Debian Codegraph @

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

raw
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/vbinary.h"
14
15
16 void xnn_qu8_vaddc_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 + (int32_t) *input_b * params->scalar.b_multiplier;
30 const int32_t va_multiplier = params->scalar.a_multiplier;
31 const uint32_t vshift = params->scalar.shift;
32 const int32_t voutput_min = params->scalar.output_min;
33 const int32_t voutput_max = params->scalar.output_max;
34 const int32_t voutput_zero_point = params->scalar.output_zero_point;
35
36 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
37 const int32_t va0 = input_a[0];
38 const int32_t va1 = input_a[1];
39 const int32_t va2 = input_a[2];
40 const int32_t va3 = input_a[3];
41 input_a += 4;
42
43 const int32_t vacc0 = vbias + va0 * va_multiplier;
44 const int32_t vacc1 = vbias + va1 * va_multiplier;
45 const int32_t vacc2 = vbias + va2 * va_multiplier;
46 const int32_t vacc3 = vbias + va3 * va_multiplier;
47 input_b += 4;
48
49 int32_t vout0 = math_asr_s32(vacc0, vshift);
50 int32_t vout1 = math_asr_s32(vacc1, vshift);
51 int32_t vout2 = math_asr_s32(vacc2, vshift);
52 int32_t vout3 = math_asr_s32(vacc3, vshift);
53
54 vout0 += voutput_zero_point;
55 vout1 += voutput_zero_point;
56 vout2 += voutput_zero_point;
57 vout3 += voutput_zero_point;
58
59 vout0 = math_max_s32(vout0, voutput_min);
60 vout1 = math_max_s32(vout1, voutput_min);
61 vout2 = math_max_s32(vout2, voutput_min);
62 vout3 = math_max_s32(vout3, voutput_min);
63
64 vout0 = math_min_s32(vout0, voutput_max);
65 vout1 = math_min_s32(vout1, voutput_max);
66 vout2 = math_min_s32(vout2, voutput_max);
67 vout3 = math_min_s32(vout3, voutput_max);
68
69 output[0] = (uint8_t) vout0;
70 output[1] = (uint8_t) vout1;
71 output[2] = (uint8_t) vout2;
72 output[3] = (uint8_t) vout3;
73 output += 4;
74 }
75 if XNN_UNLIKELY(batch != 0) {
76 do {
77 const int32_t va = *input_a++;
78 const int32_t vacc = vbias + va * va_multiplier;
79
80 int32_t vout = math_asr_s32(vacc, vshift);
81 vout = vout + voutput_zero_point;
82 vout = math_max_s32(vout, voutput_min);
83 vout = math_min_s32(vout, voutput_max);
84 *output++ = (uint8_t) vout;
85
86 batch -= sizeof(uint8_t);
87 } while (batch != 0);
88 }
89 }