1*b2055c35SXin Li // Copyright 2018 Google Inc. All Rights Reserved.
2*b2055c35SXin Li //
3*b2055c35SXin Li // Use of this source code is governed by a BSD-style license
4*b2055c35SXin Li // that can be found in the COPYING file in the root of the source
5*b2055c35SXin Li // tree. An additional intellectual property rights grant can be found
6*b2055c35SXin Li // in the file PATENTS. All contributing project authors may
7*b2055c35SXin Li // be found in the AUTHORS file in the root of the source tree.
8*b2055c35SXin Li // -----------------------------------------------------------------------------
9*b2055c35SXin Li
10*b2055c35SXin Li #ifndef WEBP_DSP_QUANT_H_
11*b2055c35SXin Li #define WEBP_DSP_QUANT_H_
12*b2055c35SXin Li
13*b2055c35SXin Li #include <string.h>
14*b2055c35SXin Li
15*b2055c35SXin Li #include "src/dsp/dsp.h"
16*b2055c35SXin Li #include "src/webp/types.h"
17*b2055c35SXin Li
18*b2055c35SXin Li #if defined(WEBP_USE_NEON) && !defined(WEBP_ANDROID_NEON) && \
19*b2055c35SXin Li !defined(WEBP_HAVE_NEON_RTCD)
20*b2055c35SXin Li #include <arm_neon.h>
21*b2055c35SXin Li
22*b2055c35SXin Li #define IsFlat IsFlat_NEON
23*b2055c35SXin Li
horizontal_add_uint32x4(const uint32x4_t a)24*b2055c35SXin Li static uint32_t horizontal_add_uint32x4(const uint32x4_t a) {
25*b2055c35SXin Li #if WEBP_AARCH64
26*b2055c35SXin Li return vaddvq_u32(a);
27*b2055c35SXin Li #else
28*b2055c35SXin Li const uint64x2_t b = vpaddlq_u32(a);
29*b2055c35SXin Li const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
30*b2055c35SXin Li vreinterpret_u32_u64(vget_high_u64(b)));
31*b2055c35SXin Li return vget_lane_u32(c, 0);
32*b2055c35SXin Li #endif
33*b2055c35SXin Li }
34*b2055c35SXin Li
IsFlat(const int16_t * levels,int num_blocks,int thresh)35*b2055c35SXin Li static WEBP_INLINE int IsFlat(const int16_t* levels, int num_blocks,
36*b2055c35SXin Li int thresh) {
37*b2055c35SXin Li const int16x8_t tst_ones = vdupq_n_s16(-1);
38*b2055c35SXin Li uint32x4_t sum = vdupq_n_u32(0);
39*b2055c35SXin Li int i;
40*b2055c35SXin Li
41*b2055c35SXin Li for (i = 0; i < num_blocks; ++i) {
42*b2055c35SXin Li // Set DC to zero.
43*b2055c35SXin Li const int16x8_t a_0 = vsetq_lane_s16(0, vld1q_s16(levels), 0);
44*b2055c35SXin Li const int16x8_t a_1 = vld1q_s16(levels + 8);
45*b2055c35SXin Li
46*b2055c35SXin Li const uint16x8_t b_0 = vshrq_n_u16(vtstq_s16(a_0, tst_ones), 15);
47*b2055c35SXin Li const uint16x8_t b_1 = vshrq_n_u16(vtstq_s16(a_1, tst_ones), 15);
48*b2055c35SXin Li
49*b2055c35SXin Li sum = vpadalq_u16(sum, b_0);
50*b2055c35SXin Li sum = vpadalq_u16(sum, b_1);
51*b2055c35SXin Li
52*b2055c35SXin Li levels += 16;
53*b2055c35SXin Li }
54*b2055c35SXin Li return thresh >= (int)horizontal_add_uint32x4(sum);
55*b2055c35SXin Li }
56*b2055c35SXin Li
57*b2055c35SXin Li #else
58*b2055c35SXin Li
59*b2055c35SXin Li #define IsFlat IsFlat_C
60*b2055c35SXin Li
IsFlat(const int16_t * levels,int num_blocks,int thresh)61*b2055c35SXin Li static WEBP_INLINE int IsFlat(const int16_t* levels, int num_blocks,
62*b2055c35SXin Li int thresh) {
63*b2055c35SXin Li int score = 0;
64*b2055c35SXin Li while (num_blocks-- > 0) { // TODO(skal): refine positional scoring?
65*b2055c35SXin Li int i;
66*b2055c35SXin Li for (i = 1; i < 16; ++i) { // omit DC, we're only interested in AC
67*b2055c35SXin Li score += (levels[i] != 0);
68*b2055c35SXin Li if (score > thresh) return 0;
69*b2055c35SXin Li }
70*b2055c35SXin Li levels += 16;
71*b2055c35SXin Li }
72*b2055c35SXin Li return 1;
73*b2055c35SXin Li }
74*b2055c35SXin Li
75*b2055c35SXin Li #endif // defined(WEBP_USE_NEON) && !defined(WEBP_ANDROID_NEON) &&
76*b2055c35SXin Li // !defined(WEBP_HAVE_NEON_RTCD)
77*b2055c35SXin Li
IsFlatSource16(const uint8_t * src)78*b2055c35SXin Li static WEBP_INLINE int IsFlatSource16(const uint8_t* src) {
79*b2055c35SXin Li const uint32_t v = src[0] * 0x01010101u;
80*b2055c35SXin Li int i;
81*b2055c35SXin Li for (i = 0; i < 16; ++i) {
82*b2055c35SXin Li if (memcmp(src + 0, &v, 4) || memcmp(src + 4, &v, 4) ||
83*b2055c35SXin Li memcmp(src + 8, &v, 4) || memcmp(src + 12, &v, 4)) {
84*b2055c35SXin Li return 0;
85*b2055c35SXin Li }
86*b2055c35SXin Li src += BPS;
87*b2055c35SXin Li }
88*b2055c35SXin Li return 1;
89*b2055c35SXin Li }
90*b2055c35SXin Li
91*b2055c35SXin Li #endif // WEBP_DSP_QUANT_H_
92