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
xnn_f32_dwconv_minmax_ukernel_up16x3__fma3(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_dwconv_minmax_ukernel_up16x3__fma3(
18 size_t channels,
19 size_t output_width,
20 const float** input,
21 const float* weights,
22 float* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const float* zero,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const __m256 vmax = _mm256_load_ps(params->avx.max);
33 const __m256 vmin = _mm256_load_ps(params->avx.min);
34 do {
35 const float* i0 = input[0];
36 assert(i0 != NULL);
37 if XNN_UNPREDICTABLE(i0 != zero) {
38 i0 = (const float*) ((uintptr_t) i0 + input_offset);
39 }
40 const float* i1 = input[1];
41 assert(i1 != NULL);
42 if XNN_UNPREDICTABLE(i1 != zero) {
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 }
45 const float* i2 = input[2];
46 assert(i2 != NULL);
47 if XNN_UNPREDICTABLE(i2 != zero) {
48 i2 = (const float*) ((uintptr_t) i2 + input_offset);
49 }
50 input = (const float**) ((uintptr_t) input + input_stride);
51
52 size_t c = channels;
53 const float* w = weights;
54 for (; c >= 16; c -= 16) {
55 __m256 vacc01234567p0 = _mm256_load_ps(w);
56 __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8);
57
58
59 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
60 const __m256 vi0x89ABCDEF = _mm256_loadu_ps(i0 + 8);
61 i0 += 16;
62
63 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
64 const __m256 vk0x89ABCDEF = _mm256_load_ps(w + 24);
65 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
66 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0);
67
68 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
69 const __m256 vi1x89ABCDEF = _mm256_loadu_ps(i1 + 8);
70 i1 += 16;
71
72 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
73 const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40);
74 vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);
75 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0);
76
77 const __m256 vi2x01234567 = _mm256_loadu_ps(i2);
78 const __m256 vi2x89ABCDEF = _mm256_loadu_ps(i2 + 8);
79 i2 += 16;
80
81 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
82 const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56);
83 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
84 vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0);
85
86 w += 64;
87
88
89 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
90 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
91 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
92 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);
93
94 _mm256_storeu_ps(output, vacc01234567);
95 _mm256_storeu_ps(output + 8, vacc89ABCDEF);
96 output += 16;
97 }
98 for (; c >= 8; c -= 8) {
99 __m256 vacc01234567p0 = _mm256_load_ps(w);
100
101 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
102 i0 += 8;
103
104 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
105 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
106
107 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
108 i1 += 8;
109
110 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
111 vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);
112
113 const __m256 vi2x01234567 = _mm256_loadu_ps(i2);
114 i2 += 8;
115
116 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
117 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
118
119 w += 8;
120
121
122 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
123 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
124
125 _mm256_storeu_ps(output, vacc01234567);
126 output += 8;
127 }
128 if XNN_UNLIKELY(c != 0) {
129 assert(c >= 1);
130 assert(c <= 7);
131 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ¶ms->avx.mask_table[7 - c]);
132
133 __m256 vacc01234567p0 = _mm256_load_ps(w);
134
135 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask);
136 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
137 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
138
139 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask);
140 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
141 vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);
142
143 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask);
144 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
145 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
146
147
148 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
149 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
150
151 __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567);
152 if (c & 4) {
153 _mm_storeu_ps(output, vacc0123);
154 vacc0123 = _mm256_extractf128_ps(vacc01234567, 1);
155 output += 4;
156 }
157 if (c & 2) {
158 _mm_storel_pi((__m64*) output, vacc0123);
159 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
160 output += 2;
161 }
162 if (c & 1) {
163 _mm_store_ss(output, vacc0123);
164 output += 1;
165 }
166 }
167
168 output = (float*) ((uintptr_t) output + output_increment);
169 } while (--output_width != 0);
170 }
171