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__avx_acc2(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__avx_acc2(
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_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567));
66 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi0x89ABCDEF, vk0x89ABCDEF));
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 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
75 __m256 vacc89ABCDEFp1 = _mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF);
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_add_ps(vacc01234567p0, _mm256_mul_ps(vi2x01234567, vk2x01234567));
84 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, _mm256_mul_ps(vi2x89ABCDEF, vk2x89ABCDEF));
85
86 w += 64;
87
88 // Add up all accumulators to vacc0123456789ABCDEFp0
89 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
90 vacc89ABCDEFp0 = _mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1);
91
92 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
93 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
94 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
95 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);
96
97 _mm256_storeu_ps(output, vacc01234567);
98 _mm256_storeu_ps(output + 8, vacc89ABCDEF);
99 output += 16;
100 }
101 for (; c >= 8; c -= 8) {
102 __m256 vacc01234567p0 = _mm256_load_ps(w);
103
104 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
105 i0 += 8;
106
107 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
108 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567));
109
110 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
111 i1 += 8;
112
113 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
114 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
115
116 const __m256 vi2x01234567 = _mm256_loadu_ps(i2);
117 i2 += 8;
118
119 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
120 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi2x01234567, vk2x01234567));
121
122 w += 8;
123
124 // Add up all accumulators to vacc01234567p0
125 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
126
127 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
128 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
129
130 _mm256_storeu_ps(output, vacc01234567);
131 output += 8;
132 }
133 if XNN_UNLIKELY(c != 0) {
134 assert(c >= 1);
135 assert(c <= 7);
136 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ¶ms->avx.mask_table[7 - c]);
137
138 __m256 vacc01234567p0 = _mm256_load_ps(w);
139
140 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask);
141 const __m256 vk0x01234567 = _mm256_load_ps(w + 16);
142 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567));
143
144 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask);
145 const __m256 vk1x01234567 = _mm256_load_ps(w + 32);
146 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
147
148 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask);
149 const __m256 vk2x01234567 = _mm256_load_ps(w + 48);
150 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi2x01234567, vk2x01234567));
151
152 // Add up all accumulators to vacc01234567p0
153 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
154
155 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
156 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
157
158 __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567);
159 if (c & 4) {
160 _mm_storeu_ps(output, vacc0123);
161 vacc0123 = _mm256_extractf128_ps(vacc01234567, 1);
162 output += 4;
163 }
164 if (c & 2) {
165 _mm_storel_pi((__m64*) output, vacc0123);
166 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
167 output += 2;
168 }
169 if (c & 1) {
170 _mm_store_ss(output, vacc0123);
171 output += 1;
172 }
173 }
174
175 output = (float*) ((uintptr_t) output + output_increment);
176 } while (--output_width != 0);
177 }
178