• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-clamp/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/clamp.h>
15 #include <xnnpack/common.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_clamp_ukernel__avx_x8(size_t n,const float * x,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_clamp_ukernel__avx_x8(
21     size_t n,
22     const float* x,
23     float* y,
24     const union xnn_f32_minmax_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 vy_min = _mm256_broadcast_ps((const __m128*) params->sse.min);
32   const __m256 vy_max = _mm256_broadcast_ps((const __m128*) params->sse.max);
33 
34   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35     __m256 vacc01234567 = _mm256_loadu_ps(x);
36     x += 8;
37 
38     vacc01234567 = _mm256_max_ps(vacc01234567, vy_min);
39 
40     vacc01234567 = _mm256_min_ps(vacc01234567, vy_max);
41 
42     _mm256_storeu_ps(y, vacc01234567);
43     y += 8;
44   }
45   if XNN_UNLIKELY(n != 0) {
46     assert(n >= 1 * sizeof(float));
47     assert(n <= 7 * sizeof(float));
48     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
49 
50     __m256 vacc = _mm256_maskload_ps(x, vmask);
51     vacc = _mm256_max_ps(vacc, vy_min);
52     vacc = _mm256_min_ps(vacc, vy_max);
53 
54     // _mm256_maskstore_ps(y, vmask, vacc) could be used here, but triggers msan failures (probably an msan bug).
55     __m128 vacc_lo = _mm256_castps256_ps128(vacc);
56     if (n & (4 * sizeof(float))) {
57       _mm_storeu_ps(y, vacc_lo);
58       vacc_lo = _mm256_extractf128_ps(vacc, 1);
59       y += 4;
60     }
61     if (n & (2 * sizeof(float))) {
62       _mm_storel_pi((__m64*) y, vacc_lo);
63       vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
64       y += 2;
65     }
66     if (n & (1 * sizeof(float))) {
67       _mm_store_ss(y, vacc_lo);
68     }
69   }
70 }
71