1 // Copyright 2019 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_f32_rmax_ukernel__neon(size_t n,const float * x,float * y)14 void xnn_f32_rmax_ukernel__neon(
15 size_t n,
16 const float* x,
17 float* y)
18 {
19 assert(n != 0);
20 assert(n % sizeof(float) == 0);
21
22 float32x4_t vmax0 = vld1q_dup_f32(x);
23 float32x4_t vmax1 = vmax0;
24 float32x4_t vmax2 = vmax0;
25 float32x4_t vmax3 = vmax0;
26 for (; n >= 64; n -= 64) {
27 const float32x4_t vx0 = vld1q_f32(x); x += 4;
28 const float32x4_t vx1 = vld1q_f32(x); x += 4;
29 const float32x4_t vx2 = vld1q_f32(x); x += 4;
30 const float32x4_t vx3 = vld1q_f32(x); x += 4;
31
32 vmax0 = vmaxq_f32(vmax0, vx0);
33 vmax1 = vmaxq_f32(vmax1, vx1);
34 vmax2 = vmaxq_f32(vmax2, vx2);
35 vmax3 = vmaxq_f32(vmax3, vx3);
36 }
37 float32x4_t vmax = vmaxq_f32(vmaxq_f32(vmax0, vmax1), vmaxq_f32(vmax2, vmax3));
38 for (; n >= 16; n -= 16) {
39 const float32x4_t vx = vld1q_f32(x); x += 4;
40 vmax = vmaxq_f32(vmax, vx);
41 }
42 #if XNN_ARCH_ARM64
43 float32x2_t vmax_lo = vget_low_f32(vpmaxq_f32(vmax, vmax));
44 #else
45 float32x2_t vmax_lo = vmax_f32(vget_low_f32(vmax), vget_high_f32(vmax));
46 #endif
47 if XNN_UNLIKELY(n != 0) {
48 do {
49 const float32x2_t vx = vld1_dup_f32(x); x += 1;
50 vmax_lo = vmax_f32(vmax_lo, vx);
51 n -= 4;
52 } while (n != 0);
53 }
54 #if XNN_ARCH_ARM64
55 *y = vmaxv_f32(vmax_lo);
56 #else
57 vst1_lane_f32(y, vpmax_f32(vmax_lo, vmax_lo), 0);
58 #endif
59 }
60