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_relu_ukernel_3x8s4__wasmsimd(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_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_relu_ukernel_3x8s4__wasmsimd(
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_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 3);
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
53 do {
54 v128_t vacc0x0123 = wasm_v128_load(w + 0);
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 k = kc;
63 while (k >= 4 * sizeof(float)) {
64 v128_t va0 = wasm_v128_load(a0);
65 a0 += 4;
66 v128_t va1 = wasm_v128_load(a1);
67 a1 += 4;
68 v128_t va2 = wasm_v128_load(a2);
69 a2 += 4;
70
71
72 const v128_t vb0123c0 = wasm_v128_load(w + 0);
73 const v128_t vb4567c0 = wasm_v128_load(w + 4);
74
75 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
76 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
77 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
78 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
79 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
80 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
81
82 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
83 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
84 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
85
86 const v128_t vb0123c1 = wasm_v128_load(w + 8);
87 const v128_t vb4567c1 = wasm_v128_load(w + 12);
88
89 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
90 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
91 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
92 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
93 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
94 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
95
96 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
97 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
98 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
99
100 const v128_t vb0123c2 = wasm_v128_load(w + 16);
101 const v128_t vb4567c2 = wasm_v128_load(w + 20);
102
103 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
104 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
105 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
106 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
107 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
108 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
109
110 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
111 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
112 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
113
114 const v128_t vb0123c3 = wasm_v128_load(w + 24);
115 const v128_t vb4567c3 = wasm_v128_load(w + 28);
116
117 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
118 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
119 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
120 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
121 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
122 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
123
124
125 w += 32;
126 k -= 4 * sizeof(float);
127 }
128 if XNN_UNLIKELY(k != 0) {
129 v128_t va0 = wasm_v128_load(a0);
130 a0 = (const float*) ((uintptr_t) a0 + k);
131 v128_t va1 = wasm_v128_load(a1);
132 a1 = (const float*) ((uintptr_t) a1 + k);
133 v128_t va2 = wasm_v128_load(a2);
134 a2 = (const float*) ((uintptr_t) a2 + k);
135
136 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
137
138 const v128_t vb0123c0 = wasm_v128_load(w + 0);
139 const v128_t vb4567c0 = wasm_v128_load(w + 4);
140
141 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
142 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
143 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
144 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
145 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
146 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
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
152 const v128_t vb0123c1 = wasm_v128_load(w + 8);
153 const v128_t vb4567c1 = wasm_v128_load(w + 12);
154
155 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
156 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
157 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
158 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
159 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
160 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
161
162 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
163 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
164 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
165
166 const v128_t vb0123c2 = wasm_v128_load(w + 16);
167 const v128_t vb4567c2 = wasm_v128_load(w + 20);
168
169 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
170 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
171 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
172 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
173 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
174 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
175
176 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
177 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
178 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
179
180 const v128_t vb0123c3 = wasm_v128_load(w + 24);
181 const v128_t vb4567c3 = wasm_v128_load(w + 28);
182
183 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
184 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
185 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
186 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
187 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
188 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
189
190
191 w += 32;
192 }
193
194 const v128_t vzero = wasm_i32x4_const_splat(0);
195 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
196 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
197 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
198 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
199 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
200 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
201
202 if XNN_LIKELY(nc >= 8) {
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 a2 = (const float*) ((uintptr_t) a2 - kc);
214 a1 = (const float*) ((uintptr_t) a1 - kc);
215 a0 = (const float*) ((uintptr_t) a0 - kc);
216
217 nc -= 8;
218 } else {
219 if (nc & 4) {
220 wasm_v128_store(c2, vacc2x0123);
221 wasm_v128_store(c1, vacc1x0123);
222 wasm_v128_store(c0, vacc0x0123);
223
224 vacc2x0123 = vacc2x4567;
225 vacc1x0123 = vacc1x4567;
226 vacc0x0123 = vacc0x4567;
227
228 c2 += 4;
229 c1 += 4;
230 c0 += 4;
231 }
232 if (nc & 2) {
233 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
234 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
235 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
236
237 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
238 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
239 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
240
241 c2 += 2;
242 c1 += 2;
243 c0 += 2;
244 }
245 if (nc & 1) {
246 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
247 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
248 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
249 }
250
251 nc = 0;
252 }
253 } while (nc != 0);
254 }
255