1 /*
2 * Copyright (c) 2023, Alliance for Open Media. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <arm_neon.h>
12 #include <assert.h>
13
14 #include "config/aom_config.h"
15 #include "config/aom_dsp_rtcd.h"
16 #include "aom/aom_integer.h"
17 #include "aom_dsp/arm/aom_neon_sve_bridge.h"
18 #include "aom_dsp/arm/mem_neon.h"
19 #include "aom_ports/mem.h"
20
aom_vector_var_sve(const int16_t * ref,const int16_t * src,int bwl)21 int aom_vector_var_sve(const int16_t *ref, const int16_t *src, int bwl) {
22 assert(bwl >= 2 && bwl <= 5);
23 int width = 4 << bwl;
24
25 int64x2_t sse_s64[2] = { vdupq_n_s64(0), vdupq_n_s64(0) };
26 int16x8_t v_mean[2] = { vdupq_n_s16(0), vdupq_n_s16(0) };
27
28 do {
29 int16x8_t r0 = vld1q_s16(ref);
30 int16x8_t s0 = vld1q_s16(src);
31
32 // diff: dynamic range [-510, 510] 10 (signed) bits.
33 int16x8_t diff0 = vsubq_s16(r0, s0);
34 // v_mean: dynamic range 16 * diff -> [-8160, 8160], 14 (signed) bits.
35 v_mean[0] = vaddq_s16(v_mean[0], diff0);
36
37 // v_sse: dynamic range 2 * 16 * diff^2 -> [0, 8,323,200], 24 (signed) bits.
38 sse_s64[0] = aom_sdotq_s16(sse_s64[0], diff0, diff0);
39
40 int16x8_t r1 = vld1q_s16(ref + 8);
41 int16x8_t s1 = vld1q_s16(src + 8);
42
43 // diff: dynamic range [-510, 510] 10 (signed) bits.
44 int16x8_t diff1 = vsubq_s16(r1, s1);
45 // v_mean: dynamic range 16 * diff -> [-8160, 8160], 14 (signed) bits.
46 v_mean[1] = vaddq_s16(v_mean[1], diff1);
47
48 // v_sse: dynamic range 2 * 16 * diff^2 -> [0, 8,323,200], 24 (signed) bits.
49 sse_s64[1] = aom_sdotq_s16(sse_s64[1], diff1, diff1);
50
51 ref += 16;
52 src += 16;
53 width -= 16;
54 } while (width != 0);
55
56 // Dynamic range [0, 65280], 16 (unsigned) bits.
57 const uint32_t mean_abs = abs(vaddlvq_s16(vaddq_s16(v_mean[0], v_mean[1])));
58 const int64_t sse = vaddvq_s64(vaddq_s64(sse_s64[0], sse_s64[1]));
59
60 // (mean_abs * mean_abs): dynamic range 32 (unsigned) bits.
61 return (int)(sse - ((mean_abs * mean_abs) >> (bwl + 2)));
62 }
63