1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemm_ukernel_1x8s4__wasmsimd(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_default_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_ukernel_1x8s4__wasmsimd(
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_default_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
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
41 do {
42 v128_t vacc0x0123 = wasm_v128_load(w + 0);
43 v128_t vacc0x4567 = wasm_v128_load(w + 4);
44 w += 8;
45
46 size_t k = kc;
47 while (k >= 4 * sizeof(float)) {
48 v128_t va0 = wasm_v128_load(a0);
49 a0 += 4;
50
51
52 const v128_t vb0123c0 = wasm_v128_load(w + 0);
53 const v128_t vb4567c0 = wasm_v128_load(w + 4);
54
55 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
56 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
57
58 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
59
60 const v128_t vb0123c1 = wasm_v128_load(w + 8);
61 const v128_t vb4567c1 = wasm_v128_load(w + 12);
62
63 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
64 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
65
66 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
67
68 const v128_t vb0123c2 = wasm_v128_load(w + 16);
69 const v128_t vb4567c2 = wasm_v128_load(w + 20);
70
71 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
72 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
73
74 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
75
76 const v128_t vb0123c3 = wasm_v128_load(w + 24);
77 const v128_t vb4567c3 = wasm_v128_load(w + 28);
78
79 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
80 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
81
82
83 w += 32;
84 k -= 4 * sizeof(float);
85 }
86 if XNN_UNLIKELY(k != 0) {
87 v128_t va0 = wasm_v128_load(a0);
88 a0 = (const float*) ((uintptr_t) a0 + k);
89
90 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
91
92 const v128_t vb0123c0 = wasm_v128_load(w + 0);
93 const v128_t vb4567c0 = wasm_v128_load(w + 4);
94
95 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
96 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
97
98 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
99
100 const v128_t vb0123c1 = wasm_v128_load(w + 8);
101 const v128_t vb4567c1 = wasm_v128_load(w + 12);
102
103 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
104 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
105
106 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
107
108 const v128_t vb0123c2 = wasm_v128_load(w + 16);
109 const v128_t vb4567c2 = wasm_v128_load(w + 20);
110
111 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
112 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
113
114 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
115
116 const v128_t vb0123c3 = wasm_v128_load(w + 24);
117 const v128_t vb4567c3 = wasm_v128_load(w + 28);
118
119 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
120 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
121
122
123 w += 32;
124 }
125
126
127 if XNN_LIKELY(nc >= 8) {
128 wasm_v128_store(c0, vacc0x0123);
129 wasm_v128_store(c0 + 4, vacc0x4567);
130 c0 = (float*) ((uintptr_t) c0 + cn_stride);
131
132 a0 = (const float*) ((uintptr_t) a0 - kc);
133
134 nc -= 8;
135 } else {
136 if (nc & 4) {
137 wasm_v128_store(c0, vacc0x0123);
138
139 vacc0x0123 = vacc0x4567;
140
141 c0 += 4;
142 }
143 if (nc & 2) {
144 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
145
146 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
147
148 c0 += 2;
149 }
150 if (nc & 1) {
151 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
152 }
153
154 nc = 0;
155 }
156 } while (nc != 0);
157 }
158