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_up24x9__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_up24x9__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 >= 24; c -= 24) {
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 v128_t vaccGHIJ = wasm_v128_load((const void*) ((uintptr_t) w + 16 * sizeof(int32_t)));
88 v128_t vaccKLMN = wasm_v128_load((const void*) ((uintptr_t) w + 20 * sizeof(int32_t)));
89
90
91 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
92 const v128_t vk0x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 0 * sizeof(int8_t)));
93 const v128_t vi0x89ABCDEF = wasm_i16x8_load8x8(i0 + 8);
94 const v128_t vk0x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 8 * sizeof(int8_t)));
95 const v128_t vi0xGHIJKLMN = wasm_i16x8_load8x8(i0 + 16);
96 const v128_t vk0xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 16 * sizeof(int8_t)));
97 i0 += 24;
98
99 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
100 v128_t vprod89ABCDEF = wasm_i16x8_mul(vi0x89ABCDEF, vk0x89ABCDEF);
101 v128_t vprodGHIJKLMN = wasm_i16x8_mul(vi0xGHIJKLMN, vk0xGHIJKLMN);
102
103
104 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
105 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 24 * sizeof(int8_t)));
106 const v128_t vi1x89ABCDEF = wasm_i16x8_load8x8(i1 + 8);
107 const v128_t vk1x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 32 * sizeof(int8_t)));
108 const v128_t vi1xGHIJKLMN = wasm_i16x8_load8x8(i1 + 16);
109 const v128_t vk1xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 40 * sizeof(int8_t)));
110 i1 += 24;
111
112 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
113 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi1x89ABCDEF, vk1x89ABCDEF));
114 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi1xGHIJKLMN, vk1xGHIJKLMN));
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 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
119 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
120 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
121 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
122
123 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
124 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 48 * sizeof(int8_t)));
125 const v128_t vi2x89ABCDEF = wasm_i16x8_load8x8(i2 + 8);
126 const v128_t vk2x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 56 * sizeof(int8_t)));
127 const v128_t vi2xGHIJKLMN = wasm_i16x8_load8x8(i2 + 16);
128 const v128_t vk2xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 64 * sizeof(int8_t)));
129 i2 += 24;
130
131 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
132 vprod89ABCDEF = wasm_i16x8_mul(vi2x89ABCDEF, vk2x89ABCDEF);
133 vprodGHIJKLMN = wasm_i16x8_mul(vi2xGHIJKLMN, vk2xGHIJKLMN);
134
135
136 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
137 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 72 * sizeof(int8_t)));
138 const v128_t vi3x89ABCDEF = wasm_i16x8_load8x8(i3 + 8);
139 const v128_t vk3x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 80 * sizeof(int8_t)));
140 const v128_t vi3xGHIJKLMN = wasm_i16x8_load8x8(i3 + 16);
141 const v128_t vk3xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 88 * sizeof(int8_t)));
142 i3 += 24;
143
144 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
145 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi3x89ABCDEF, vk3x89ABCDEF));
146 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi3xGHIJKLMN, vk3xGHIJKLMN));
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 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
151 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
152 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
153 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
154
155 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
156 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 96 * sizeof(int8_t)));
157 const v128_t vi4x89ABCDEF = wasm_i16x8_load8x8(i4 + 8);
158 const v128_t vk4x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 104 * sizeof(int8_t)));
