• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vhswish/avx.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 <immintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f32_vhswish_ukernel__fma3_x8(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vhswish_ukernel__fma3_x8(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26 
27   const __m256 vsixth = _mm256_load_ps(params->avx.sixth);
28   const __m256 vhalf = _mm256_load_ps(params->avx.half);
29   const __m256 vone = _mm256_load_ps(params->avx.one);
30   const __m256 vzero = _mm256_setzero_ps();
31 
32   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
33     const __m256 vx = _mm256_loadu_ps(x);
34     x += 8;
35     __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);
36     vacc = _mm256_max_ps(vacc, vzero);
37     vacc = _mm256_min_ps(vacc, vone);
38     vacc = _mm256_mul_ps(vacc, vx);
39     _mm256_storeu_ps(y, vacc);
40     y += 8;
41   }
42   if XNN_UNLIKELY(n != 0) {
43     assert(n >= 1 * sizeof(float));
44     assert(n <= 7 * sizeof(float));
45     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
46 
47     const __m256 vx = _mm256_maskload_ps(x, vmask);
48     __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);
49     vacc = _mm256_max_ps(vacc, vzero);
50     vacc = _mm256_min_ps(vacc, vone);
51     vacc = _mm256_mul_ps(vacc, vx);
52 
53     __m128 vacc_lo = _mm256_castps256_ps128(vacc);
54     if (n & (4 * sizeof(float))) {
55       _mm_storeu_ps(y, vacc_lo);
56       vacc_lo = _mm256_extractf128_ps(vacc, 1);
57       y += 4;
58     }
59     if (n & (2 * sizeof(float))) {
60       _mm_storel_pi((__m64*) y, vacc_lo);
61       vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
62       y += 2;
63     }
64     if (n & (1 * sizeof(float))) {
65       _mm_store_ss(y, vacc_lo);
66     }
67   }
68 }
69