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_gemm_ukernel_6x8__wasmsimd_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 union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_ukernel_6x8__wasmsimd_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 union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 6);
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 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 a4 = a3;
62 c4 = c3;
63 }
64 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
65 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr != 6) {
67 a5 = a4;
68 c5 = c4;
69 }
70
71 do {
72 v128_t vacc0x0123 = wasm_v128_load(w + 0);
73 v128_t vacc0x4567 = wasm_v128_load(w + 4);
74 v128_t vacc1x0123 = vacc0x0123;
75 v128_t vacc1x4567 = vacc0x4567;
76 v128_t vacc2x0123 = vacc0x0123;
77 v128_t vacc2x4567 = vacc0x4567;
78 v128_t vacc3x0123 = vacc0x0123;
79 v128_t vacc3x4567 = vacc0x4567;
80 v128_t vacc4x0123 = vacc0x0123;
81 v128_t vacc4x4567 = vacc0x4567;
82 v128_t vacc5x0123 = vacc0x0123;
83 v128_t vacc5x4567 = vacc0x4567;
84 w += 8;
85
86 size_t k = kc;
87 while (k >= 4 * sizeof(float)) {
88 const v128_t va0 = wasm_v128_load(a0);
89 a0 += 4;
90 const v128_t va1 = wasm_v128_load(a1);
91 a1 += 4;
92 const v128_t va2 = wasm_v128_load(a2);
93 a2 += 4;
94 const v128_t va3 = wasm_v128_load(a3);
95 a3 += 4;
96 const v128_t va4 = wasm_v128_load(a4);
97 a4 += 4;
98 const v128_t va5 = wasm_v128_load(a5);
99 a5 += 4;
100
101 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
102 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
103 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
104 const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
105 const v128_t va4c0 = wasm_v32x4_shuffle(va4, va4, 0, 0, 0, 0);
106 const v128_t va5c0 = wasm_v32x4_shuffle(va5, va5, 0, 0, 0, 0);
107
108 const v128_t vb0123c0 = wasm_v128_load(w + 0);
109 const v128_t vb4567c0 = wasm_v128_load(w + 4);
110
111 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
112 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
113 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
114 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
115 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c0, vb0123c0));
116 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c0, vb0123c0));
117 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
118 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
119 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
120 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
121 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c0, vb4567c0));
122 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c0, vb4567c0));
123 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
124 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
125 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
126 const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
127 const v128_t va4c1 = wasm_v32x4_shuffle(va4, va4, 1, 1, 1, 1);
128 const v128_t va5c1 = wasm_v32x4_shuffle(va5, va5, 1, 1, 1, 1);
129
130 const v128_t vb0123c1 = wasm_v128_load(w + 8);
131 const v128_t vb4567c1 = wasm_v128_load(w + 12);
132
133 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
134 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
135 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
136 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
137 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c1, vb0123c1));
138 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c1, vb0123c1));
139 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
140 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
141 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
142 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
143 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c1, vb4567c1));
144 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c1, vb4567c1));
145 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
146 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
147 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
148 const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
149 const v128_t va4c2 = wasm_v32x4_shuffle(va4, va4, 2, 2, 2, 2);
150 const v128_t va5c2 = wasm_v32x4_shuffle(va5, va5, 2, 2, 2, 2);
151
152 const v128_t vb0123c2 = wasm_v128_load(w + 16);
153 const v128_t vb4567c2 = wasm_v128_load(w + 20);
154
155 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
156 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
157 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
158 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
159 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c2, vb0123c2));
160 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c2, vb0123c2));
161 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
162 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
163 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
164 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
165 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c2, vb4567c2));
166 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c2, vb4567c2));