159 const v128_t vi4xGHIJKLMN = wasm_i16x8_load8x8(i4 + 16);
160 const v128_t vk4xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 112 * sizeof(int8_t)));
161 i4 += 24;
162
163 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
164 vprod89ABCDEF = wasm_i16x8_mul(vi4x89ABCDEF, vk4x89ABCDEF);
165 vprodGHIJKLMN = wasm_i16x8_mul(vi4xGHIJKLMN, vk4xGHIJKLMN);
166
167
168 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
169 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 120 * sizeof(int8_t)));
170 const v128_t vi5x89ABCDEF = wasm_i16x8_load8x8(i5 + 8);
171 const v128_t vk5x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 128 * sizeof(int8_t)));
172 const v128_t vi5xGHIJKLMN = wasm_i16x8_load8x8(i5 + 16);
173 const v128_t vk5xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 136 * sizeof(int8_t)));
174 i5 += 24;
175
176 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
177 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi5x89ABCDEF, vk5x89ABCDEF));
178 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi5xGHIJKLMN, vk5xGHIJKLMN));
179
180 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
181 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
182 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
183 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
184 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
185 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
186
187 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
188 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 144 * sizeof(int8_t)));
189 const v128_t vi6x89ABCDEF = wasm_i16x8_load8x8(i6 + 8);
190 const v128_t vk6x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 152 * sizeof(int8_t)));
191 const v128_t vi6xGHIJKLMN = wasm_i16x8_load8x8(i6 + 16);
192 const v128_t vk6xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 160 * sizeof(int8_t)));
193 i6 += 24;
194
195 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
196 vprod89ABCDEF = wasm_i16x8_mul(vi6x89ABCDEF, vk6x89ABCDEF);
197 vprodGHIJKLMN = wasm_i16x8_mul(vi6xGHIJKLMN, vk6xGHIJKLMN);
198
199
200 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
201 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 168 * sizeof(int8_t)));
202 const v128_t vi7x89ABCDEF = wasm_i16x8_load8x8(i7 + 8);
203 const v128_t vk7x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 176 * sizeof(int8_t)));
204 const v128_t vi7xGHIJKLMN = wasm_i16x8_load8x8(i7 + 16);
205 const v128_t vk7xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 184 * sizeof(int8_t)));
206 i7 += 24;
207
208 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
209 vprod89ABCDEF = wasm_i16x8_add(vprod89ABCDEF, wasm_i16x8_mul(vi7x89ABCDEF, vk7x89ABCDEF));
210 vprodGHIJKLMN = wasm_i16x8_add(vprodGHIJKLMN, wasm_i16x8_mul(vi7xGHIJKLMN, vk7xGHIJKLMN));
211
212 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
213 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
214 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
215 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
216 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
217 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
218
219 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
220 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 192 * sizeof(int8_t)));
221 const v128_t vi8x89ABCDEF = wasm_i16x8_load8x8(i8 + 8);
222 const v128_t vk8x89ABCDEF = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 200 * sizeof(int8_t)));
223 const v128_t vi8xGHIJKLMN = wasm_i16x8_load8x8(i8 + 16);
224 const v128_t vk8xGHIJKLMN = wasm_i16x8_load8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 208 * sizeof(int8_t)));
225 i8 += 24;
226
227 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
228 vprod89ABCDEF = wasm_i16x8_mul(vi8x89ABCDEF, vk8x89ABCDEF);
229 vprodGHIJKLMN = wasm_i16x8_mul(vi8xGHIJKLMN, vk8xGHIJKLMN);
230
231 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
232 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
