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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/vmul.h>
16
17
xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4(
19 size_t n,
20 const int8_t* input_a,
21 const int8_t* input_b,
22 int8_t* output,
23 const union xnn_qs8_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 float vscale = params->fp32_scalar.scale;
27 const float voutput_min_less_zero_point = params->fp32_scalar.output_min_less_zero_point;
28 const float voutput_max_less_zero_point = params->fp32_scalar.output_max_less_zero_point;
29 const float vmagic_bias = params->fp32_scalar.magic_bias;
30 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar.magic_bias_less_output_zero_point;
31
32 const int32_t vb = (int32_t) *input_b - params->fp32_scalar.b_zero_point;
33 for (; n >= 4 * sizeof(int8_t); n -= 4 * sizeof(int8_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 vacc0 = va0 * vb;
41 const int32_t vacc1 = va1 * vb;
42 const int32_t vacc2 = va2 * vb;
43 const int32_t vacc3 = va3 * vb;
44
45 float vfpacc0 = (float) vacc0 * vscale;
46 float vfpacc1 = (float) vacc1 * vscale;
47 float vfpacc2 = (float) vacc2 * vscale;
48 float vfpacc3 = (float) vacc3 * vscale;
49
50 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
51 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
52 vfpacc2 = math_max_f32(vfpacc2, voutput_min_less_zero_point);
53 vfpacc3 = math_max_f32(vfpacc3, voutput_min_less_zero_point);
54
55 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
56 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
57 vfpacc2 = math_min_f32(vfpacc2, voutput_max_less_zero_point);
58 vfpacc3 = math_min_f32(vfpacc3, voutput_max_less_zero_point);
59
60 vfpacc0 += vmagic_bias;
61 vfpacc1 += vmagic_bias;
62 vfpacc2 += vmagic_bias;
63 vfpacc3 += vmagic_bias;
64
65 const int32_t vout0 = (int32_t) fp32_to_bits(vfpacc0) - vmagic_bias_less_output_zero_point;
66 const int32_t vout1 = (int32_t) fp32_to_bits(vfpacc1) - vmagic_bias_less_output_zero_point;
67 const int32_t vout2 = (int32_t) fp32_to_bits(vfpacc2) - vmagic_bias_less_output_zero_point;
68 const int32_t vout3 = (int32_t) fp32_to_bits(vfpacc3) - vmagic_bias_less_output_zero_point;
69
70 output[0] = (int8_t) vout0;
71 output[1] = (int8_t) vout1;
72 output[2] = (int8_t) vout2;
73 output[3] = (int8_t) vout3;
74 output += 4;
75 }
76 if XNN_UNLIKELY(n != 0) {
77 do {
78 const int32_t va = (int32_t) *input_a++ - va_zero_point;
79 const int32_t vacc = va * vb;
80
81 float vfpacc = (float) vacc * vscale;
82 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
83 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
84 vfpacc += vmagic_bias;
85 const int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point;
86 *output++ = (int8_t) vout;
87
88 n -= sizeof(int8_t);
89 } while (n != 0);
90 }
91 }
92