1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/wasmsimd-splat.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_6x8__wasmsimd_splat(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_6x8__wasmsimd_splat(
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)])
30 {
31 assert(mr != 0);
32 assert(mr <= 6);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (6 * 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 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 4) {
58 c4 = c3;
59 }
60 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
61 if XNN_UNPREDICTABLE(mr != 6) {
62 c5 = c4;
63 }
64
65 do {
66 v128_t vacc0x0123 = wasm_v128_load(w);
67 v128_t vacc0x4567 = wasm_v128_load(w + 4);
68 v128_t vacc1x0123 = vacc0x0123;
69 v128_t vacc1x4567 = vacc0x4567;
70 v128_t vacc2x0123 = vacc0x0123;
71 v128_t vacc2x4567 = vacc0x4567;
72 v128_t vacc3x0123 = vacc0x0123;
73 v128_t vacc3x4567 = vacc0x4567;
74 v128_t vacc4x0123 = vacc0x0123;
75 v128_t vacc4x4567 = vacc0x4567;
76 v128_t vacc5x0123 = vacc0x0123;
77 v128_t vacc5x4567 = vacc0x4567;
78 w += 8;
79
80 size_t p = ks;
81 do {
82 const float* restrict a0 = a[0];
83 assert(a0 != NULL);
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const float*) ((uintptr_t) a0 + a_offset);
86 }
87 const float* restrict a1 = a[1];
88 assert(a1 != NULL);
89 if XNN_UNPREDICTABLE(a1 != zero) {
90 a1 = (const float*) ((uintptr_t) a1 + a_offset);
91 }
92 const float* restrict a2 = a[2];
93 assert(a2 != NULL);
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const float*) ((uintptr_t) a2 + a_offset);
96 }
97 const float* restrict a3 = a[3];
98 assert(a3 != NULL);
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const float*) ((uintptr_t) a3 + a_offset);
101 }
102 const float* restrict a4 = a[4];
103 assert(a4 != NULL);
104 if XNN_UNPREDICTABLE(a4 != zero) {
105 a4 = (const float*) ((uintptr_t) a4 + a_offset);
106 }
107 const float* restrict a5 = a[5];
108 assert(a5 != NULL);
109 if XNN_UNPREDICTABLE(a5 != zero) {
110 a5 = (const float*) ((uintptr_t) a5 + a_offset);
111 }
112 a += 6;
113
114 size_t k = kc;
115 while (k >= 4 * sizeof(float)) {
116 const v128_t va0 = wasm_v128_load(a0);
117 a0 += 4;
118 const v128_t va1 = wasm_v128_load(a1);
119 a1 += 4;
120 const v128_t va2 = wasm_v128_load(a2);
121 a2 += 4;
122 const v128_t va3 = wasm_v128_load(a3);
123 a3 += 4;
124 const v128_t va4 = wasm_v128_load(a4);
125 a4 += 4;
126 const v128_t va5 = wasm_v128_load(a5);
127 a5 += 4;
128
129 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
130 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
131 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
132 const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
133 const v128_t va4c0 = wasm_v32x4_shuffle(va4, va4, 0, 0, 0, 0);
134 const v128_t va5c0 = wasm_v32x4_shuffle(va5, va5, 0, 0, 0, 0);
135
136 const v128_t vb0123c0 = wasm_v128_load(w + 0);
137 const v128_t vb4567c0 = wasm_v128_load(w + 4);
138
139 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
140 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
141 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
142 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
143 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c0, vb0123c0));
144 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c0, vb0123c0));
145 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
146 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
147 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
148 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
149 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c0, vb4567c0));
150 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c0, vb4567c0));
151 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
152 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
153 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
154 const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
155 const v128_t va4c1 = wasm_v32x4_shuffle(va4, va4, 1, 1, 1, 1);
156 const v128_t va5c1 = wasm_v32x4_shuffle(va5, va5, 1, 1, 1, 1);
157
158 const v128_t vb0123c1 = wasm_v128_load(w + 8);
159 const v128_t vb4567c1 = wasm_v128_load(w + 12);
160
161 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
162 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
163 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
164 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
165 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c1, vb0123c1));
166 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c1, vb0123c1));
167 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
168 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
169 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
170 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
171 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c1, vb4567c1));
172 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c1, vb4567c1));
173 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
174 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
175 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
176 const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
177 const v128_t va4c2 = wasm_v32x4_shuffle(va4, va4, 2, 2, 2, 2);
178 const v128_t va5c2 = wasm_v32x4_shuffle(va5, va5, 2, 2, 2, 2);
179
180 const v128_t vb0123c2 = wasm_v128_load(w + 16);
181 const v128_t vb4567c2 = wasm_v128_load(w + 20);
182
183 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
184 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
185 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
186 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
187 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c2, vb0123c2));
188 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c2, vb0123c2));
189 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
190 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
191 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
192 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