233 vacc89AB = wasm_i32x4_add(vacc89AB, wasm_i32x4_extend_low_i16x8(vprod89ABCDEF));
234 vaccCDEF = wasm_i32x4_add(vaccCDEF, wasm_i32x4_extend_high_i16x8(vprod89ABCDEF));
235 vaccGHIJ = wasm_i32x4_add(vaccGHIJ, wasm_i32x4_extend_low_i16x8(vprodGHIJKLMN));
236 vaccKLMN = wasm_i32x4_add(vaccKLMN, wasm_i32x4_extend_high_i16x8(vprodGHIJKLMN));
237
238
239 w = (const void*) ((uintptr_t) w + 24 * sizeof(int32_t) + 216 * sizeof(int8_t));
240
241 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
242 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
243 vacc89AB = wasm_f32x4_convert_i32x4(vacc89AB);
244 vaccCDEF = wasm_f32x4_convert_i32x4(vaccCDEF);
245 vaccGHIJ = wasm_f32x4_convert_i32x4(vaccGHIJ);
246 vaccKLMN = wasm_f32x4_convert_i32x4(vaccKLMN);
247
248 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
249 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
250 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
251 vacc89AB = wasm_f32x4_mul(vacc89AB, vscale);
252 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vscale);
253 vaccGHIJ = wasm_f32x4_mul(vaccGHIJ, vscale);
254 vaccKLMN = wasm_f32x4_mul(vaccKLMN, vscale);
255
256 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
257 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
258 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
259 vacc89AB = wasm_f32x4_add(vacc89AB, vmagic_bias);
260 vaccCDEF = wasm_f32x4_add(vaccCDEF, vmagic_bias);
261 vaccGHIJ = wasm_f32x4_add(vaccGHIJ, vmagic_bias);
262 vaccKLMN = wasm_f32x4_add(vaccKLMN, vmagic_bias);
263
264 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
265 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
266 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
267 vacc89AB = wasm_i32x4_max(vacc89AB, vmagic_min);
268 vaccCDEF = wasm_i32x4_max(vaccCDEF, vmagic_min);
269 vaccGHIJ = wasm_i32x4_max(vaccGHIJ, vmagic_min);
270 vaccKLMN = wasm_i32x4_max(vaccKLMN, vmagic_min);
271
272 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
273 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
274 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
275 vacc89AB = wasm_i32x4_sub(vacc89AB, vmagic_bias_less_output_zero_point);
276 vaccCDEF = wasm_i32x4_sub(vaccCDEF, vmagic_bias_less_output_zero_point);
277 vaccGHIJ = wasm_i32x4_sub(vaccGHIJ, vmagic_bias_less_output_zero_point);
278 vaccKLMN = wasm_i32x4_sub(vaccKLMN, vmagic_bias_less_output_zero_point);
279
280 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
281 v128_t vout89ABCDEF = wasm_i16x8_narrow_i32x4(vacc89AB, vaccCDEF);
282 v128_t voutGHIJKLMN = wasm_i16x8_narrow_i32x4(vaccGHIJ, vaccKLMN);
283
284 v128_t vout0123456789ABCDEF = wasm_i8x16_narrow_i16x8(vout01234567, vout89ABCDEF);
285 v128_t voutGHIJKLMNGHIJKLMN = wasm_i8x16_narrow_i16x8(voutGHIJKLMN, voutGHIJKLMN);
286
287 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
288 vout0123456789ABCDEF = wasm_i8x16_min(vout0123456789ABCDEF, voutput_max);
289 voutGHIJKLMNGHIJKLMN = wasm_i8x16_min(voutGHIJKLMNGHIJKLMN, voutput_max);
290
291 wasm_v128_store(output, vout0123456789ABCDEF);
292 *((double*) (output + 16)) = wasm_f64x2_extract_lane(voutGHIJKLMNGHIJKLMN, 0);
293 output += 24;
294 }
295 if XNN_UNLIKELY(c != 0) {
296 const int8_t* k = (const int8_t*) ((uintptr_t) w + 24 * sizeof(int32_t));
297 do {
298 v128_t vacc0123 = wasm_v128_load(w);
299 v128_t vacc4567 = wasm_v128_load((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
300
301
302 const v128_t vi0x01234567 = wasm_i16x8_load8x8(i0);
303 const v128_t vk0x01234567 = wasm_i16x8_load8x8(k);
304 i0 += 8;
305
306 v128_t vprod01234567 = wasm_i16x8_mul(vi0x01234567, vk0x01234567);
307
308
309 const v128_t vi1x01234567 = wasm_i16x8_load8x8(i1);
310 const v128_t vk1x01234567 = wasm_i16x8_load8x8((const void*) (k + 24));
311 i1 += 8;
312
313 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi1x01234567, vk1x01234567));
314
315 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
316 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
317
318 const v128_t vi2x01234567 = wasm_i16x8_load8x8(i2);
