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_qu8_dwconv_minmax_fp32_ukernel_up1x9__wasm_fmagic(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__wasm_fmagic(
18 size_t channels,
19 size_t output_width,
20 const uint8_t** input,
21 const void* weights,
22 uint8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const uint8_t* zero,
27 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float vscale = params->fp32_scalar_fmagic.scale;
33 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
34 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
35 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
36 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
37 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
38 do {
39 const uint8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const uint8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const uint8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const uint8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const uint8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const uint8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const uint8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const uint8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const uint8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 input = (const uint8_t**) ((uintptr_t) input + input_stride);
85
86 size_t c = channels;
87 const void* w = weights;
88 do {
89 int32_t vacc = unaligned_load_s32(w);
90
91 const int32_t vi0 = (int32_t) (uint32_t) *i0++;
92 const int32_t vk0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[0] - vkernel_zero_point;
93 vacc += vi0 * vk0;
94 const int32_t vi1 = (int32_t) (uint32_t) *i1++;
95 const int32_t vk1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[1] - vkernel_zero_point;
96 vacc += vi1 * vk1;
97 const int32_t vi2 = (int32_t) (uint32_t) *i2++;
98 const int32_t vk2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[2] - vkernel_zero_point;
99 vacc += vi2 * vk2;
100 const int32_t vi3 = (int32_t) (uint32_t) *i3++;
101 const int32_t vk3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[3] - vkernel_zero_point;
102 vacc += vi3 * vk3;
103 const int32_t vi4 = (int32_t) (uint32_t) *i4++;
104 const int32_t vk4 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[4] - vkernel_zero_point;
105 vacc += vi4 * vk4;
106 const int32_t vi5 = (int32_t) (uint32_t) *i5++;
107 const int32_t vk5 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[5] - vkernel_zero_point;
108 vacc += vi5 * vk5;
109 const int32_t vi6 = (int32_t) (uint32_t) *i6++;
110 const int32_t vk6 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[6] - vkernel_zero_point;
111 vacc += vi6 * vk6;
112 const int32_t vi7 = (int32_t) (uint32_t) *i7++;
113 const int32_t vk7 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[7] - vkernel_zero_point;
114 vacc += vi7 * vk7;
115 const int32_t vi8 = (int32_t) (uint32_t) *i8++;
116 const int32_t vk8 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[8] - vkernel_zero_point;
117 vacc += vi8 * vk8;
118
119 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 9 * sizeof(uint8_t));
120
121 float vfpacc = (float) vacc * vscale;
122
123 vfpacc = __builtin_wasm_max_f32(vfpacc, voutput_min_less_zero_point);
124 vfpacc = __builtin_wasm_min_f32(vfpacc, voutput_max_less_zero_point);
125 vfpacc += vmagic_bias;
126 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
127
128 *output++ = (uint8_t) vout;
129 } while (--c != 0);
130
131 output = (uint8_t*) ((uintptr_t) output + output_increment);
132 } while (--output_width != 0);
133 }
134