• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsqrt/avx-sqrt.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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_vsqrt_ukernel__avx_sqrt_x16(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsqrt_ukernel__avx_sqrt_x16(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26 
27   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
28     const __m256 vx01234567 = _mm256_loadu_ps(x);
29     const __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
30     x += 16;
31 
32     const __m256 vy01234567 = _mm256_sqrt_ps(vx01234567);
33     const __m256 vy89ABCDEF = _mm256_sqrt_ps(vx89ABCDEF);
34 
35     _mm256_storeu_ps(y, vy01234567);
36     _mm256_storeu_ps(y + 8, vy89ABCDEF);
37     y += 16;
38   }
39   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
40     const __m256 vx = _mm256_loadu_ps(x);
41     x += 8;
42     const __m256 vy = _mm256_sqrt_ps(vx);
43     _mm256_storeu_ps(y, vy);
44     y += 8;
45   }
46   if XNN_UNLIKELY(n != 0) {
47     assert(n >= 1 * sizeof(float));
48     assert(n <= 7 * sizeof(float));
49     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
50 
51     const __m256 vx = _mm256_maskload_ps(x, vmask);
52     const __m256 vy = _mm256_sqrt_ps(vx);
53 
54     __m128 vy_lo = _mm256_castps256_ps128(vy);
55     if (n & (4 * sizeof(float))) {
56       _mm_storeu_ps(y, vy_lo);
57       vy_lo = _mm256_extractf128_ps(vy, 1);
58       y += 4;
59     }
60     if (n & (2 * sizeof(float))) {
61       _mm_storel_pi((__m64*) y, vy_lo);
62       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
63       y += 2;
64     }
65     if (n & (1 * sizeof(float))) {
66       _mm_store_ss(y, vy_lo);
67     }
68   }
69 }
70