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_gemminc_minmax_ukernel_3x8s4__wasmrelaxedsimd(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_3x8s4__wasmrelaxedsimd(
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 <= 3);
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
55 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
56 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
57 do {
58 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
59 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
60 v128_t vacc1x0123 = wasm_v128_load(acc + 8);
61 v128_t vacc1x4567 = wasm_v128_load(acc + 12);
62 v128_t vacc2x0123 = wasm_v128_load(acc + 16);
63 v128_t vacc2x4567 = wasm_v128_load(acc + 20);
64 acc += 24;
65
66 size_t k = kc;
67 while (k >= 4 * sizeof(float)) {
68 v128_t va0 = wasm_v128_load(a0);
69 a0 += 4;
70 v128_t va1 = wasm_v128_load(a1);
71 a1 += 4;
72 v128_t va2 = wasm_v128_load(a2);
73 a2 += 4;
74
75
76 const v128_t vb0123c0 = wasm_v128_load(w + 0);
77 const v128_t vb4567c0 = wasm_v128_load(w + 4);
78
79 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
80 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c0));
81 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c0));
82 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
83 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c0));
84 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c0));
85
86 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
87 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
88 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
89
90 const v128_t vb0123c1 = wasm_v128_load(w + 8);
91 const v128_t vb4567c1 = wasm_v128_load(w + 12);
92
93 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
94 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c1));
95 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c1));
96 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
97 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c1));
98 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c1));
99
100 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
101 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
102 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
103
104 const v128_t vb0123c2 = wasm_v128_load(w + 16);
105 const v128_t vb4567c2 = wasm_v128_load(w + 20);
106
107 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
108 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c2));
109 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c2));
110 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
111 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c2));
112 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c2));
113
114 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
115 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
116 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
117
118 const v128_t vb0123c3 = wasm_v128_load(w + 24);
119 const v128_t vb4567c3 = wasm_v128_load(w + 28);
120
121 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
122 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123c3));
123 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123c3));
124 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
125 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567c3));
126 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567c3));
127
128
129 w += 32;
130 k -= 4 * sizeof(float);
131 }
132 if XNN_UNLIKELY(k != 0) {
133 v128_t va0 = wasm_v128_load(a0);
134 a0 = (const float*) ((uintptr_t) a0 + k);
135 v128_t va1 = wasm_v128_load(a1);
136 a1 = (const float*) ((uintptr_t) a1 + k);
137 v128_t va2 = wasm_v128_load(a2);
138 a2 = (const float*) ((uintptr_t) a2 + k);
139
140 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
141
142 const v128_t vb0123c0 = wasm_v128_load(w + 0);
143 const v128_t vb4567c0 = wasm_v128_load(w + 4);
144
145 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
146 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
147 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
148 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
149 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
150 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
151
152 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
153 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
154 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
155
156 const v128_t vb0123c1 = wasm_v128_load(w + 8);
157 const v128_t vb4567c1 = wasm_v128_load(w + 12);
158
159 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
160 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
161 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
162 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
163 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
164 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
165
166 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
167 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
168 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
169
170 const v128_t vb0123c2 = wasm_v128_load(w + 16);
171 const v128_t vb4567c2 = wasm_v128_load(w + 20);
172
173 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
174 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
175 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
176 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
177 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
178 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
179
180 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
181 va1 = wasm_v32x4_shuffle(va1, va1, 1, 2, 3, 0);
182 va2 = wasm_v32x4_shuffle(va2, va2, 1, 2, 3, 0);
183
184 const v128_t vb0123c3 = wasm_v128_load(w + 24);
185 const v128_t vb4567c3 = wasm_v128_load(w + 28);
186
187 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
188 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
189 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
190 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
191 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(wasm_v128_andnot(va1, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
192 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(wasm_v128_andnot(va2, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
193
194
195 w += 32;
196 }
197
198 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
199 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
200 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
201 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
202 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
203 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
204
205 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
206 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
207 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
208 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
209 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
210 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
211
212 if XNN_LIKELY(nc >= 8) {
213 wasm_v128_store(c2, vacc2x0123);
214 wasm_v128_store(c2 + 4, vacc2x4567);
215 c2 = (float*) ((uintptr_t) c2 + cn_stride);
216 wasm_v128_store(c1, vacc1x0123);
217 wasm_v128_store(c1 + 4, vacc1x4567);
218 c1 = (float*) ((uintptr_t) c1 + cn_stride);
219 wasm_v128_store(c0, vacc0x0123);
220 wasm_v128_store(c0 + 4, vacc0x4567);
221 c0 = (float*) ((uintptr_t) c0 + cn_stride);
222
223 a2 = (const float*) ((uintptr_t) a2 - kc);
224 a1 = (const float*) ((uintptr_t) a1 - kc);
225 a0 = (const float*) ((uintptr_t) a0 - kc);
226
227 nc -= 8;
228 } else {
229 if (nc & 4) {
230 wasm_v128_store(c2, vacc2x0123);
231 wasm_v128_store(c1, vacc1x0123);
232 wasm_v128_store(c0, vacc0x0123);
233
234 vacc2x0123 = vacc2x4567;
235 vacc1x0123 = vacc1x4567;
236 vacc0x0123 = vacc0x4567;
237
238 c2 += 4;
239 c1 += 4;
240 c0 += 4;
241 }
242 if (nc & 2) {
243 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
244 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
245 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
246
247 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
248 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
249 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
250
251 c2 += 2;
252 c1 += 2;
253 c0 += 2;
254 }
255 if (nc & 1) {
256 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
257 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
258 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
259 }
260
261 nc = 0;
262 }
263 } while (nc != 0);
264 }
265