1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_relu_ukernel_3x8__wasmsimd_splat(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_3x8__wasmsimd_splat(
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)])
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 const v128_t va0 = wasm_v128_load(a0);
84 a0 += 4;
85 const v128_t va1 = wasm_v128_load(a1);
86 a1 += 4;
87 const v128_t va2 = wasm_v128_load(a2);
88 a2 += 4;
89
90 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
91 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
92 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
93
94 const v128_t vb0123c0 = wasm_v128_load(w + 0);
95 const v128_t vb4567c0 = wasm_v128_load(w + 4);
96
97 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
98 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
99 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
100 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
101 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
102 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
103 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
104 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
105 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
106
107 const v128_t vb0123c1 = wasm_v128_load(w + 8);
108 const v128_t vb4567c1 = wasm_v128_load(w + 12);
109
110 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
111 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
112 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
113 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
114 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
115 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
116 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
117 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
118 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
119
120 const v128_t vb0123c2 = wasm_v128_load(w + 16);
121 const v128_t vb4567c2 = wasm_v128_load(w + 20);
122
123 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
124 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
125 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
126 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
127 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
128 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, 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
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(va0c3, vb0123c3));
137 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
138 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
139 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
140 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
141 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
142
143 w += 32;
144 k -= 4 * sizeof(float);
145 }
146 if XNN_UNLIKELY(k != 0) {
147 do {
148 const v128_t vb0123 = wasm_v128_load(w);
149 const v128_t vb4567 = wasm_v128_load(w + 4);
150 w += 8;
151
152 const v128_t va0 = wasm_v128_load32_splat(a0);
153 a0 += 1;
154 const v128_t va1 = wasm_v128_load32_splat(a1);
155 a1 += 1;
156 const v128_t va2 = wasm_v128_load32_splat(a2);
157 a2 += 1;
158
159 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
160 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
161 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
162 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
163 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
164 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
165 k -= sizeof(float);
166 } while (k != 0);
167 }
168 p -= 3 * sizeof(void*);
169 } while (p != 0);
170
171 const v128_t vzero = wasm_i32x4_const_splat(0);
172 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
173 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vzero);
174 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vzero);
175 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
176 vacc1x4567 = wasm_i32x4_max(vacc1x4567, vzero);
177 vacc2x4567 = wasm_i32x4_max(vacc2x4567, vzero);
178
179 if XNN_LIKELY(nc >= 8) {
180 wasm_v128_store(c2, vacc2x0123);
181 wasm_v128_store(c2 + 4, vacc2x4567);
182 c2 = (float*) ((uintptr_t) c2 + cn_stride);
183 wasm_v128_store(c1, vacc1x0123);
184 wasm_v128_store(c1 + 4, vacc1x4567);
185 c1 = (float*) ((uintptr_t) c1 + cn_stride);
186 wasm_v128_store(c0, vacc0x0123);
187 wasm_v128_store(c0 + 4, vacc0x4567);
188 c0 = (float*) ((uintptr_t) c0 + cn_stride);
189
190 a = (const float**restrict) ((uintptr_t) a - ks);
191 nc -= 8;
192 } else {
193 if (nc & 4) {
194 wasm_v128_store(c2, vacc2x0123);
195 wasm_v128_store(c1, vacc1x0123);
196 wasm_v128_store(c0, vacc0x0123);
197
198 vacc2x0123 = vacc2x4567;
199 vacc1x0123 = vacc1x4567;
200 vacc0x0123 = vacc0x4567;
201
202 c2 += 4;
203 c1 += 4;
204 c0 += 4;
205 }
206 if (nc & 2) {
207 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
208 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
209 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
210
211 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
212 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
213 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
214
215 c2 += 2;
216 c1 += 2;
217 c0 += 2;
218 }
219 if (nc & 1) {
220 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
221 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
222 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
223 }
224
225 nc = 0;
226 }
227 } while (nc != 0);
228 }
229