blob: 82a9fc370c4b6c452111a94f1045ffd5e55518be [file] [log] [blame] [edit]
/*
* Copyright (c) 2019, Alliance for Open Media. 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 "config/aom_dsp_rtcd.h"
#include "aom/aom_integer.h"
#include "aom_dsp/arm/sum_neon.h"
#include "av1/common/arm/mem_neon.h"
#include "av1/common/arm/transpose_neon.h"
int aom_satd_lp_neon(const int16_t *coeff, int length) {
const int16x4_t zero = vdup_n_s16(0);
int32x4_t accum = vdupq_n_s32(0);
do {
const int16x8_t src0 = vld1q_s16(coeff);
const int16x8_t src8 = vld1q_s16(coeff + 8);
accum = vabal_s16(accum, vget_low_s16(src0), zero);
accum = vabal_s16(accum, vget_high_s16(src0), zero);
accum = vabal_s16(accum, vget_low_s16(src8), zero);
accum = vabal_s16(accum, vget_high_s16(src8), zero);
length -= 16;
coeff += 16;
} while (length != 0);
{
// satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
const int64x2_t s0 = vpaddlq_s32(accum); // cascading summation of 'accum'.
const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
vreinterpret_s32_s64(vget_high_s64(s0)));
const int satd = vget_lane_s32(s1, 0);
return satd;
}
}
void aom_int_pro_row_neon(int16_t hbuf[16], const uint8_t *ref,
const int ref_stride, const int height) {
int i;
const uint8_t *idx = ref;
uint16x8_t vec0 = vdupq_n_u16(0);
uint16x8_t vec1 = vec0;
uint8x16_t tmp;
for (i = 0; i < height; ++i) {
tmp = vld1q_u8(idx);
idx += ref_stride;
vec0 = vaddw_u8(vec0, vget_low_u8(tmp));
vec1 = vaddw_u8(vec1, vget_high_u8(tmp));
}
if (128 == height) {
vec0 = vshrq_n_u16(vec0, 6);
vec1 = vshrq_n_u16(vec1, 6);
} else if (64 == height) {
vec0 = vshrq_n_u16(vec0, 5);
vec1 = vshrq_n_u16(vec1, 5);
} else if (32 == height) {
vec0 = vshrq_n_u16(vec0, 4);
vec1 = vshrq_n_u16(vec1, 4);
} else if (16 == height) {
vec0 = vshrq_n_u16(vec0, 3);
vec1 = vshrq_n_u16(vec1, 3);
}
vst1q_s16(hbuf, vreinterpretq_s16_u16(vec0));
hbuf += 8;
vst1q_s16(hbuf, vreinterpretq_s16_u16(vec1));
}
int16_t aom_int_pro_col_neon(const uint8_t *ref, const int width) {
const uint8_t *idx;
uint16x8_t sum = vdupq_n_u16(0);
for (idx = ref; idx < (ref + width); idx += 16) {
uint8x16_t vec = vld1q_u8(idx);
sum = vaddq_u16(sum, vpaddlq_u8(vec));
}
#if defined(__aarch64__)
return (int16_t)vaddvq_u16(sum);
#else
const uint32x4_t a = vpaddlq_u16(sum);
const uint64x2_t b = vpaddlq_u32(a);
const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
vreinterpret_u32_u64(vget_high_u64(b)));
return (int16_t)vget_lane_u32(c, 0);
#endif
}
// coeff: 16 bits, dynamic range [-32640, 32640].
// length: value range {16, 64, 256, 1024}.
int aom_satd_neon(const tran_low_t *coeff, int length) {
const int32x4_t zero = vdupq_n_s32(0);
int32x4_t accum = zero;
do {
const int32x4_t src0 = vld1q_s32(&coeff[0]);
const int32x4_t src8 = vld1q_s32(&coeff[4]);
const int32x4_t src16 = vld1q_s32(&coeff[8]);
const int32x4_t src24 = vld1q_s32(&coeff[12]);
accum = vabaq_s32(accum, src0, zero);
accum = vabaq_s32(accum, src8, zero);
accum = vabaq_s32(accum, src16, zero);
accum = vabaq_s32(accum, src24, zero);
length -= 16;
coeff += 16;
} while (length != 0);
// satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
#ifdef __aarch64__
return vaddvq_s32(accum);
#else
return horizontal_add_s32x4(accum);
#endif // __aarch64__
}
int aom_vector_var_neon(const int16_t *ref, const int16_t *src, const int bwl) {
int32x4_t v_mean = vdupq_n_s32(0);
int32x4_t v_sse = v_mean;
int16x8_t v_ref, v_src;
int16x4_t v_low;
int i, width = 4 << bwl;
for (i = 0; i < width; i += 8) {
v_ref = vld1q_s16(&ref[i]);
v_src = vld1q_s16(&src[i]);
const int16x8_t diff = vsubq_s16(v_ref, v_src);
// diff: dynamic range [-510, 510], 10 bits.
v_mean = vpadalq_s16(v_mean, diff);
v_low = vget_low_s16(diff);
v_sse = vmlal_s16(v_sse, v_low, v_low);
#if defined(__aarch64__)
v_sse = vmlal_high_s16(v_sse, diff, diff);
#else
const int16x4_t v_high = vget_high_s16(diff);
v_sse = vmlal_s16(v_sse, v_high, v_high);
#endif
}
#if defined(__aarch64__)
const int mean = vaddvq_s32(v_mean);
const uint32_t sse = (uint32_t)vaddvq_s32(v_sse);
#else
const int mean = horizontal_add_s32x4(v_mean);
const uint32_t sse = (uint32_t)horizontal_add_s32x4(v_sse);
#endif
// (mean * mean): dynamic range 32 bits - can be stored in uint32_t
const uint32_t meansq = (uint32_t)abs(mean) * (uint32_t)abs(mean);
const int var = sse - (int)(meansq >> (bwl + 2));
return var;
}