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_minmax_ukernel_5x8__wasmsimd_arm_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (5 * 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
61 const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min);
62 const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max);
63 do {
64 v128_t vacc0x0123 = wasm_v128_load(w);
65 v128_t vacc0x4567 = wasm_v128_load(w + 4);
66 v128_t vacc1x0123 = vacc0x0123;
67 v128_t vacc1x4567 = vacc0x4567;
68 v128_t vacc2x0123 = vacc0x0123;
69 v128_t vacc2x4567 = vacc0x4567;
70 v128_t vacc3x0123 = vacc0x0123;
71 v128_t vacc3x4567 = vacc0x4567;
72 v128_t vacc4x0123 = vacc0x0123;
73 v128_t vacc4x4567 = vacc0x4567;
74 w += 8;
75
76 size_t p = ks;
77 do {
78 const float* restrict a0 = a[0];
79 assert(a0 != NULL);
80 if XNN_UNPREDICTABLE(a0 != zero) {
81 a0 = (const float*) ((uintptr_t) a0 + a_offset);
82 }
83 const float* restrict a1 = a[1];
84 assert(a1 != NULL);
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const float*) ((uintptr_t) a1 + a_offset);
87 }
88 const float* restrict a2 = a[2];
89 assert(a2 != NULL);
90 if XNN_UNPREDICTABLE(a2 != zero) {
91 a2 = (const float*) ((uintptr_t) a2 + a_offset);
92 }
93 const float* restrict a3 = a[3];
94 assert(a3 != NULL);
95 if XNN_UNPREDICTABLE(a3 != zero) {
96 a3 = (const float*) ((uintptr_t) a3 + a_offset);
97 }
98 const float* restrict a4 = a[4];
99 assert(a4 != NULL);
100 if XNN_UNPREDICTABLE(a4 != zero) {
101 a4 = (const float*) ((uintptr_t) a4 + a_offset);
102 }
103 a += 5;
104
105 size_t k = kc;
106 while (k >= 4 * sizeof(float)) {
107 const v128_t va0 = wasm_v128_load(a0);
108 a0 += 4;
109 const v128_t va1 = wasm_v128_load(a1);
110 a1 += 4;
111 const v128_t va2 = wasm_v128_load(a2);
112 a2 += 4;
113 const v128_t va3 = wasm_v128_load(a3);
114 a3 += 4;
115 const v128_t va4 = wasm_v128_load(a4);
116 a4 += 4;
117
118 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
119 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
120 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
121 const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
122 const v128_t va4c0 = wasm_v32x4_shuffle(va4, va4, 0, 0, 0, 0);
123
124 const v128_t vb0123c0 = wasm_v128_load(w + 0);
125 const v128_t vb4567c0 = wasm_v128_load(w + 4);
126
127 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
128 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
129 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
130 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
131 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c0, vb0123c0));
132 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
133 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
134 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
135 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
136 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c0, vb4567c0));
137 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
138 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
139 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
140 const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
141 const v128_t va4c1 = wasm_v32x4_shuffle(va4, va4, 1, 1, 1, 1);
142
143 const v128_t vb0123c1 = wasm_v128_load(w + 8);
144 const v128_t vb4567c1 = wasm_v128_load(w + 12);
145
146 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
147 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
148 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
149 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
150 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c1, vb0123c1));
151 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
152 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
153 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
154 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
155 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c1, vb4567c1));
156 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
157 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
158 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
159 const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
160 const v128_t va4c2 = wasm_v32x4_shuffle(va4, va4, 2, 2, 2, 2);
161
162 const v128_t vb0123c2 = wasm_v128_load(w + 16);
163 const v128_t vb4567c2 = wasm_v128_load(w + 20);
164
165 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
166 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
167 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
168 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
169 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c2, vb0123c2));
170 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
171 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
172 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
173 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
174 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c2, vb4567c2));
175 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
176 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
177 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
178 const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
