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_5x8__wasmsimd_x86_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_5x8__wasmsimd_x86_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 <= 5);
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
67 do {
68 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
69 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
70 v128_t vacc1x0123 = wasm_v128_load(acc + 8);
71 v128_t vacc1x4567 = wasm_v128_load(acc + 12);
72 v128_t vacc2x0123 = wasm_v128_load(acc + 16);
73 v128_t vacc2x4567 = wasm_v128_load(acc + 20);
74 v128_t vacc3x0123 = wasm_v128_load(acc + 24);
75 v128_t vacc3x4567 = wasm_v128_load(acc + 28);
76 v128_t vacc4x0123 = wasm_v128_load(acc + 32);
77 v128_t vacc4x4567 = wasm_v128_load(acc + 36);
78 acc += 40;
79
80 size_t k = kc;
81 while (k >= 4 * sizeof(float)) {
82 const v128_t va0 = wasm_v128_load(a0);
83 a0 += 4;
84 const v128_t va1 = wasm_v128_load(a1);
85 a1 += 4;
86 const v128_t va2 = wasm_v128_load(a2);
87 a2 += 4;
88 const v128_t va3 = wasm_v128_load(a3);
89 a3 += 4;
90 const v128_t va4 = wasm_v128_load(a4);
91 a4 += 4;
92
93 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
94 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
95 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
96 const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
97 const v128_t va4c0 = wasm_v32x4_shuffle(va4, va4, 0, 0, 0, 0);
98
99 const v128_t vb0123c0 = wasm_v128_load(w + 0);
100 const v128_t vb4567c0 = wasm_v128_load(w + 4);
101
102 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
103 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
104 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
105 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
106 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c0, vb0123c0));
107 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
108 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
109 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
110 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
111 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c0, vb4567c0));
112 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
113 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
114 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
115 const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
116 const v128_t va4c1 = wasm_v32x4_shuffle(va4, va4, 1, 1, 1, 1);
117
118 const v128_t vb0123c1 = wasm_v128_load(w + 8);
119 const v128_t vb4567c1 = wasm_v128_load(w + 12);
120
121 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
122 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
123 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
124 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
125 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c1, vb0123c1));
126 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
127 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
128 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
129 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
130 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c1, vb4567c1));
131 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
132 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
133 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
134 const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
135 const v128_t va4c2 = wasm_v32x4_shuffle(va4, va4, 2, 2, 2, 2);
136
137 const v128_t vb0123c2 = wasm_v128_load(w + 16);
138 const v128_t vb4567c2 = wasm_v128_load(w + 20);
139
140 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
141 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
142 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
143 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
144 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c2, vb0123c2));
145 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
146 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
147 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
148 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
149 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c2, vb4567c2));
150 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
151 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
152 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
153 const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
154 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3);
155
156 const v128_t vb0123c3 = wasm_v128_load(w + 24);
157 const v128_t vb4567c3 = wasm_v128_load(w + 28);
158
159 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
160 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
161 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
162 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
163 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3));
