1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/wasmsimd-s4.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/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_4x8s4__wasmrelaxedsimd(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_4x8s4__wasmrelaxedsimd(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
58 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
59 do {
60 v128_t vacc0x0123 = wasm_v128_load(w);
61 v128_t vacc0x4567 = wasm_v128_load(w + 4);
62 v128_t vacc1x0123 = vacc0x0123;
63 v128_t vacc1x4567 = vacc0x4567;
64 v128_t vacc2x0123 = vacc0x0123;
65 v128_t vacc2x4567 = vacc0x4567;
66 v128_t vacc3x0123 = vacc0x0123;
67 v128_t vacc3x4567 = vacc0x4567;
68 w += 8;
69
70 size_t p = ks;
71 do {
72 const float* restrict a0 = a[0];
73 assert(a0 != NULL);
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const float*) ((uintptr_t) a0 + a_offset);
76 }
77 const float* restrict a1 = a[1];
78 assert(a1 != NULL);
79 if XNN_UNPREDICTABLE(a1 != zero) {
80 a1 = (const float*) ((uintptr_t) a1 + a_offset);
81 }
82 const float* restrict a2 = a[2];
83 assert(a2 != NULL);
84 if XNN_UNPREDICTABLE(a2 != zero) {
85 a2 = (const float*) ((uintptr_t) a2 + a_offset);
86 }
87 const float* restrict a3 = a[3];
88 assert(a3 != NULL);
89 if XNN_UNPREDICTABLE(a3 != zero) {
90 a3 = (const float*) ((uintptr_t) a3 + a_offset);
91 }
92 a += 4;
93
94 size_t k = kc;
95 while (k >= 4 * sizeof(float)) {
96 v128_t va0 = wasm_v128_load(a0);
97 a0 += 4;
98 v128_t va1 = wasm_v128_load(a1);
99 a1 += 4;
100 v128_t va2 = wasm_v128_load(a2);
101 a2 += 4;
102 v128_t va3 = wasm_v128_load(a3);
103 a3 += 4;
104
105
106 const v128_t vb0123c0 = wasm_v128_load(w + 0);
107 const v128_t vb4567c0 = wasm_v128_load(w + 4);
108
109 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
110 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
111 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
112 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
113 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
114 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
115 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
116 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
117
118 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
119 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
120 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
121 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
122
123 const v128_t vb0123c1 = wasm_v128_load(w + 8);
124 const v128_t vb4567c1 = wasm_v128_load(w + 12);
125
126 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
127 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
128 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
129 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
130 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
131 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
132 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
133 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
134
135 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
136 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
137 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
138 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
139
140 const v128_t vb0123c2 = wasm_v128_load(w + 16);
141 const v128_t vb4567c2 = wasm_v128_load(w + 20);
142
143 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
144 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
145 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
146 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
147 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
148 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
149 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
150 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
151
152 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
153 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
154 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
155 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
156
157 const v128_t vb0123c3 = wasm_v128_load(w + 24);
158 const v128_t vb4567c3 = wasm_v128_load(w + 28);
159
160 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
161 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
162 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
163 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
164 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
165 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
166 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
167 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
168
169
170 w += 32;
171 k -= 4 * sizeof(float);
172 }
173 if XNN_UNLIKELY(k != 0) {
174 v128_t va0 = wasm_v128_load(a0);
175 a0 = (const float*) ((uintptr_t) a0 + k);
176 v128_t va1 = wasm_v128_load(a1);
177 a1 = (const float*) ((uintptr_t) a1 + k);
178 v128_t va2 = wasm_v128_load(a2);
179 a2 = (const float*) ((uintptr_t) a2 + k);
180 v128_t va3 = wasm_v128_load(a3);
181 a3 = (const float*) ((uintptr_t) a3 + k);
182
183 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
184
185 const v128_t vb0123c0 = wasm_v128_load(w + 0);
186 const v128_t vb4567c0 = wasm_v128_load(w + 4);
187
188 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
189 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
190 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
191 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
192 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
193 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
194 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
195 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
196
197 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
198 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
199 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
200 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
201
202 const v128_t vb0123c1 = wasm_v128_load(w + 8);
203 const v128_t vb4567c1 = wasm_v128_load(w + 12);
204
205 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
206 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
207 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
208 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
209 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
210 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
211 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
212 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
213
214 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
215 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
216 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
217 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
218
219 const v128_t vb0123c2 = wasm_v128_load(w + 16);
220 const v128_t vb4567c2 = wasm_v128_load(w + 20);
221
222 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
223 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
224 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
225 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
226 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
227 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
228 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
229 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
230
231 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
232 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
233 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
234 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
235
236 const v128_t vb0123c3 = wasm_v128_load(w + 24);
237 const v128_t vb4567c3 = wasm_v128_load(w + 28);
238
239 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
240 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
241 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
242 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
243 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
244 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
245 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
246 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
247
248
249 w += 32;
250 }
251 p -= 4 * sizeof(void*);
252 } while (p != 0);
253
254 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
255 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
256 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
257 vacc3x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x0123);
258 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
259 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
260 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
261 vacc3x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x4567);
262
263 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
264 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
265 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
266 vacc3x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x0123);
267 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
268 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
269 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
270 vacc3x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x4567);
271
272 if XNN_LIKELY(nc >= 8) {
273 wasm_v128_store(c3, vacc3x0123);
274 wasm_v128_store(c3 + 4, vacc3x4567);
275 c3 = (float*) ((uintptr_t) c3 + cn_stride);
276 wasm_v128_store(c2, vacc2x0123);
277 wasm_v128_store(c2 + 4, vacc2x4567);
278 c2 = (float*) ((uintptr_t) c2 + cn_stride);
279 wasm_v128_store(c1, vacc1x0123);
280 wasm_v128_store(c1 + 4, vacc1x4567);
281 c1 = (float*) ((uintptr_t) c1 + cn_stride);
282 wasm_v128_store(c0, vacc0x0123);
283 wasm_v128_store(c0 + 4, vacc0x4567);
284 c0 = (float*) ((uintptr_t) c0 + cn_stride);
285
286 a = (const float**restrict) ((uintptr_t) a - ks);
287 nc -= 8;
288 } else {
289 if (nc & 4) {
290 wasm_v128_store(c3, vacc3x0123);
291 wasm_v128_store(c2, vacc2x0123);
292 wasm_v128_store(c1, vacc1x0123);
293 wasm_v128_store(c0, vacc0x0123);
294
295 vacc3x0123 = vacc3x4567;
296 vacc2x0123 = vacc2x4567;
297 vacc1x0123 = vacc1x4567;
298 vacc0x0123 = vacc0x4567;
299
300 c3 += 4;
301 c2 += 4;
302 c1 += 4;
303 c0 += 4;
304 }
305 if (nc & 2) {
306 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
307 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
308 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
309 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
310
311 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
312 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
313 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
314 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
315
316 c3 += 2;
317 c2 += 2;
318 c1 += 2;
319 c0 += 2;
320 }
321 if (nc & 1) {
322 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
323 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
324 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
325 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
326 }
327
328 nc = 0;
329 }
330 } while (nc != 0);
331 }
332