• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-igemm/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/igemm.h>
15 
16 
xnn_f32_igemm_relu_ukernel_4x8__wasmsimd_splat(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_relu_ukernel_4x8__wasmsimd_splat(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const float**restrict a,
23     const float*restrict w,
24     float*restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const float* zero,
29     const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31   assert(mr != 0);
32   assert(mr <= 4);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(float) == 0);
36   assert(ks != 0);
37   assert(ks % (4 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(float) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   float* c0 = c;
44   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52   float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53   if XNN_UNPREDICTABLE(mr != 4) {
54     c3 = c2;
55   }
56 
57   do {
58     v128_t vacc0x0123 = wasm_v128_load(w);
59     v128_t vacc0x4567 = wasm_v128_load(w + 4);
60     v128_t vacc1x0123 = vacc0x0123;
61     v128_t vacc1x4567 = vacc0x4567;
62     v128_t vacc2x0123 = vacc0x0123;
63     v128_t vacc2x4567 = vacc0x4567;
64     v128_t vacc3x0123 = vacc0x0123;
65     v128_t vacc3x4567 = vacc0x4567;
66     w += 8;
67 
68     size_t p = ks;
69     do {
70       const float* restrict a0 = a[0];
71       assert(a0 != NULL);
72       if XNN_UNPREDICTABLE(a0 != zero) {
73         a0 = (const float*) ((uintptr_t) a0 + a_offset);
74       }
75       const float* restrict a1 = a[1];
76       assert(a1 != NULL);
77       if XNN_UNPREDICTABLE(a1 != zero) {
78         a1 = (const float*) ((uintptr_t) a1 + a_offset);
79       }
80       const float* restrict a2 = a[2];
81       assert(a2 != NULL);
82       if XNN_UNPREDICTABLE(a2 != zero) {
83         a2 = (const float*) ((uintptr_t) a2 + a_offset);
84       }
85       const float* restrict a3 = a[3];
86       assert(a3 != NULL);
87       if XNN_UNPREDICTABLE(a3 != zero) {
88         a3 = (const float*) ((uintptr_t) a3 + a_offset);
89       }
90       a += 4;
91 
92       size_t k = kc;
93       while (k >= 4 * sizeof(float)) {
94         const v128_t va0 = wasm_v128_load(a0);
95         a0 += 4;
96         const v128_t va1 = wasm_v128_load(a1);
97         a1 += 4;
98         const v128_t va2 = wasm_v128_load(a2);
99         a2 += 4;
100         const v128_t va3 = wasm_v128_load(a3);
101         a3 += 4;
102 
103         const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
104         const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
105         const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
106         const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 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         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
116         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
117         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
118         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
119         const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
120         const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
121         const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
122         const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
123 
124         const v128_t vb0123c1 = wasm_v128_load(w + 8);
125         const v128_t vb4567c1 = wasm_v128_load(w + 12);
126 
127         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
128         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
129         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
130         vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
131         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
132         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
133         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
134         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
135         const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
136         const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
137         const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
138         const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
139 
140         const v128_t vb0123c2 = wasm_v128_load(w + 16);
141         const v128_t vb4567c2 = wasm_v128_load(w + 20);
142 
143         vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
144         vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
145         vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
146         vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
147         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
148         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
149         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
150         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
151         const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
152         const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
153         const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
154         const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 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         vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
164         vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
165         vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
166         vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
167 
168         w += 32;
169         k -= 4 * sizeof(float);
170       }
171       if XNN_UNLIKELY(k != 0) {
172         do {
173           const v128_t vb0123 = wasm_v128_load(w);
174           const v128_t vb4567 = wasm_v128_load(w + 4);
175           w += 8;
176 
177           const v128_t va0 = wasm_v128_load32_splat(a0);
178           a0 += 1;
179           const v128_t va1 = wasm_v128_load32_splat(a1);
180           a1 += 1;
181           const v128_t va2 = wasm_v128_load32_splat(a2);
182           a2 += 1;
183           const v128_t va3 = wasm_v128_load32_splat(a3);
184           a3 += 1;
185 
186           vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
187           vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
188           vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
189           vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
190           vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
191           vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
192           vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
193           vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
194           k -= sizeof(float);
195         } while (k != 0);
196       }
197       p -= 4 * sizeof(void*);
198     } while (p != 0);
199 
200     const v128_t vzero = wasm_i32x4_const_splat(0);
201     vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
202     vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
203     vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
204     vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
205     vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
206     vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
207     vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
208     vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
209 
210     if XNN_LIKELY(nc >= 8) {
211       wasm_v128_store(c3, vacc3x0123);
212       wasm_v128_store(c3 + 4, vacc3x4567);
213       c3 = (float*) ((uintptr_t) c3 + cn_stride);
214       wasm_v128_store(c2, vacc2x0123);
215       wasm_v128_store(c2 + 4, vacc2x4567);
216       c2 = (float*) ((uintptr_t) c2 + cn_stride);
217       wasm_v128_store(c1, vacc1x0123);
218       wasm_v128_store(c1 + 4, vacc1x4567);
219       c1 = (float*) ((uintptr_t) c1 + cn_stride);
220       wasm_v128_store(c0, vacc0x0123);
221       wasm_v128_store(c0 + 4, vacc0x4567);
222       c0 = (float*) ((uintptr_t) c0 + cn_stride);
223 
224       a = (const float**restrict) ((uintptr_t) a - ks);
225       nc -= 8;
226     } else {
227       if (nc & 4) {
228         wasm_v128_store(c3, vacc3x0123);
229         wasm_v128_store(c2, vacc2x0123);
230         wasm_v128_store(c1, vacc1x0123);
231         wasm_v128_store(c0, vacc0x0123);
232 
233         vacc3x0123 = vacc3x4567;
234         vacc2x0123 = vacc2x4567;
235         vacc1x0123 = vacc1x4567;
236         vacc0x0123 = vacc0x4567;
237 
238         c3 += 4;
239         c2 += 4;
240         c1 += 4;
241         c0 += 4;
242       }
243       if (nc & 2) {
244         *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
245         *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
246         *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
247         *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
248 
249         vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
250         vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
251         vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
252         vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
253 
254         c3 += 2;
255         c2 += 2;
256         c1 += 2;
257         c0 += 2;
258       }
259       if (nc & 1) {
260         *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
261         *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
262         *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
263         *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
264       }
265 
266       nc = 0;
267     }
268   } while (nc != 0);
269 }
270