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_gemm_relu_ukernel_5x8__wasmrelaxedsimd_fma_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 union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_relu_ukernel_5x8__wasmrelaxedsimd_fma_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 union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 5);
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 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 a4 = a3;
62 c4 = c3;
63 }
64
65 do {
66 v128_t vacc0x0123 = wasm_v128_load(w + 0);
67 v128_t vacc0x4567 = wasm_v128_load(w + 4);
68 v128_t vacc1x0123 = vacc0x0123;
69 v128_t vacc1x4567 = vacc0x4567;
70 v128_t vacc2x0123 = vacc0x0123;
71 v128_t vacc2x4567 = vacc0x4567;
72 v128_t vacc3x0123 = vacc0x0123;
73 v128_t vacc3x4567 = vacc0x4567;
74 v128_t vacc4x0123 = vacc0x0123;
75 v128_t vacc4x4567 = vacc0x4567;
76 w += 8;
77
78 size_t k = kc;
79 while (k >= 4 * sizeof(float)) {
80 const v128_t va0 = wasm_v128_load(a0);
81 a0 += 4;
82 const v128_t va1 = wasm_v128_load(a1);
83 a1 += 4;
84 const v128_t va2 = wasm_v128_load(a2);
85 a2 += 4;
86 const v128_t va3 = wasm_v128_load(a3);
87 a3 += 4;
88 const v128_t va4 = wasm_v128_load(a4);
89 a4 += 4;
90
91 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
92 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
93 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
94 const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
95 const v128_t va4c0 = wasm_v32x4_shuffle(va4, va4, 0, 0, 0, 0);
96
97 const v128_t vb0123c0 = wasm_v128_load(w + 0);
98 const v128_t vb4567c0 = wasm_v128_load(w + 4);
99
100 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
101 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
102 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
103 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
104 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c0, vb0123c0));
105 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
106 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
107 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
108 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
109 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c0, vb4567c0));
110 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
111 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
112 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
113 const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
114 const v128_t va4c1 = wasm_v32x4_shuffle(va4, va4, 1, 1, 1, 1);
115
116 const v128_t vb0123c1 = wasm_v128_load(w + 8);
117 const v128_t vb4567c1 = wasm_v128_load(w + 12);
118
119 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
120 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
121 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
122 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
123 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c1, vb0123c1));
124 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
125 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
126 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
127 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
128 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c1, vb4567c1));
129 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
130 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
131 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
132 const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
133 const v128_t va4c2 = wasm_v32x4_shuffle(va4, va4, 2, 2, 2, 2);
134
135 const v128_t vb0123c2 = wasm_v128_load(w + 16);
136 const v128_t vb4567c2 = wasm_v128_load(w + 20);
137
138 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
139 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
140 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
141 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
142 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c2, vb0123c2));
143 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
144 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
145 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
146 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
147 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c2, vb4567c2));
148 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
149 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
150 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
151 const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
152 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3);
153
154 const v128_t vb0123c3 = wasm_v128_load(w + 24);
155 const v128_t vb4567c3 = wasm_v128_load(w + 28);
156
157 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
158 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
