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_up16x9__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_up16x9__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 >= 16; c -= 16) {
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 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
87 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
88
89
90 const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
91 const v128_t vk0x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(uint8_t)));
92 const v128_t vi0x89ABCDEF = wasm_u16x8_load8x8(i0 + 8);
93 const v128_t vk0x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(uint8_t)));
94 i0 += 16;
95
96 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
97 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
98
99 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
100 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
101 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
102 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
103
104 const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
105 const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(uint8_t)));
106 const v128_t vi1x89ABCDEF = wasm_u16x8_load8x8(i1 + 8);
107 const v128_t vk1x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(uint8_t)));
108 v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
109 v128_t vsumx89ABCDEF = wasm_i16x8_add(vi0x89ABCDEF, vi1x89ABCDEF);
110 i1 += 16;
111
112 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
113 vprod89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF);
114
115 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
116 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
117 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
118 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
119
120 const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
121 const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(uint8_t)));
122 const v128_t vi2x89ABCDEF = wasm_u16x8_load8x8(i2 + 8);
123 const v128_t vk2x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(uint8_t)));
124 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
125 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi2x89ABCDEF);
126 i2 += 16;
127
128 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
129 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
130
131 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
132 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
133 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
134 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
135
136 const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
137 const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(uint8_t)));
138 const v128_t vi3x89ABCDEF = wasm_u16x8_load8x8(i3 + 8);
139 const v128_t vk3x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(uint8_t)));
140 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
141 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi3x89ABCDEF);
142 i3 += 16;
143
144 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
145 vprod89ABCDEF = wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF);
146
147 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
148 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
149 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
150 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
151
152 const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
153 const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(uint8_t)));
154 const v128_t vi4x89ABCDEF = wasm_u16x8_load8x8(i4 + 8);
155 const v128_t vk4x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(uint8_t)));
156 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
157 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi4x89ABCDEF);
158 i4 += 16;
159
160 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
161 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
162
163 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
164 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
165 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
166 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
167
168 const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
169 const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(uint8_t)));
170 const v128_t vi5x89ABCDEF = wasm_u16x8_load8x8(i5 + 8);
171 const v128_t vk5x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(uint8_t)));
172 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
173 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi5x89ABCDEF);
174 i5 += 16;
175
176 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
177 vprod89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF);
178
179 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
180 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
181 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
182 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
183
184 const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
185 const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(uint8_t)));
186 const v128_t vi6x89ABCDEF = wasm_u16x8_load8x8(i6 + 8);
187 const v128_t vk6x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(uint8_t)));
188 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
189 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi6x89ABCDEF);
190 i6 += 16;
191
192 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
193 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
194
195 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
196 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
197 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
198 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
199
200 const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
201 const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(uint8_t)));
202 const v128_t vi7x89ABCDEF = wasm_u16x8_load8x8(i7 + 8);
203 const v128_t vk7x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(uint8_t)));
204 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
205 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi7x89ABCDEF);
206 i7 += 16;
207
208 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
209 vprod89ABCDEF = wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF);
210
211 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
212 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
213 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
214 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
215
216 const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
217 const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(uint8_t)));
218 const v128_t vi8x89ABCDEF = wasm_u16x8_load8x8(i8 + 8);
219 const v128_t vk8x89ABCDEF = wasm_u16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(uint8_t)));
220 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
221 vsumx89ABCDEF = wasm_i16x8_add(vsumx89ABCDEF, vi8x89ABCDEF);
222 i8 += 16;
223
224 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
225 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
226
227 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
228 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
229 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_u32x4_extend_low_u16x8(vprod89ABCDEF));
230 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_u32x4_extend_high_u16x8(vprod89ABCDEF));
231
232 vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
233 vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
234 vacc89AB = wasm_i32x4_sub(vacc89AB, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx89ABCDEF), vkernel_zero_point));
235 vaccCDEF = wasm_i32x4_sub(vaccCDEF, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx89ABCDEF), vkernel_zero_point));
236
