1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/wasmsimd-s4.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_4x8s4__wasmsimd_arm(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_4x8s4__wasmsimd_arm(
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 <= 4);
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
61 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
62 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
63 do {
64 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
65 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
66 v128_t vacc1x0123 = wasm_v128_load(acc + 8);
67 v128_t vacc1x4567 = wasm_v128_load(acc + 12);
68 v128_t vacc2x0123 = wasm_v128_load(acc + 16);
69 v128_t vacc2x4567 = wasm_v128_load(acc + 20);
70 v128_t vacc3x0123 = wasm_v128_load(acc + 24);
71 v128_t vacc3x4567 = wasm_v128_load(acc + 28);
72 acc += 32;
73
74 size_t k = kc;
75 while (k >= 4 * sizeof(float)) {
76 v128_t va0 = wasm_v128_load(a0);
77 a0 += 4;
78 v128_t va1 = wasm_v128_load(a1);
79 a1 += 4;
80 v128_t va2 = wasm_v128_load(a2);
81 a2 += 4;
82 v128_t va3 = wasm_v128_load(a3);
83 a3 += 4;
84
85
86 const v128_t vb0123c0 = wasm_v128_load(w + 0);
87 const v128_t vb4567c0 = wasm_v128_load(w + 4);
88
89 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
90 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
91 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
92 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
93 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
94 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
95 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
96 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
97
98 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
99 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
100 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
101 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
102
103 const v128_t vb0123c1 = wasm_v128_load(w + 8);
104 const v128_t vb4567c1 = wasm_v128_load(w + 12);
105
106 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
107 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
108 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
109 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
110 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
111 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
112 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
113 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
114
115 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
116 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
117 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
118 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
119
120 const v128_t vb0123c2 = wasm_v128_load(w + 16);
121 const v128_t vb4567c2 = wasm_v128_load(w + 20);
122
123 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
124 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
125 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
126 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
127 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
128 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
129 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
130 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
131
132 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
133 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
134 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
135 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
136
137 const v128_t vb0123c3 = wasm_v128_load(w + 24);
138 const v128_t vb4567c3 = wasm_v128_load(w + 28);
139
140 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
141 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
142 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
143 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
144 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
145 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
146 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
147 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
148
149
150 w += 32;
151 k -= 4 * sizeof(float);
152 }
153 if XNN_UNLIKELY(k != 0) {
154 v128_t va0 = wasm_v128_load(a0);
155 a0 = (const float*) ((uintptr_t) a0 + k);
156 v128_t va1 = wasm_v128_load(a1);
157 a1 = (const float*) ((uintptr_t) a1 + k);
158 v128_t va2 = wasm_v128_load(a2);
159 a2 = (const float*) ((uintptr_t) a2 + k);
160 v128_t va3 = wasm_v128_load(a3);
161 a3 = (const float*) ((uintptr_t) a3 + k);
162
163 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
164
165 const v128_t vb0123c0 = wasm_v128_load(w + 0);
166 const v128_t vb4567c0 = wasm_v128_load(w + 4);
167
168 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
169 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
170 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
171 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
172 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
173 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
174 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
175 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
176
177 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
178 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
179 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
180 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
181
182 const v128_t vb0123c1 = wasm_v128_load(w + 8);
183 const v128_t vb4567c1 = wasm_v128_load(w + 12);
184
185 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
186 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
187 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
188 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
189 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
190 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
191 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
192 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
193
194 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
195 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
196 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
197 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
198
199 const v128_t vb0123c2 = wasm_v128_load(w + 16);
200 const v128_t vb4567c2 = wasm_v128_load(w + 20);
201
202 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
203 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
204 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
205 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
206 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
207 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
208 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
209 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
210
211 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
212 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
213 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
214 va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
215
216 const v128_t vb0123c3 = wasm_v128_load(w + 24);
217 const v128_t vb4567c3 = wasm_v128_load(w + 28);
218
219 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
220 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
221 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
222 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
223 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
224 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
225 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
226 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(wasm_v128_andnot(va3, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
227
228
229 w += 32;
230 }
231
232 vacc0x0123 = wasm_f32x4_max(vmin, vacc0x0123);
233 vacc1x0123 = wasm_f32x4_max(vmin, vacc1x0123);
234 vacc2x0123 = wasm_f32x4_max(vmin, vacc2x0123);
235 vacc3x0123 = wasm_f32x4_max(vmin, vacc3x0123);
236 vacc0x4567 = wasm_f32x4_max(vmin, vacc0x4567);
237 vacc1x4567 = wasm_f32x4_max(vmin, vacc1x4567);
238 vacc2x4567 = wasm_f32x4_max(vmin, vacc2x4567);
239 vacc3x4567 = wasm_f32x4_max(vmin, vacc3x4567);
240
241 vacc0x0123 = wasm_f32x4_min(vmax, vacc0x0123);
242 vacc1x0123 = wasm_f32x4_min(vmax, vacc1x0123);
243 vacc2x0123 = wasm_f32x4_min(vmax, vacc2x0123);
244 vacc3x0123 = wasm_f32x4_min(vmax, vacc3x0123);
245 vacc0x4567 = wasm_f32x4_min(vmax, vacc0x4567);
246 vacc1x4567 = wasm_f32x4_min(vmax, vacc1x4567);
247 vacc2x4567 = wasm_f32x4_min(vmax, vacc2x4567);
248 vacc3x4567 = wasm_f32x4_min(vmax, vacc3x4567);
249
250 if XNN_LIKELY(nc >= 8) {
251 wasm_v128_store(c3, vacc3x0123);
252 wasm_v128_store(c3 + 4, vacc3x4567);
253 c3 = (float*) ((uintptr_t) c3 + cn_stride);
254 wasm_v128_store(c2, vacc2x0123);
255 wasm_v128_store(c2 + 4, vacc2x4567);
256 c2 = (float*) ((uintptr_t) c2 + cn_stride);
257 wasm_v128_store(c1, vacc1x0123);
258 wasm_v128_store(c1 + 4, vacc1x4567);
259 c1 = (float*) ((uintptr_t) c1 + cn_stride);
260 wasm_v128_store(c0, vacc0x0123);
261 wasm_v128_store(c0 + 4, vacc0x4567);
262 c0 = (float*) ((uintptr_t) c0 + cn_stride);
263
264 a3 = (const float*) ((uintptr_t) a3 - kc);
265 a2 = (const float*) ((uintptr_t) a2 - kc);
266 a1 = (const float*) ((uintptr_t) a1 - kc);
267 a0 = (const float*) ((uintptr_t) a0 - kc);
268
269 nc -= 8;
270 } else {
271 if (nc & 4) {
272 wasm_v128_store(c3, vacc3x0123);
273 wasm_v128_store(c2, vacc2x0123);
274 wasm_v128_store(c1, vacc1x0123);
275 wasm_v128_store(c0, vacc0x0123);
276
277 vacc3x0123 = vacc3x4567;
278 vacc2x0123 = vacc2x4567;
279 vacc1x0123 = vacc1x4567;
280 vacc0x0123 = vacc0x4567;
281
282 c3 += 4;
283 c2 += 4;
284 c1 += 4;
285 c0 += 4;
286 }
287 if (nc & 2) {
288 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
289 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
290 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
291 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
292
293 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
294 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
295 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
296 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
297
298 c3 += 2;
299 c2 += 2;
300 c1 += 2;
301 c0 += 2;
302 }
303 if (nc & 1) {
304 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
305 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
306 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
307 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
308 }
309
310 nc = 0;
311 }
312 } while (nc != 0);
313 }
314