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 <xnnpack/math.h>
13 #include <xnnpack/vmul.h>
14
15
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)])16 void xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x2(
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 int32_t vb_zero_point = params->fp32_scalar.b_zero_point;
25 const float vscale = params->fp32_scalar.scale;
26 const float voutput_min_less_zero_point = params->fp32_scalar.output_min_less_zero_point;
27 const float voutput_max_less_zero_point = params->fp32_scalar.output_max_less_zero_point;
28 const float vmagic_bias = params->fp32_scalar.magic_bias;
29 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar.magic_bias_less_output_zero_point;
30
31 for (; n >= 2 * sizeof(uint8_t); n -= 2 * 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 input_a += 2;
35
36 const int32_t vb0 = input_b[0] - vb_zero_point;
37 const int32_t vb1 = input_b[1] - vb_zero_point;
38 input_b += 2;
39
40 const int32_t vacc0 = va0 * vb0;
41 const int32_t vacc1 = va1 * vb1;
42
43 float vfpacc0 = (float) vacc0 * vscale;
44 float vfpacc1 = (float) vacc1 * vscale;
45
46 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
47 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
48
49 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
50 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
51
52 vfpacc0 += vmagic_bias;
53 vfpacc1 += vmagic_bias;
54
55 const int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point;
56 const int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point;
57
58 output[0] = (uint8_t) vout0;
59 output[1] = (uint8_t) vout1;
60 output += 2;
61 }
62 if XNN_UNLIKELY(n != 0) {
63 const int32_t va = (int32_t) *input_a - va_zero_point;
64 const int32_t vb = (int32_t) *input_b - vb_zero_point;
65 const int32_t vacc = va * vb;
66
67 float vfpacc = (float) vacc * vscale;
68 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
69 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
70 vfpacc += vmagic_bias;
71 const int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
72 *output = (uint8_t) vout;
73 }
74 }
75