237 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(uint8_t));
238
239 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
240 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
241 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
242 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
243
244 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
245 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
246 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
247 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
248 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
249
250 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
251 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
252 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
253 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
254 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
255
256 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
257 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
258 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
259 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
260 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
261
262 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
263 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
264 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
265 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
266 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
267
268 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
269 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
270
271 v128_t vout0123456789ABCDEF = wasm_u8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
272
273 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
274 vout0123456789ABCDEF = wasm_u8x16_min(vout0123456789ABCDEF, voutput_max);
275
276 wasm_v128_store(output, vout0123456789ABCDEF);
277 output += 16;
278 }
279 if XNN_UNLIKELY(c != 0) {
280 const uint8_t* k = (const uint8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
281 do {
282 v128_t vacc0123 = wasm_v128_load(w);
283 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
284
285
286 const v128_t vi0x01234567 = wasm_u16x8_load8x8(i0);
287 const v128_t vk0x01234567 = wasm_u16x8_load8x8(k);
288 i0 += 8;
289
290 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
291
292 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
293 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
294
295 const v128_t vi1x01234567 = wasm_u16x8_load8x8(i1);
296 const v128_t vk1x01234567 = wasm_u16x8_load8x8((const void*) (k + 16));
297 v128_t vsumx01234567 = wasm_i16x8_add(vi0x01234567, vi1x01234567);
298 i1 += 8;
299
300 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
301
302 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
303 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
304
305 const v128_t vi2x01234567 = wasm_u16x8_load8x8(i2);
306 const v128_t vk2x01234567 = wasm_u16x8_load8x8((const void*) (k + 32));
307 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi2x01234567);
308 i2 += 8;
309
310 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
311
312 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
313 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
314
315 const v128_t vi3x01234567 = wasm_u16x8_load8x8(i3);
316 const v128_t vk3x01234567 = wasm_u16x8_load8x8((const void*) (k + 48));
317 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi3x01234567);
318 i3 += 8;
319
320 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
321
322 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
323 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
324
325 const v128_t vi4x01234567 = wasm_u16x8_load8x8(i4);
326 const v128_t vk4x01234567 = wasm_u16x8_load8x8((const void*) (k + 64));
327 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi4x01234567);
328 i4 += 8;
329
330 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
331
332 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
333 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
334
335 const v128_t vi5x01234567 = wasm_u16x8_load8x8(i5);
336 const v128_t vk5x01234567 = wasm_u16x8_load8x8((const void*) (k + 80));
337 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi5x01234567);
338 i5 += 8;
339
340 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
341
342 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
343 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
344
345 const v128_t vi6x01234567 = wasm_u16x8_load8x8(i6);
346 const v128_t vk6x01234567 = wasm_u16x8_load8x8((const void*) (k + 96));
347 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi6x01234567);
348 i6 += 8;
349
350 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
351
352 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
353 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
354
355 const v128_t vi7x01234567 = wasm_u16x8_load8x8(i7);
356 const v128_t vk7x01234567 = wasm_u16x8_load8x8((const void*) (k + 112));
357 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi7x01234567);
358 i7 += 8;
359
360 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
361
362 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
363 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
364
365 const v128_t vi8x01234567 = wasm_u16x8_load8x8(i8);
366 const v128_t vk8x01234567 = wasm_u16x8_load8x8((const void*) (k + 128));
367 vsumx01234567 = wasm_i16x8_add(vsumx01234567, vi8x01234567);
368 i8 += 8;
369
370 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
371
372 vacc0123 = wasm_i32x4_add(vacc0123, wasm_u32x4_extend_low_u16x8(vprod01234567));
373 vacc4567 = wasm_i32x4_add(vacc4567, wasm_u32x4_extend_high_u16x8(vprod01234567));
374
375 k += 8;
376
377 vacc0123 = wasm_i32x4_sub(vacc0123, wasm_i32x4_mul(wasm_u32x4_extend_low_u16x8(vsumx01234567), vkernel_zero_point));
378 vacc4567 = wasm_i32x4_sub(vacc4567, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(vsumx01234567), vkernel_zero_point));
379
380 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
381 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
382
383 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
384 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
385 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
386
387 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
388 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
389 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
390
391 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
392 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
393 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
394
395 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
396 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
397 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
398
399 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
400 v128_t vout0123456701234567 = wasm_u8x16_narrow_i16x8(vout01234567, vout01234567);
401
402 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
403 vout0123456701234567 = wasm_u8x16_min(vout0123456701234567, voutput_max);
404
405 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
406
407 if XNN_LIKELY(c >= 8) {
408 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
409 output += 8;
410 c -= 8;
411 } else {
412 if (c & 4) {
413 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
414 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
415 output += 4;
416 }
417 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
418 if (c & 2) {
419 *((uint16_t*) output) = (uint16_t) vout0123;
420 vout0123 >>= 16;
421 output += 2;
422 }
423 if (c & 1) {
424 *output = (uint8_t) vout0123;
425 output += 1;
426 }
427 c = 0;
428 }
429 } while (c != 0);
430 }
431
432 output = (uint8_t*) ((uintptr_t) output + output_increment);
433 } while (--output_width != 0);
434 }
435