1 /*
2 * Copyright 2021 Google LLC
3 *
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
6 */
7
8 #include <algorithm>
9 #include <cmath>
10 #include <cstdio>
11 #include <cstdint>
12
13 #include "experimental/lowp-basic/QMath.h"
14
15 struct Stats {
16 int diff_8_bits = 0;
17 int max_diff = 0;
18 int min_diff = 0;
19 int64_t total = 0;
20
logStats21 void log(int16_t golden, int16_t candidate) {
22 int diff = candidate - golden;
23 max_diff = std::max(max_diff, diff);
24 min_diff = std::min(min_diff, diff);
25 diff_8_bits += candidate != golden;
26 total++;
27 }
28
printStats29 void print() const {
30 printf("8-bit diff: %d - %g%%\n", diff_8_bits, 100.0 * diff_8_bits / total);
31 printf("differences min: %d max: %d\n", min_diff, max_diff);
32 printf("total: %lld\n", total);
33 }
34 };
35
golden_lerp(float t,int16_t a,int16_t b)36 static float golden_lerp(float t, int16_t a, int16_t b) {
37 return (1.0f - t) * a + t * b;
38 }
39
40 template <int logPixelScale>
saturating_lerp(float t,int16_t a,int16_t b)41 static int16_t saturating_lerp(float t, int16_t a, int16_t b) {
42 const int16_t half = 1 << (logPixelScale - 1);
43 Q15 qt(floor(t * 32768.f + 0.5f));
44 Q15 qa(a << logPixelScale);
45 Q15 qb(b << logPixelScale);
46
47 Q15 answer = simulate_neon_vqrdmulhq_s16(qt, qb - qa) + qa;
48 return (answer[0] + half) >> logPixelScale;
49 }
50
51 template <int logPixelScale>
ssse3_lerp(float t,int16_t a,int16_t b)52 static int16_t ssse3_lerp(float t, int16_t a, int16_t b) {
53 const int16_t half = 1 << (logPixelScale - 1);
54 Q15 qt(floor(t * 32768.f + 0.5f));
55 Q15 qa(a << logPixelScale);
56 Q15 qb(b << logPixelScale);
57
58 Q15 answer = simulate_ssse3_mm_mulhrs_epi16(qt, qb - qa) + qa;
59 return (answer[0] + half) >> logPixelScale;
60 }
61
full_res_lerp(float t,int16_t a,int16_t b)62 static int16_t full_res_lerp(float t, int16_t a, int16_t b) {
63 int32_t ft(floor(t * 65536.0f + 0.5f));
64
65 int32_t temp = ft * (b - a) + a * 65536;
66 int32_t rounded = temp + 32768;
67 return rounded >> 16;
68 }
69
70 // Change of parameters on t from [0, 1) to [-1, 1). This cuts the number if differences in half.
71 template <int logPixelScale>
balanced_lerp(float t,int16_t a,int16_t b)72 static int16_t balanced_lerp(float t, int16_t a, int16_t b) {
73 const int16_t half = 1 << logPixelScale;
74 // t on [-1, 1).
75 Q15 qt (floor(t * 65536.0f - 32768.0f + 0.5f));
76 // need to pick logPixelScale to scale by addition 1/2.
77 Q15 qw ((b - a) << logPixelScale);
78 Q15 qm ((a + b) << logPixelScale);
79 Q15 answer = simulate_ssse3_mm_mulhrs_epi16(qt, qw) + qm;
80 // Extra shift to divide by 2.
81 return (answer[0] + half) >> (logPixelScale + 1);
82 }
83
84 template <typename Lerp>
check_lerp(Lerp lerp)85 static Stats check_lerp(Lerp lerp) {
86 Stats stats;
87 for (float t = 0; t < 1.0f - 1.0f / 65536.0f ; t += 1.0f/65536.0f)
88 for (int a = 255; a >= 0; a--)
89 for (int b = 255; b >= 0; b--) {
90 float l = golden_lerp(t, a, b);
91 int16_t golden = floor(l + 0.5f);
92 int16_t candidate = lerp(t, a, b);
93 stats.log(golden, candidate);
94 }
95 return stats;
96 }
97
main()98 int main() {
99 Stats stats;
100
101 printf("\nUsing full_res_lerp...\n");
102 stats = check_lerp(full_res_lerp);
103 stats.print();
104
105 printf("\nUsing vqrdmulhq_s16...\n");
106 stats = check_lerp(saturating_lerp<7>);
107 stats.print();
108
109 printf("\nUsing mm_mulhrs_epi16...\n");
110 stats = check_lerp(ssse3_lerp<7>);
111 stats.print();
112
113 printf("\nInterval [-1, 1) mm_mulhrs_epi16...\n");
114 stats = check_lerp(balanced_lerp<7>);
115 stats.print();
116
117 printf("Done.");
118 return 0;
119 }
120
121