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 <arm_neon.h>
9
10 #include <xnnpack/unpool.h>
11
12
xnn_x32_unpool_ukernel__neon(size_t kernel_elements,size_t channels,uint32_t fill,const uint32_t * input,const uint32_t * index,uint32_t ** output)13 void xnn_x32_unpool_ukernel__neon(
14 size_t kernel_elements,
15 size_t channels,
16 uint32_t fill,
17 const uint32_t* input,
18 const uint32_t* index,
19 uint32_t** output)
20 {
21 // Pre-initialize outputs with constant.
22 const uint32x4_t vfill = vdupq_n_u32(fill);
23 uint32_t** os = output;
24 do {
25 uint32_t* o = *os++;
26 size_t c = channels;
27 for (; c >= 4; c -= 4) {
28 vst1q_u32(o, vfill); o += 4;
29 }
30 if (c != 0) {
31 if (c & 2) {
32 vst1_u32(o, vget_low_u32(vfill)); o += 2;
33 }
34 if (c & 1) {
35 vst1q_lane_u32(o, vfill, 0);
36 }
37 }
38 } while (--kernel_elements != 0);
39
40 // Copy indexed elements to output.
41 size_t offset = 0;
42 do {
43 const uint32_t i = *index++;
44 *((uint32_t*) ((uintptr_t) output[i] + offset)) = *input++;
45 offset += sizeof(uint32_t);
46 } while (--channels != 0);
47 }
48