1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-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/dwconv.h>
13 #include <xnnpack/math.h>
14 #include <xnnpack/unaligned.h>
15
16
xnn_qc8_dwconv_minmax_fp32_ukernel_up1x3__scalar_fmagic(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up1x3__scalar_fmagic(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
33 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
34 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
35 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
36 do {
37 const int8_t* i0 = input[0];
38 assert(i0 != NULL);
39 if XNN_UNPREDICTABLE(i0 != zero) {
40 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41 }
42 const int8_t* i1 = input[1];
43 assert(i1 != NULL);
44 if XNN_UNPREDICTABLE(i1 != zero) {
45 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46 }
47 const int8_t* i2 = input[2];
48 assert(i2 != NULL);
49 if XNN_UNPREDICTABLE(i2 != zero) {
50 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51 }
52 input = (const int8_t**) ((uintptr_t) input + input_stride);
53
54 size_t c = channels;
55 const void* w = weights;
56 do {
57 int32_t vacc = unaligned_load_s32(w);
58
59 const int32_t vi0 = (int32_t) *i0++;
60 const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0];
61 vacc += vi0 * vk0;
62 const int32_t vi1 = (int32_t) *i1++;
63 const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1];
64 vacc += vi1 * vk1;
65 const int32_t vi2 = (int32_t) *i2++;
66 const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2];
67 vacc += vi2 * vk2;
68
69 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 3 * sizeof(int8_t));
70
71 const float vscale = unaligned_load_f32(w);
72 w = (const void*) ((const float*) w + 1);
73 float vfpacc = (float) vacc * vscale;
74
75 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
76 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
77 vfpacc += vmagic_bias;
78 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
79
80 *output++ = (int8_t) vout;
81 } while (--c != 0);
82
83 output = (int8_t*) ((uintptr_t) output + output_increment);
84 } while (--output_width != 0);
85 }
86