xref: /aosp_15_r20/external/XNNPACK/src/f16-vlrelu/gen/vlrelu-neonfp16arith-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vlrelu/neonfp16arith.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f16_vlrelu_ukernel__neonfp16arith_x16(size_t batch,const void * input,void * output,const union xnn_f16_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vlrelu_ukernel__neonfp16arith_x16(
19     size_t batch,
20     const void* input,
21     void* output,
22     const union xnn_f16_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   assert(batch != 0);
25   assert(batch % sizeof(__fp16) == 0);
26 
27   const float16x8_t vslope = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.slope));
28   const __fp16* i = (const __fp16*) input;
29   __fp16* o = (__fp16*) output;
30   for (; batch >= 16 * sizeof(__fp16); batch -= 16 * sizeof(__fp16)) {
31     const float16x8_t vx01234567 = vld1q_f16(i); i += 8;
32     const float16x8_t vx89ABCDEF = vld1q_f16(i); i += 8;
33 
34     float16x8_t vacc01234567 = vmulq_f16(vx01234567, vslope);
35     const uint16x8_t vmask01234567 = vcltq_s16(vreinterpretq_s16_f16(vx01234567), vmovq_n_s16(0));
36     float16x8_t vacc89ABCDEF = vmulq_f16(vx89ABCDEF, vslope);
37     const uint16x8_t vmask89ABCDEF = vcltq_s16(vreinterpretq_s16_f16(vx89ABCDEF), vmovq_n_s16(0));
38 
39     vacc01234567 = vbslq_f16(vmask01234567, vacc01234567, vx01234567);
40     vacc89ABCDEF = vbslq_f16(vmask89ABCDEF, vacc89ABCDEF, vx89ABCDEF);
41 
42     vst1q_f16(o, vacc01234567); o += 8;
43     vst1q_f16(o, vacc89ABCDEF); o += 8;
44   }
45   for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
46     const float16x8_t vx = vld1q_f16(i); i += 8;
47     float16x8_t vacc = vmulq_f16(vx, vslope);
48     const uint16x8_t vmask = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0));
49     vacc = vbslq_f16(vmask, vacc, vx);
50     vst1q_f16(o, vacc); o += 8;
51   }
52   if XNN_UNLIKELY(batch != 0) {
53     const float16x8_t vx = vld1q_f16(i);
54     float16x8_t vacc = vmulq_f16(vx, vslope);
55     const uint16x8_t vmask = vcltq_s16(vreinterpretq_s16_f16(vx), vmovq_n_s16(0));
56     vacc = vbslq_f16(vmask, vacc, vx);
57 
58     float16x4_t vacc_lo = vget_low_f16(vacc);
59     if (batch & (4 * sizeof(__fp16))) {
60       vst1_f16(o, vacc_lo); o += 4;
61       vacc_lo = vget_high_f16(vacc);
62     }
63     if (batch & (2 * sizeof(__fp16))) {
64       vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2;
65       vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
66     }
67     if (batch & (1 * sizeof(__fp16))) {
68       vst1_lane_f16(o, vacc_lo, 0);
69     }
70   }
71 }
72