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_x2(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_x2(
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 >= 2 * sizeof(uint8_t); n -= 2 * 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 input_a += 2;
37
38 const int32_t vb0 = input_b[0] - vb_zero_point;
39 const int32_t vb1 = input_b[1] - vb_zero_point;
40 input_b += 2;
41
42 const int32_t vacc0 = va0 * vb0;
43 const int32_t vacc1 = va1 * vb1;
44
45 float vfpacc0 = (float) vacc0 * vscale;
46 float vfpacc1 = (float) vacc1 * 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
51 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
52 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
53
54 vfpacc0 += vmagic_bias;
55 vfpacc1 += vmagic_bias;
56
57 const int32_t vout0 = (int32_t) fp32_to_bits(vfpacc0) - vmagic_bias_less_output_zero_point;
58 const int32_t vout1 = (int32_t) fp32_to_bits(vfpacc1) - vmagic_bias_less_output_zero_point;
59
60 output[0] = (uint8_t) vout0;
61 output[1] = (uint8_t) vout1;
62 output += 2;
63 }
64 if XNN_UNLIKELY(n != 0) {
65 const int32_t va = (int32_t) *input_a - va_zero_point;
66 const int32_t vb = (int32_t) *input_b - vb_zero_point;
67 const int32_t vacc = va * vb;
68
69 float vfpacc = (float) vacc * vscale;
70 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
71 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
72 vfpacc += vmagic_bias;
73 const int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point;
74 *output = (uint8_t) vout;
75 }
76 }
77