xref: /aosp_15_r20/external/XNNPACK/src/xx-fill/neon-x64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 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/fill.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker 
xnn_xx_fill_ukernel__neon_x64(size_t rows,size_t channels,void * output,size_t output_stride,const uint32_t fill_pattern)13*4bdc9457SAndroid Build Coastguard Worker void xnn_xx_fill_ukernel__neon_x64(
14*4bdc9457SAndroid Build Coastguard Worker     size_t rows,
15*4bdc9457SAndroid Build Coastguard Worker     size_t channels,
16*4bdc9457SAndroid Build Coastguard Worker     void* output,
17*4bdc9457SAndroid Build Coastguard Worker     size_t output_stride,
18*4bdc9457SAndroid Build Coastguard Worker     const uint32_t fill_pattern)
19*4bdc9457SAndroid Build Coastguard Worker {
20*4bdc9457SAndroid Build Coastguard Worker   assert(rows != 0);
21*4bdc9457SAndroid Build Coastguard Worker   assert(channels != 0);
22*4bdc9457SAndroid Build Coastguard Worker 
23*4bdc9457SAndroid Build Coastguard Worker   const size_t output_increment = output_stride - channels;
24*4bdc9457SAndroid Build Coastguard Worker 
25*4bdc9457SAndroid Build Coastguard Worker   const uint8x16_t vfill_pattern = vreinterpretq_u8_u32(vdupq_n_u32(fill_pattern));
26*4bdc9457SAndroid Build Coastguard Worker   do {
27*4bdc9457SAndroid Build Coastguard Worker     size_t c = channels;
28*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 64 * sizeof(uint8_t); c -= 64 * sizeof(uint8_t)) {
29*4bdc9457SAndroid Build Coastguard Worker       vst1q_u8(output, vfill_pattern); output = ((uint8_t*) output + 16);
30*4bdc9457SAndroid Build Coastguard Worker       vst1q_u8(output, vfill_pattern); output = ((uint8_t*) output + 16);
31*4bdc9457SAndroid Build Coastguard Worker       vst1q_u8(output, vfill_pattern); output = ((uint8_t*) output + 16);
32*4bdc9457SAndroid Build Coastguard Worker       vst1q_u8(output, vfill_pattern); output = ((uint8_t*) output + 16);
33*4bdc9457SAndroid Build Coastguard Worker     }
34*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
35*4bdc9457SAndroid Build Coastguard Worker       vst1q_u8(output, vfill_pattern); output = ((uint8_t*) output + 16);
36*4bdc9457SAndroid Build Coastguard Worker     }
37*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNLIKELY(c != 0) {
38*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (8 * sizeof(uint8_t))) {
39*4bdc9457SAndroid Build Coastguard Worker         vst1_u8(output, vget_low_u8(vfill_pattern)); output = ((uint8_t*) output + 8);
40*4bdc9457SAndroid Build Coastguard Worker       }
41*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (4 * sizeof(uint8_t))) {
42*4bdc9457SAndroid Build Coastguard Worker         vst1q_lane_u32(output, vreinterpretq_u32_u8(vfill_pattern), 0); output = ((uint8_t*) output + 4);
43*4bdc9457SAndroid Build Coastguard Worker       }
44*4bdc9457SAndroid Build Coastguard Worker       uint8x8_t vfill_subpattern = vget_low_u8(vfill_pattern);
45*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (2 * sizeof(uint8_t))) {
46*4bdc9457SAndroid Build Coastguard Worker         vst1_lane_u16(output, vreinterpret_u16_u8(vfill_subpattern), 0); output = ((uint8_t*) output + 2);
47*4bdc9457SAndroid Build Coastguard Worker         vfill_subpattern = vext_u8(vfill_subpattern, vfill_subpattern, 2);
48*4bdc9457SAndroid Build Coastguard Worker       }
49*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (1 * sizeof(uint8_t))) {
50*4bdc9457SAndroid Build Coastguard Worker         vst1_lane_u8(output, vfill_subpattern, 0); output = ((uint8_t*) output + 1);
51*4bdc9457SAndroid Build Coastguard Worker       }
52*4bdc9457SAndroid Build Coastguard Worker     }
53*4bdc9457SAndroid Build Coastguard Worker     output = (void*) ((uintptr_t) output + output_increment);
54*4bdc9457SAndroid Build Coastguard Worker   } while (--rows != 0);
55*4bdc9457SAndroid Build Coastguard Worker }
56