1 // Copyright 2020 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <wasm_simd128.h>
9
10 #include <xnnpack/zip.h>
11
12
xnn_x32_zip_x4_ukernel__wasmsimd(size_t n,const uint32_t * input,uint32_t * output)13 void xnn_x32_zip_x4_ukernel__wasmsimd(
14 size_t n,
15 const uint32_t* input,
16 uint32_t* output)
17 {
18 assert(n != 0);
19 assert(n % sizeof(uint32_t) == 0);
20
21 const float* x = (const float*) input;
22 const float* y = (const float*) ((uintptr_t) x + n);
23 const float* z = (const float*) ((uintptr_t) y + n);
24 const float* w = (const float*) ((uintptr_t) z + n);
25 float* o = (float*) output;
26
27 while (n >= 4 * sizeof(uint32_t)) {
28 const v128_t vx = wasm_v128_load(x);
29 x += 4;
30 const v128_t vy = wasm_v128_load(y);
31 y += 4;
32 const v128_t vz = wasm_v128_load(z);
33 z += 4;
34 const v128_t vw = wasm_v128_load(w);
35 w += 4;
36
37 const v128_t vxy_lo = wasm_v32x4_shuffle(vx, vy, 0, 4, 1, 5);
38 const v128_t vxy_hi = wasm_v32x4_shuffle(vx, vy, 2, 6, 3, 7);
39 const v128_t vzw_lo = wasm_v32x4_shuffle(vz, vw, 0, 4, 1, 5);
40 const v128_t vzw_hi = wasm_v32x4_shuffle(vz, vw, 2, 6, 3, 7);
41
42 const v128_t vxyzw0 = wasm_v32x4_shuffle(vxy_lo, vzw_lo, 0, 1, 4, 5);
43 const v128_t vxyzw1 = wasm_v32x4_shuffle(vxy_lo, vzw_lo, 2, 3, 6, 7);
44 const v128_t vxyzw2 = wasm_v32x4_shuffle(vxy_hi, vzw_hi, 0, 1, 4, 5);
45 const v128_t vxyzw3 = wasm_v32x4_shuffle(vxy_hi, vzw_hi, 2, 3, 6, 7);
46
47 wasm_v128_store(o, vxyzw0);
48 wasm_v128_store(o + 4, vxyzw1);
49 wasm_v128_store(o + 8, vxyzw2);
50 wasm_v128_store(o + 12, vxyzw3);
51 o += 16;
52 n -= 4 * sizeof(uint32_t);
53 }
54 if XNN_UNLIKELY(n != 0) {
55 if (n & (2 * sizeof(uint32_t))) {
56 const double vx = *((const double*) x);
57 x += 2;
58 const double vy = *((const double*) y);
59 y += 2;
60 const double vz = *((const double*) z);
61 z += 2;
62 const double vw = *((const double*) w);
63 w += 2;
64
65 const v128_t vxy = wasm_f64x2_make(vx, vy);
66 const v128_t vzw = wasm_f64x2_make(vz, vw);
67
68 const v128_t vxyzw_lo = wasm_v32x4_shuffle(vxy, vzw, 0, 2, 4, 6);
69 const v128_t vxyzw_hi = wasm_v32x4_shuffle(vxy, vzw, 1, 3, 5, 7);
70
71 wasm_v128_store(o, vxyzw_lo);
72 wasm_v128_store(o + 4, vxyzw_hi);
73 o += 8;
74 }
75 if (n & (1 * sizeof(uint32_t))) {
76 const float vx = *x;
77 const float vy = *y;
78 const float vz = *z;
79 const float vw = *w;
80 o[0] = vx;
81 o[1] = vy;
82 o[2] = vz;
83 o[3] = vw;
84 }
85 }
86 }
87