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_up16x3__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_up16x3__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 >= 16; c -= 16) {
58 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
59 __m256 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 8)));
60
61
62 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
63 const __m256 vi0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 8)));
64 i0 += 16;
65
66 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
67 const __m256 vk0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 24)));
68 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
69 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
70
71 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
72 const __m256 vi1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 8)));
73 i1 += 16;
74
75 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 32)));
76 const __m256 vk1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 40)));
77 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
78 __m256 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF), _MM_FROUND_NO_EXC));
79
80 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
81 const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 8)));
82 i2 += 16;
83
84 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 48)));
85 const __m256 vk2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56)));
86 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
87 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
88
89 w += 64;
90
91 // Add up all accumulators to vacc0123456789ABCDEFp0
92 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
93 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
94
95 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
96 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
97 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
98 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);
99
100 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
101 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vacc89ABCDEF, _MM_FROUND_NO_EXC));
102 o += 16;
103 }
104 for (; c >= 8; c -= 8) {
105 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
106
107 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
108 i0 += 8;
109
110 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
111 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
112
113 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
114 i1 += 8;
115
116 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
117 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
118
119 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
120 i2 += 8;
121
122 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
123 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
124
125 w += 8;
126
127 // Add up all accumulators to vacc01234567p0
128 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
129
130 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
131 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
132
133 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
134 o += 8;
135 }
136 if XNN_UNLIKELY(c != 0) {
137 assert(c >= 1);
138 assert(c <= 7);
139
140 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
141
142 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
143
144 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
145 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
146
147 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
148
149 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
150 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
151
152 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
153
154 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
155 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
156
157 // Add up all accumulators to vacc01234567p0
158 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
159
160 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
161 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
162
163 __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC);
164 if (c & 4) {
165 _mm_storel_epi64((__m128i*) o, vh01234567);
166 vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
167 o += 4;
168 }
169 if (c & 2) {
170 _mm_storeu_si32(o, vh01234567);
171 vh01234567 = _mm_srli_epi64(vh01234567, 32);
172 o += 2;
173 }
174 if (c & 1) {
175 *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
176 o += 1;
177 }
178 }
179
180 o = (uint16_t*) ((uintptr_t) o + output_increment);
181 } while (--output_width != 0);
182 }
183