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_up8x3__fma3_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_up8x3__fma3_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 >= 8; c -= 8) {
55 __m256 vacc01234567p0 = _mm256_load_ps(w);
56
57
58 const __m256 vi0x01234567 = _mm256_loadu_ps(i0);
59 i0 += 8;
60
61 const __m256 vk0x01234567 = _mm256_load_ps(w + 8);
62 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
63
64 const __m256 vi1x01234567 = _mm256_loadu_ps(i1);
65 i1 += 8;
66
67 const __m256 vk1x01234567 = _mm256_load_ps(w + 16);
68 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
69
70 const __m256 vi2x01234567 = _mm256_loadu_ps(i2);
71 i2 += 8;
72
73 const __m256 vk2x01234567 = _mm256_load_ps(w + 24);
74 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
75
76 w += 32;
77
78 // Add up all accumulators to vacc01234567p0
79 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
80
81 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
82 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
83
84 _mm256_storeu_ps(output, vacc01234567);
85 output += 8;
86 }
87 if XNN_UNLIKELY(c != 0) {
88 assert(c >= 1);
89 assert(c <= 7);
90 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ¶ms->avx.mask_table[7 - c]);
91
92 __m256 vacc01234567p0 = _mm256_load_ps(w);
93
94 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask);
95 const __m256 vk0x01234567 = _mm256_load_ps(w + 8);
96 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);
97
98 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask);
99 const __m256 vk1x01234567 = _mm256_load_ps(w + 16);
100 __m256 vacc01234567p1 = _mm256_mul_ps(vi1x01234567, vk1x01234567);
101
102 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask);
103 const __m256 vk2x01234567 = _mm256_load_ps(w + 24);
104 vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);
105
106 // Add up all accumulators to vacc01234567p0
107 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);
108
109 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
110 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
111
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