1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vunary/avx.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_vsqr_ukernel__avx_x8(size_t n,const float * x,float * y,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsqr_ukernel__avx_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
30 const __m256 vx01234567 = _mm256_loadu_ps(x);
31 x += 8;
32
33 const __m256 vy01234567 = _mm256_mul_ps(vx01234567, vx01234567);
34
35 _mm256_storeu_ps(y, vy01234567);
36 y += 8;
37 }
38 if XNN_UNLIKELY(n != 0) {
39 assert(n >= 1 * sizeof(float));
40 assert(n <= 7 * sizeof(float));
41 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
42
43 const __m256 vx = _mm256_maskload_ps(x, vmask);
44 const __m256 vy = _mm256_mul_ps(vx, vx);
45
46 __m128 vy_lo = _mm256_castps256_ps128(vy);
47 if (n & (4 * sizeof(float))) {
48 _mm_storeu_ps(y, vy_lo);
49 vy_lo = _mm256_extractf128_ps(vy, 1);
50 y += 4;
51 }
52 if (n & (2 * sizeof(float))) {
53 _mm_storel_pi((__m64*) y, vy_lo);
54 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
55 y += 2;
56 }
57 if (n & (1 * sizeof(float))) {
58 _mm_store_ss(y, vy_lo);
59 }
60 }
61 }
62