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_gemm_minmax_ukernel_5x8s4__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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_minmax_ukernel_5x8s4__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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(mr != 0);
30   assert(mr <= 5);
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 
65   const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
66   const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
67   do {
68     v128_t vacc0x0123 = wasm_v128_load(w + 0);
69     v128_t vacc0x4567 = wasm_v128_load(w + 4);
70     v128_t vacc1x0123 = vacc0x0123;
71     v128_t vacc1x4567 = vacc0x4567;
72     v128_t vacc2x0123 = vacc0x0123;
73     v128_t vacc2x4567 = vacc0x4567;
74     v128_t vacc3x0123 = vacc0x0123;
75     v128_t vacc3x4567 = vacc0x4567;
76     v128_t vacc4x0123 = vacc0x0123;
77     v128_t vacc4x4567 = vacc0x4567;
78     w += 8;
79 
80     size_t k = kc;
81     while (k >= 4 * sizeof(float)) {
82       v128_t va0 = wasm_v128_load(a0);
83       a0 += 4;
84       v128_t va1 = wasm_v128_load(a1);
85       a1 += 4;
86       v128_t va2 = wasm_v128_load(a2);
87       a2 += 4;
88       v128_t va3 = wasm_v128_load(a3);
89       a3 += 4;
90       v128_t va4 = wasm_v128_load(a4);
91       a4 += 4;
92 
93 
94       const v128_t vb0123c0 = wasm_v128_load(w + 0);
95       const v128_t vb4567c0 = wasm_v128_load(w + 4);
96 
97       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
98       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
99       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
100       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c0));
101       vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c0));
102       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
103       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
104       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
105       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c0));
106       vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c0));
107 
108       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
109       va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
110       va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
111       va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
112       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
113 
114       const v128_t vb0123c1 = wasm_v128_load(w + 8);
115       const v128_t vb4567c1 = wasm_v128_load(w + 12);
116 
117       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
118       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
119       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
120       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c1));
121       vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c1));
122       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
123       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
124       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
125       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c1));
126       vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c1));
127 
128       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
129       va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
130       va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
131       va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
132       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
133 
134       const v128_t vb0123c2 = wasm_v128_load(w + 16);
135       const v128_t vb4567c2 = wasm_v128_load(w + 20);
136 
137       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
138       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
139       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
140       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c2));
141       vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c2));
142       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
143       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
144       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
145       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c2));
146       vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c2));
147 
148       va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
149       va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
150       va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
151       va3 = wasm_v32x4_shuffle(va3, va3, 1, 2, 3, 0);
152       va4 = wasm_v32x4_shuffle(va4, va4, 1, 2, 3, 0);
153 
154       const v128_t vb0123c3 = wasm_v128_load(w + 24);
155       const v128_t vb4567c3 = wasm_v128_load(w + 28);
156 
157       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
158       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
159       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
160       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123c3));
161       vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123c3));
162       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
163       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
164       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
165       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567c3));
166       vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567c3));
