1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/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_vaddc_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_vaddc_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 + (int32_t) *input_b * params->scalar.b_multiplier;
24 const int32_t va_multiplier = params->scalar.a_multiplier;
25 const uint32_t vshift = params->scalar.shift;
26 const int32_t voutput_min_less_zero_point = params->scalar.output_min_less_zero_point;
27 const int32_t voutput_max_less_zero_point = params->scalar.output_max_less_zero_point;
28 const int32_t voutput_zero_point = params->scalar.output_zero_point;
29
30 for (; n >= 4 * sizeof(int8_t); n -= 4 * sizeof(int8_t)) {
31 const int32_t va0 = input_a[0];
32 const int32_t va1 = input_a[1];
33 const int32_t va2 = input_a[2];
34 const int32_t va3 = input_a[3];
35 input_a += 4;
36
37 const int32_t vacc0 = vbias + va0 * va_multiplier;
38 const int32_t vacc1 = vbias + va1 * va_multiplier;
39 const int32_t vacc2 = vbias + va2 * va_multiplier;
40 const int32_t vacc3 = vbias + va3 * va_multiplier;
41 input_b += 4;
42
43 int32_t vout0 = math_asr_s32(vacc0, vshift);
44 int32_t vout1 = math_asr_s32(vacc1, vshift);
45 int32_t vout2 = math_asr_s32(vacc2, vshift);
46 int32_t vout3 = math_asr_s32(vacc3, vshift);
47
48 vout0 = math_max_s32(vout0, voutput_min_less_zero_point);
49 vout1 = math_max_s32(vout1, voutput_min_less_zero_point);
50 vout2 = math_max_s32(vout2, voutput_min_less_zero_point);
51 vout3 = math_max_s32(vout3, voutput_min_less_zero_point);
52
53 vout0 = math_min_s32(vout0, voutput_max_less_zero_point);
54 vout1 = math_min_s32(vout1, voutput_max_less_zero_point);
55 vout2 = math_min_s32(vout2, voutput_max_less_zero_point);
56 vout3 = math_min_s32(vout3, voutput_max_less_zero_point);
57
58 vout0 += voutput_zero_point;
59 vout1 += voutput_zero_point;
60 vout2 += voutput_zero_point;
61 vout3 += voutput_zero_point;
62
63 output[0] = (int8_t) vout0;
64 output[1] = (int8_t) vout1;
65 output[2] = (int8_t) vout2;
66 output[3] = (int8_t) vout3;
67 output += 4;
68 }
69 if XNN_UNLIKELY(n != 0) {
70 do {
71 const int32_t va = *input_a++;
72 const int32_t vacc = vbias + va * va_multiplier;
73
74 int32_t vout = math_asr_s32(vacc, vshift);
75 vout = math_max_s32(vout, voutput_min_less_zero_point);
76 vout = math_min_s32(vout, voutput_max_less_zero_point);
77 *output++ = (int8_t) (vout + voutput_zero_point);
78
79 n -= sizeof(int8_t);
80 } while (n != 0);
81 }
82 }
83