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_relu_ukernel_4x8s4__wasmsimd(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_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_relu_ukernel_4x8s4__wasmsimd(
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_relu_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 do {
58 v128_t vacc0x0123 = wasm_v128_load(w);
59 v128_t vacc0x4567 = wasm_v128_load(w + 4);
60 v128_t vacc1x0123 = vacc0x0123;
61 v128_t vacc1x4567 = vacc0x4567;
62 v128_t vacc2x0123 = vacc0x0123;
63 v128_t vacc2x4567 = vacc0x4567;
64 v128_t vacc3x0123 = vacc0x0123;
65 v128_t vacc3x4567 = vacc0x4567;
66 w += 8;
67
68 size_t p = ks;
69 do {
70 const float* restrict a0 = a[0];
71 assert(a0 != NULL);
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const float*) ((uintptr_t) a0 + a_offset);
74 }
75 const float* restrict a1 = a[1];
76 assert(a1 != NULL);
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const float*) ((uintptr_t) a1 + a_offset);
79 }
80 const float* restrict a2 = a[2];
81 assert(a2 != NULL);
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const float*) ((uintptr_t) a2 + a_offset);
84 }
85 const float* restrict a3 = a[3];
86 assert(a3 != NULL);
87 if XNN_UNPREDICTABLE(a3 != zero) {
88 a3 = (const float*) ((uintptr_t) a3 + a_offset);
89 }
90 a += 4;
91
92 size_t k = kc;
93 while (k >= 4 * sizeof(float)) {
94 v128_t va0 = wasm_v128_load(a0);
95 a0 += 4;
96 v128_t va1 = wasm_v128_load(a1);
97 a1 += 4;
98 v128_t va2 = wasm_v128_load(a2);
99 a2 += 4;
100 v128_t va3 = wasm_v128_load(a3);
101 a3 += 4;
102
103
104 const v128_t vb0123c0 = wasm_v128_load(w + 0);
105 const v128_t vb4567c0 = wasm_v128_load(w + 4);
106
107 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
108 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
109 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
110 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
111 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
112 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
113 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
114 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
115
116 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
117 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
118 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
119 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
120
121 const v128_t vb0123c1 = wasm_v128_load(w + 8);
122 const v128_t vb4567c1 = wasm_v128_load(w + 12);
123
124 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
125 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
126 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
127 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
128 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
129 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
130 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
131 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
132
133 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
134 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
135 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
136 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
137
138 const v128_t vb0123c2 = wasm_v128_load(w + 16);
139 const v128_t vb4567c2 = wasm_v128_load(w + 20);
140
141 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
142 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
143 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
144 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
145 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
146 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
147 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
148 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
149
150 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
151 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
152 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
153 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
154
155 const v128_t vb0123c3 = wasm_v128_load(w + 24);
156 const v128_t vb4567c3 = wasm_v128_load(w + 28);
157
158 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
159 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
160 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
161 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
162 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
163 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
164 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
165 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
166
167
168 w += 32;
169 k -= 4 * sizeof(float);
170 }
171 if XNN_UNLIKELY(k != 0) {
172 v128_t va0 = wasm_v128_load(a0);
173 a0 = (const float*) ((uintptr_t) a0 + k);
174 v128_t va1 = wasm_v128_load(a1);
175 a1 = (const float*) ((uintptr_t) a1 + k);
176 v128_t va2 = wasm_v128_load(a2);
177 a2 = (const float*) ((uintptr_t) a2 + k);
178 v128_t va3 = wasm_v128_load(a3);
179 a3 = (const float*) ((uintptr_t) a3 + k);
180
181 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
182
183 const v128_t vb0123c0 = wasm_v128_load(w + 0);
184 const v128_t vb4567c0 = wasm_v128_load(w + 4);
185
186 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
187 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
188 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
189 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
190 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
191 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
192 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
193 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
194
195 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
196 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
197 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
198 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
199
200 const v128_t vb0123c1 = wasm_v128_load(w + 8);
201 const v128_t vb4567c1 = wasm_v128_load(w + 12);
202
203 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
204 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
205 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
206 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
207 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
208 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
209 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
210 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
211
212 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
213 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
214 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
215 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
216
217 const v128_t vb0123c2 = wasm_v128_load(w + 16);
218 const v128_t vb4567c2 = wasm_v128_load(w + 20);
219
220 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
221 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
222 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
223 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
224 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
225 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
226 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
227 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
228
229 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
230 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
231 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
232 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
233
234 const v128_t vb0123c3 = wasm_v128_load(w + 24);
235 const v128_t vb4567c3 = wasm_v128_load(w + 28);
236
237 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
238 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
239 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
240 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
241 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
242 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
243 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
244 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
245
246
247 w += 32;
248 }
249 p -= 4 * sizeof(void*);
250 } while (p != 0);
251
252 const v128_t vzero = wasm_i32x4_const_splat(0);
253 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
254 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
255 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
256 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
257 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
258 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
259 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
260 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
261
262 if XNN_LIKELY(nc >= 8) {
263 wasm_v128_store(c3, vacc3x0123);
264 wasm_v128_store(c3 + 4, vacc3x4567);
265 c3 = (float*) ((uintptr_t) c3 + cn_stride);
266 wasm_v128_store(c2, vacc2x0123);
267 wasm_v128_store(c2 + 4, vacc2x4567);
268 c2 = (float*) ((uintptr_t) c2 + cn_stride);
269 wasm_v128_store(c1, vacc1x0123);
270 wasm_v128_store(c1 + 4, vacc1x4567);
271 c1 = (float*) ((uintptr_t) c1 + cn_stride);
272 wasm_v128_store(c0, vacc0x0123);
273 wasm_v128_store(c0 + 4, vacc0x4567);
274 c0 = (float*) ((uintptr_t) c0 + cn_stride);
275
276 a = (const float**restrict) ((uintptr_t) a - ks);
277 nc -= 8;
278 } else {
279 if (nc & 4) {
280 wasm_v128_store(c3, vacc3x0123);
281 wasm_v128_store(c2, vacc2x0123);
282 wasm_v128_store(c1, vacc1x0123);
283 wasm_v128_store(c0, vacc0x0123);
284
285 vacc3x0123 = vacc3x4567;
286 vacc2x0123 = vacc2x4567;
287 vacc1x0123 = vacc1x4567;
288 vacc0x0123 = vacc0x4567;
289
290 c3 += 4;
291 c2 += 4;
292 c1 += 4;
293 c0 += 4;
294 }
295 if (nc & 2) {
296 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
297 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
298 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
299 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
300
301 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
302 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
303 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
304 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
305
306 c3 += 2;
307 c2 += 2;
308 c1 += 2;
309 c0 += 2;
310 }
311 if (nc & 1) {
312 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
313 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
314 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
315 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
316 }
317
318 nc = 0;
319 }
320 } while (nc != 0);
321 }
322