164 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
165 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
166 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
167 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
168 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3));
169
170 w += 32;
171 k -= 4 * sizeof(float);
172 }
173 if XNN_UNLIKELY(k != 0) {
174 do {
175 const v128_t va0 = wasm_v128_load32_splat(a0);
176 a0 += 1;
177 const v128_t va1 = wasm_v128_load32_splat(a1);
178 a1 += 1;
179 const v128_t va2 = wasm_v128_load32_splat(a2);
180 a2 += 1;
181 const v128_t va3 = wasm_v128_load32_splat(a3);
182 a3 += 1;
183 const v128_t va4 = wasm_v128_load32_splat(a4);
184 a4 += 1;
185
186 const v128_t vb0123 = wasm_v128_load(w);
187 const v128_t vb4567 = wasm_v128_load(w + 4);
188 w += 8;
189
190 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
191 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
192 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
193 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
194 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
195 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
196 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
197 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
198 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
199 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
200
201 k -= sizeof(float);
202 } while (k != 0);
203 }
204
205 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
206 vacc0x0123 = wasm_f32x4_pmax(vmin, vacc0x0123);
207 vacc1x0123 = wasm_f32x4_pmax(vmin, vacc1x0123);
208 vacc2x0123 = wasm_f32x4_pmax(vmin, vacc2x0123);
209 vacc3x0123 = wasm_f32x4_pmax(vmin, vacc3x0123);
210 vacc4x0123 = wasm_f32x4_pmax(vmin, vacc4x0123);
211 vacc0x4567 = wasm_f32x4_pmax(vmin, vacc0x4567);
212 vacc1x4567 = wasm_f32x4_pmax(vmin, vacc1x4567);
213 vacc2x4567 = wasm_f32x4_pmax(vmin, vacc2x4567);
214 vacc3x4567 = wasm_f32x4_pmax(vmin, vacc3x4567);
215 vacc4x4567 = wasm_f32x4_pmax(vmin, vacc4x4567);
216
217 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
218 vacc0x0123 = wasm_f32x4_pmin(vmax, vacc0x0123);
219 vacc1x0123 = wasm_f32x4_pmin(vmax, vacc1x0123);
220 vacc2x0123 = wasm_f32x4_pmin(vmax, vacc2x0123);
221 vacc3x0123 = wasm_f32x4_pmin(vmax, vacc3x0123);
222 vacc4x0123 = wasm_f32x4_pmin(vmax, vacc4x0123);
223 vacc0x4567 = wasm_f32x4_pmin(vmax, vacc0x4567);
224 vacc1x4567 = wasm_f32x4_pmin(vmax, vacc1x4567);
225 vacc2x4567 = wasm_f32x4_pmin(vmax, vacc2x4567);
226 vacc3x4567 = wasm_f32x4_pmin(vmax, vacc3x4567);
227 vacc4x4567 = wasm_f32x4_pmin(vmax, vacc4x4567);
228
229 if XNN_LIKELY(nc >= 8) {
230 wasm_v128_store(c4, vacc4x0123);
231 wasm_v128_store(c4 + 4, vacc4x4567);
232 c4 = (float*) ((uintptr_t) c4 + cn_stride);
233 wasm_v128_store(c3, vacc3x0123);
234 wasm_v128_store(c3 + 4, vacc3x4567);
235 c3 = (float*) ((uintptr_t) c3 + cn_stride);
236 wasm_v128_store(c2, vacc2x0123);
237 wasm_v128_store(c2 + 4, vacc2x4567);
238 c2 = (float*) ((uintptr_t) c2 + cn_stride);
239 wasm_v128_store(c1, vacc1x0123);
240 wasm_v128_store(c1 + 4, vacc1x4567);
241 c1 = (float*) ((uintptr_t) c1 + cn_stride);
242 wasm_v128_store(c0, vacc0x0123);
243 wasm_v128_store(c0 + 4, vacc0x4567);
244 c0 = (float*) ((uintptr_t) c0 + cn_stride);
245
246 a4 = (const float*) ((uintptr_t) a4 - kc);
247 a3 = (const float*) ((uintptr_t) a3 - kc);
248 a2 = (const float*) ((uintptr_t) a2 - kc);
249 a1 = (const float*) ((uintptr_t) a1 - kc);
250 a0 = (const float*) ((uintptr_t) a0 - kc);
251
252 nc -= 8;
253 } else {
254 if (nc & 4) {
255 wasm_v128_store(c4, vacc4x0123);
256 wasm_v128_store(c3, vacc3x0123);
257 wasm_v128_store(c2, vacc2x0123);
258 wasm_v128_store(c1, vacc1x0123);
259 wasm_v128_store(c0, vacc0x0123);
260
261 vacc4x0123 = vacc4x4567;
262 vacc3x0123 = vacc3x4567;
263 vacc2x0123 = vacc2x4567;
264 vacc1x0123 = vacc1x4567;
265 vacc0x0123 = vacc0x4567;
266
267 c4 += 4;
268 c3 += 4;
269 c2 += 4;
270 c1 += 4;
271 c0 += 4;
272 }
273 if (nc & 2) {
274 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
275 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
276 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
277 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
278 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
279
280 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
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 c4 += 2;
287 c3 += 2;
288 c2 += 2;
289 c1 += 2;
290 c0 += 2;
291 }
292 if (nc & 1) {
293 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
294 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
295 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
296 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
297 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
298 }
299
300 nc = 0;
301 }
302 } while (nc != 0);
303 }
304