1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmul/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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/vmul.h>
16
17
xnn_qu8_vmul_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)])18 void xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x4(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 const int32_t va_zero_point = params->fp32_scalar.a_zero_point;
26 const int32_t vb_zero_point = params->fp32_scalar.b_zero_point;
27 const float vscale = params->fp32_scalar.scale;
28 const float voutput_min_less_zero_point = params->fp32_scalar.output_min_less_zero_point;
29 const float voutput_max_less_zero_point = params->fp32_scalar.output_max_less_zero_point;
30 const float vmagic_bias = params->fp32_scalar.magic_bias;
31 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar.magic_bias_less_output_zero_point;
32
33 for (; n >= 4 * sizeof(uint8_t); n -= 4 * sizeof(uint8_t)) {
34 const int32_t va0 = input_a[0] - va_zero_point;
35 const int32_t va1 = input_a[1] - va_zero_point;
36 const int32_t va2 = input_a[2] - va_zero_point;
37 const int32_t va3 = input_a[3] - va_zero_point;
38 input_a += 4;
39
40 const int32_t vb0 = input_b[0] - vb_zero_point;
41 const int32_t vb1 = input_b[1] - vb_zero_point;
42 const int32_t vb2 = input_b[2] - vb_zero_point;
43 const int32_t vb3 = input_b[3] - vb_zero_point;
44 input_b += 4;
45
46 const int32_t vacc0 = va0 * vb0;
47 const int32_t vacc1 = va1 * vb1;
48 const int32_t vacc2 = va2 * vb2;
49 const int32_t vacc3 = va3 * vb3;
50
51 float vfpacc0 = (float) vacc0 * vscale;
52 float vfpacc1 = (float) vacc1 * vscale;
53 float vfpacc2 = (float) vacc2 * vscale;
54 float vfpacc3 = (float) vacc3 * vscale;
55
56 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
57 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
58 vfpacc2 = math_max_f32(vfpacc2, voutput_min_less_zero_point);
59 vfpacc3 = math_max_f32(vfpacc3, voutput_min_less_zero_point);
60
61 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
62 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
63 vfpacc2 = math_min_f32(vfpacc2, voutput_max_less_zero_point);
64 vfpacc3 = math_min_f32(vfpacc3, voutput_max_less_zero_point);
65
66 vfpacc0 += vmagic_bias;
67 vfpacc1 += vmagic_bias;
68 vfpacc2 += vmagic_bias;
69 vfpacc3 += vmagic_bias;
70
71 const int32_t vout0 = (int32_t) fp32_to_bits(vfpacc0) - vmagic_bias_less_output_zero_point;
72 const int32_t vout1 = (int32_t) fp32_to_bits(vfpacc1) - vmagic_bias_less_output_zero_point;
73 const int32_t vout2 = (int32_t) fp32_to_bits(vfpacc2) - vmagic_bias_less_output_zero_point;
74 const int32_t vout3 = (int32_t) fp32_to_bits(vfpacc3) - vmagic_bias_less_output_zero_point;
75
76 output[0] = (uint8_t) vout0;
77 output[1] = (uint8_t) vout1;
78 output[2] = (uint8_t) vout2;
79 output[3] = (uint8_t) vout3;
80 output += 4;
81 }
82 if XNN_UNLIKELY(n != 0) {
83 do {
84 const int32_t va = (int32_t) *input_a++ - va_zero_point;
85 const int32_t vb = (int32_t) *input_b++ - vb_zero_point;
86 const int32_t vacc = va * vb;
87
88 float vfpacc = (float) vacc * vscale;
89 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
90 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
91 vfpacc += vmagic_bias;
92 const int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point;
93 *output++ = (uint8_t) vout;
94
95 n -= sizeof(uint8_t);
96 } while (n != 0);
97 }
98 }
99