1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemminc_minmax_ukernel_6x8__wasmrelaxedsimd_splat(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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_6x8__wasmrelaxedsimd_splat(
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 float*restrict acc,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38 assert(acc != NULL);
39
40 const float* a0 = a;
41 float* c0 = c;
42 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
67 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
68 if XNN_UNPREDICTABLE(mr != 6) {
69 a5 = a4;
70 c5 = c4;
71 }
72
73 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
74 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
75 do {
76 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
77 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
78 v128_t vacc1x0123 = wasm_v128_load(acc + 8);
79 v128_t vacc1x4567 = wasm_v128_load(acc + 12);
80 v128_t vacc2x0123 = wasm_v128_load(acc + 16);
81 v128_t vacc2x4567 = wasm_v128_load(acc + 20);
82 v128_t vacc3x0123 = wasm_v128_load(acc + 24);
83 v128_t vacc3x4567 = wasm_v128_load(acc + 28);
84 v128_t vacc4x0123 = wasm_v128_load(acc + 32);
85 v128_t vacc4x4567 = wasm_v128_load(acc + 36);
86 v128_t vacc5x0123 = wasm_v128_load(acc + 40);
87 v128_t vacc5x4567 = wasm_v128_load(acc + 44);
88 acc += 48;
89
90 size_t k = kc;
91 while (k >= 4 * sizeof(float)) {
92 const v128_t va0 = wasm_v128_load(a0);
93 a0 += 4;
94 const v128_t va1 = wasm_v128_load(a1);
95 a1 += 4;
96 const v128_t va2 = wasm_v128_load(a2);
97 a2 += 4;
98 const v128_t va3 = wasm_v128_load(a3);
99 a3 += 4;
100 const v128_t va4 = wasm_v128_load(a4);
101 a4 += 4;
102 const v128_t va5 = wasm_v128_load(a5);
103 a5 += 4;
104
105 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
106 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
107 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
108 const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
109 const v128_t va4c0 = wasm_v32x4_shuffle(va4, va4, 0, 0, 0, 0);
110 const v128_t va5c0 = wasm_v32x4_shuffle(va5, va5, 0, 0, 0, 0);
111
112 const v128_t vb0123c0 = wasm_v128_load(w + 0);
113 const v128_t vb4567c0 = wasm_v128_load(w + 4);
114
115 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
116 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
117 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
118 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
119 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c0, vb0123c0));
120 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c0, vb0123c0));
121 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
122 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
123 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
124 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
125 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c0, vb4567c0));
126 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c0, vb4567c0));
127 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
128 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
129 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
130 const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
131 const v128_t va4c1 = wasm_v32x4_shuffle(va4, va4, 1, 1, 1, 1);
132 const v128_t va5c1 = wasm_v32x4_shuffle(va5, va5, 1, 1, 1, 1);
133
134 const v128_t vb0123c1 = wasm_v128_load(w + 8);
135 const v128_t vb4567c1 = wasm_v128_load(w + 12);
136
137 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
138 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
139 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
140 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
141 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c1, vb0123c1));
142 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c1, vb0123c1));
143 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
144 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
145 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
146 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
147 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c1, vb4567c1));
148 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c1, vb4567c1));
149 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
150 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
151 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
152 const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
153 const v128_t va4c2 = wasm_v32x4_shuffle(va4, va4, 2, 2, 2, 2);
154 const v128_t va5c2 = wasm_v32x4_shuffle(va5, va5, 2, 2, 2, 2);
155
156 const v128_t vb0123c2 = wasm_v128_load(w + 16);
157 const v128_t vb4567c2 = wasm_v128_load(w + 20);
158
159 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
160 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
161 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
162 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
163 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c2, vb0123c2));
164 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c2, vb0123c2));
165 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
166 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
167 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
168 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
169 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c2, vb4567c2));
170 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c2, vb4567c2));
171 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
172 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
173 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
174 const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
175 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3);
176 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3);
177
178 const v128_t vb0123c3 = wasm_v128_load(w + 24);
179 const v128_t vb4567c3 = wasm_v128_load(w + 28);
180
181 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
182 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
183 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
184 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
185 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3));
