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_qs8_dwconv_minmax_fp32_ukernel_up2x9__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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__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_qs8_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 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 input = (const int8_t**) ((uintptr_t) input + input_stride);
84
85 size_t c = channels;
86 const void* w = weights;
87 for (; c >= 2; c -= 2) {
88 int32_t vacc0 = unaligned_indexed_load_s32(w, 0);
89 int32_t vacc1 = unaligned_indexed_load_s32(w, 1);
90
91
92 const int32_t vi0x0 = (int32_t) i0[0];
93 const int32_t vi0x1 = (int32_t) i0[1];
94 i0 += 2;
95
96 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
97 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1];
98
99 vacc0 += vi0x0 * vk0x0;
100 vacc1 += vi0x1 * vk0x1;
101
102 const int32_t vi1x0 = (int32_t) i1[0];
103 const int32_t vi1x1 = (int32_t) i1[1];
104 i1 += 2;
105
106 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
107 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3];
108
109 vacc0 += vi1x0 * vk1x0;
110 vacc1 += vi1x1 * vk1x1;
111
112 const int32_t vi2x0 = (int32_t) i2[0];
113 const int32_t vi2x1 = (int32_t) i2[1];
114 i2 += 2;
115
116 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
117 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5];
118
119 vacc0 += vi2x0 * vk2x0;
120 vacc1 += vi2x1 * vk2x1;
121
122 const int32_t vi3x0 = (int32_t) i3[0];
123 const int32_t vi3x1 = (int32_t) i3[1];
124 i3 += 2;
125
126 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
127 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7];
128
129 vacc0 += vi3x0 * vk3x0;
130 vacc1 += vi3x1 * vk3x1;
131
132 const int32_t vi4x0 = (int32_t) i4[0];
133 const int32_t vi4x1 = (int32_t) i4[1];
134 i4 += 2;
135
136 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
137 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9];
138
139 vacc0 += vi4x0 * vk4x0;
140 vacc1 += vi4x1 * vk4x1;
141
142 const int32_t vi5x0 = (int32_t) i5[0];
143 const int32_t vi5x1 = (int32_t) i5[1];
144 i5 += 2;
145
146 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
147 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11];
148
149 vacc0 += vi5x0 * vk5x0;
150 vacc1 += vi5x1 * vk5x1;
151
152 const int32_t vi6x0 = (int32_t) i6[0];
153 const int32_t vi6x1 = (int32_t) i6[1];
154 i6 += 2;
155
156 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
157 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13];
158
159 vacc0 += vi6x0 * vk6x0;
160 vacc1 += vi6x1 * vk6x1;
161
162 const int32_t vi7x0 = (int32_t) i7[0];
163 const int32_t vi7x1 = (int32_t) i7[1];
164 i7 += 2;
165
166 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
167 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15];
168
169 vacc0 += vi7x0 * vk7x0;
170 vacc1 += vi7x1 * vk7x1;
171
172 const int32_t vi8x0 = (int32_t) i8[0];
173 const int32_t vi8x1 = (int32_t) i8[1];
174 i8 += 2;
175
176 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
177 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17];
178
179 vacc0 += vi8x0 * vk8x0;
180 vacc1 += vi8x1 * vk8x1;
181
182 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 18 * sizeof(int8_t));
183
184 float vfpacc0 = (float) vacc0;
185 float vfpacc1 = (float) vacc1;
186
187 vfpacc0 *= vscale;
188 vfpacc1 *= vscale;
189
190 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
191 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
192
193 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
194 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
195
196 vfpacc0 += vmagic_bias;
197 vfpacc1 += vmagic_bias;
198
199 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point;
200 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point;
201
202 output[0] = (int8_t) vout0;
203 output[1] = (int8_t) vout1;
204 output += 2;
205 }
206 if XNN_UNLIKELY(c != 0) {
207 int32_t vacc = unaligned_load_s32(w);
208
209 const int32_t vi0 = (int32_t) *i0;
210 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
211 vacc += vi0 * vk0;
212 const int32_t vi1 = (int32_t) *i1;
213 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
214 vacc += vi1 * vk1;
215 const int32_t vi2 = (int32_t) *i2;
216 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
217 vacc += vi2 * vk2;
218 const int32_t vi3 = (int32_t) *i3;
219 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
220 vacc += vi3 * vk3;
221 const int32_t vi4 = (int32_t) *i4;
222 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
223 vacc += vi4 * vk4;
224 const int32_t vi5 = (int32_t) *i5;
225 const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
226 vacc += vi5 * vk5;
227 const int32_t vi6 = (int32_t) *i6;
228 const int32_t vk6 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
229 vacc += vi6 * vk6;
230 const int32_t vi7 = (int32_t) *i7;
231 const int32_t vk7 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
232 vacc += vi7 * vk7;
233 const int32_t vi8 = (int32_t) *i8;
234 const int32_t vk8 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
235 vacc += vi8 * vk8;
236
237 float vfpacc = (float) vacc * vscale;
238
239 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
240 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
241 vfpacc += vmagic_bias;
242 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
243
244 *output++ = (int8_t) vout;
245 }
246
247 output = (int8_t*) ((uintptr_t) output + output_increment);
248 } while (--output_width != 0);
249 }
250