1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_ukernel_1x8s4__wasmsimd(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_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_ukernel_1x8s4__wasmsimd(
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_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
45 do {
46 v128_t vacc0x0123 = wasm_v128_load(w);
47 v128_t vacc0x4567 = wasm_v128_load(w + 4);
48 w += 8;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 while (k >= 4 * sizeof(float)) {
61 v128_t va0 = wasm_v128_load(a0);
62 a0 += 4;
63
64
65 const v128_t vb0123c0 = wasm_v128_load(w + 0);
66 const v128_t vb4567c0 = wasm_v128_load(w + 4);
67
68 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
69 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
70
71 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
72
73 const v128_t vb0123c1 = wasm_v128_load(w + 8);
74 const v128_t vb4567c1 = wasm_v128_load(w + 12);
75
76 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
77 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
78
79 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
80
81 const v128_t vb0123c2 = wasm_v128_load(w + 16);
82 const v128_t vb4567c2 = wasm_v128_load(w + 20);
83
84 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
85 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
86
87 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
88
89 const v128_t vb0123c3 = wasm_v128_load(w + 24);
90 const v128_t vb4567c3 = wasm_v128_load(w + 28);
91
92 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
93 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
94
95
96 w += 32;
97 k -= 4 * sizeof(float);
98 }
99 if XNN_UNLIKELY(k != 0) {
100 v128_t va0 = wasm_v128_load(a0);
101 a0 = (const float*) ((uintptr_t) a0 + k);
102
103 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
104
105 const v128_t vb0123c0 = wasm_v128_load(w + 0);
106 const v128_t vb4567c0 = wasm_v128_load(w + 4);
107
108 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
109 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
110
111 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
112
113 const v128_t vb0123c1 = wasm_v128_load(w + 8);
114 const v128_t vb4567c1 = wasm_v128_load(w + 12);
115
116 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
117 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
118
119 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
120
121 const v128_t vb0123c2 = wasm_v128_load(w + 16);
122 const v128_t vb4567c2 = wasm_v128_load(w + 20);
123
124 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
125 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
126
127 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
128
129 const v128_t vb0123c3 = wasm_v128_load(w + 24);
130 const v128_t vb4567c3 = wasm_v128_load(w + 28);
131
132 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
133 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
134
135
136 w += 32;
137 }
138 p -= 1 * sizeof(void*);
139 } while (p != 0);
140
141
142 if XNN_LIKELY(nc >= 8) {
143 wasm_v128_store(c0, vacc0x0123);
144 wasm_v128_store(c0 + 4, vacc0x4567);
145 c0 = (float*) ((uintptr_t) c0 + cn_stride);
146
147 a = (const float**restrict) ((uintptr_t) a - ks);
148 nc -= 8;
149 } else {
150 if (nc & 4) {
151 wasm_v128_store(c0, vacc0x0123);
152
153 vacc0x0123 = vacc0x4567;
154
155 c0 += 4;
156 }
157 if (nc & 2) {
158 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
159
160 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
161
162 c0 += 2;
163 }
164 if (nc & 1) {
165 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
166 }
167
168 nc = 0;
169 }
170 } while (nc != 0);
171 }
172