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_up2x9__scalar_imagic(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_up2x9__scalar_imagic(
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_imagic.scale;
33 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
34 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
35 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
36 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
37 const int32_t vkernel_zero_point = params->fp32_scalar_imagic.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 for (; c >= 2; c -= 2) {
89 int32_t vacc0 = unaligned_indexed_load_s32(w, 0);
90 int32_t vacc1 = unaligned_indexed_load_s32(w, 1);
91
92
93 const int32_t vi0x0 = (int32_t) (uint32_t) i0[0];
94 const int32_t vi0x1 = (int32_t) (uint32_t) i0[1];
95 i0 += 2;
96
97 const int32_t vk0x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point;
98 const int32_t vk0x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1] - vkernel_zero_point;
99
100 vacc0 += vi0x0 * vk0x0;
101 vacc1 += vi0x1 * vk0x1;
102
103 const int32_t vi1x0 = (int32_t) (uint32_t) i1[0];
104 const int32_t vi1x1 = (int32_t) (uint32_t) i1[1];
105 i1 += 2;
106
107 const int32_t vk1x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point;
108 const int32_t vk1x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3] - vkernel_zero_point;
109
110 vacc0 += vi1x0 * vk1x0;
111 vacc1 += vi1x1 * vk1x1;
112
113 const int32_t vi2x0 = (int32_t) (uint32_t) i2[0];
114 const int32_t vi2x1 = (int32_t) (uint32_t) i2[1];
115 i2 += 2;
116
117 const int32_t vk2x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point;
118 const int32_t vk2x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5] - vkernel_zero_point;
119
120 vacc0 += vi2x0 * vk2x0;
121 vacc1 += vi2x1 * vk2x1;
122
123 const int32_t vi3x0 = (int32_t) (uint32_t) i3[0];
124 const int32_t vi3x1 = (int32_t) (uint32_t) i3[1];
125 i3 += 2;
126
127 const int32_t vk3x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point;
128 const int32_t vk3x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7] - vkernel_zero_point;
129
130 vacc0 += vi3x0 * vk3x0;
131 vacc1 += vi3x1 * vk3x1;
132
133 const int32_t vi4x0 = (int32_t) (uint32_t) i4[0];
134 const int32_t vi4x1 = (int32_t) (uint32_t) i4[1];
135 i4 += 2;
136
137 const int32_t vk4x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point;
138 const int32_t vk4x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9] - vkernel_zero_point;
139
140 vacc0 += vi4x0 * vk4x0;
141 vacc1 += vi4x1 * vk4x1;
142
143 const int32_t vi5x0 = (int32_t) (uint32_t) i5[0];
144 const int32_t vi5x1 = (int32_t) (uint32_t) i5[1];
145 i5 += 2;
146
147 const int32_t vk5x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10] - vkernel_zero_point;
148 const int32_t vk5x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11] - vkernel_zero_point;
149
150 vacc0 += vi5x0 * vk5x0;
151 vacc1 += vi5x1 * vk5x1;
152
153 const int32_t vi6x0 = (int32_t) (uint32_t) i6[0];
154 const int32_t vi6x1 = (int32_t) (uint32_t) i6[1];
155 i6 += 2;
156
157 const int32_t vk6x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12] - vkernel_zero_point;
158 const int32_t vk6x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13] - vkernel_zero_point;
159
160 vacc0 += vi6x0 * vk6x0;
161 vacc1 += vi6x1 * vk6x1;
162
163 const int32_t vi7x0 = (int32_t) (uint32_t) i7[0];
164 const int32_t vi7x1 = (int32_t) (uint32_t) i7[1];
165 i7 += 2;
166
167 const int32_t vk7x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14] - vkernel_zero_point;
168 const int32_t vk7x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15] - vkernel_zero_point;
169
170 vacc0 += vi7x0 * vk7x0;
171 vacc1 += vi7x1 * vk7x1;
172
173 const int32_t vi8x0 = (int32_t) (uint32_t) i8[0];
174 const int32_t vi8x1 = (int32_t) (uint32_t) i8[1];
175 i8 += 2;
176
177 const int32_t vk8x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16] - vkernel_zero_point;
178 const int32_t vk8x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17] - vkernel_zero_point;
179
180 vacc0 += vi8x0 * vk8x0;
181 vacc1 += vi8x1 * vk8x1;
182
183 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 18 * sizeof(uint8_t));
184
185 float vfpacc0 = (float) vacc0;
186 float vfpacc1 = (float) vacc1;
187
188 vfpacc0 *= vscale;
189 vfpacc1 *= vscale;
190
191 vfpacc0 += vmagic_bias;
192 vfpacc1 += vmagic_bias;
193
194 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0);
195 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1);
196
197 vout0 = math_max_s32(vout0, vmagic_min);
198 vout1 = math_max_s32(vout1, vmagic_min);
199
200 vout0 = math_min_s32(vout0, vmagic_max);
201 vout1 = math_min_s32(vout1, vmagic_max);
202
203 vout0 -= vmagic_bias_less_zero_point;
204 vout1 -= vmagic_bias_less_zero_point;
205
206 output[0] = (uint8_t) vout0;
207 output[1] = (uint8_t) vout1;
208 output += 2;
209 }
210 if XNN_UNLIKELY(c != 0) {
211 int32_t vacc = unaligned_load_s32(w);
212
213 const int32_t vi0 = (int32_t) (uint32_t) *i0;
214 const int32_t vk0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0] - vkernel_zero_point;
215 vacc += vi0 * vk0;
216 const int32_t vi1 = (int32_t) (uint32_t) *i1;
217 const int32_t vk1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2] - vkernel_zero_point;
218 vacc += vi1 * vk1;
219 const int32_t vi2 = (int32_t) (uint32_t) *i2;
220 const int32_t vk2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4] - vkernel_zero_point;
221 vacc += vi2 * vk2;
222 const int32_t vi3 = (int32_t) (uint32_t) *i3;
223 const int32_t vk3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6] - vkernel_zero_point;
224 vacc += vi3 * vk3;
225 const int32_t vi4 = (int32_t) (uint32_t) *i4;
226 const int32_t vk4 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8] - vkernel_zero_point;
227 vacc += vi4 * vk4;
228 const int32_t vi5 = (int32_t) (uint32_t) *i5;
229 const int32_t vk5 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10] - vkernel_zero_point;
230 vacc += vi5 * vk5;
231 const int32_t vi6 = (int32_t) (uint32_t) *i6;
232 const int32_t vk6 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12] - vkernel_zero_point;
233 vacc += vi6 * vk6;
234 const int32_t vi7 = (int32_t) (uint32_t) *i7;
235 const int32_t vk7 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14] - vkernel_zero_point;
236 vacc += vi7 * vk7;
237 const int32_t vi8 = (int32_t) (uint32_t) *i8;
238 const int32_t vk8 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16] - vkernel_zero_point;
239 vacc += vi8 * vk8;
240
241 float vfpacc = (float) vacc * vscale;
242
243 vfpacc += vmagic_bias;
244 int32_t vout = (int32_t) float_as_uint32(vfpacc);
245 vout = math_max_s32(vout, vmagic_min);
246 vout = math_min_s32(vout, vmagic_max);
247 vout -= vmagic_bias_less_zero_point;
248
249 *output++ = (uint8_t) vout;
250 }
251
252 output = (uint8_t*) ((uintptr_t) output + output_increment);
253 } while (--output_width != 0);
254 }
255