1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vlrelu/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_vlrelu_ukernel__avx_x16(size_t n,const float * x,float * y,const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vlrelu_ukernel__avx_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26
27 const __m256 vslope = _mm256_load_ps(params->avx.slope);
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 __m256 vacc01234567 = _mm256_mul_ps(vx01234567, vslope);
34 __m256 vacc89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vslope);
35
36 vacc01234567 = _mm256_blendv_ps(vx01234567, vacc01234567, vx01234567);
37 vacc89ABCDEF = _mm256_blendv_ps(vx89ABCDEF, vacc89ABCDEF, vx89ABCDEF);
38
39 _mm256_storeu_ps(y, vacc01234567);
40 _mm256_storeu_ps(y + 8, vacc89ABCDEF);
41 y += 16;
42 }
43 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
44 const __m256 vx = _mm256_loadu_ps(x);
45 x += 8;
46 __m256 vacc = _mm256_mul_ps(vx, vslope);
47 vacc = _mm256_blendv_ps(vx, vacc, vx);
48 _mm256_storeu_ps(y, vacc);
49 y += 8;
50 }
51 if XNN_UNLIKELY(n != 0) {
52 assert(n >= 1 * sizeof(float));
53 assert(n <= 7 * sizeof(float));
54 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
55
56 const __m256 vx = _mm256_maskload_ps(x, vmask);
57 __m256 vacc = _mm256_mul_ps(vx, vslope);
58 vacc = _mm256_blendv_ps(vx, vacc, vx);
59
60 __m128 vacc_lo = _mm256_castps256_ps128(vacc);
61 if (n & (4 * sizeof(float))) {
62 _mm_storeu_ps(y, vacc_lo);
63 vacc_lo = _mm256_extractf128_ps(vacc, 1);
64 y += 4;
65 }
66 if (n & (2 * sizeof(float))) {
67 _mm_storel_pi((__m64*) y, vacc_lo);
68 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
69 y += 2;
70 }
71 if (n & (1 * sizeof(float))) {
72 _mm_store_ss(y, vacc_lo);
73 }
74 }
75 }
76