167 
168 
169       w += 32;
170       k -= 4 * sizeof(float);
171     }
172     if XNN_UNLIKELY(k != 0) {
173       do {
174         const v128_t va0 = wasm_v32x4_load_splat(a0);
175         a0 += 1;
176         const v128_t va1 = wasm_v32x4_load_splat(a1);
177         a1 += 1;
178         const v128_t va2 = wasm_v32x4_load_splat(a2);
179         a2 += 1;
180         const v128_t va3 = wasm_v32x4_load_splat(a3);
181         a3 += 1;
182         const v128_t va4 = wasm_v32x4_load_splat(a4);
183         a4 += 1;
184 
185         const v128_t vb0123 = wasm_v128_load(w);
186         const v128_t vb4567 = wasm_v128_load(w + 4);
187         w += 8;
188 
189         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
190         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
191         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
192         vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
193         vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4, vb0123));
194         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
195         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
196         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
197         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
198         vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4, vb4567));
199 
200         k -= sizeof(float);
201       } while (k != 0);
202     }
203 
204     vacc0x0123 = wasm_f32x4_max(vacc0x0123, vmin);
205     vacc1x0123 = wasm_f32x4_max(vacc1x0123, vmin);
206     vacc2x0123 = wasm_f32x4_max(vacc2x0123, vmin);
207     vacc3x0123 = wasm_f32x4_max(vacc3x0123, vmin);
208     vacc4x0123 = wasm_f32x4_max(vacc4x0123, vmin);
209     vacc0x4567 = wasm_f32x4_max(vacc0x4567, vmin);
210     vacc1x4567 = wasm_f32x4_max(vacc1x4567, vmin);
211     vacc2x4567 = wasm_f32x4_max(vacc2x4567, vmin);
212     vacc3x4567 = wasm_f32x4_max(vacc3x4567, vmin);
213     vacc4x4567 = wasm_f32x4_max(vacc4x4567, vmin);
214 
215     vacc0x0123 = wasm_f32x4_min(vacc0x0123, vmax);
216     vacc1x0123 = wasm_f32x4_min(vacc1x0123, vmax);
217     vacc2x0123 = wasm_f32x4_min(vacc2x0123, vmax);
218     vacc3x0123 = wasm_f32x4_min(vacc3x0123, vmax);
219     vacc4x0123 = wasm_f32x4_min(vacc4x0123, vmax);
220     vacc0x4567 = wasm_f32x4_min(vacc0x4567, vmax);
221     vacc1x4567 = wasm_f32x4_min(vacc1x4567, vmax);
222     vacc2x4567 = wasm_f32x4_min(vacc2x4567, vmax);
223     vacc3x4567 = wasm_f32x4_min(vacc3x4567, vmax);
224     vacc4x4567 = wasm_f32x4_min(vacc4x4567, vmax);
225 
226     if XNN_LIKELY(nc >= 8) {
227       wasm_v128_store(c4, vacc4x0123);
228       wasm_v128_store(c4 + 4, vacc4x4567);
229       c4 = (float*) ((uintptr_t) c4 + cn_stride);
230       wasm_v128_store(c3, vacc3x0123);
231       wasm_v128_store(c3 + 4, vacc3x4567);
232       c3 = (float*) ((uintptr_t) c3 + cn_stride);
233       wasm_v128_store(c2, vacc2x0123);
234       wasm_v128_store(c2 + 4, vacc2x4567);
235       c2 = (float*) ((uintptr_t) c2 + cn_stride);
236       wasm_v128_store(c1, vacc1x0123);
237       wasm_v128_store(c1 + 4, vacc1x4567);
238       c1 = (float*) ((uintptr_t) c1 + cn_stride);
239       wasm_v128_store(c0, vacc0x0123);
240       wasm_v128_store(c0 + 4, vacc0x4567);
241       c0 = (float*) ((uintptr_t) c0 + cn_stride);
242 
243       a4 = (const float*) ((uintptr_t) a4 - kc);
244       a3 = (const float*) ((uintptr_t) a3 - kc);
245       a2 = (const float*) ((uintptr_t) a2 - kc);
246       a1 = (const float*) ((uintptr_t) a1 - kc);
247       a0 = (const float*) ((uintptr_t) a0 - kc);
248 
249       nc -= 8;
250     } else {
251       if (nc & 4) {
252         wasm_v128_store(c4, vacc4x0123);
253         wasm_v128_store(c3, vacc3x0123);
254         wasm_v128_store(c2, vacc2x0123);
255         wasm_v128_store(c1, vacc1x0123);
256         wasm_v128_store(c0, vacc0x0123);
257 
258         vacc4x0123 = vacc4x4567;
259         vacc3x0123 = vacc3x4567;
260         vacc2x0123 = vacc2x4567;
261         vacc1x0123 = vacc1x4567;
262         vacc0x0123 = vacc0x4567;
263 
264         c4 += 4;
265         c3 += 4;
266         c2 += 4;
267         c1 += 4;
268         c0 += 4;
269       }
270       if (nc & 2) {
271         *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
272         *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
273         *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
274         *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
275         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
276 
277         vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
278         vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
279         vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
280         vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
281         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
282 
283         c4 += 2;
284         c3 += 2;
285         c2 += 2;
286         c1 += 2;
287         c0 += 2;
288       }
289       if (nc & 1) {
290         *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
291         *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
292         *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
293         *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
294         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
295       }
296 
297       nc = 0;
298     }
299   } while (nc != 0);
300 }
301