186 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3));
187 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
188 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
189 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
190 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
191 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3));
192 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3));
193
194 w += 32;
195 k -= 4 * sizeof(float);
196 }
197 if XNN_UNLIKELY(k != 0) {
198 do {
199 const v128_t va0 = wasm_v128_load32_splat(a0);
200 a0 += 1;
201 const v128_t va1 = wasm_v128_load32_splat(a1);
202 a1 += 1;
203 const v128_t va2 = wasm_v128_load32_splat(a2);
204 a2 += 1;
205 const v128_t va3 = wasm_v128_load32_splat(a3);
206 a3 += 1;
207 const v128_t va4 = wasm_v128_load32_splat(a4);
208 a4 += 1;
209 const v128_t va5 = wasm_v128_load32_splat(a5);
210 a5 += 1;
211
212 const v128_t vb0123 = wasm_v128_load(w);
213 const v128_t vb4567 = wasm_v128_load(w + 4);
214 w += 8;
215
216 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
217 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
218 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
219 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
220 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
221 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123));
222 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
223 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
224 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
225 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
226 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
227 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567));
228
229 k -= sizeof(float);
230 } while (k != 0);
231 }
232
233 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
234 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
235 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
236 vacc3x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x0123);
237 vacc4x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x0123);
238 vacc5x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc5x0123);
239 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
240 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
241 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
242 vacc3x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x4567);
243 vacc4x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc4x4567);
244 vacc5x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc5x4567);
245
246 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
247 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
248 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
249 vacc3x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x0123);
250 vacc4x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x0123);
251 vacc5x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc5x0123);
252 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
253 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
254 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
255 vacc3x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x4567);
256 vacc4x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc4x4567);
257 vacc5x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc5x4567);
258
259 if XNN_LIKELY(nc >= 8) {
260 wasm_v128_store(c5, vacc5x0123);
261 wasm_v128_store(c5 + 4, vacc5x4567);
262 c5 = (float*) ((uintptr_t) c5 + cn_stride);
263 wasm_v128_store(c4, vacc4x0123);
264 wasm_v128_store(c4 + 4, vacc4x4567);
265 c4 = (float*) ((uintptr_t) c4 + cn_stride);
266 wasm_v128_store(c3, vacc3x0123);
267 wasm_v128_store(c3 + 4, vacc3x4567);
268 c3 = (float*) ((uintptr_t) c3 + cn_stride);
269 wasm_v128_store(c2, vacc2x0123);
270 wasm_v128_store(c2 + 4, vacc2x4567);
271 c2 = (float*) ((uintptr_t) c2 + cn_stride);
272 wasm_v128_store(c1, vacc1x0123);
273 wasm_v128_store(c1 + 4, vacc1x4567);
274 c1 = (float*) ((uintptr_t) c1 + cn_stride);
275 wasm_v128_store(c0, vacc0x0123);
276 wasm_v128_store(c0 + 4, vacc0x4567);
277 c0 = (float*) ((uintptr_t) c0 + cn_stride);
278
279 a5 = (const float*) ((uintptr_t) a5 - kc);
280 a4 = (const float*) ((uintptr_t) a4 - kc);
281 a3 = (const float*) ((uintptr_t) a3 - kc);
282 a2 = (const float*) ((uintptr_t) a2 - kc);
283 a1 = (const float*) ((uintptr_t) a1 - kc);
284 a0 = (const float*) ((uintptr_t) a0 - kc);
285
286 nc -= 8;
287 } else {
288 if (nc & 4) {
289 wasm_v128_store(c5, vacc5x0123);
290 wasm_v128_store(c4, vacc4x0123);
291 wasm_v128_store(c3, vacc3x0123);
292 wasm_v128_store(c2, vacc2x0123);
293 wasm_v128_store(c1, vacc1x0123);
294 wasm_v128_store(c0, vacc0x0123);
295
296 vacc5x0123 = vacc5x4567;
297 vacc4x0123 = vacc4x4567;
298 vacc3x0123 = vacc3x4567;
299 vacc2x0123 = vacc2x4567;
300 vacc1x0123 = vacc1x4567;
301 vacc0x0123 = vacc0x4567;
302
303 c5 += 4;
304 c4 += 4;
305 c3 += 4;
306 c2 += 4;
307 c1 += 4;
308 c0 += 4;
309 }
310 if (nc & 2) {
311 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
312 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
313 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
314 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
315 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
316 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
317
318 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
319 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
320 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
321 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
322 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
323 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
324
325 c5 += 2;
326 c4 += 2;
327 c3 += 2;
328 c2 += 2;
329 c1 += 2;
330 c0 += 2;
331 }
332 if (nc & 1) {
333 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
334 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
335 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
336 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
337 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
338 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
339 }
340
341 nc = 0;
342 }
343 } while (nc != 0);
344 }
345