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_2x16(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_2x16(
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
38 const size_t input_increment = input_stride * 2 - channels;
39 const size_t output_increment = output_stride * 2 - channels;
40
41 do {
42 if XNN_UNPREDICTABLE(rows < 2) {
43 i1 = i0;
44 o1 = o0;
45 }
46
47 const float* w = weights;
48 size_t c = channels;
49 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
50 const __m256 vw01234567 = _mm256_load_ps(w);
51 const __m256 vw89ABCDEF = _mm256_load_ps(w + 8);
52 w += 16;
53
54 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
55 const __m256 vi0x89ABCDEF = _mm256_loadu_ps(i0 + 8);
56 i0 += 16;
57 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
58 const __m256 vi1x89ABCDEF = _mm256_loadu_ps(i1 + 8);
59 i1 += 16;
60
61 const __m256 vprod0x01234567 = _mm256_mul_ps(vi0x01234567, vw01234567);
62 const __m256 vprod0x89ABCDEF = _mm256_mul_ps(vi0x89ABCDEF, vw89ABCDEF);
63 const __m256 vprod1x01234567 = _mm256_mul_ps(vi1x01234567, vw01234567);
64 const __m256 vprod1x89ABCDEF = _mm256_mul_ps(vi1x89ABCDEF, vw89ABCDEF);
65
66 const __m256 vacc0x01234567 = _mm256_blendv_ps(vi0x01234567, vprod0x01234567, vi0x01234567);
67 const __m256 vacc0x89ABCDEF = _mm256_blendv_ps(vi0x89ABCDEF, vprod0x89ABCDEF, vi0x89ABCDEF);
68 const __m256 vacc1x01234567 = _mm256_blendv_ps(vi1x01234567, vprod1x01234567, vi1x01234567);
69 const __m256 vacc1x89ABCDEF = _mm256_blendv_ps(vi1x89ABCDEF, vprod1x89ABCDEF, vi1x89ABCDEF);
70
71 _mm256_storeu_ps(o0, vacc0x01234567);
72 _mm256_storeu_ps(o0 + 8, vacc0x89ABCDEF);
73 o0 += 16;
74 _mm256_storeu_ps(o1, vacc1x01234567);
75 _mm256_storeu_ps(o1 + 8, vacc1x89ABCDEF);
76 o1 += 16;
77 }
78 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
79 const __m256 vw = _mm256_load_ps(w);
80 w += 8;
81
82 const __m256 vi0 = _mm256_loadu_ps(i0);
83 i0 += 8;
84 const __m256 vi1 = _mm256_loadu_ps(i1);
85 i1 += 8;
86
87 const __m256 vprod0 = _mm256_mul_ps(vi0, vw);
88 const __m256 vprod1 = _mm256_mul_ps(vi1, vw);
89
90 const __m256 vacc0 = _mm256_blendv_ps(vi0, vprod0, vi0);
91 const __m256 vacc1 = _mm256_blendv_ps(vi1, vprod1, vi1);
92
93 _mm256_storeu_ps(o0, vacc0);
94 o0 += 8;
95 _mm256_storeu_ps(o1, vacc1);
96 o1 += 8;
97 }
98 if XNN_UNLIKELY(c != 0) {
99 assert(c >= 1 * sizeof(float));
100 assert(c <= 7 * sizeof(float));
101 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - c));
102
103 const __m256 vw = _mm256_maskload_ps(w, vmask);
104
105 const __m256 vi0 = _mm256_maskload_ps(i0, vmask);
106 i0 = (const float*) ((uintptr_t) i0 + c);
107 const __m256 vi1 = _mm256_maskload_ps(i1, vmask);
108 i1 = (const float*) ((uintptr_t) i1 + c);
109
110 const __m256 vprod0 = _mm256_mul_ps(vi0, vw);
111 const __m256 vprod1 = _mm256_mul_ps(vi1, vw);
112
113 __m256 vacc0 = _mm256_blendv_ps(vi0, vprod0, vi0);
114 __m256 vacc1 = _mm256_blendv_ps(vi1, vprod1, vi1);
115
116 __m128 vacc0_lo = _mm256_castps256_ps128(vacc0);
117 __m128 vacc1_lo = _mm256_castps256_ps128(vacc1);
118 if (c & (4 * sizeof(float))) {
119 _mm_storeu_ps(o0, vacc0_lo);
120 _mm_storeu_ps(o1, vacc1_lo);
121
122 vacc0_lo = _mm256_extractf128_ps(vacc0, 1);
123 vacc1_lo = _mm256_extractf128_ps(vacc1, 1);
124
125 o0 += 4;
126 o1 += 4;
127 }
128 if (c & (2 * sizeof(float))) {
129 _mm_storel_pi((__m64*) o0, vacc0_lo);
130 _mm_storel_pi((__m64*) o1, vacc1_lo);
131
132 vacc0_lo = _mm_movehl_ps(vacc0_lo, vacc0_lo);
133 vacc1_lo = _mm_movehl_ps(vacc1_lo, vacc1_lo);
134
135 o0 += 2;
136 o1 += 2;
137 }
138 if (c & (1 * sizeof(float))) {
139 _mm_store_ss(o0, vacc0_lo);
140 _mm_store_ss(o1, vacc1_lo);
141
142 o0 += 1;
143 o1 += 1;
144 }
145 }
146 i0 = (const float*) ((uintptr_t) i0 + input_increment);
147 o0 = (float*) ((uintptr_t) o0 + output_increment);
148 i1 = (const float*) ((uintptr_t) i1 + input_increment);
149 o1 = (float*) ((uintptr_t) o1 + output_increment);
150 rows = doz(rows, 2);
151 } while (rows != 0);
152 }
153