xref: /aosp_15_r20/external/XNNPACK/src/f16-rmax/neonfp16arith.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2022 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/common.h>
11 #include <xnnpack/rmax.h>
12 
13 
xnn_f16_rmax_ukernel__neonfp16arith(size_t batch,const void * input,void * output)14 void xnn_f16_rmax_ukernel__neonfp16arith(
15     size_t batch,
16     const void* input,
17     void* output) XNN_OOB_READS
18 {
19   assert(batch != 0);
20   assert(batch % sizeof(__fp16) == 0);
21 
22   const __fp16* i = (const __fp16*) input;
23   __fp16* o = (__fp16*) output;
24   float16x8_t vmax0 = vld1q_dup_f16(i);
25   float16x8_t vmax1 = vmax0;
26   float16x8_t vmax2 = vmax0;
27   float16x8_t vmax3 = vmax0;
28   for (; batch >= 32 * sizeof(__fp16); batch -= 32 * sizeof(__fp16)) {
29     const float16x8_t vx0 = vld1q_f16(i); i += 8;
30     const float16x8_t vx1 = vld1q_f16(i); i += 8;
31     const float16x8_t vx2 = vld1q_f16(i); i += 8;
32     const float16x8_t vx3 = vld1q_f16(i); i += 8;
33 
34     vmax0 = vmaxq_f16(vmax0, vx0);
35     vmax1 = vmaxq_f16(vmax1, vx1);
36     vmax2 = vmaxq_f16(vmax2, vx2);
37     vmax3 = vmaxq_f16(vmax3, vx3);
38   }
39   float16x8_t vmax = vmaxq_f16(vmaxq_f16(vmax0, vmax1), vmaxq_f16(vmax2, vmax3));
40   for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
41     const float16x8_t vx = vld1q_f16(i); i += 8;
42     vmax = vmaxq_f16(vmax, vx);
43   }
44   float16x4_t vmax_lo = vmax_f16(vget_low_f16(vmax), vget_high_f16(vmax));
45   if XNN_UNLIKELY(batch != 0) {
46     const float16x8_t vx = vld1q_f16(i);
47     float16x4_t vx_lo = vget_low_f16(vx);
48     if (batch & (4 * sizeof(__fp16))) {
49       vmax_lo = vmax_f16(vmax_lo, vx_lo);
50       vx_lo = vget_high_f16(vx);
51     }
52     if (batch & (2 * sizeof(__fp16))) {
53       vmax_lo = vmax_f16(vmax_lo, vext_f16(vmax_lo, vx_lo, 2));
54       vx_lo = vext_f16(vx_lo, vx_lo, 2);
55     }
56     if (batch & (1 * sizeof(__fp16))) {
57       vmax_lo = vmax_f16(vmax_lo, vext_f16(vmax_lo, vx_lo, 1));
58     }
59   }
60   #if XNN_ARCH_ARM64
61     *o = vmaxv_f16(vmax_lo);
62   #else
63     vmax_lo = vpmax_f16(vmax_lo, vmax_lo);
64     vmax_lo = vpmax_f16(vmax_lo, vmax_lo);
65     vst1_lane_f16(o, vmax_lo, 0);
66   #endif
67 }
68