1 // Auto-generated file. Do not edit!
2 // Template: src/f16-dwconv/up-fma3.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 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_f16_dwconv_minmax_ukernel_up8x3__fma3_acc2(size_t channels,size_t output_width,const void ** input,const void * weights,void * output,size_t input_stride,size_t output_increment,size_t input_offset,const void * zero,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_dwconv_minmax_ukernel_up8x3__fma3_acc2(
19 size_t channels,
20 size_t output_width,
21 const void** input,
22 const void* weights,
23 void* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const void* zero,
28 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m256 vmax = _mm256_load_ps(params->avx.max);
34 const __m256 vmin = _mm256_load_ps(params->avx.min);
35
36 uint16_t* o = (uint16_t*) output;
37 do {
38 const uint16_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const uint16_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const uint16_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
52 }
53 input = (const void**) ((uintptr_t) input + input_stride);
54
55 size_t c = channels;
56 const uint16_t* w = weights;
57 for (; c >= 8; c -= 8) {
58 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
59
60
61 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
62 i0 += 8;
63
64 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 8)));
65 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
66
67 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
68 i1 += 8;
69
70 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
71 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
72
73 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
74 i2 += 8;
75
76 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 24)));
77 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
78
79 w += 32;
80
81 // Add up all accumulators to vacc01234567p0
82 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
83
84 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
85 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
86
87 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
88 o += 8;
89 }
90 if XNN_UNLIKELY(c != 0) {
91 assert(c >= 1);
92 assert(c <= 7);
93
94 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
95
96 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
97
98 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 8)));
99 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
100
101 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
102
103 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
104 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
105
106 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
107
108 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 24)));
109 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
110
111 // Add up all accumulators to vacc01234567p0
112 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
113
114 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
115 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
116
117 __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC);
118 if (c & 4) {
119 _mm_storel_epi64((__m128i*) o, vh01234567);
120 vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
121 o += 4;
122 }
123 if (c & 2) {
124 _mm_storeu_si32(o, vh01234567);
125 vh01234567 = _mm_srli_epi64(vh01234567, 32);
126 o += 2;
127 }
128 if (c & 1) {
129 *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
130 o += 1;
131 }
132 }
133
134 o = (uint16_t*) ((uintptr_t) o + output_increment);
135 } while (--output_width != 0);
136 }
137