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_qc8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16(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_up8x9__wasmsimd_mul16(
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)]) 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 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
94 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
95
96 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
97 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
98 i1 += 8;
99
100 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
101
102 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
103 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
104
105 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
106 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
107 i2 += 8;
108
109 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
110
111 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
112 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
113
114 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
115 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
116 i3 += 8;
117
118 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
119
120 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
121 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
122
123 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
124 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
125 i4 += 8;
126
127 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
128
129 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
130 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
131
132 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
133 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
134 i5 += 8;
135
136 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
137
138 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
139 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
140
141 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
142 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
143 i6 += 8;
144
145 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
146
147 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
148 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
149
150 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
151 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
152 i7 += 8;
153
154 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
155
156 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
157 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
158
159 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
160 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
161 i8 += 8;
162
163 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
164
165 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
166 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
167
168
169 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
170
171 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
172 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
173
174 const v128_t vscale0123 = wasm_v128_load(w);
175 const v128_t vscale4567 = wasm_v128_load((const float*) w + 4);
176 w = (const void*) ((const float*) w + 8);
177
178 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
179 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
180
181 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
182 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
183 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
184
185 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
186 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
187 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
188
189 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
190 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
191 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
192
193 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
194
195 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
196
197 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
198 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
199
200 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
201 output += 8;
202 }
203 if XNN_UNLIKELY(c != 0) {
204 {
205 v128_t vacc0123 = wasm_v128_load(w);
206 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
207
208
209 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
210 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
211
212 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
213
214 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
215 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
216
217 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
218 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
219
220 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
221
222 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
223 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
224
225 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
226 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
227
228 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
229
230 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
231 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
232
233 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
234 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
235
236 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
237
238 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
239 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
240
241 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
242 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
243
244 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
245
246 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
247 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
248
249 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
250 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
251
252 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
253
254 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
255 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
256
257 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
258 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
259
260 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
261
262 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
263 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
264
265 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
266 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
267
268 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
269
270 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
271 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
272
273 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
274 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
275
276 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
277
278 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
279 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
280
281
282
283 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
284 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
285
286 const v128_t vscale0123 = wasm_v128_load((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)));
287 const v128_t vscale4567 = wasm_v128_load((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t) + 4 * sizeof(float)));
288
289 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
290 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
291
292 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
293 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
294 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
295
296 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
297 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
298 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
299
300 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
301 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
302 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
303
304 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
305 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
306
307 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
308 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
309
310
311 if (c & 4) {
312 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
313 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
314 output += 4;
315 }
316 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
317 if (c & 2) {
318 *((uint16_t*) output) = (uint16_t) vout0123;
319 vout0123 >>= 16;
320 output += 2;
321 }
322 if (c & 1) {
323 *output = (int8_t) vout0123;
324 output += 1;
325 }
326 }
327 }
328
329 output = (int8_t*) ((uintptr_t) output + output_increment);
330 } while (--output_width != 0);
331 }
332