179 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3);
180
181 const v128_t vb0123c3 = wasm_v128_load(w + 24);
182 const v128_t vb4567c3 = wasm_v128_load(w + 28);
183
184 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
185 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
186 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
187 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
188 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3));
189 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
190 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
191 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
192 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
193 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3));
194
195 w += 32;
196 k -= 4 * sizeof(float);
197 }
198 if XNN_UNLIKELY(k != 0) {
199 do {
200 const v128_t vb0123 = wasm_v128_load(w);
201 const v128_t vb4567 = wasm_v128_load(w + 4);
202 w += 8;
203
204 const v128_t va0 = wasm_v32x4_load_splat(a0);
205 a0 += 1;
206 const v128_t va1 = wasm_v32x4_load_splat(a1);
207 a1 += 1;
208 const v128_t va2 = wasm_v32x4_load_splat(a2);
209 a2 += 1;
210 const v128_t va3 = wasm_v32x4_load_splat(a3);
211 a3 += 1;
212 const v128_t va4 = wasm_v32x4_load_splat(a4);
213 a4 += 1;
214
215 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
216 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
217 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
218 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
219 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
220 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
221 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
222 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
223 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
224 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
225 k -= sizeof(float);
226 } while (k != 0);
227 }
228 p -= 5 * sizeof(void*);
229 } while (p != 0);
230
231 vacc0x0123 = wasm_f32x4_max(vacc0x0123, vmin);
232 vacc1x0123 = wasm_f32x4_max(vacc1x0123, vmin);
233 vacc2x0123 = wasm_f32x4_max(vacc2x0123, vmin);
234 vacc3x0123 = wasm_f32x4_max(vacc3x0123, vmin);
235 vacc4x0123 = wasm_f32x4_max(vacc4x0123, vmin);
236 vacc0x4567 = wasm_f32x4_max(vacc0x4567, vmin);
237 vacc1x4567 = wasm_f32x4_max(vacc1x4567, vmin);
238 vacc2x4567 = wasm_f32x4_max(vacc2x4567, vmin);
239 vacc3x4567 = wasm_f32x4_max(vacc3x4567, vmin);
240 vacc4x4567 = wasm_f32x4_max(vacc4x4567, vmin);
241
242 vacc0x0123 = wasm_f32x4_min(vacc0x0123, vmax);
243 vacc1x0123 = wasm_f32x4_min(vacc1x0123, vmax);
244 vacc2x0123 = wasm_f32x4_min(vacc2x0123, vmax);
245 vacc3x0123 = wasm_f32x4_min(vacc3x0123, vmax);
246 vacc4x0123 = wasm_f32x4_min(vacc4x0123, vmax);
247 vacc0x4567 = wasm_f32x4_min(vacc0x4567, vmax);
248 vacc1x4567 = wasm_f32x4_min(vacc1x4567, vmax);
249 vacc2x4567 = wasm_f32x4_min(vacc2x4567, vmax);
250 vacc3x4567 = wasm_f32x4_min(vacc3x4567, vmax);
251 vacc4x4567 = wasm_f32x4_min(vacc4x4567, vmax);
252
253 if XNN_LIKELY(nc >= 8) {
254 wasm_v128_store(c4, vacc4x0123);
255 wasm_v128_store(c4 + 4, vacc4x4567);
256 c4 = (float*) ((uintptr_t) c4 + cn_stride);
257 wasm_v128_store(c3, vacc3x0123);
258 wasm_v128_store(c3 + 4, vacc3x4567);
259 c3 = (float*) ((uintptr_t) c3 + cn_stride);
260 wasm_v128_store(c2, vacc2x0123);
261 wasm_v128_store(c2 + 4, vacc2x4567);
262 c2 = (float*) ((uintptr_t) c2 + cn_stride);
263 wasm_v128_store(c1, vacc1x0123);
264 wasm_v128_store(c1 + 4, vacc1x4567);
265 c1 = (float*) ((uintptr_t) c1 + cn_stride);
266 wasm_v128_store(c0, vacc0x0123);
267 wasm_v128_store(c0 + 4, vacc0x4567);
268 c0 = (float*) ((uintptr_t) c0 + cn_stride);
269
270 a = (const float**restrict) ((uintptr_t) a - ks);
271 nc -= 8;
272 } else {
273 if (nc & 4) {
274 wasm_v128_store(c4, vacc4x0123);
275 wasm_v128_store(c3, vacc3x0123);
276 wasm_v128_store(c2, vacc2x0123);
277 wasm_v128_store(c1, vacc1x0123);
278 wasm_v128_store(c0, vacc0x0123);
279
280 vacc4x0123 = vacc4x4567;
281 vacc3x0123 = vacc3x4567;
282 vacc2x0123 = vacc2x4567;
283 vacc1x0123 = vacc1x4567;
284 vacc0x0123 = vacc0x4567;
285
286 c4 += 4;
287 c3 += 4;
288 c2 += 4;
289 c1 += 4;
290 c0 += 4;
291 }
292 if (nc & 2) {
293 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
294 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
295 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
296 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
297 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
298
299 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
300 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
301 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
302 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
303 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
304
305 c4 += 2;
306 c3 += 2;
307 c2 += 2;
308 c1 += 2;
309 c0 += 2;
310 }
311 if (nc & 1) {
312 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
313 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
314 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
315 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
316 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
317 }
318
319 nc = 0;
320 }
321 } while (nc != 0);
322 }
323