193 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c2, vb4567c2));
194 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c2, vb4567c2));
195 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
196 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
197 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
198 const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
199 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3);
200 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3);
201
202 const v128_t vb0123c3 = wasm_v128_load(w + 24);
203 const v128_t vb4567c3 = wasm_v128_load(w + 28);
204
205 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
206 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
207 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
208 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
209 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3));
210 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3));
211 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
212 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
213 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
214 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
215 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3));
216 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3));
217
218 w += 32;
219 k -= 4 * sizeof(float);
220 }
221 if XNN_UNLIKELY(k != 0) {
222 do {
223 const v128_t vb0123 = wasm_v128_load(w);
224 const v128_t vb4567 = wasm_v128_load(w + 4);
225 w += 8;
226
227 const v128_t va0 = wasm_v128_load32_splat(a0);
228 a0 += 1;
229 const v128_t va1 = wasm_v128_load32_splat(a1);
230 a1 += 1;
231 const v128_t va2 = wasm_v128_load32_splat(a2);
232 a2 += 1;
233 const v128_t va3 = wasm_v128_load32_splat(a3);
234 a3 += 1;
235 const v128_t va4 = wasm_v128_load32_splat(a4);
236 a4 += 1;
237 const v128_t va5 = wasm_v128_load32_splat(a5);
238 a5 += 1;
239
240 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
241 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
242 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
243 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
244 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
245 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
246 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
247 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
248 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
249 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
250 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123));
251 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567));
252 k -= sizeof(float);
253 } while (k != 0);
254 }
255 p -= 6 * sizeof(void*);
256 } while (p != 0);
257
258 const v128_t vzero = wasm_i32x4_const_splat(0);
259 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
260 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
261 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
262 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
263 vacc4x0123 = wasm_i32x4_max(vacc4x0123, vzero);
264 vacc5x0123 = wasm_i32x4_max(vacc5x0123, vzero);
265 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
266 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
267 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
268 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
269 vacc4x4567 = wasm_i32x4_max(vacc4x4567, vzero);
270 vacc5x4567 = wasm_i32x4_max(vacc5x4567, vzero);
271
272 if XNN_LIKELY(nc >= 8) {
273 wasm_v128_store(c5, vacc5x0123);
274 wasm_v128_store(c5 + 4, vacc5x4567);
275 c5 = (float*) ((uintptr_t) c5 + cn_stride);
276 wasm_v128_store(c4, vacc4x0123);
277 wasm_v128_store(c4 + 4, vacc4x4567);
278 c4 = (float*) ((uintptr_t) c4 + cn_stride);
279 wasm_v128_store(c3, vacc3x0123);
280 wasm_v128_store(c3 + 4, vacc3x4567);
281 c3 = (float*) ((uintptr_t) c3 + cn_stride);
282 wasm_v128_store(c2, vacc2x0123);
283 wasm_v128_store(c2 + 4, vacc2x4567);
284 c2 = (float*) ((uintptr_t) c2 + cn_stride);
285 wasm_v128_store(c1, vacc1x0123);
286 wasm_v128_store(c1 + 4, vacc1x4567);
287 c1 = (float*) ((uintptr_t) c1 + cn_stride);
288 wasm_v128_store(c0, vacc0x0123);
289 wasm_v128_store(c0 + 4, vacc0x4567);
290 c0 = (float*) ((uintptr_t) c0 + cn_stride);
291
292 a = (const float**restrict) ((uintptr_t) a - ks);
293 nc -= 8;
294 } else {
295 if (nc & 4) {
296 wasm_v128_store(c5, vacc5x0123);
297 wasm_v128_store(c4, vacc4x0123);
298 wasm_v128_store(c3, vacc3x0123);
299 wasm_v128_store(c2, vacc2x0123);
300 wasm_v128_store(c1, vacc1x0123);
301 wasm_v128_store(c0, vacc0x0123);
302
303 vacc5x0123 = vacc5x4567;
304 vacc4x0123 = vacc4x4567;
305 vacc3x0123 = vacc3x4567;
306 vacc2x0123 = vacc2x4567;
307 vacc1x0123 = vacc1x4567;
308 vacc0x0123 = vacc0x4567;
309
310 c5 += 4;
311 c4 += 4;
312 c3 += 4;
313 c2 += 4;
314 c1 += 4;
315 c0 += 4;
316 }
317 if (nc & 2) {
318 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
319 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
320 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
321 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
322 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
323 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
324
325 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
326 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
327 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
328 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
329 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
330 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
331
332 c5 += 2;
333 c4 += 2;
334 c3 += 2;
335 c2 += 2;
336 c1 += 2;
337 c0 += 2;
338 }
339 if (nc & 1) {
340 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
341 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
342 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
343 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
344 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
345 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
346 }
347
348 nc = 0;
349 }
350 } while (nc != 0);
351 }
352