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_qu8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16(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_up8x9__wasmsimd_mul16(
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)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const v128_t vkernel_zero_point = wasm_u32x4_load16x4(params->fp32_wasmsimd.kernel_zero_point);
33 do {
34 const uint8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const uint8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const uint8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 const uint8_t* i3 = input[3];
50 assert(i3 != NULL);
51 if XNN_UNPREDICTABLE(i3 != zero) {
52 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
53 }
54 const uint8_t* i4 = input[4];
55 assert(i4 != NULL);
56 if XNN_UNPREDICTABLE(i4 != zero) {
57 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
58 }
59 const uint8_t* i5 = input[5];
60 assert(i5 != NULL);
61 if XNN_UNPREDICTABLE(i5 != zero) {
62 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
63 }
64 const uint8_t* i6 = input[6];
65 assert(i6 != NULL);
66 if XNN_UNPREDICTABLE(i6 != zero) {
67 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
68 }
69 const uint8_t* i7 = input[7];
70 assert(i7 != NULL);
71 if XNN_UNPREDICTABLE(i7 != zero) {
72 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
73 }
74 const uint8_t* i8 = input[8];
75 assert(i8 != NULL);
76 if XNN_UNPREDICTABLE(i8 != zero) {
77 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
78 }
79 input = (const uint8_t**) ((uintptr_t) input + input_stride);
80
81 size_t c = channels;
82 const void* w = weights;
83 for (; c >= 8; c -= 8) {
84 v128_t vacc0123 = wasm_v128_load(w);
85 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
86
87
88 const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
89 const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
90 i0 += 8;
91
92 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
93
94 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
95 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
96
97 const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
98 const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
99 v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
100 i1 += 8;
101
102 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
103
104 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
105 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
106
107 const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
108 const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
109 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
110 i2 += 8;
111
112 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
113
114 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
115 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
116
117 const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
118 const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
119 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
120 i3 += 8;
121
122 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
123
124 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
125 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
126
127 const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
128 const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
129 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
130 i4 += 8;
131
132 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
133
134 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
135 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
136
137 const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
138 const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
139 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
140 i5 += 8;
141
142 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
143
144 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
145 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
146
147 const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
148 const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
149 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
150 i6 += 8;
151
152 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
153
154 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
155 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
156
157 const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
158 const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
159 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
160 i7 += 8;
161
162 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
163
164 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
165 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
166
167 const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
168 const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
169 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
170 i8 += 8;
171
172 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
173
174 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
175 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
176
177 vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
178 vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
179
180 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t));
181
182 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
183 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
184
185 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
186 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
187 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
188
189 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
190 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
191 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
192
193 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
194 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
195 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
196
197 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
198 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
199 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
200
201 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
202
203 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
204
205 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
206 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
207
208 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
209 output += 8;
210 }
211 if XNN_UNLIKELY(c != 0) {
212 {
213 v128_t vacc0123 = wasm_v128_load(w);
214 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
215
216
217 const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
218 const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
219
220 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
221
222 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
223 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
224
225 const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
226 const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
227 v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
228
229 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
230
231 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
232 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
233
234 const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
235 const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
236 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
237
238 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
239
240 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
241 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
242
243 const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
244 const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
245 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
246
247 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
248
249 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
250 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
251
252 const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
253 const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
254 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
255
256 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
257
258 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
259 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
260
261 const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
262 const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
263 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
264
265 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
266
267 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
268 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
269
270 const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
271 const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
272 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
273
274 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
275
276 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
277 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
278
279 const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
280 const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
281 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
282
283 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
284
285 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
286 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
287
288 const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
289 const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
290 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
291
292 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
293
294 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
295 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
296
297
298 vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
299 vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
300
301 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
302 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
303
304 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
305 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
306 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
307
308 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
309 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
310 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
311
312 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
313 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
314 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
315
316 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
317 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
318 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
319
320 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
321 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
322
323 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
324 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
325
326
327 if (c & 4) {
328 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
329 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
330 output += 4;
331 }
332 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
333 if (c & 2) {
334 *((uint16_t*) output) = (uint16_t) vout0123;
335 vout0123 >>= 16;
336 output += 2;
337 }
338 if (c & 1) {
339 *output = (uint8_t) vout0123;
340 output += 1;
341 }
342 }
343 }
344
345 output = (uint8_t*) ((uintptr_t) output + output_increment);
346 } while (--output_width != 0);
347 }
348