo, diff_s32_lo, 31);
diff_s32_hi = vsraq_n_s32(diff_s32_hi, diff_s32_hi, 31);
int32x4_t round_s32_lo = vrshrq_n_s32(diff_s32_lo, 12);
int32x4_t round_s32_hi = vrshrq_n_s32(diff_s32_hi, 12);
*sumv = vrsraq_n_s32(*sumv, diff_s32_lo, 12);
*sumv = vrsraq_n_s32(*sumv, diff_s32_hi, 12);
*ssev = vmlaq_s32(*ssev, round_s32_lo, round_s32_lo);
*ssev = vmlaq_s32(*ssev, round_s32_hi, round_s32_hi);
}
#if AOM_ARCH_AARCH64
// Use tbl for doing a double-width zero extension from 8->32 bits since we can
// do this in one instruction rather than two (indices out of range (255 here)
// are set to zero by tbl).
DECLARE_ALIGNED(16, static const uint8_t, obmc_variance_permute_idx[]) = {
0, 255, 255, 255, 1, 255, 255, 255, 2, 255, 255, 255, 3, 255, 255, 255,
4, 255, 255, 255, 5, 255, 255, 255, 6, 255, 255, 255, 7, 255, 255, 255,
8, 255, 255, 255, 9, 255, 255, 255, 10, 255, 255, 255, 11, 255, 255, 255,
12, 255, 255, 255, 13, 255, 255, 255, 14, 255, 255, 255, 15, 255, 255, 255
};
static inline void obmc_variance_8x1_s32_neon(
int32x4_t pre_lo, int32x4_t pre_hi, const int32_t *wsrc,
const int32_t *mask, int32x4_t *ssev, int32x4_t *sumv) {
int32x4_t wsrc_lo = vld1q_s32(&wsrc[0]);
int32x4_t wsrc_hi = vld1q_s32(&wsrc[4]);
int32x4_t mask_lo = vld1q_s32(&mask[0]);
int32x4_t mask_hi = vld1q_s32(&mask[4]);
int32x4_t diff_lo = vmlsq_s32(wsrc_lo, pre_lo, mask_lo);
int32x4_t diff_hi = vmlsq_s32(wsrc_hi, pre_hi, mask_hi);
// ROUND_POWER_OF_TWO_SIGNED(value, 12) rounds to nearest with ties away from
// zero, however vrshrq_n_s32 rounds to nearest with ties rounded up. This
// difference only affects the bit patterns at the rounding breakpoints
// exactly, so we can add -1 to all negative numbers to move the breakpoint
// one value across and into the correct rounding region.
diff_lo = vsraq_n_s32(diff_lo, diff_lo, 31);
diff_hi = vsraq_n_s32(diff_hi, diff_hi, 31);
int32x4_t round_lo = vrshrq_n_s32(diff_lo, 12);
int32x4_t round_hi = vrshrq_n_s32(diff_hi, 12);
*sumv = vrsraq_n_s32(*sumv, diff_lo, 12);
*sumv = vrsraq_n_s32(*sumv, diff_hi, 12);
*ssev = vmlaq_s32(*ssev, round_lo, round_lo);
*ssev = vmlaq_s32(*ssev, round_hi, round_hi);
}
static inline void obmc_variance_large_neon(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int width,
int height, unsigned *sse,
int *sum) {
assert(width % 16 == 0);
// Use tbl for doing a double-width zero extension from 8->32 bits since we
// can do this in one instruction rather than two.
uint8x16_t pre_idx0 = vld1q_u8(&obmc_variance_permute_idx[0]);
uint8x16_t pre_idx1 = vld1q_u8(&obmc_variance_permute_idx[16]);
uint8x16_t pre_idx2 = vld1q_u8(&obmc_variance_permute_idx[32]);
uint8x16_t pre_idx3 = vld1q_u8(&obmc_variance_permute_idx[48]);
int32x4_t ssev = vdupq_n_s32(0);
int32x4_t sumv = vdupq_n_s32(0);
int h = height;
do {
int w = width;
do {
uint8x16_t pre_u8 = vld1q_u8(pre);
int32x4_t pre_s32_lo = vreinterpretq_s32_u8(vqtbl1q_u8(pre_u8, pre_idx0));
int32x4_t pre_s32_hi = vreinterpretq_s32_u8(vqtbl1q_u8(pre_u8, pre_idx1));
obmc_variance_8x1_s32_neon(pre_s32_lo, pre_s32_hi, &wsrc[0], &mask[0],
&ssev, &sumv);
pre_s32_lo = vreinterpretq_s32_u8(vqtbl1q_u8(pre_u8, pre_idx2));
pre_s32_hi = vreinterpretq_s32_u8(vqtbl1q_u8(pre_u8, pre_idx3));
obmc_variance_8x1_s32_neon(pre_s32_lo, pre_s32_hi, &wsrc[8], &mask[8],
&ssev, &sumv);
wsrc += 16;
mask += 16;
pre += 16;
w -= 16;
} while (w != 0);
pre += pre_stride - width;
} while (--h != 0);
*sse = horizontal_add_s32x4(ssev);
*sum = horizontal_add_s32x4(sumv);
}
#else // !AOM_ARCH_AARCH64
static inline void obmc_variance_large_neon(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int width,
int height, unsigned *sse,
int *sum) {
// Non-aarch64 targets do not have a 128-bit tbl instruction, so use the
// widening version of the core kernel instead.
