1 // Auto-generated file. Do not edit!
2 // Template: src/f32-hswish/avx512f.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/intrinsics-polyfill.h>
16 #include <xnnpack/hswish.h>
17
18
xnn_f32_hswish_ukernel__avx512f_x16(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_hswish_ukernel__avx512f_x16(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const __m512 vsixth = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.sixth));
29 const __m512 vhalf = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.half));
30 const __m512 vone = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.one));
31 const __m512 vzero = _mm512_setzero_ps();
32
33 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
34 const __m512 vx0123456789ABCDEF = _mm512_loadu_ps(x);
35 x += 16;
36
37 __m512 vacc0123456789ABCDEF = _mm512_fmadd_ps(vx0123456789ABCDEF, vsixth, vhalf);
38
39 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEF, vzero);
40
41 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vone);
42
43 vacc0123456789ABCDEF = _mm512_mul_ps(vacc0123456789ABCDEF, vx0123456789ABCDEF);
44
45 _mm512_storeu_ps(y, vacc0123456789ABCDEF);
46 y += 16;
47 }
48 if XNN_UNLIKELY(n != 0) {
49 assert(n >= 1 * sizeof(float));
50 assert(n <= 15 * sizeof(float));
51 // Prepare mask for valid 32-bit elements (depends on n).
52 n >>= 2 /* log2(sizeof(float)) */;
53 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
54
55 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
56 __m512 vacc = _mm512_fmadd_ps(vx, vsixth, vhalf);
57 vacc = _mm512_max_ps(vacc, vzero);
58 vacc = _mm512_min_ps(vacc, vone);
59 vacc = _mm512_mul_ps(vacc, vx);
60 _mm512_mask_storeu_ps(y, vmask, vacc);
61 }
62 }
63