• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vbinary/vop-neon.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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/vbinary.h>
16 
17 
xnn_f32_vmul_minmax_ukernel__neon_x8(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vmul_minmax_ukernel__neon_x8(
19     size_t n,
20     const float* a,
21     const float* b,
22     float* y,
23     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27   assert(a != NULL);
28   assert(b != NULL);
29   assert(y != NULL);
30 
31   const float32x4_t vy_min = vld1q_dup_f32(&params->scalar.min);
32   const float32x4_t vy_max = vld1q_dup_f32(&params->scalar.max);
33 
34   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35     const float32x4_t va0123 = vld1q_f32(a); a += 4;
36     const float32x4_t vb0123 = vld1q_f32(b); b += 4;
37     const float32x4_t va4567 = vld1q_f32(a); a += 4;
38     const float32x4_t vb4567 = vld1q_f32(b); b += 4;
39 
40     float32x4_t vy0123 = vmulq_f32(va0123, vb0123);
41     float32x4_t vy4567 = vmulq_f32(va4567, vb4567);
42 
43 
44     vy0123 = vmaxq_f32(vy0123, vy_min);
45     vy4567 = vmaxq_f32(vy4567, vy_min);
46 
47     vy0123 = vminq_f32(vy0123, vy_max);
48     vy4567 = vminq_f32(vy4567, vy_max);
49 
50     vst1q_f32(y, vy0123); y += 4;
51     vst1q_f32(y, vy4567); y += 4;
52   }
53   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
54     const float32x4_t va0123 = vld1q_f32(a); a += 4;
55     const float32x4_t vb0123 = vld1q_f32(b); b += 4;
56 
57     float32x4_t vy0123 = vmulq_f32(va0123, vb0123);
58     vy0123 = vmaxq_f32(vy0123, vy_min);
59     vy0123 = vminq_f32(vy0123, vy_max);
60     vst1q_f32(y, vy0123); y += 4;
61   }
62   if XNN_UNLIKELY(n != 0) {
63     const float32x4_t va0123 = vld1q_f32(a);
64     const float32x4_t vb0123 = vld1q_f32(b);
65 
66     float32x4_t vy0123 = vmulq_f32(va0123, vb0123);
67     vy0123 = vmaxq_f32(vy0123, vy_min);
68     vy0123 = vminq_f32(vy0123, vy_max);
69 
70     float32x2_t vy01 = vget_low_f32(vy0123);
71     if (n & (2 * sizeof(float))) {
72       vst1_f32(y, vy01); y += 2;
73       vy01 = vget_high_f32(vy0123);
74     }
75     if (n & (1 * sizeof(float))) {
76       vst1_lane_f32(y, vy01, 0);
77     }
78   }
79 }
80