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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
30 const __m256 vx01234567 = _mm256_loadu_ps(x);
31 const __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
32 x += 16;
33
34 const __m256 vy01234567 = _mm256_mul_ps(vx01234567, vx01234567);
35 const __m256 vy89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vx89ABCDEF);
36
37 _mm256_storeu_ps(y, vy01234567);
38 _mm256_storeu_ps(y + 8, vy89ABCDEF);
39 y += 16;
40 }
41 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
42 const __m256 vx = _mm256_loadu_ps(x);
43 x += 8;
44 const __m256 vy = _mm256_mul_ps(vx, vx);
45 _mm256_storeu_ps(y, vy);
46 y += 8;
47 }
48 if XNN_UNLIKELY(n != 0) {
49 assert(n >= 1 * sizeof(float));
50 assert(n <= 7 * sizeof(float));
51 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
52
53 const __m256 vx = _mm256_maskload_ps(x, vmask);
54 const __m256 vy = _mm256_mul_ps(vx, vx);
55
56 __m128 vy_lo = _mm256_castps256_ps128(vy);
57 if (n & (4 * sizeof(float))) {
58 _mm_storeu_ps(y, vy_lo);
59 vy_lo = _mm256_extractf128_ps(vy, 1);
60 y += 4;
61 }
62 if (n & (2 * sizeof(float))) {
63 _mm_storel_pi((__m64*) y, vy_lo);
64 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
65 y += 2;
66 }
67 if (n & (1 * sizeof(float))) {
68 _mm_store_ss(y, vy_lo);
69 }
70 }
71 }
72