xref: /aosp_15_r20/external/XNNPACK/src/f32-vsqrt/gen/avx-sqrt-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
28     const __m256 vx = _mm256_loadu_ps(x);
29     x += 8;
30     const __m256 vy = _mm256_sqrt_ps(vx);
31     _mm256_storeu_ps(y, vy);
32     y += 8;
33   }
34   if XNN_UNLIKELY(n != 0) {
35     assert(n >= 1 * sizeof(float));
36     assert(n <= 7 * sizeof(float));
37     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
38 
39     const __m256 vx = _mm256_maskload_ps(x, vmask);
40     const __m256 vy = _mm256_sqrt_ps(vx);
41 
42     __m128 vy_lo = _mm256_castps256_ps128(vy);
43     if (n & (4 * sizeof(float))) {
44       _mm_storeu_ps(y, vy_lo);
45       vy_lo = _mm256_extractf128_ps(vy, 1);
46       y += 4;
47     }
48     if (n & (2 * sizeof(float))) {
49       _mm_storel_pi((__m64*) y, vy_lo);
50       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
51       y += 2;
52     }
53     if (n & (1 * sizeof(float))) {
54       _mm_store_ss(y, vy_lo);
55     }
56   }
57 }
58