1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/vadd.h>
14
15
xnn_qs8_vadd_minmax_ukernel__scalar_x4(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qs8_vadd_minmax_ukernel__scalar_x4(
17 size_t n,
18 const int8_t* input_a,
19 const int8_t* input_b,
20 int8_t* output,
21 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23 const int32_t vbias = params->scalar.bias;
24 const int32_t va_multiplier = params->scalar.a_multiplier;
25 const int32_t vb_multiplier = params->scalar.b_multiplier;
26 const uint32_t vshift = params->scalar.shift;
27 const int32_t voutput_min_less_zero_point = params->scalar.output_min_less_zero_point;
28 const int32_t voutput_max_less_zero_point = params->scalar.output_max_less_zero_point;
29 const int32_t voutput_zero_point = params->scalar.output_zero_point;
30
31 for (; n >= 4 * sizeof(int8_t); n -= 4 * sizeof(int8_t)) {
32 const int32_t va0 = input_a[0];
33 const int32_t va1 = input_a[1];
34 const int32_t va2 = input_a[2];
35 const int32_t va3 = input_a[3];
36 input_a += 4;
37
38 const int32_t vb0 = input_b[0];
39 int32_t vacc0 = vbias + va0 * va_multiplier;
40 const int32_t vb1 = input_b[1];
41 int32_t vacc1 = vbias + va1 * va_multiplier;
42 const int32_t vb2 = input_b[2];
43 int32_t vacc2 = vbias + va2 * va_multiplier;
44 const int32_t vb3 = input_b[3];
45 int32_t vacc3 = vbias + va3 * va_multiplier;
46 input_b += 4;
47
48 vacc0 += vb0 * vb_multiplier;
49 vacc1 += vb1 * vb_multiplier;
50 vacc2 += vb2 * vb_multiplier;
51 vacc3 += vb3 * vb_multiplier;
52
53 int32_t vout0 = math_asr_s32(vacc0, vshift);
54 int32_t vout1 = math_asr_s32(vacc1, vshift);
55 int32_t vout2 = math_asr_s32(vacc2, vshift);
56 int32_t vout3 = math_asr_s32(vacc3, vshift);
57
58 vout0 = math_max_s32(vout0, voutput_min_less_zero_point);
59 vout1 = math_max_s32(vout1, voutput_min_less_zero_point);
60 vout2 = math_max_s32(vout2, voutput_min_less_zero_point);
61 vout3 = math_max_s32(vout3, voutput_min_less_zero_point);
62
63 vout0 = math_min_s32(vout0, voutput_max_less_zero_point);
64 vout1 = math_min_s32(vout1, voutput_max_less_zero_point);
65 vout2 = math_min_s32(vout2, voutput_max_less_zero_point);
66 vout3 = math_min_s32(vout3, voutput_max_less_zero_point);
67
68 vout0 += voutput_zero_point;
69 vout1 += voutput_zero_point;
70 vout2 += voutput_zero_point;
71 vout3 += voutput_zero_point;
72
73 output[0] = (int8_t) vout0;
74 output[1] = (int8_t) vout1;
75 output[2] = (int8_t) vout2;
76 output[3] = (int8_t) vout3;
77 output += 4;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 do {
81 const int32_t va = *input_a++;
82 const int32_t vb = *input_b++;
83 const int32_t vacc = vbias + va * va_multiplier + vb * vb_multiplier;
84
85 int32_t vout = math_asr_s32(vacc, vshift);
86 vout = math_max_s32(vout, voutput_min_less_zero_point);
87 vout = math_min_s32(vout, voutput_max_less_zero_point);
88 *output++ = (int8_t) (vout + voutput_zero_point);
89
90 n -= sizeof(int8_t);
91 } while (n != 0);
92 }
93 }
94