1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-hswish/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/hswish.h>
16 
17 
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19 
xnn_f32_hswish_ukernel__fma3_x8(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_hswish_ukernel__fma3_x8(
21     size_t n,
22     const float* x,
23     float* y,
24     const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26   assert(n != 0);
27   assert(n % sizeof(float) == 0);
28 
29   const __m256 vsixth = _mm256_broadcast_ps((const __m128*) params->sse.sixth);
30   const __m256 vhalf = _mm256_broadcast_ps((const __m128*) params->sse.half);
31   const __m256 vone = _mm256_broadcast_ps((const __m128*) params->sse.one);
32   const __m256 vzero = _mm256_setzero_ps();
33 
34   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35     const __m256 vx01234567 = _mm256_loadu_ps(x);
36     x += 8;
37 
38     __m256 vacc01234567 = _mm256_fmadd_ps(vx01234567, vsixth, vhalf);
39 
40     vacc01234567 = _mm256_max_ps(vacc01234567, vzero);
41 
42     vacc01234567 = _mm256_min_ps(vacc01234567, vone);
43 
44     vacc01234567 = _mm256_mul_ps(vacc01234567, vx01234567);
45 
46     _mm256_storeu_ps(y, vacc01234567);
47     y += 8;
48   }
49   if XNN_UNLIKELY(n != 0) {
50     assert(n >= 1 * sizeof(float));
51     assert(n <= 7 * sizeof(float));
52     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
53 
54     const __m256 vx = _mm256_maskload_ps(x, vmask);
55     __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);
56     vacc = _mm256_max_ps(vacc, vzero);
57     vacc = _mm256_min_ps(vacc, vone);
58     vacc = _mm256_mul_ps(vacc, vx);
59 
60     // _mm256_maskstore_ps(y, vmask, vacc) could be used here, but triggers msan failures (probably an msan bug).
61     __m128 vacc_lo = _mm256_castps256_ps128(vacc);
62     if (n & (4 * sizeof(float))) {
63       _mm_storeu_ps(y, vacc_lo);
64       vacc_lo = _mm256_extractf128_ps(vacc, 1);
65       y += 4;
66     }
67     if (n & (2 * sizeof(float))) {
68       _mm_storel_pi((__m64*) y, vacc_lo);
69       vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
70       y += 2;
71     }
72     if (n & (1 * sizeof(float))) {
73       _mm_store_ss(y, vacc_lo);
74     }
75   }
76 }
77