319 const v128_t vk2x01234567 = wasm_i16x8_load8x8((const void*) (k + 48));
320 i2 += 8;
321
322 vprod01234567 = wasm_i16x8_mul(vi2x01234567, vk2x01234567);
323
324
325 const v128_t vi3x01234567 = wasm_i16x8_load8x8(i3);
326 const v128_t vk3x01234567 = wasm_i16x8_load8x8((const void*) (k + 72));
327 i3 += 8;
328
329 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi3x01234567, vk3x01234567));
330
331 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
332 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
333
334 const v128_t vi4x01234567 = wasm_i16x8_load8x8(i4);
335 const v128_t vk4x01234567 = wasm_i16x8_load8x8((const void*) (k + 96));
336 i4 += 8;
337
338 vprod01234567 = wasm_i16x8_mul(vi4x01234567, vk4x01234567);
339
340
341 const v128_t vi5x01234567 = wasm_i16x8_load8x8(i5);
342 const v128_t vk5x01234567 = wasm_i16x8_load8x8((const void*) (k + 120));
343 i5 += 8;
344
345 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi5x01234567, vk5x01234567));
346
347 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
348 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
349
350 const v128_t vi6x01234567 = wasm_i16x8_load8x8(i6);
351 const v128_t vk6x01234567 = wasm_i16x8_load8x8((const void*) (k + 144));
352 i6 += 8;
353
354 vprod01234567 = wasm_i16x8_mul(vi6x01234567, vk6x01234567);
355
356
357 const v128_t vi7x01234567 = wasm_i16x8_load8x8(i7);
358 const v128_t vk7x01234567 = wasm_i16x8_load8x8((const void*) (k + 168));
359 i7 += 8;
360
361 vprod01234567 = wasm_i16x8_add(vprod01234567, wasm_i16x8_mul(vi7x01234567, vk7x01234567));
362
363 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
364 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
365
366 const v128_t vi8x01234567 = wasm_i16x8_load8x8(i8);
367 const v128_t vk8x01234567 = wasm_i16x8_load8x8((const void*) (k + 192));
368 i8 += 8;
369
370 vprod01234567 = wasm_i16x8_mul(vi8x01234567, vk8x01234567);
371
372 vacc0123 = wasm_i32x4_add(vacc0123, wasm_i32x4_extend_low_i16x8(vprod01234567));
373 vacc4567 = wasm_i32x4_add(vacc4567, wasm_i32x4_extend_high_i16x8(vprod01234567));
374
375 k += 8;
376
377
378 vacc0123 = wasm_f32x4_convert_i32x4(vacc0123);
379 vacc4567 = wasm_f32x4_convert_i32x4(vacc4567);
380
381 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
382 vacc0123 = wasm_f32x4_mul(vacc0123, vscale);
383 vacc4567 = wasm_f32x4_mul(vacc4567, vscale);
384
385 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
386 vacc0123 = wasm_f32x4_add(vacc0123, vmagic_bias);
387 vacc4567 = wasm_f32x4_add(vacc4567, vmagic_bias);
388
389 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
390 vacc0123 = wasm_i32x4_max(vacc0123, vmagic_min);
391 vacc4567 = wasm_i32x4_max(vacc4567, vmagic_min);
392
393 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
394 vacc0123 = wasm_i32x4_sub(vacc0123, vmagic_bias_less_output_zero_point);
395 vacc4567 = wasm_i32x4_sub(vacc4567, vmagic_bias_less_output_zero_point);
396
397 v128_t vout01234567 = wasm_i16x8_narrow_i32x4(vacc0123, vacc4567);
398 v128_t vout0123456701234567 = wasm_i8x16_narrow_i16x8(vout01234567, vout01234567);
399
400 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
401 vout0123456701234567 = wasm_i8x16_min(vout0123456701234567, voutput_max);
402
403 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
404
405 if XNN_LIKELY(c >= 8) {
406 *((double*) output) = wasm_f64x2_extract_lane(vout0123456701234567, 0);
407 output += 8;
408 c -= 8;
409 } else {
410 if (c & 4) {
411 *((float*) output) = wasm_f32x4_extract_lane(vout0123456701234567, 0);
412 vout0123456701234567 = wasm_u64x2_shr(vout0123456701234567, 32);
413 output += 4;
414 }
415 uint32_t vout0123 = wasm_i32x4_extract_lane(vout0123456701234567, 0);
416 if (c & 2) {
417 *((uint16_t*) output) = (uint16_t) vout0123;
418 vout0123 >>= 16;
419 output += 2;
420 }
421 if (c & 1) {
422 *output = (int8_t) vout0123;
423 output += 1;
424 }
425 c = 0;
426 }
427 } while (c != 0);
428 }
429
430 output = (int8_t*) ((uintptr_t) output + output_increment);
431 } while (--output_width != 0);
432 }
433