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_qu8_vadd_minmax_ukernel__scalar_x2(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_vadd_minmax_ukernel__scalar_x2(
17 size_t n,
18 const uint8_t* input_a,
19 const uint8_t* input_b,
20 uint8_t* output,
21 const union xnn_qu8_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 >= 2 * sizeof(uint8_t); n -= 2 * sizeof(uint8_t)) {
32 const int32_t va0 = input_a[0];
33 const int32_t va1 = input_a[1];
34 input_a += 2;
35
36 const int32_t vb0 = input_b[0];
37 int32_t vacc0 = vbias + va0 * va_multiplier;
38 const int32_t vb1 = input_b[1];
39 int32_t vacc1 = vbias + va1 * va_multiplier;
40 input_b += 2;
41
42 vacc0 += vb0 * vb_multiplier;
43 vacc1 += vb1 * vb_multiplier;
44
45 int32_t vout0 = math_asr_s32(vacc0, vshift);
46 int32_t vout1 = math_asr_s32(vacc1, 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
51 vout0 = math_min_s32(vout0, voutput_max_less_zero_point);
52 vout1 = math_min_s32(vout1, voutput_max_less_zero_point);
53
54 vout0 += voutput_zero_point;
55 vout1 += voutput_zero_point;
56
57 output[0] = (uint8_t) vout0;
58 output[1] = (uint8_t) vout1;
59 output += 2;
60 }
61 if XNN_UNLIKELY(n != 0) {
62 const int32_t va = *input_a;
63 const int32_t vb = *input_b;
64 const int32_t vacc = vbias + va * va_multiplier + vb * vb_multiplier;
65
66 int32_t vout = math_asr_s32(vacc, vshift);
67 vout = math_max_s32(vout, voutput_min_less_zero_point);
68 vout = math_min_s32(vout, voutput_max_less_zero_point);
69 *output++ = (uint8_t) (vout + voutput_zero_point);
70 }
71 }
72