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_4x8__wasmsimd_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_4x8__wasmsimd_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 <= 4);
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
59 do {
60 v128_t vacc0x0123 = wasm_v128_load(w + 0);
61 v128_t vacc0x4567 = wasm_v128_load(w + 4);
62 v128_t vacc1x0123 = vacc0x0123;
63 v128_t vacc1x4567 = vacc0x4567;
64 v128_t vacc2x0123 = vacc0x0123;
65 v128_t vacc2x4567 = vacc0x4567;
66 v128_t vacc3x0123 = vacc0x0123;
67 v128_t vacc3x4567 = vacc0x4567;
68 w += 8;
69
70 size_t k = kc;
71 while (k >= 4 * sizeof(float)) {
72 const v128_t va0 = wasm_v128_load(a0);
73 a0 += 4;
74 const v128_t va1 = wasm_v128_load(a1);
75 a1 += 4;
76 const v128_t va2 = wasm_v128_load(a2);
77 a2 += 4;
78 const v128_t va3 = wasm_v128_load(a3);
79 a3 += 4;
80
81 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
82 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
83 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
84 const v128_t va3c0 = wasm_v32x4_shuffle(va3, va3, 0, 0, 0, 0);
85
86 const v128_t vb0123c0 = wasm_v128_load(w + 0);
87 const v128_t vb4567c0 = wasm_v128_load(w + 4);
88
89 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
90 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
91 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
92 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c0, vb0123c0));
93 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
94 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
95 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
96 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c0, vb4567c0));
97 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
98 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
99 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
100 const v128_t va3c1 = wasm_v32x4_shuffle(va3, va3, 1, 1, 1, 1);
101
102 const v128_t vb0123c1 = wasm_v128_load(w + 8);
103 const v128_t vb4567c1 = wasm_v128_load(w + 12);
104
105 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
106 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
107 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
108 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c1, vb0123c1));
109 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
110 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
111 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
112 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c1, vb4567c1));
113 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
114 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
115 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
116 const v128_t va3c2 = wasm_v32x4_shuffle(va3, va3, 2, 2, 2, 2);
117
118 const v128_t vb0123c2 = wasm_v128_load(w + 16);
119 const v128_t vb4567c2 = wasm_v128_load(w + 20);
120
121 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
122 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
123 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
124 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c2, vb0123c2));
125 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
126 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
127 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
128 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c2, vb4567c2));
129 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
130 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
131 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
132 const v128_t va3c3 = wasm_v32x4_shuffle(va3, va3, 3, 3, 3, 3);
133
134 const v128_t vb0123c3 = wasm_v128_load(w + 24);
135 const v128_t vb4567c3 = wasm_v128_load(w + 28);
136
137 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
138 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
139 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
140 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3c3, vb0123c3));
141 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
142 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
143 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
144 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3c3, vb4567c3));
145
146 w += 32;
147 k -= 4 * sizeof(float);
148 }
149 if XNN_UNLIKELY(k != 0) {
150 do {
151 const v128_t va0 = wasm_v128_load32_splat(a0);
152 a0 += 1;
153 const v128_t va1 = wasm_v128_load32_splat(a1);
154 a1 += 1;
155 const v128_t va2 = wasm_v128_load32_splat(a2);
156 a2 += 1;
157 const v128_t va3 = wasm_v128_load32_splat(a3);
158 a3 += 1;
159
160 const v128_t vb0123 = wasm_v128_load(w);
161 const v128_t vb4567 = wasm_v128_load(w + 4);
162 w += 8;
163
164 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
165 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
166 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
167 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(va3, vb0123));
168 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
169 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
170 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
171 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(va3, vb4567));
172
173 k -= sizeof(float);
174 } while (k != 0);
175 }
176
177 const v128_t vzero = wasm_i32x4_const_splat(0);
178 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
179 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
180 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
181 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vzero);
182 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
183 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
184 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
185 vacc3x4567 = wasm_i32x4_max(vacc3x4567, vzero);
186
187 if XNN_LIKELY(nc >= 8) {
188 wasm_v128_store(c3, vacc3x0123);
189 wasm_v128_store(c3 + 4, vacc3x4567);
190 c3 = (float*) ((uintptr_t) c3 + cn_stride);
191 wasm_v128_store(c2, vacc2x0123);
192 wasm_v128_store(c2 + 4, vacc2x4567);
193 c2 = (float*) ((uintptr_t) c2 + cn_stride);
194 wasm_v128_store(c1, vacc1x0123);
195 wasm_v128_store(c1 + 4, vacc1x4567);
196 c1 = (float*) ((uintptr_t) c1 + cn_stride);
197 wasm_v128_store(c0, vacc0x0123);
198 wasm_v128_store(c0 + 4, vacc0x4567);
199 c0 = (float*) ((uintptr_t) c0 + cn_stride);
200
201 a3 = (const float*) ((uintptr_t) a3 - kc);
202 a2 = (const float*) ((uintptr_t) a2 - kc);
203 a1 = (const float*) ((uintptr_t) a1 - kc);
204 a0 = (const float*) ((uintptr_t) a0 - kc);
205
206 nc -= 8;
207 } else {
208 if (nc & 4) {
209 wasm_v128_store(c3, vacc3x0123);
210 wasm_v128_store(c2, vacc2x0123);
211 wasm_v128_store(c1, vacc1x0123);
212 wasm_v128_store(c0, vacc0x0123);
213
214 vacc3x0123 = vacc3x4567;
215 vacc2x0123 = vacc2x4567;
216 vacc1x0123 = vacc1x4567;
217 vacc0x0123 = vacc0x4567;
218
219 c3 += 4;
220 c2 += 4;
221 c1 += 4;
222 c0 += 4;
223 }
224 if (nc & 2) {
225 *((double*) c3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
226 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
227 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
228 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
229
230 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
231 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
232 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
233 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
234
235 c3 += 2;
236 c2 += 2;
237 c1 += 2;
238 c0 += 2;
239 }
240 if (nc & 1) {
241 *c3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
242 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
243 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
244 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
245 }
246
247 nc = 0;
248 }
249 } while (nc != 0);
250 }
251