159 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
160 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
161 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3));
162 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
163 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
164 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
165 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
166 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3));
167
168 w += 32;
169 k -= 4 * sizeof(float);
170 }
171 if XNN_UNLIKELY(k != 0) {
172 do {
173 const v128_t va0 = wasm_v128_load32_splat(a0);
174 a0 += 1;
175 const v128_t va1 = wasm_v128_load32_splat(a1);
176 a1 += 1;
177 const v128_t va2 = wasm_v128_load32_splat(a2);
178 a2 += 1;
179 const v128_t va3 = wasm_v128_load32_splat(a3);
180 a3 += 1;
181 const v128_t va4 = wasm_v128_load32_splat(a4);
182 a4 += 1;
183
184 const v128_t vb0123 = wasm_v128_load(w);
185 const v128_t vb4567 = wasm_v128_load(w + 4);
186 w += 8;
187
188 vacc0x0123 = __builtin_wasm_fma_f32x4(vacc0x0123, va0, vb0123);
189 vacc1x0123 = __builtin_wasm_fma_f32x4(vacc1x0123, va1, vb0123);
190 vacc2x0123 = __builtin_wasm_fma_f32x4(vacc2x0123, va2, vb0123);
191 vacc3x0123 = __builtin_wasm_fma_f32x4(vacc3x0123, va3, vb0123);
192 vacc4x0123 = __builtin_wasm_fma_f32x4(vacc4x0123, va4, vb0123);
193 vacc0x4567 = __builtin_wasm_fma_f32x4(vacc0x4567, va0, vb4567);
194 vacc1x4567 = __builtin_wasm_fma_f32x4(vacc1x4567, va1, vb4567);
195 vacc2x4567 = __builtin_wasm_fma_f32x4(vacc2x4567, va2, vb4567);
196 vacc3x4567 = __builtin_wasm_fma_f32x4(vacc3x4567, va3, vb4567);
197 vacc4x4567 = __builtin_wasm_fma_f32x4(vacc4x4567, va4, vb4567);
198
199 k -= sizeof(float);
200 } while (k != 0);
201 }
202
203 const v128_t vzero = wasm_i32x4_const_splat(0);
204 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
205 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
206 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
207 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
208 vacc4x0123 = wasm_i32x4_max(vacc4x0123, vzero);
209 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
210 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
211 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
212 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
213 vacc4x4567 = wasm_i32x4_max(vacc4x4567, vzero);
214
215 if XNN_LIKELY(nc >= 8) {
216 wasm_v128_store(c4, vacc4x0123);
217 wasm_v128_store(c4 + 4, vacc4x4567);
218 c4 = (float*) ((uintptr_t) c4 + cn_stride);
219 wasm_v128_store(c3, vacc3x0123);
220 wasm_v128_store(c3 + 4, vacc3x4567);
221 c3 = (float*) ((uintptr_t) c3 + cn_stride);
222 wasm_v128_store(c2, vacc2x0123);
223 wasm_v128_store(c2 + 4, vacc2x4567);
224 c2 = (float*) ((uintptr_t) c2 + cn_stride);
225 wasm_v128_store(c1, vacc1x0123);
226 wasm_v128_store(c1 + 4, vacc1x4567);
227 c1 = (float*) ((uintptr_t) c1 + cn_stride);
228 wasm_v128_store(c0, vacc0x0123);
229 wasm_v128_store(c0 + 4, vacc0x4567);
230 c0 = (float*) ((uintptr_t) c0 + cn_stride);
231
232 a4 = (const float*) ((uintptr_t) a4 - kc);
233 a3 = (const float*) ((uintptr_t) a3 - kc);
234 a2 = (const float*) ((uintptr_t) a2 - kc);
235 a1 = (const float*) ((uintptr_t) a1 - kc);
236 a0 = (const float*) ((uintptr_t) a0 - kc);
237
238 nc -= 8;
239 } else {
240 if (nc & 4) {
241 wasm_v128_store(c4, vacc4x0123);
242 wasm_v128_store(c3, vacc3x0123);
243 wasm_v128_store(c2, vacc2x0123);
244 wasm_v128_store(c1, vacc1x0123);
245 wasm_v128_store(c0, vacc0x0123);
246
247 vacc4x0123 = vacc4x4567;
248 vacc3x0123 = vacc3x4567;
249 vacc2x0123 = vacc2x4567;
250 vacc1x0123 = vacc1x4567;
251 vacc0x0123 = vacc0x4567;
252
253 c4 += 4;
254 c3 += 4;
255 c2 += 4;
256 c1 += 4;
257 c0 += 4;
258 }
259 if (nc & 2) {
260 *((double*) c4) = wasm_f64x2_extract_lane(vacc4x0123, 0);
261 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
262 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
263 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
264 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
265
266 vacc4x0123 = wasm_v32x4_shuffle(vacc4x0123, vacc4x0123, 2, 3, 2, 3);
267 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
268 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
269 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
270 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
271
272 c4 += 2;
273 c3 += 2;
274 c2 += 2;
275 c1 += 2;
276 c0 += 2;
277 }
278 if (nc & 1) {
279 *c4 = wasm_f32x4_extract_lane(vacc4x0123, 0);
280 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
281 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
282 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
283 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
284 }
285
286 nc = 0;
287 }
288 } while (nc != 0);
289 }
290