• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vrnd/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/math.h>
16 #include <xnnpack/vunary.h>
17 
18 
xnn_f32_vrndne_ukernel__avx_x16(size_t n,const float * x,float * y,const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vrndne_ukernel__avx_x16(
20     size_t n,
21     const float* x,
22     float* y,
23     const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27 
28   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
29     const __m256 vx01234567 = _mm256_loadu_ps(x);
30     const __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
31     x += 16;
32 
33     const __m256 vy01234567 = _mm256_round_ps(vx01234567, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
34     const __m256 vy89ABCDEF = _mm256_round_ps(vx89ABCDEF, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
35 
36     _mm256_storeu_ps(y, vy01234567);
37     _mm256_storeu_ps(y + 8, vy89ABCDEF);
38     y += 16;
39   }
40   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
41     const __m256 vx = _mm256_loadu_ps(x);
42     x += 8;
43 
44     const __m256 vy = _mm256_round_ps(vx, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
45 
46     _mm256_storeu_ps(y, vy);
47     y += 8;
48   }
49   if XNN_UNLIKELY(n != 0) {
50     assert(n >= 1 * sizeof(float));
51     assert(n <= 7 * sizeof(float));
52     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx.mask_table[7] - n));
53 
54     const __m256 vx = _mm256_maskload_ps(x, vmask);
55     const __m256 vy = _mm256_round_ps(vx, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
56 
57     __m128 vy_lo = _mm256_castps256_ps128(vy);
58     if (n & (4 * sizeof(float))) {
59       _mm_storeu_ps(y, vy_lo);
60       vy_lo = _mm256_extractf128_ps(vy, 1);
61       y += 4;
62     }
63     if (n & (2 * sizeof(float))) {
64       _mm_storel_pi((__m64*) y, vy_lo);
65       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
66       y += 2;
67     }
68     if (n & (1 * sizeof(float))) {
69       _mm_store_ss(y, vy_lo);
70     }
71   }
72 }
73