• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_4x8__relaxedwasmsimd_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_4x8__relaxedwasmsimd_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 <= 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       const v128_t va0 = wasm_v128_load(a0);
77       a0 += 4;
78       const v128_t va1 = wasm_v128_load(a1);
79       a1 += 4;
80       const v128_t va2 = wasm_v128_load(a2);
81       a2 += 4;
82       const v128_t va3 = wasm_v128_load(a3);
83       a3 += 4;
84 
85       const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
86       const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
87       const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
88       const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
89 
90       const v128_t vb0123c0 = wasm_v128_load(w + 0);
91       const v128_t vb4567c0 = wasm_v128_load(w + 4);
92 
93       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
94       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
95       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
96       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
97       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
98       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
99       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
100       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
101       const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
102       const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
103       const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
104       const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
105 
106       const v128_t vb0123c1 = wasm_v128_load(w + 8);
107       const v128_t vb4567c1 = wasm_v128_load(w + 12);
108 
109       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
110       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
111       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
112       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
113       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
114       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
115       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
116       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
117       const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
118       const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
119       const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
120       const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
121 
122       const v128_t vb0123c2 = wasm_v128_load(w + 16);
123       const v128_t vb4567c2 = wasm_v128_load(w + 20);
124 
125       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
126       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
127       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
128       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
129       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
130       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
131       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
132       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
133       const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
134       const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
135       const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
136       const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
137 
138       const v128_t vb0123c3 = wasm_v128_load(w + 24);
139       const v128_t vb4567c3 = wasm_v128_load(w + 28);
140 
141       vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
142       vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
143       vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
144       vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
145       vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
146       vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
147       vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
148       vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
149 
150       w += 32;
151       k -= 4 * sizeof(float);
152     }
153     if XNN_UNLIKELY(k != 0) {
154       do {
155         const v128_t va0 = wasm_v128_load32_splat(a0);
156         a0 += 1;
157         const v128_t va1 = wasm_v128_load32_splat(a1);
158         a1 += 1;
159         const v128_t va2 = wasm_v128_load32_splat(a2);
160         a2 += 1;
161         const v128_t va3 = wasm_v128_load32_splat(a3);
162         a3 += 1;
163 
164         const v128_t vb0123 = wasm_v128_load(w);
165         const v128_t vb4567 = wasm_v128_load(w + 4);
166         w += 8;
167 
168         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
169         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
170         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
171         vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
172         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
173         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
174         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
175         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
176 
177         k -= sizeof(float);
178       } while (k != 0);
179     }
180 
181     vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
182     vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
183     vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
184     vacc3x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x0123);
185     vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
186     vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
187     vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
188     vacc3x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc3x4567);
189 
190     vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
191     vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
192     vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
193     vacc3x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x0123);
194     vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
195     vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
196     vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
197     vacc3x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc3x4567);
198 
199     if XNN_LIKELY(nc >= 8) {
200       wasm_v128_store(c3, vacc3x0123);
201       wasm_v128_store(c3 + 4, vacc3x4567);
202       c3 = (float*) ((uintptr_t) c3 + cn_stride);
203       wasm_v128_store(c2, vacc2x0123);
204       wasm_v128_store(c2 + 4, vacc2x4567);
205       c2 = (float*) ((uintptr_t) c2 + cn_stride);
206       wasm_v128_store(c1, vacc1x0123);
207       wasm_v128_store(c1 + 4, vacc1x4567);
208       c1 = (float*) ((uintptr_t) c1 + cn_stride);
209       wasm_v128_store(c0, vacc0x0123);
210       wasm_v128_store(c0 + 4, vacc0x4567);
211       c0 = (float*) ((uintptr_t) c0 + cn_stride);
212 
213       a3 = (const float*) ((uintptr_t) a3 - kc);
214       a2 = (const float*) ((uintptr_t) a2 - kc);
215       a1 = (const float*) ((uintptr_t) a1 - kc);
216       a0 = (const float*) ((uintptr_t) a0 - kc);
217 
218       nc -= 8;
219     } else {
220       if (nc & 4) {
221         wasm_v128_store(c3, vacc3x0123);
222         wasm_v128_store(c2, vacc2x0123);
223         wasm_v128_store(c1, vacc1x0123);
224         wasm_v128_store(c0, vacc0x0123);
225 
226         vacc3x0123 = vacc3x4567;
227         vacc2x0123 = vacc2x4567;
228         vacc1x0123 = vacc1x4567;
229         vacc0x0123 = vacc0x4567;
230 
231         c3 += 4;
232         c2 += 4;
233         c1 += 4;
234         c0 += 4;
235       }
236       if (nc & 2) {
237         *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
238         *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
239         *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
240         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
241 
242         vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
243         vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
244         vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
245         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
246 
247         c3 += 2;
248         c2 += 2;
249         c1 += 2;
250         c0 += 2;
251       }
252       if (nc & 1) {
253         *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
254         *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
255         *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
256         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
257       }
258 
259       nc = 0;
260     }
261   } while (nc != 0);
262 }
263