1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_1x8__wasmsimd_x86_splat(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_1x8__wasmsimd_x86_splat(
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)])
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 const v128_t va0 = wasm_v128_load(a0);
64 a0 += 4;
65
66 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
67
68 const v128_t vb0123c0 = wasm_v128_load(w + 0);
69 const v128_t vb4567c0 = wasm_v128_load(w + 4);
70
71 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
72 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
73 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
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(va0c1, vb0123c1));
79 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
80 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
81
82 const v128_t vb0123c2 = wasm_v128_load(w + 16);
83 const v128_t vb4567c2 = wasm_v128_load(w + 20);
84
85 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
86 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
87 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
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(va0c3, vb0123c3));
93 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
94
95 w += 32;
96 k -= 4 * sizeof(float);
97 }
98 if XNN_UNLIKELY(k != 0) {
99 do {
100 const v128_t vb0123 = wasm_v128_load(w);
101 const v128_t vb4567 = wasm_v128_load(w + 4);
102 w += 8;
103
104 const v128_t va0 = wasm_v128_load32_splat(a0);
105 a0 += 1;
106
107 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
108 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
109 k -= sizeof(float);
110 } while (k != 0);
111 }
112 p -= 1 * sizeof(void*);
113 } while (p != 0);
114
115 vacc0x0123 = wasm_f32x4_pmax(vmin, vacc0x0123);
116 vacc0x4567 = wasm_f32x4_pmax(vmin, vacc0x4567);
117
118 vacc0x0123 = wasm_f32x4_pmin(vmax, vacc0x0123);
119 vacc0x4567 = wasm_f32x4_pmin(vmax, vacc0x4567);
120
121 if XNN_LIKELY(nc >= 8) {
122 wasm_v128_store(c0, vacc0x0123);
123 wasm_v128_store(c0 + 4, vacc0x4567);
124 c0 = (float*) ((uintptr_t) c0 + cn_stride);
125
126 a = (const float**restrict) ((uintptr_t) a - ks);
127 nc -= 8;
128 } else {
129 if (nc & 4) {
130 wasm_v128_store(c0, vacc0x0123);
131
132 vacc0x0123 = vacc0x4567;
133
134 c0 += 4;
135 }
136 if (nc & 2) {
137 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
138
139 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
140
141 c0 += 2;
142 }
143 if (nc & 1) {
144 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
145 }
146
147 nc = 0;
148 }
149 } while (nc != 0);
150 }
151