xref: /aosp_15_r20/external/XNNPACK/src/x32-zip/x2-neon.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker 
6*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
7*4bdc9457SAndroid Build Coastguard Worker 
8*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/zip.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker 
xnn_x32_zip_x2_ukernel__neon(size_t n,const uint32_t * input,uint32_t * output)13*4bdc9457SAndroid Build Coastguard Worker void xnn_x32_zip_x2_ukernel__neon(
14*4bdc9457SAndroid Build Coastguard Worker     size_t n,
15*4bdc9457SAndroid Build Coastguard Worker     const uint32_t* input,
16*4bdc9457SAndroid Build Coastguard Worker     uint32_t* output)
17*4bdc9457SAndroid Build Coastguard Worker {
18*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
19*4bdc9457SAndroid Build Coastguard Worker   assert(n % 4 == 0);
20*4bdc9457SAndroid Build Coastguard Worker 
21*4bdc9457SAndroid Build Coastguard Worker   const uint32_t* x = input;
22*4bdc9457SAndroid Build Coastguard Worker   const uint32_t* y = (const uint32_t*) ((uintptr_t) x + n);
23*4bdc9457SAndroid Build Coastguard Worker   uint32_t* o = output;
24*4bdc9457SAndroid Build Coastguard Worker 
25*4bdc9457SAndroid Build Coastguard Worker   while (n >= 16) {
26*4bdc9457SAndroid Build Coastguard Worker     uint32x4x2_t vxy;
27*4bdc9457SAndroid Build Coastguard Worker     vxy.val[0] = vld1q_u32(x); x += 4;
28*4bdc9457SAndroid Build Coastguard Worker     vxy.val[1] = vld1q_u32(y); y += 4;
29*4bdc9457SAndroid Build Coastguard Worker     vst2q_u32(o, vxy); o += 8;
30*4bdc9457SAndroid Build Coastguard Worker     n -= 16;
31*4bdc9457SAndroid Build Coastguard Worker   }
32*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
33*4bdc9457SAndroid Build Coastguard Worker     if (n & 8) {
34*4bdc9457SAndroid Build Coastguard Worker       uint32x2x2_t vxy;
35*4bdc9457SAndroid Build Coastguard Worker       vxy.val[0] = vld1_u32(x); x += 2;
36*4bdc9457SAndroid Build Coastguard Worker       vxy.val[1] = vld1_u32(y); y += 2;
37*4bdc9457SAndroid Build Coastguard Worker       vst2_u32(o, vxy); o += 4;
38*4bdc9457SAndroid Build Coastguard Worker     }
39*4bdc9457SAndroid Build Coastguard Worker     if (n & 4) {
40*4bdc9457SAndroid Build Coastguard Worker       uint32x2_t vxy = vld1_dup_u32(x);
41*4bdc9457SAndroid Build Coastguard Worker       vxy = vld1_lane_u32(y, vxy, 1);
42*4bdc9457SAndroid Build Coastguard Worker       vst1_u32(o, vxy);
43*4bdc9457SAndroid Build Coastguard Worker     }
44*4bdc9457SAndroid Build Coastguard Worker   }
45*4bdc9457SAndroid Build Coastguard Worker }
46