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_gemminc_minmax_ukernel_3x8__wasmrelaxedsimd_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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_3x8__wasmrelaxedsimd_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 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 const v128_t va0 = wasm_v128_load(a0);
69 a0 += 4;
70 const v128_t va1 = wasm_v128_load(a1);
71 a1 += 4;
72 const v128_t va2 = wasm_v128_load(a2);
73 a2 += 4;
74
75 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
76 const v128_t va1c0 = wasm_v32x4_shuffle(va1, va1, 0, 0, 0, 0);
77 const v128_t va2c0 = wasm_v32x4_shuffle(va2, va2, 0, 0, 0, 0);
78
79 const v128_t vb0123c0 = wasm_v128_load(w + 0);
80 const v128_t vb4567c0 = wasm_v128_load(w + 4);
81
82 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
83 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c0, vb0123c0));
84 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c0, vb0123c0));
85 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
86 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c0, vb4567c0));
87 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c0, vb4567c0));
88 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
89 const v128_t va1c1 = wasm_v32x4_shuffle(va1, va1, 1, 1, 1, 1);
90 const v128_t va2c1 = wasm_v32x4_shuffle(va2, va2, 1, 1, 1, 1);
91
92 const v128_t vb0123c1 = wasm_v128_load(w + 8);
93 const v128_t vb4567c1 = wasm_v128_load(w + 12);
94
95 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
96 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c1, vb0123c1));
97 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c1, vb0123c1));
98 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
99 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c1, vb4567c1));
100 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c1, vb4567c1));
101 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
102 const v128_t va1c2 = wasm_v32x4_shuffle(va1, va1, 2, 2, 2, 2);
103 const v128_t va2c2 = wasm_v32x4_shuffle(va2, va2, 2, 2, 2, 2);
104
105 const v128_t vb0123c2 = wasm_v128_load(w + 16);
106 const v128_t vb4567c2 = wasm_v128_load(w + 20);
107
108 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
109 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c2, vb0123c2));
110 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c2, vb0123c2));
111 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
112 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c2, vb4567c2));
113 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c2, vb4567c2));
114 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
115 const v128_t va1c3 = wasm_v32x4_shuffle(va1, va1, 3, 3, 3, 3);
116 const v128_t va2c3 = wasm_v32x4_shuffle(va2, va2, 3, 3, 3, 3);
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(va0c3, vb0123c3));
122 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1c3, vb0123c3));
123 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2c3, vb0123c3));
124 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
125 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1c3, vb4567c3));
126 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2c3, vb4567c3));
127
128 w += 32;
129 k -= 4 * sizeof(float);
130 }
131 if XNN_UNLIKELY(k != 0) {
132 do {
133 const v128_t va0 = wasm_v128_load32_splat(a0);
134 a0 += 1;
135 const v128_t va1 = wasm_v128_load32_splat(a1);
136 a1 += 1;
137 const v128_t va2 = wasm_v128_load32_splat(a2);
138 a2 += 1;
139
140 const v128_t vb0123 = wasm_v128_load(w);
141 const v128_t vb4567 = wasm_v128_load(w + 4);
142 w += 8;
143
144 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
145 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(va1, vb0123));
146 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(va2, vb0123));
147 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
148 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(va1, vb4567));
149 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(va2, vb4567));
150
151 k -= sizeof(float);
152 } while (k != 0);
153 }
154
155 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
156 vacc1x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x0123);
157 vacc2x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x0123);
158 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
159 vacc1x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc1x4567);
160 vacc2x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc2x4567);
161
162 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
163 vacc1x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x0123);
164 vacc2x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x0123);
165 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
166 vacc1x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc1x4567);
167 vacc2x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc2x4567);
168
169 if XNN_LIKELY(nc >= 8) {
170 wasm_v128_store(c2, vacc2x0123);
171 wasm_v128_store(c2 + 4, vacc2x4567);
172 c2 = (float*) ((uintptr_t) c2 + cn_stride);
173 wasm_v128_store(c1, vacc1x0123);
174 wasm_v128_store(c1 + 4, vacc1x4567);
175 c1 = (float*) ((uintptr_t) c1 + cn_stride);
176 wasm_v128_store(c0, vacc0x0123);
177 wasm_v128_store(c0 + 4, vacc0x4567);
178 c0 = (float*) ((uintptr_t) c0 + cn_stride);
179
180 a2 = (const float*) ((uintptr_t) a2 - kc);
181 a1 = (const float*) ((uintptr_t) a1 - kc);
182 a0 = (const float*) ((uintptr_t) a0 - kc);
183
184 nc -= 8;
185 } else {
186 if (nc & 4) {
187 wasm_v128_store(c2, vacc2x0123);
188 wasm_v128_store(c1, vacc1x0123);
189 wasm_v128_store(c0, vacc0x0123);
190
191 vacc2x0123 = vacc2x4567;
192 vacc1x0123 = vacc1x4567;
193 vacc0x0123 = vacc0x4567;
194
195 c2 += 4;
196 c1 += 4;
197 c0 += 4;
198 }
199 if (nc & 2) {
200 *((double*) c2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
201 *((double*) c1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
202 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
203
204 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
205 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
206 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
207
208 c2 += 2;
209 c1 += 2;
210 c0 += 2;
211 }
212 if (nc & 1) {
213 *c2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
214 *c1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
215 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
216 }
217
218 nc = 0;
219 }
220 } while (nc != 0);
221 }
222