assert(width % 16 == 0);
int32x4_t ssev = vdupq_n_s32(0);
int32x4_t sumv = vdupq_n_s32(0);
int h = height;
do {
int w = width;
do {
uint8x16_t pre_u8 = vld1q_u8(pre);
int16x8_t pre_s16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pre_u8)));
obmc_variance_8x1_s16_neon(pre_s16, &wsrc[0], &mask[0], &ssev, &sumv);
pre_s16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pre_u8)));
obmc_variance_8x1_s16_neon(pre_s16, &wsrc[8], &mask[8], &ssev, &sumv);
wsrc += 16;
mask += 16;
pre += 16;
w -= 16;
} while (w != 0);
pre += pre_stride - width;
} while (--h != 0);
*sse = horizontal_add_s32x4(ssev);
*sum = horizontal_add_s32x4(sumv);
}
#endif // AOM_ARCH_AARCH64
static inline void obmc_variance_neon_128xh(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int h,
unsigned *sse, int *sum) {
obmc_variance_large_neon(pre, pre_stride, wsrc, mask, 128, h, sse, sum);
}
static inline void obmc_variance_neon_64xh(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int h,
unsigned *sse, int *sum) {
obmc_variance_large_neon(pre, pre_stride, wsrc, mask, 64, h, sse, sum);
}
static inline void obmc_variance_neon_32xh(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int h,
unsigned *sse, int *sum) {
obmc_variance_large_neon(pre, pre_stride, wsrc, mask, 32, h, sse, sum);
}
static inline void obmc_variance_neon_16xh(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int h,
unsigned *sse, int *sum) {
obmc_variance_large_neon(pre, pre_stride, wsrc, mask, 16, h, sse, sum);
}
static inline void obmc_variance_neon_8xh(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int h,
unsigned *sse, int *sum) {
int32x4_t ssev = vdupq_n_s32(0);
int32x4_t sumv = vdupq_n_s32(0);
do {
uint8x8_t pre_u8 = vld1_u8(pre);
int16x8_t pre_s16 = vreinterpretq_s16_u16(vmovl_u8(pre_u8));
obmc_variance_8x1_s16_neon(pre_s16, wsrc, mask, &ssev, &sumv);
pre += pre_stride;
wsrc += 8;
mask += 8;
} while (--h != 0);
*sse = horizontal_add_s32x4(ssev);
*sum = horizontal_add_s32x4(sumv);
}
static inline void obmc_variance_neon_4xh(const uint8_t *pre, int pre_stride,
const int32_t *wsrc,
const int32_t *mask, int h,
unsigned *sse, int *sum) {
assert(h % 2 == 0);
int32x4_t ssev = vdupq_n_s32(0);
int32x4_t sumv = vdupq_n_s32(0);
do {
uint8x8_t pre_u8 = load_unaligned_u8(pre, pre_stride);
int16x8_t pre_s16 = vreinterpretq_s16_u16(vmovl_u8(pre_u8));
obmc_variance_8x1_s16_neon(pre_s16, wsrc, mask, &ssev, &sumv);
pre += 2 * pre_stride;
wsrc += 8;
mask += 8;
h -= 2;
} while (h != 0);
*sse = horizontal_add_s32x4(ssev);
*sum = horizontal_add_s32x4(sumv);
}
#define OBMC_VARIANCE_WXH_NEON(W, H) \
unsigned aom_obmc_variance##W##x##H##_neon( \
const uint8_t *pre, int pre_stride, const int32_t *wsrc, \
const int32_t *mask, unsigned *sse) { \
int sum; \
obmc_variance_neon_##W##xh(pre, pre_stride, wsrc, mask, H, sse, &sum); \
return *sse - (unsigned)(((int64_t)sum * sum) / (W * H)); \
}
OBMC_VARIANCE_WXH_NEON(4, 4)
OBMC_VARIANCE_WXH_NEON(4, 8)
OBMC_VARIANCE_WXH_NEON(8, 4)
OBMC_VARIANCE_WXH_NEON(8, 8)
OBMC_VARIANCE_WXH_NEON(8, 16)
OBMC_VARIANCE_WXH_NEON(16, 8)
OBMC_VARIANCE_WXH_NEON(16, 16)
OBMC_VARIANCE_WXH_NEON(16, 32)
OBMC_VARIANCE_WXH_NEON(32, 16)
OBMC_VARIANCE_WXH_NEON(32, 32)
OBMC_VARIANCE_WXH_NEON(32, 64)
OBMC_VARIANCE_WXH_NEON(64, 32)
OBMC_VARIANCE_WXH_NEON(64, 64)
OBMC_VARIANCE_WXH_NEON(64, 128)
OBMC_VARIANCE_WXH_NEON(128, 64)
OBMC_VARIANCE_WXH_NEON(128, 128)
OBMC_VARIANCE_WXH_NEON(4, 16)
OBMC_VARIANCE_WXH_NEON(16, 4)
OBMC_VARIANCE_WXH_NEON(8, 32)
OBMC_VARIANCE_WXH_NEON(32, 8)
OBMC_VARIANCE_WXH_NEON(16, 64)
OBMC_VARIANCE_WXH_NEON(64, 16)