167 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
168 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
169 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
170 const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
171 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3);
172 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3);
173
174 const v128_t vb0123c3 = wasm_v128_load(w + 24);
175 const v128_t vb4567c3 = wasm_v128_load(w + 28);
176
177 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
178 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
179 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
180 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
181 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3));
182 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3));
183 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
184 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
185 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
186 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
187 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3));
188 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3));
189
190 w += 32;
191 k -= 4 * sizeof(float);
192 }
193 if XNN_UNLIKELY(k != 0) {
194 do {
195 const v128_t va0 = wasm_v128_load32_splat(a0);
196 a0 += 1;
197 const v128_t va1 = wasm_v128_load32_splat(a1);
198 a1 += 1;
199 const v128_t va2 = wasm_v128_load32_splat(a2);
200 a2 += 1;
201 const v128_t va3 = wasm_v128_load32_splat(a3);
202 a3 += 1;
203 const v128_t va4 = wasm_v128_load32_splat(a4);
204 a4 += 1;
205 const v128_t va5 = wasm_v128_load32_splat(a5);
206 a5 += 1;
207
208 const v128_t vb0123 = wasm_v128_load(w);
209 const v128_t vb4567 = wasm_v128_load(w + 4);
210 w += 8;
211
212 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
213 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
214 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
215 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
216 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
217 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5, vb0123));
218 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
219 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
220 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
221 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
222 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
223 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5, vb4567));
224
225 k -= sizeof(float);
226 } while (k != 0);
227 }
228
229
230 if XNN_LIKELY(nc >= 8) {
231 wasm_v128_store(c5, vacc5x0123);
232 wasm_v128_store(c5 + 4, vacc5x4567);
233 c5 = (float*) ((uintptr_t) c5 + cn_stride);
234 wasm_v128_store(c4, vacc4x0123);
235 wasm_v128_store(c4 + 4, vacc4x4567);
236 c4 = (float*) ((uintptr_t) c4 + cn_stride);
237 wasm_v128_store(c3, vacc3x0123);
238 wasm_v128_store(c3 + 4, vacc3x4567);
239 c3 = (float*) ((uintptr_t) c3 + cn_stride);
240 wasm_v128_store(c2, vacc2x0123);
241 wasm_v128_store(c2 + 4, vacc2x4567);
242 c2 = (float*) ((uintptr_t) c2 + cn_stride);
243 wasm_v128_store(c1, vacc1x0123);
244 wasm_v128_store(c1 + 4, vacc1x4567);
245 c1 = (float*) ((uintptr_t) c1 + cn_stride);
246 wasm_v128_store(c0, vacc0x0123);
247 wasm_v128_store(c0 + 4, vacc0x4567);
248 c0 = (float*) ((uintptr_t) c0 + cn_stride);
249
250 a5 = (const float*) ((uintptr_t) a5 - kc);
251 a4 = (const float*) ((uintptr_t) a4 - kc);
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(c5, vacc5x0123);
261 wasm_v128_store(c4, vacc4x0123);
262 wasm_v128_store(c3, vacc3x0123);
263 wasm_v128_store(c2, vacc2x0123);
264 wasm_v128_store(c1, vacc1x0123);
265 wasm_v128_store(c0, vacc0x0123);
266
267 vacc5x0123 = vacc5x4567;
268 vacc4x0123 = vacc4x4567;
269 vacc3x0123 = vacc3x4567;
270 vacc2x0123 = vacc2x4567;
271 vacc1x0123 = vacc1x4567;
272 vacc0x0123 = vacc0x4567;
273
274 c5 += 4;
275 c4 += 4;
276 c3 += 4;
277 c2 += 4;
278 c1 += 4;
279 c0 += 4;
280 }
281 if (nc & 2) {
282 *((double*) c5) = wasm_f64x2_extract_lane(vacc5x0123, 0);
283 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
284 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
285 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
286 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
287 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
288
289 vacc5x0123 = wasm_v32x4_shuffle(vacc5x0123, vacc5x0123, 2, 3, 2, 3);
290 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
291 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
292 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
293 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
294 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
295
296 c5 += 2;
297 c4 += 2;
298 c3 += 2;
299 c2 += 2;
300 c1 += 2;
301 c0 += 2;
302 }
303 if (nc & 1) {
304 *c5 = wasm_f32x4_extract_lane(vacc5x0123, 0);
305 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
306 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
307 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
308 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
309 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
310 }
311
312 nc = 0;
313 }
314 } while (nc != 0);
315 }
316