1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/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/vmul.h>
14
15
xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4(
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_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
22 {
23 const int32_t va_zero_point = params->fp32_scalar.a_zero_point;
24 const float vscale = params->fp32_scalar.scale;
25 const float voutput_min_less_zero_point = params->fp32_scalar.output_min_less_zero_point;
26 const float voutput_max_less_zero_point = params->fp32_scalar.output_max_less_zero_point;
27 const float vmagic_bias = params->fp32_scalar.magic_bias;
28 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar.magic_bias_less_output_zero_point;
29
30 const int32_t vb = (int32_t) *input_b - params->fp32_scalar.b_zero_point;
31 for (; n >= 4 * sizeof(uint8_t); n -= 4 * sizeof(uint8_t)) {
32 const int32_t va0 = input_a[0] - va_zero_point;
33 const int32_t va1 = input_a[1] - va_zero_point;
34 const int32_t va2 = input_a[2] - va_zero_point;
35 const int32_t va3 = input_a[3] - va_zero_point;
36 input_a += 4;
37
38 const int32_t vacc0 = va0 * vb;
39 const int32_t vacc1 = va1 * vb;
40 const int32_t vacc2 = va2 * vb;
41 const int32_t vacc3 = va3 * vb;
42
43 float vfpacc0 = (float) vacc0 * vscale;
44 float vfpacc1 = (float) vacc1 * vscale;
45 float vfpacc2 = (float) vacc2 * vscale;
46 float vfpacc3 = (float) vacc3 * vscale;
47
48 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
49 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
50 vfpacc2 = math_max_f32(vfpacc2, voutput_min_less_zero_point);
51 vfpacc3 = math_max_f32(vfpacc3, voutput_min_less_zero_point);
52
53 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
54 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
55 vfpacc2 = math_min_f32(vfpacc2, voutput_max_less_zero_point);
56 vfpacc3 = math_min_f32(vfpacc3, voutput_max_less_zero_point);
57
58 vfpacc0 += vmagic_bias;
59 vfpacc1 += vmagic_bias;
60 vfpacc2 += vmagic_bias;
61 vfpacc3 += vmagic_bias;
62
63 const int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point;
64 const int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point;
65 const int32_t vout2 = (int32_t) float_as_uint32(vfpacc2) - vmagic_bias_less_output_zero_point;
66 const int32_t vout3 = (int32_t) float_as_uint32(vfpacc3) - vmagic_bias_less_output_zero_point;
67
68 output[0] = (uint8_t) vout0;
69 output[1] = (uint8_t) vout1;
70 output[2] = (uint8_t) vout2;
71 output[3] = (uint8_t) vout3;
72 output += 4;
73 }
74 if XNN_UNLIKELY(n != 0) {
75 do {
76 const int32_t va = (int32_t) *input_a++ - va_zero_point;
77 const int32_t vacc = va * vb;
78
79 float vfpacc = (float) vacc * vscale;
80 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
81 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
82 vfpacc += vmagic_bias;
83 const int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
84 *output++ = (uint8_t) vout;
85
86 n -= sizeof(uint8_t);
87 } while (n != 0);
88 }
89 }
90