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
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_vneg_ukernel__avx_x16(size_t n,const float * x,float * y,const union xnn_f32_neg_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vneg_ukernel__avx_x16(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_neg_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28 assert(x != NULL);
29 assert(y != NULL);
30
31 const __m256 vsign_mask = _mm256_broadcast_ps((const __m128*) params->sse.sign_mask);
32 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33 const __m256 vx01234567 = _mm256_loadu_ps(x);
34 const __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
35 x += 16;
36
37 const __m256 vy01234567 = _mm256_xor_ps(vx01234567, vsign_mask);
38 const __m256 vy89ABCDEF = _mm256_xor_ps(vx89ABCDEF, vsign_mask);
39
40 _mm256_storeu_ps(y, vy01234567);
41 _mm256_storeu_ps(y + 8, vy89ABCDEF);
42 y += 16;
43 }
44 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
45 const __m256 vx = _mm256_loadu_ps(x);
46 x += 8;
47 const __m256 vy = _mm256_xor_ps(vx, vsign_mask);
48 _mm256_storeu_ps(y, vy);
49 y += 8;
50 }
51 if XNN_UNLIKELY(n != 0) {
52 assert(n >= 1 * sizeof(float));
53 assert(n <= 7 * sizeof(float));
54 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
55
56 const __m256 vx = _mm256_maskload_ps(x, vmask);
57 const __m256 vy = _mm256_xor_ps(vx, vsign_mask);
58
59 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
60 __m128 vy_lo = _mm256_castps256_ps128(vy);
61 if (n & (4 * sizeof(float))) {
62 _mm_storeu_ps(y, vy_lo);
63 vy_lo = _mm256_extractf128_ps(vy, 1);
64 y += 4;
65 }
66 if (n & (2 * sizeof(float))) {
67 _mm_storel_pi((__m64*) y, vy_lo);
68 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
69 y += 2;
70 }
71 if (n & (1 * sizeof(float))) {
72 _mm_store_ss(y, vy_lo);
73 }
74 }
75 }
76