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_relu_ukernel_1x8__wasmsimd_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_relu_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_relu_ukernel_1x8__wasmsimd_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_relu_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 do {
42 v128_t vacc0x0123 = wasm_v128_load(w + 0);
43 v128_t vacc0x4567 = wasm_v128_load(w + 4);
44 w += 8;
45
46 size_t k = kc;
47 while (k >= 4 * sizeof(float)) {
48 const v128_t va0 = wasm_v128_load(a0);
49 a0 += 4;
50
51 const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);
52
53 const v128_t vb0123c0 = wasm_v128_load(w + 0);
54 const v128_t vb4567c0 = wasm_v128_load(w + 4);
55
56 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c0, vb0123c0));
57 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c0, vb4567c0));
58 const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);
59
60 const v128_t vb0123c1 = wasm_v128_load(w + 8);
61 const v128_t vb4567c1 = wasm_v128_load(w + 12);
62
63 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c1, vb0123c1));
64 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c1, vb4567c1));
65 const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);
66
67 const v128_t vb0123c2 = wasm_v128_load(w + 16);
68 const v128_t vb4567c2 = wasm_v128_load(w + 20);
69
70 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c2, vb0123c2));
71 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c2, vb4567c2));
72 const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);
73
74 const v128_t vb0123c3 = wasm_v128_load(w + 24);
75 const v128_t vb4567c3 = wasm_v128_load(w + 28);
76
77 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0c3, vb0123c3));
78 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0c3, vb4567c3));
79
80 w += 32;
81 k -= 4 * sizeof(float);
82 }
83 if XNN_UNLIKELY(k != 0) {
84 do {
85 const v128_t va0 = wasm_v32x4_load_splat(a0);
86 a0 += 1;
87
88 const v128_t vb0123 = wasm_v128_load(w);
89 const v128_t vb4567 = wasm_v128_load(w + 4);
90 w += 8;
91
92 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(va0, vb0123));
93 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(va0, vb4567));
94
95 k -= sizeof(float);
96 } while (k != 0);
97 }
98
99 const v128_t vzero = wasm_f32x4_splat(0.0f);
100 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vzero);
101 vacc0x4567 = wasm_i32x4_max(vacc0x4567, vzero);
102
103 if XNN_LIKELY(nc >= 8) {
104 wasm_v128_store(c0, vacc0x0123);
105 wasm_v128_store(c0 + 4, vacc0x4567);
106 c0 = (float*) ((uintptr_t) c0 + cn_stride);
107
108 a0 = (const float*) ((uintptr_t) a0 - kc);
109
110 nc -= 8;
111 } else {
112 if (nc & 4) {
113 wasm_v128_store(c0, vacc0x0123);
114
115 vacc0x0123 = vacc0x4567;
116
117 c0 += 4;
118 }
119 if (nc & 2) {
120 *((double*) c0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
121
122 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
123
124 c0 += 2;
125 }
126 if (nc & 1) {
127 *c0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
128 }
129
130 nc = 0;
131 }
132 } while (nc != 0);
133 }
134