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_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_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 union xnn_f32_minmax_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 const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
42 const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
43 do {
44 v128_t vacc0x0123 = wasm_v128_load(w + 0);
45 v128_t vacc0x4567 = wasm_v128_load(w + 4);
46 w += 8;
47
48 size_t k = kc;
49 while (k >= 4 * sizeof(float)) {
50 const v128_t va0 = wasm_v128_load(a0);
51 a0 += 4;
52
53 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
54
55 const v128_t vb0123c0 = wasm_v128_load(w + 0);
56 const v128_t vb4567c0 = wasm_v128_load(w + 4);
57
58 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
59 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
60 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
61
62 const v128_t vb0123c1 = wasm_v128_load(w + 8);
63 const v128_t vb4567c1 = wasm_v128_load(w + 12);
64
65 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
66 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
67 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
68
69 const v128_t vb0123c2 = wasm_v128_load(w + 16);
70 const v128_t vb4567c2 = wasm_v128_load(w + 20);
71
72 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
73 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
74 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
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(va0c3, vb0123c3));
80 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
81
82 w += 32;
83 k -= 4 * sizeof(float);
84 }
85 if XNN_UNLIKELY(k != 0) {
86 do {
87 const v128_t va0 = wasm_v128_load32_splat(a0);
88 a0 += 1;
89
90 const v128_t vb0123 = wasm_v128_load(w);
91 const v128_t vb4567 = wasm_v128_load(w + 4);
92 w += 8;
93
94 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
95 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
96
97 k -= sizeof(float);
98 } while (k != 0);
99 }
100
101 vacc0x0123 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x0123);
102 vacc0x4567 = __builtin_wasm_relaxed_max_f32x4(vmin, vacc0x4567);
103
104 vacc0x0123 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x0123);
105 vacc0x4567 = __builtin_wasm_relaxed_min_f32x4(vmax, vacc0x4567);
106
107 if XNN_LIKELY(nc >= 8) {
108 wasm_v128_store(c0, vacc0x0123);
109 wasm_v128_store(c0 + 4, vacc0x4567);
110 c0 = (float*) ((uintptr_t) c0 + cn_stride);
111
112 a0 = (const float*) ((uintptr_t) a0 - kc);
113
114 nc -= 8;
115 } else {
116 if (nc & 4) {
117 wasm_v128_store(c0, vacc0x0123);
118
119 vacc0x0123 = vacc0x4567;
120
121 c0 += 4;
122 }
123 if (nc & 2) {
124 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
125
126 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
127
128 c0 += 2;
129 }
130 if (nc & 1) {
131 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
132 }
133
134 nc = 0;
135 }
136 } while (nc != 0);
137 }
138