1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemminc_minmax_ukernel_1x8__wasmsimd_x86_splat(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_1x8__wasmsimd_x86_splat(
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 const v128_t va0 = wasm_v128_load(a0);
53 a0 += 4;
54
55 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
56
57 const v128_t vb0123c0 = wasm_v128_load(w + 0);
58 const v128_t vb4567c0 = wasm_v128_load(w + 4);
59
60 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
61 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
62 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
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(va0c1, vb0123c1));
68 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
69 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
70
71 const v128_t vb0123c2 = wasm_v128_load(w + 16);
72 const v128_t vb4567c2 = wasm_v128_load(w + 20);
73
74 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
75 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
76 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
77
78 const v128_t vb0123c3 = wasm_v128_load(w + 24);
79 const v128_t vb4567c3 = wasm_v128_load(w + 28);
80
81 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
82 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
83
84 w += 32;
85 k -= 4 * sizeof(float);
86 }
87 if XNN_UNLIKELY(k != 0) {
88 do {
89 const v128_t va0 = wasm_v128_load32_splat(a0);
90 a0 += 1;
91
92 const v128_t vb0123 = wasm_v128_load(w);
93 const v128_t vb4567 = wasm_v128_load(w + 4);
94 w += 8;
95
96 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
97 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
98
99 k -= sizeof(float);
100 } while (k != 0);
101 }
102
103 vacc0x0123 = wasm_f32x4_pmax(vmin, vacc0x0123);
104 vacc0x4567 = wasm_f32x4_pmax(vmin, vacc0x4567);
105
106 vacc0x0123 = wasm_f32x4_pmin(vmax, vacc0x0123);
107 vacc0x4567 = wasm_f32x4_pmin(vmax, vacc0x4567);
108
109 if XNN_LIKELY(nc >= 8) {
110 wasm_v128_store(c0, vacc0x0123);
111 wasm_v128_store(c0 + 4, vacc0x4567);
112 c0 = (float*) ((uintptr_t) c0 + cn_stride);
113
114 a0 = (const float*) ((uintptr_t) a0 - kc);
115
116 nc -= 8;
117 } else {
118 if (nc & 4) {
119 wasm_v128_store(c0, vacc0x0123);
120
121 vacc0x0123 = vacc0x4567;
122
123 c0 += 4;
124 }
125 if (nc & 2) {
126 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
127
128 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
129
130 c0 += 2;
131 }
132 if (nc & 1) {
133 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
134 }
135
136 nc = 0;
137 }
138 } while (nc != 0);
139 }
140