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/fill.h>
11
12
xnn_x32_fill_ukernel__neon(size_t rows,size_t channels,uint32_t * output,size_t output_stride,const uint32_t * fill_value)13 void xnn_x32_fill_ukernel__neon(
14 size_t rows,
15 size_t channels,
16 uint32_t* output,
17 size_t output_stride,
18 const uint32_t* fill_value)
19 {
20 assert(rows != 0);
21 assert(channels != 0);
22 assert(channels % sizeof(uint32_t) == 0);
23 assert(fill_value != NULL);
24
25 const size_t output_increment = output_stride - channels;
26
27 const uint32x4_t vfill = vld1q_dup_u32(fill_value);
28 do {
29 size_t c = channels;
30 for (; c >= 16 * sizeof(uint32_t); c -= 16 * sizeof(uint32_t)) {
31 vst1q_u32(output, vfill); output += 4;
32 vst1q_u32(output, vfill); output += 4;
33 vst1q_u32(output, vfill); output += 4;
34 vst1q_u32(output, vfill); output += 4;
35 }
36 for (; c >= 4 * sizeof(uint32_t); c -= 4 * sizeof(uint32_t)) {
37 vst1q_u32(output, vfill); output += 4;
38 }
39 if XNN_UNLIKELY(c != 0) {
40 if XNN_LIKELY(c & (2 * sizeof(uint32_t))) {
41 vst1_u32(output, vget_low_u32(vfill)); output += 2;
42 }
43 if XNN_LIKELY(c & (1 * sizeof(uint32_t))) {
44 vst1q_lane_u32(output, vfill, 0); output += 1;
45 }
46 }
47 output = (void*) ((uintptr_t) output + output_increment);
48 } while (--rows != 0);
49 }
50