1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemm_relu_ukernel_4x8s4__wasmsimd(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_relu_ukernel_4x8s4__wasmsimd(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58
59 do {
60 v128_t vacc0x0123 = wasm_v128_load(w + 0);
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 k = kc;
71 while (k >= 4 * sizeof(float)) {
72 v128_t va0 = wasm_v128_load(a0);
73 a0 += 4;
74 v128_t va1 = wasm_v128_load(a1);
75 a1 += 4;
76 v128_t va2 = wasm_v128_load(a2);
77 a2 += 4;
78 v128_t va3 = wasm_v128_load(a3);
79 a3 += 4;
80
81
82 const v128_t vb0123c0 = wasm_v128_load(w + 0);
83 const v128_t vb4567c0 = wasm_v128_load(w + 4);
84
85 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
86 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
87 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
88 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
89 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
90 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
91 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
92 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
93
94 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
95 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
96 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
97 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
98
99 const v128_t vb0123c1 = wasm_v128_load(w + 8);
100 const v128_t vb4567c1 = wasm_v128_load(w + 12);
101
102 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
103 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
104 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
105 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
106 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
107 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
108 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
109 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
110
111 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
112 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
113 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
114 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
115
116 const v128_t vb0123c2 = wasm_v128_load(w + 16);
117 const v128_t vb4567c2 = wasm_v128_load(w + 20);
118
119 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
120 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
121 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
122 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
123 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
124 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
125 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
126 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
127
128 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
129 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
130 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
131 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
132
133 const v128_t vb0123c3 = wasm_v128_load(w + 24);
134 const v128_t vb4567c3 = wasm_v128_load(w + 28);
135
136 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
137 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
138 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
139 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
140 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
141 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
142 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
143 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
144
145
146 w += 32;
147 k -= 4 * sizeof(float);
148 }
149 if XNN_UNLIKELY(k != 0) {
150 v128_t va0 = wasm_v128_load(a0);
151 a0 = (const float*) ((uintptr_t) a0 + k);
152 v128_t va1 = wasm_v128_load(a1);
153 a1 = (const float*) ((uintptr_t) a1 + k);
154 v128_t va2 = wasm_v128_load(a2);
155 a2 = (const float*) ((uintptr_t) a2 + k);
156 v128_t va3 = wasm_v128_load(a3);
157 a3 = (const float*) ((uintptr_t) a3 + k);
158
159 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
160
161 const v128_t vb0123c0 = wasm_v128_load(w + 0);
162 const v128_t vb4567c0 = wasm_v128_load(w + 4);
163
164 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
165 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
166 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
167 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
168 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
169 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
170 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
171 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
172
173 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
174 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
175 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
176 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
177
178 const v128_t vb0123c1 = wasm_v128_load(w + 8);
179 const v128_t vb4567c1 = wasm_v128_load(w + 12);
180
181 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
182 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
183 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
184 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
185 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
186 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
187 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
188 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
189
190 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
191 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
192 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
193 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
194
195 const v128_t vb0123c2 = wasm_v128_load(w + 16);
196 const v128_t vb4567c2 = wasm_v128_load(w + 20);
197
198 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
199 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
200 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
201 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
202 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
203 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
204 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
205 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
206
207 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
208 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
209 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
210 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
211
212 const v128_t vb0123c3 = wasm_v128_load(w + 24);
213 const v128_t vb4567c3 = wasm_v128_load(w + 28);
214
215 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
216 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
217 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
218 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
219 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
220 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
221 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
222 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
223
224
225 w += 32;
226 }
227
228 const v128_t vzero = wasm_i32x4_const_splat(0);
229 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
230 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
231 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
232 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
233 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
234 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
235 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
236 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
237
238 if XNN_LIKELY(nc >= 8) {
239 wasm_v128_store(c3, vacc3x0123);
240 wasm_v128_store(c3 + 4, vacc3x4567);
241 c3 = (float*) ((uintptr_t) c3 + cn_stride);
242 wasm_v128_store(c2, vacc2x0123);
243 wasm_v128_store(c2 + 4, vacc2x4567);
244 c2 = (float*) ((uintptr_t) c2 + cn_stride);
245 wasm_v128_store(c1, vacc1x0123);
246 wasm_v128_store(c1 + 4, vacc1x4567);
247 c1 = (float*) ((uintptr_t) c1 + cn_stride);
248 wasm_v128_store(c0, vacc0x0123);
249 wasm_v128_store(c0 + 4, vacc0x4567);
250 c0 = (float*) ((uintptr_t) c0 + cn_stride);
251
252 a3 = (const float*) ((uintptr_t) a3 - kc);
253 a2 = (const float*) ((uintptr_t) a2 - kc);
254 a1 = (const float*) ((uintptr_t) a1 - kc);
255 a0 = (const float*) ((uintptr_t) a0 - kc);
256
257 nc -= 8;
258 } else {
259 if (nc & 4) {
260 wasm_v128_store(c3, vacc3x0123);
261 wasm_v128_store(c2, vacc2x0123);
262 wasm_v128_store(c1, vacc1x0123);
263 wasm_v128_store(c0, vacc0x0123);
264
265 vacc3x0123 = vacc3x4567;
266 vacc2x0123 = vacc2x4567;
267 vacc1x0123 = vacc1x4567;
268 vacc0x0123 = vacc0x4567;
269
270 c3 += 4;
271 c2 += 4;
272 c1 += 4;
273 c0 += 4;
274 }
275 if (nc & 2) {
276 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
277 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
278 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
279 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
280
281 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
282 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
283 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
284 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
285
286 c3 += 2;
287 c2 += 2;
288 c1 += 2;
289 c0 += 2;
290 }
291 if (nc & 1) {
292 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
293 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
294 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
295 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
296 }
297
298 nc = 0;
299 }
300 } while (nc != 0);
301 }
302