Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-scalar-u4.c

raw
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/vbinary.h"
14
15
16 void xnn_qu8_vadd_minmax_ukernel__scalar_u4(
17 size_t batch,
18 const uint8_t* input_a,
19 const uint8_t* input_b,
20 uint8_t* output,
21 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23 assert(batch != 0);
24 assert(batch % sizeof(uint8_t) == 0);
25 assert(input_a != NULL);
26 assert(input_b != NULL);
27 assert(output != NULL);
28
29 const int32_t vbias = params->scalar.bias;
30 const int32_t va_multiplier = params->scalar.a_multiplier;
31 const int32_t vb_multiplier = params->scalar.b_multiplier;
32 const uint32_t vshift = params->scalar.shift;
33 const int32_t voutput_min = params->scalar.output_min;
34 const int32_t voutput_max = params->scalar.output_max;
35 const int32_t voutput_zero_point = params->scalar.output_zero_point;
36
37 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
38 const int32_t va0 = input_a[0];
39 const int32_t va1 = input_a[1];
40 const int32_t va2 = input_a[2];
41 const int32_t va3 = input_a[3];
42 input_a += 4;
43
44 const int32_t vb0 = input_b[0];
45 int32_t vacc0 = vbias + va0 * va_multiplier;
46 const int32_t vb1 = input_b[1];
47 int32_t vacc1 = vbias + va1 * va_multiplier;
48 const int32_t vb2 = input_b[2];
49 int32_t vacc2 = vbias + va2 * va_multiplier;
50 const int32_t vb3 = input_b[3];
51 int32_t vacc3 = vbias + va3 * va_multiplier;
52 input_b += 4;
53
54 vacc0 += vb0 * vb_multiplier;
55 vacc1 += vb1 * vb_multiplier;
56 vacc2 += vb2 * vb_multiplier;
57 vacc3 += vb3 * vb_multiplier;
58
59 int32_t vout0 = math_asr_s32(vacc0, vshift);
60 int32_t vout1 = math_asr_s32(vacc1, vshift);
61 int32_t vout2 = math_asr_s32(vacc2, vshift);
62 int32_t vout3 = math_asr_s32(vacc3, vshift);
63
64 vout0 += voutput_zero_point;
65 vout1 += voutput_zero_point;
66 vout2 += voutput_zero_point;
67 vout3 += voutput_zero_point;
68
69 vout0 = math_max_s32(vout0, voutput_min);
70 vout1 = math_max_s32(vout1, voutput_min);
71 vout2 = math_max_s32(vout2, voutput_min);
72 vout3 = math_max_s32(vout3, voutput_min);
73
74 vout0 = math_min_s32(vout0, voutput_max);
75 vout1 = math_min_s32(vout1, voutput_max);
76 vout2 = math_min_s32(vout2, voutput_max);
77 vout3 = math_min_s32(vout3, voutput_max);
78
79 output[0] = (uint8_t) vout0;
80 output[1] = (uint8_t) vout1;
81 output[2] = (uint8_t) vout2;
82 output[3] = (uint8_t) vout3;
83 output += 4;
84 }
85 if XNN_UNLIKELY(batch != 0) {
86 do {
87 const int32_t va = *input_a++;
88 const int32_t vb = *input_b++;
89 const int32_t vacc = vbias + va * va_multiplier + vb * vb_multiplier;
90
91 int32_t vout = math_asr_s32(vacc, vshift);
92 vout += voutput_zero_point;
93 vout = math_max_s32(vout, voutput_min);
94 vout = math_min_s32(vout, voutput_max);
95 *output++ = (uint8_t) vout;
96
97 batch -= sizeof(uint8_t);
98 } while (batch != 0);
99 }
100 }