1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/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/math.h>
15 #include <xnnpack/prelu.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_prelu_ukernel__avx_2x8(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride)20 void xnn_f32_prelu_ukernel__avx_2x8(
21 size_t rows,
22 size_t channels,
23 const float*restrict input,
24 size_t input_stride,
25 const float*restrict weights,
26 float*restrict output,
27 size_t output_stride)
28 {
29 assert(rows != 0);
30 assert(channels != 0);
31 assert(channels % sizeof(float) == 0);
32
33 const float* i0 = input;
34 float* o0 = output;
35 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
36 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
37 if XNN_UNPREDICTABLE(rows < 2) {
38 i1 = i0;
39 o1 = o0;
40 }
41
42 const size_t input_increment = input_stride * 2 - channels;
43 const size_t output_increment = output_stride * 2 - channels;
44
45 do {
46 const float* w = weights;
47 size_t c = channels;
48 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
49 const __m256 vw01234567 = _mm256_load_ps(w);
50 w += 8;
51
52 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
53 i0 += 8;
54 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
55 i1 += 8;
56
57 const __m256 vprod0x01234567 = _mm256_mul_ps(vi0x01234567, vw01234567);
58 const __m256 vprod1x01234567 = _mm256_mul_ps(vi1x01234567, vw01234567);
59
60 const __m256 vacc0x01234567 = _mm256_blendv_ps(vi0x01234567, vprod0x01234567, vi0x01234567);
61 const __m256 vacc1x01234567 = _mm256_blendv_ps(vi1x01234567, vprod1x01234567, vi1x01234567);
62
63 _mm256_storeu_ps(o0, vacc0x01234567);
64 o0 += 8;
65 _mm256_storeu_ps(o1, vacc1x01234567);
66 o1 += 8;
67 }
68 if XNN_UNLIKELY(c != 0) {
69 assert(c >= 1 * sizeof(float));
70 assert(c <= 7 * sizeof(float));
71 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - c));
72
73 const __m256 vw = _mm256_maskload_ps(w, vmask);
74
75 const __m256 vi0 = _mm256_maskload_ps(i0, vmask);
76 i0 = (const float*) ((uintptr_t) i0 + c);
77 const __m256 vi1 = _mm256_maskload_ps(i1, vmask);
78 i1 = (const float*) ((uintptr_t) i1 + c);
79
80 const __m256 vprod0 = _mm256_mul_ps(vi0, vw);
81 const __m256 vprod1 = _mm256_mul_ps(vi1, vw);
82
83 __m256 vacc0 = _mm256_blendv_ps(vi0, vprod0, vi0);
84 __m256 vacc1 = _mm256_blendv_ps(vi1, vprod1, vi1);
85
86 // _mm256_maskstore_ps(o1, vmask, vacc1) could be used here, but triggers msan failures (probably an msan bug).
87 __m128 vacc0_lo = _mm256_castps256_ps128(vacc0);
88 __m128 vacc1_lo = _mm256_castps256_ps128(vacc1);
89 if (c & (4 * sizeof(float))) {
90 _mm_storeu_ps(o0, vacc0_lo);
91 _mm_storeu_ps(o1, vacc1_lo);
92
93 vacc0_lo = _mm256_extractf128_ps(vacc0, 1);
94 vacc1_lo = _mm256_extractf128_ps(vacc1, 1);
95
96 o0 += 4;
97 o1 += 4;
98 }
99 if (c & (2 * sizeof(float))) {
100 _mm_storel_pi((__m64*) o0, vacc0_lo);
101 _mm_storel_pi((__m64*) o1, vacc1_lo);
102
103 vacc0_lo = _mm_movehl_ps(vacc0_lo, vacc0_lo);
104 vacc1_lo = _mm_movehl_ps(vacc1_lo, vacc1_lo);
105
106 o0 += 2;
107 o1 += 2;
108 }
109 if (c & (1 * sizeof(float))) {
110 _mm_store_ss(o0, vacc0_lo);
111 _mm_store_ss(o1, vacc1_lo);
112
113 o0 += 1;
114 o1 += 1;
115 }
116 }
117 i0 = (const float*) ((uintptr_t) i0 + input_increment);
118 o0 = (float*) ((uintptr_t) o0 + output_increment);
119 i1 = (const float*) ((uintptr_t) i1 + input_increment);
120 o1 = (float*) ((uintptr_t) o1 + output_increment);
121 if XNN_UNPREDICTABLE(rows < 4) {
122 i1 = i0;
123 o1 = o0;
124 }
125 rows = doz(rows, 2);
126 } while (rows != 0);
127 }
128