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_up16x9__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_up16x9__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 >= 16; c -= 16) {
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 v128_t vacc89AB = wasm_v128_load((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
86 v128_t vaccCDEF = wasm_v128_load((const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
87
88
89 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
90 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
91 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
92 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
93 i0 += 16;
94
95 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
96 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
97
98 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
99 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
100 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
101 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
102
103 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
104 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
105 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
106 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
107 i1 += 16;
108
109 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
110 vprod89ABCDEF = wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF);
111
112 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
113 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
114 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
115 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
116
117 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
118 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
119 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
120 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
121 i2 += 16;
122
123 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
124 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
125
126 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
127 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
128 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
129 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
130
131 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
132 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
133 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
134 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)));
135 i3 += 16;
136
137 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
138 vprod89ABCDEF = wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF);
139
140 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
141 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
142 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
143 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
144
145 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
146 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)));
147 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
148 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)));
149 i4 += 16;
150
151 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
152 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
153
154 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
155 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
156 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
157 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
158
159 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
160 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)));
161 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
162 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)));
163 i5 += 16;
164
165 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
166 vprod89ABCDEF = wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF);
167
168 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
169 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
170 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
171 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
172
173 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
174 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)));
175 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
176 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)));
177 i6 += 16;
178
179 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
180 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
181
182 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
183 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
184 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
185 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
186
187 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
188 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)));
189 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
190 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)));
191 i7 += 16;
192
193 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
194 vprod89ABCDEF = wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF);
195
196 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
197 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
198 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
199 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
200
201 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
202 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)));
203 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
204 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)));
205 i8 += 16;
206
207 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
208 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
209
210 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
211 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
212 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
213 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
214
215
216 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
217
218 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
219 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
220 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
221 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
222
223 const v128_t vscale0123 = wasm_v128_load(w);
224 const v128_t vscale4567 = wasm_v128_load((const float*) w + 4);
225 const v128_t vscale89AB = wasm_v128_load((const float*) w + 8);
226 const v128_t vscaleCDEF = wasm_v128_load((const float*) w + 12);
227 w = (const void*) ((const float*) w + 16);
228
229 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
230 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
231 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale89AB);
232 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscaleCDEF);
233
234 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
235 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
236 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
237 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
238 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
239
240 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
241 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
242 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
243 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
244 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
245
246 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
247 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
248 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
249 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
250 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
251
252 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
253 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
254
255 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
256
257 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
258 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
259
260 wasm_v128_store(output, vout0123456789ABCDEF);
261 output += 16;
262 }
263 if XNN_UNLIKELY(c != 0) {
264 const int8_t* k = (const int8_t*) ((uintptr_t) w + 16 * sizeof(int32_t));
265 do {
266 v128_t vacc0123 = wasm_v128_load(w);
267 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
268
269
270 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
271 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
272 i0 += 8;
273
274 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
275
276 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
277 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
278
279 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
280 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 16));
281 i1 += 8;
282
283 vprod01234567 = wasm_i16x8_mul(vi1x01234567, vk1x01234567);
284
285 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
286 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
287
288 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
289 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 32));
290 i2 += 8;
291
292 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
293
294 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
295 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
296
297 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
298 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
299 i3 += 8;
300
301 vprod01234567 = wasm_i16x8_mul(vi3x01234567, vk3x01234567);
302
303 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
304 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
305
306 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
307 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 64));
308 i4 += 8;
309
310 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
311
312 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
313 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
314
315 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
316 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 80));
317 i5 += 8;
318
319 vprod01234567 = wasm_i16x8_mul(vi5x01234567, vk5x01234567);
320
321 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
322 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
323
324 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
325 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
326 i6 += 8;
327
328 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
329
330 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
331 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
332
333 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
334 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 112));
335 i7 += 8;
336
337 vprod01234567 = wasm_i16x8_mul(vi7x01234567, vk7x01234567);
338
339 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
340 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
341
342 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
343 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 128));
344 i8 += 8;
345
346 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
347
348 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
349 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
350
351 k += 8;
352
353
354 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
355 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
356
357 const v128_t vscale0123 = wasm_v128_load((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
358 const v128_t vscale4567 = wasm_v128_load((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t) + 4 * sizeof(float)));
359
360 vacc0123 = wasm_f32x4_mul(vacc0123, vscale0123);
361 vacc4567 = wasm_f32x4_mul(vacc4567, vscale4567);
362
363 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
364 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
365 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
366
367 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
368 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
369 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
370
371 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
372 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
373 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
374
375 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
376 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
377
378 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
379 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
380
381 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
382
383 if XNN_LIKELY(c >= 8) {
384 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
385 output += 8;
386 c -= 8;
387 } else {
388 if (c & 4) {
389 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
390 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
391 output += 4;
392 }
393 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
394 if (c & 2) {
395 *((uint16_t*) output) = (uint16_t) vout0123;
396 vout0123 >>= 16;
397 output += 2;
398 }
399 if (c & 1) {
400 *output = (int8_t) vout0123;
401 output += 1;
402 }
403 c = 0;
404 }
405 } while (c != 0);
406 }
407
408 output = (int8_t*) ((uintptr_t) output + output_increment);
409 } while (--output_width != 0);
410 }
411