1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-wasmsimd-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <wasm_simd128.h>
13
14 #include <xnnpack/dwconv.h>
15
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16_add16(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_up8x9__wasmsimd_mul16_add16(
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)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 do {
33 const int8_t* i0 = input[0];
34 assert(i0 != NULL);
35 if XNN_UNPREDICTABLE(i0 != zero) {
36 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37 }
38 const int8_t* i1 = input[1];
39 assert(i1 != NULL);
40 if XNN_UNPREDICTABLE(i1 != zero) {
41 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42 }
43 const int8_t* i2 = input[2];
44 assert(i2 != NULL);
45 if XNN_UNPREDICTABLE(i2 != zero) {
46 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47 }
48 const int8_t* i3 = input[3];
49 assert(i3 != NULL);
50 if XNN_UNPREDICTABLE(i3 != zero) {
51 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52 }
53 const int8_t* i4 = input[4];
54 assert(i4 != NULL);
55 if XNN_UNPREDICTABLE(i4 != zero) {
56 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57 }
58 const int8_t* i5 = input[5];
59 assert(i5 != NULL);
60 if XNN_UNPREDICTABLE(i5 != zero) {
61 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62 }
63 const int8_t* i6 = input[6];
64 assert(i6 != NULL);
65 if XNN_UNPREDICTABLE(i6 != zero) {
66 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67 }
68 const int8_t* i7 = input[7];
69 assert(i7 != NULL);
70 if XNN_UNPREDICTABLE(i7 != zero) {
71 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72 }
73 const int8_t* i8 = input[8];
74 assert(i8 != NULL);
75 if XNN_UNPREDICTABLE(i8 != zero) {
76 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77 }
78 input = (const int8_t**) ((uintptr_t) input + input_stride);
79
80 size_t c = channels;
81 const void* w = weights;
82 for (; c >= 8; c -= 8) {
83 v128_t vacc0123 = wasm_v128_load(w);
84 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
85
86
87 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
88 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
89 i0 += 8;
90
91 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
92
93
94 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
95 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
96 i1 += 8;
97
98 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
99
100 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
101 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
102
103 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
104 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
105 i2 += 8;
106
107 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
108
109
110 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
111 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
112 i3 += 8;
113
114 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
115
116 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
117 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
118
119 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
120 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
121 i4 += 8;
122
123 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
124
125
126 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
127 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
128 i5 += 8;
129
130 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
131
132 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
133 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
134
135 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
136 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
137 i6 += 8;
138
139 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
140
141
142 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
143 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
144 i7 += 8;
145
146 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
147
148 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
149 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
150
151 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
152 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
153 i8 += 8;
154
155 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
156
157 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
158 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
159
160
161 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
162
163 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
164 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
165
166 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
167 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
168 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
169
170 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
171 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
172 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
173
174 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
175 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
176 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
177
178 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
179 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
180 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
181
182 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
183
184 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
185
186 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
187 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
188
189 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
190 output += 8;
191 }
192 if XNN_UNLIKELY(c != 0) {
193 {
194 v128_t vacc0123 = wasm_v128_load(w);
195 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
196
197
198 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
199 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
200
201 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
202
203
204 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
205 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
206
207 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
208
209 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
210 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
211
212 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
213 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
214
215 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
216
217
218 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
219 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
220
221 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
222
223 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
224 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
225
226 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
227 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
228
229 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
230
231
232 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
233 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
234
235 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
236
237 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
238 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
239
240 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
241 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
242
243 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
244
245
246 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
247 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
248
249 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
250
251 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
252 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
253
254 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
255 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
256
257 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
258
259 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
260 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
261
262
263
264 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
265 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
266
267 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
268 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
269 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
270
271 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
272 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
273 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
274
275 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
276 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
277 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
278
279 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
280 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
281 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
282
283 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
284 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
285
286 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
287 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
288
289
290 if (c & 4) {
291 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
292 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
293 output += 4;
294 }
295 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
296 if (c & 2) {
297 *((uint16_t*) output) = (uint16_t) vout0123;
298 vout0123 >>= 16;
299 output += 2;
300 }
301 if (c & 1) {
302 *output = (int8_t) vout0123;
303 output += 1;
304 }
305 }
306 }
307
308 output = (int8_t*) ((uintptr_t) output + output_increment);
309 } while (--output_width != 0);
310 }
311