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_gemminc_minmax_ukernel_1x8s4__wasmsimd_arm(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_1x8s4__wasmsimd_arm(
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 <= 1);
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
43 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
44 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
45 do {
46 v128_t vacc0x0123 = wasm_v128_load(acc + 0);
47 v128_t vacc0x4567 = wasm_v128_load(acc + 4);
48 acc += 8;
49
50 size_t k = kc;
51 while (k >= 4 * sizeof(float)) {
52 v128_t va0 = wasm_v128_load(a0);
53 a0 += 4;
54
55
56 const v128_t vb0123c0 = wasm_v128_load(w + 0);
57 const v128_t vb4567c0 = wasm_v128_load(w + 4);
58
59 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
60 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
61
62 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
63
64 const v128_t vb0123c1 = wasm_v128_load(w + 8);
65 const v128_t vb4567c1 = wasm_v128_load(w + 12);
66
67 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
68 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
69
70 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
71
72 const v128_t vb0123c2 = wasm_v128_load(w + 16);
73 const v128_t vb4567c2 = wasm_v128_load(w + 20);
74
75 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
76 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
77
78 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
79
80 const v128_t vb0123c3 = wasm_v128_load(w + 24);
81 const v128_t vb4567c3 = wasm_v128_load(w + 28);
82
83 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
84 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
85
86
87 w += 32;
88 k -= 4 * sizeof(float);
89 }
90 if XNN_UNLIKELY(k != 0) {
91 v128_t va0 = wasm_v128_load(a0);
92 a0 = (const float*) ((uintptr_t) a0 + k);
93
94 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
95
96 const v128_t vb0123c0 = wasm_v128_load(w + 0);
97 const v128_t vb4567c0 = wasm_v128_load(w + 4);
98
99 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
100 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
101
102 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
103
104 const v128_t vb0123c1 = wasm_v128_load(w + 8);
105 const v128_t vb4567c1 = wasm_v128_load(w + 12);
106
107 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
108 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
109
110 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
111
112 const v128_t vb0123c2 = wasm_v128_load(w + 16);
113 const v128_t vb4567c2 = wasm_v128_load(w + 20);
114
115 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
116 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
117
118 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
119
120 const v128_t vb0123c3 = wasm_v128_load(w + 24);
121 const v128_t vb4567c3 = wasm_v128_load(w + 28);
122
123 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
124 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
125
126
127 w += 32;
128 }
129
130 vacc0x0123 = wasm_f32x4_max(vmin, vacc0x0123);
131 vacc0x4567 = wasm_f32x4_max(vmin, vacc0x4567);
132
133 vacc0x0123 = wasm_f32x4_min(vmax, vacc0x0123);
134 vacc0x4567 = wasm_f32x4_min(vmax, vacc0x4567);
135
136 if XNN_LIKELY(nc >= 8) {
137 wasm_v128_store(c0, vacc0x0123);
138 wasm_v128_store(c0 + 4, vacc0x4567);
139 c0 = (float*) ((uintptr_t) c0 + cn_stride);
140
141 a0 = (const float*) ((uintptr_t) a0 - kc);
142
143 nc -= 8;
144 } else {
145 if (nc & 4) {
146 wasm_v128_store(c0, vacc0x0123);
147
148 vacc0x0123 = vacc0x4567;
149
150 c0 += 4;
151 }
152 if (nc & 2) {
153 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
154
155 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
156
157 c0 += 2;
158 }
159 if (nc & 1) {
160 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
161 }
162
163 nc = 0;
164 }
165 } while (nc != 0);
166 }
167