1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-avx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/dwconv.h>
15
16
17 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
18
xnn_f32_dwconv_ukernel_up8x4__fma3(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,const union xnn_f32_output_params params[restrict static1])19 void xnn_f32_dwconv_ukernel_up8x4__fma3(
20 size_t channels,
21 size_t output_width,
22 const float** input,
23 const float* weights,
24 float* output,
25 size_t input_stride,
26 size_t output_increment,
27 const union xnn_f32_output_params params[restrict static 1])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
33 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 const float* i1 = input[1];
38 assert(i1 != NULL);
39 const float* i2 = input[2];
40 assert(i2 != NULL);
41 const float* i3 = input[3];
42 assert(i3 != NULL);
43 input = (const float**) ((uintptr_t) input + input_stride);
44
45 size_t c = channels;
46 const float* w = weights;
47 for (; c >= 8; c -= 8) {
48 __m256 vacc01234567p0 = _mm256_load_ps(w);
49
50
51 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
52 i0 += 8;
53
54 const __m256 vk0x01234567 = _mm256_load_ps(w + 8);
55 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
56
57 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
58 i1 += 8;
59
60 const __m256 vk1x01234567 = _mm256_load_ps(w + 16);
61 vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);
62
63 const __m256 vi2x01234567 = _mm256_loadu_ps(i2);
64 i2 += 8;
65
66 const __m256 vk2x01234567 = _mm256_load_ps(w + 24);
67 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
68
69 const __m256 vi3x01234567 = _mm256_loadu_ps(i3);
70 i3 += 8;
71
72 const __m256 vk3x01234567 = _mm256_load_ps(w + 32);
73 vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);
74
75 w += 40;
76
77
78 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
79 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
80
81 _mm256_storeu_ps(output, vacc01234567);
82 output += 8;
83 }
84 if XNN_UNLIKELY(c != 0) {
85 assert(c >= 1);
86 assert(c <= 7);
87 __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[7 - c]);
88
89 __m256 vacc01234567p0 = _mm256_load_ps(w);
90
91 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask);
92 const __m256 vk0x01234567 = _mm256_load_ps(w + 8);
93 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
94
95 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask);
96 const __m256 vk1x01234567 = _mm256_load_ps(w + 16);
97 vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);
98
99 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask);
100 const __m256 vk2x01234567 = _mm256_load_ps(w + 24);
101 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
102
103 const __m256 vi3x01234567 = _mm256_maskload_ps(i3, vmask);
104 const __m256 vk3x01234567 = _mm256_load_ps(w + 32);
105 vacc01234567p0 = _mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p0);
106
107
108 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
109 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
110
111 // _mm256_maskstore_ps(output, vmask, vacc01234567); output += c; could be used here, but triggers msan failures (probably an msan bug).
112 __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567);
113 if (c & 4) {
114 _mm_storeu_ps(output, vacc0123);
115 vacc0123 = _mm256_extractf128_ps(vacc01234567, 1);
116 output += 4;
117 }
118 if (c & 2) {
119 _mm_storel_pi((__m64*) output, vacc0123);
120 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
121 output += 2;
122 }
123 if (c & 1) {
124 _mm_store_ss(output, vacc0123);
125 output += 1;
126 }
127 }
128
129 output = (float*) ((uintptr_t) output + output_increment);
130 } while (--output_width != 0);
131 }
132