xref: /aosp_15_r20/external/XNNPACK/src/qu8-vlrelu/gen/vlrelu-scalar-andxor-x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vlrelu/scalar-andxor.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <xnnpack/math.h>
13 #include <xnnpack/vlrelu.h>
14 
15 
xnn_qu8_vlrelu_ukernel__scalar_andxor_x4(size_t n,const uint8_t * x,uint8_t * y,const union xnn_qu8_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_vlrelu_ukernel__scalar_andxor_x4(
17     size_t n,
18     const uint8_t* x,
19     uint8_t* y,
20     const union xnn_qu8_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)])
21 {
22   const int32_t vinput_zero_point = params->scalar_andxor.input_zero_point;
23   const int32_t vmultiplier_diff = params->scalar_andxor.multiplier_diff;
24   const int32_t vmultiplier_base = params->scalar_andxor.multiplier_base;
25   const int32_t vbias = params->scalar_andxor.bias;
26   for (; n >= 4 * sizeof(uint8_t); n -= 4 * sizeof(uint8_t)) {
27     int32_t vacc0 = (int32_t) x[0];
28     int32_t vacc1 = (int32_t) x[1];
29     int32_t vacc2 = (int32_t) x[2];
30     int32_t vacc3 = (int32_t) x[3];
31     x += 4;
32 
33     vacc0 -= vinput_zero_point;
34     vacc1 -= vinput_zero_point;
35     vacc2 -= vinput_zero_point;
36     vacc3 -= vinput_zero_point;
37 
38     int32_t vmultiplier0 = math_asr_s32(vacc0, 31);
39     int32_t vmultiplier1 = math_asr_s32(vacc1, 31);
40     int32_t vmultiplier2 = math_asr_s32(vacc2, 31);
41     int32_t vmultiplier3 = math_asr_s32(vacc3, 31);
42 
43     vmultiplier0 &= vmultiplier_diff;
44     vmultiplier1 &= vmultiplier_diff;
45     vmultiplier2 &= vmultiplier_diff;
46     vmultiplier3 &= vmultiplier_diff;
47 
48     vmultiplier0 ^= vmultiplier_base;
49     vmultiplier1 ^= vmultiplier_base;
50     vmultiplier2 ^= vmultiplier_base;
51     vmultiplier3 ^= vmultiplier_base;
52 
53     vacc0 = vbias + vacc0 * vmultiplier0;
54     vacc1 = vbias + vacc1 * vmultiplier1;
55     vacc2 = vbias + vacc2 * vmultiplier2;
56     vacc3 = vbias + vacc3 * vmultiplier3;
57 
58     int32_t vout0 = math_asr_s32(vacc0, 8);
59     int32_t vout1 = math_asr_s32(vacc1, 8);
60     int32_t vout2 = math_asr_s32(vacc2, 8);
61     int32_t vout3 = math_asr_s32(vacc3, 8);
62 
63     vout0 = math_max_s32(vout0, 0);
64     vout1 = math_max_s32(vout1, 0);
65     vout2 = math_max_s32(vout2, 0);
66     vout3 = math_max_s32(vout3, 0);
67 
68     vout0 = math_min_s32(vout0, 255);
69     vout1 = math_min_s32(vout1, 255);
70     vout2 = math_min_s32(vout2, 255);
71     vout3 = math_min_s32(vout3, 255);
72 
73     y[0] = (uint8_t) vout0;
74     y[1] = (uint8_t) vout1;
75     y[2] = (uint8_t) vout2;
76     y[3] = (uint8_t) vout3;
77     y += 4;
78   }
79   if XNN_UNLIKELY(n != 0) {
80     do {
81       int32_t vacc = (int32_t) *x++ - vinput_zero_point;
82       const int32_t vmultiplier = vmultiplier_base ^ (vmultiplier_diff & math_asr_s32(vacc, 31));
83       vacc = vbias + vacc * vmultiplier;
84 
85       int32_t vout = math_asr_s32(vacc, 8);
86       vout = math_max_s32(vout, 0);
87       vout = math_min_s32(vout, 255);
88       *y++ = (uint8_t) vout;
89 
90       n -= sizeof(uint8_t);
91     } while (n != 0);
92   }
93 }
94