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_minmax_ukernel_1x8s4__wasmrelaxedsimd(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_1x8s4__wasmrelaxedsimd(
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_minmax_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 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
46 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
47 do {
48 v128_t vacc0x0123 = wasm_v128_load(w);
49 v128_t vacc0x4567 = wasm_v128_load(w + 4);
50 w += 8;
51
52 size_t p = ks;
53 do {
54 const float* restrict a0 = a[0];
55 assert(a0 != NULL);
56 if XNN_UNPREDICTABLE(a0 != zero) {
57 a0 = (const float*) ((uintptr_t) a0 + a_offset);
58 }
59 a += 1;
60
61 size_t k = kc;
62 while (k >= 4 * sizeof(float)) {
63 v128_t va0 = wasm_v128_load(a0);
64 a0 += 4;
65
66
67 const v128_t vb0123c0 = wasm_v128_load(w + 0);
68 const v128_t vb4567c0 = wasm_v128_load(w + 4);
69
70 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c0));
71 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c0));
72
73 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
74
75 const v128_t vb0123c1 = wasm_v128_load(w + 8);
76 const v128_t vb4567c1 = wasm_v128_load(w + 12);
77
78 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c1));
79 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c1));
80
81 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
82
83 const v128_t vb0123c2 = wasm_v128_load(w + 16);
84 const v128_t vb4567c2 = wasm_v128_load(w + 20);
85
86 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c2));
87 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c2));
88
89 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
90
91 const v128_t vb0123c3 = wasm_v128_load(w + 24);
92 const v128_t vb4567c3 = wasm_v128_load(w + 28);
93
94 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123c3));
95 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567c3));
96
97
98 w += 32;
99 k -= 4 * sizeof(float);
100 }
101 if XNN_UNLIKELY(k != 0) {
102 v128_t va0 = wasm_v128_load(a0);
103 a0 = (const float*) ((uintptr_t) a0 + k);
104
105 const v128_t vzero = wasm_f32x4_const_splat(0.0f);
106
107 const v128_t vb0123c0 = wasm_v128_load(w + 0);
108 const v128_t vb4567c0 = wasm_v128_load(w + 4);
109
110 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c0, vzero)), vb0123c0));
111 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c0, vzero)), vb4567c0));
112
113 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
114
115 const v128_t vb0123c1 = wasm_v128_load(w + 8);
116 const v128_t vb4567c1 = wasm_v128_load(w + 12);
117
118 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c1, vzero)), vb0123c1));
119 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c1, vzero)), vb4567c1));
120
121 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
122
123 const v128_t vb0123c2 = wasm_v128_load(w + 16);
124 const v128_t vb4567c2 = wasm_v128_load(w + 20);
125
126 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c2, vzero)), vb0123c2));
127 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c2, vzero)), vb4567c2));
128
129 va0 = wasm_v32x4_shuffle(va0, va0, 1, 2, 3, 0);
130
131 const v128_t vb0123c3 = wasm_v128_load(w + 24);
132 const v128_t vb4567c3 = wasm_v128_load(w + 28);
133
134 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb0123c3, vzero)), vb0123c3));
135 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(wasm_v128_andnot(va0, wasm_f32x4_eq(vb4567c3, vzero)), vb4567c3));
136
137
138 w += 32;
139 }
140 p -= 1 * sizeof(void*);
141 } while (p != 0);
142
143 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
144 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
145
146 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
147 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
148
149 if XNN_LIKELY(nc >= 8) {
150 wasm_v128_store(c0, vacc0x0123);
151 wasm_v128_store(c0 + 4, vacc0x4567);
152 c0 = (float*) ((uintptr_t) c0 + cn_stride);
153
154 a = (const float**restrict) ((uintptr_t) a - ks);
155 nc -= 8;
156 } else {
157 if (nc & 4) {
158 wasm_v128_store(c0, vacc0x0123);
159
160 vacc0x0123 = vacc0x4567;
161
162 c0 += 4;
163 }
164 if (nc & 2) {
165 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
166
167 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
168
169 c0 += 2;
170 }
171 if (nc & 1) {
172 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
173 }
174
175 nc = 0;
176 }
177 } while (nc != 0);
178 }
179