1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_relu_ukernel_3x8s4__wasmsimd(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_3x8s4__wasmsimd(
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)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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
53 do {
54 v128_t vacc0x0123 = wasm_v128_load(w);
55 v128_t vacc0x4567 = wasm_v128_load(w + 4);
56 v128_t vacc1x0123 = vacc0x0123;
57 v128_t vacc1x4567 = vacc0x4567;
58 v128_t vacc2x0123 = vacc0x0123;
59 v128_t vacc2x4567 = vacc0x4567;
60 w += 8;
61
62 size_t p = ks;
63 do {
64 const float* restrict a0 = a[0];
65 assert(a0 != NULL);
66 if XNN_UNPREDICTABLE(a0 != zero) {
67 a0 = (const float*) ((uintptr_t) a0 + a_offset);
68 }
69 const float* restrict a1 = a[1];
70 assert(a1 != NULL);
71 if XNN_UNPREDICTABLE(a1 != zero) {
72 a1 = (const float*) ((uintptr_t) a1 + a_offset);
73 }
74 const float* restrict a2 = a[2];
75 assert(a2 != NULL);
76 if XNN_UNPREDICTABLE(a2 != zero) {
77 a2 = (const float*) ((uintptr_t) a2 + a_offset);
78 }
79 a += 3;
80
81 size_t k = kc;
82 while (k >= 4 * sizeof(float)) {
83 v128_t va0 = wasm_v128_load(a0);
84 a0 += 4;
85 v128_t va1 = wasm_v128_load(a1);
86 a1 += 4;
87 v128_t va2 = wasm_v128_load(a2);
88 a2 += 4;
89
90
91 const v128_t vb0123c0 = wasm_v128_load(w + 0);
92 const v128_t vb4567c0 = wasm_v128_load(w + 4);
93
94 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
95 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
96 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
97 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
98 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
99 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
100
101 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
102 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
103 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
104
105 const v128_t vb0123c1 = wasm_v128_load(w + 8);
106 const v128_t vb4567c1 = wasm_v128_load(w + 12);
107
108 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
109 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
110 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
111 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
112 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
113 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, 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
119 const v128_t vb0123c2 = wasm_v128_load(w + 16);
120 const v128_t vb4567c2 = wasm_v128_load(w + 20);
121
122 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
123 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
124 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
125 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
126 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
127 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
128
129 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
130 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
131 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
132
133 const v128_t vb0123c3 = wasm_v128_load(w + 24);
134 const v128_t vb4567c3 = wasm_v128_load(w + 28);
135
136 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
137 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
138 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
139 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
140 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
141 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
142
143
144 w += 32;
145 k -= 4 * sizeof(float);
146 }
147 if XNN_UNLIKELY(k != 0) {
148 v128_t va0 = wasm_v128_load(a0);
149 a0 = (const float*) ((uintptr_t) a0 + k);
150 v128_t va1 = wasm_v128_load(a1);
151 a1 = (const float*) ((uintptr_t) a1 + k);
152 v128_t va2 = wasm_v128_load(a2);
153 a2 = (const float*) ((uintptr_t) a2 + k);
154
155 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
156
157 const v128_t vb0123c0 = wasm_v128_load(w + 0);
158 const v128_t vb4567c0 = wasm_v128_load(w + 4);
159
160 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
161 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
162 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
163 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
164 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
165 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
166
167 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
168 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
169 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
170
171 const v128_t vb0123c1 = wasm_v128_load(w + 8);
172 const v128_t vb4567c1 = wasm_v128_load(w + 12);
173
174 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
175 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
176 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
177 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
178 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
179 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
180
181 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
182 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
183 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
184
185 const v128_t vb0123c2 = wasm_v128_load(w + 16);
186 const v128_t vb4567c2 = wasm_v128_load(w + 20);
187
188 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
189 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
190 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
191 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
192 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
193 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
194
195 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
196 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
197 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
198
199 const v128_t vb0123c3 = wasm_v128_load(w + 24);
200 const v128_t vb4567c3 = wasm_v128_load(w + 28);
201
202 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
203 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
204 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
205 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
206 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
207 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
208
209
210 w += 32;
211 }
212 p -= 3 * sizeof(void*);
213 } while (p != 0);
214
215 const v128_t vzero = wasm_i32x4_const_splat(0);
216 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
217 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
218 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
219 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
220 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
221 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
222
223 if XNN_LIKELY(nc >= 8) {
224 wasm_v128_store(c2, vacc2x0123);
225 wasm_v128_store(c2 + 4, vacc2x4567);
226 c2 = (float*) ((uintptr_t) c2 + cn_stride);
227 wasm_v128_store(c1, vacc1x0123);
228 wasm_v128_store(c1 + 4, vacc1x4567);
229 c1 = (float*) ((uintptr_t) c1 + cn_stride);
230 wasm_v128_store(c0, vacc0x0123);
231 wasm_v128_store(c0 + 4, vacc0x4567);
232 c0 = (float*) ((uintptr_t) c0 + cn_stride);
233
234 a = (const float**restrict) ((uintptr_t) a - ks);
235 nc -= 8;
236 } else {
237 if (nc & 4) {
238 wasm_v128_store(c2, vacc2x0123);
239 wasm_v128_store(c1, vacc1x0123);
240 wasm_v128_store(c0, vacc0x0123);
241
242 vacc2x0123 = vacc2x4567;
243 vacc1x0123 = vacc1x4567;
244 vacc0x0123 = vacc0x4567;
245
246 c2 += 4;
247 c1 += 4;
248 c0 += 4;
249 }
250 if (nc & 2) {
251 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
252 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
253 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
254
255 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
256 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
257 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
258
259 c2 += 2;
260 c1 += 2;
261 c0 += 2;
262 }
263 if (nc & 1) {
264 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
265 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
266 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
267 }
268
269 nc = 0;
270 }
271 } while (nc != 0);
272 }
273