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_up8x9__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_up8x9__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 const uint16_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const uint16_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const uint16_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const uint16_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const uint16_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const uint16_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
82 }
83 input = (const void**) ((uintptr_t) input + input_stride);
84
85 size_t c = channels;
86 const uint16_t* w = weights;
87 for (; c >= 8; c -= 8) {
88 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
89
90
91 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
92 i0 += 8;
93
94 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 8)));
95 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
96
97 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
98 i1 += 8;
99
100 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 16)));
101 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
102
103 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
104 i2 += 8;
105
106 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 24)));
107 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
108
109 const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
110 i3 += 8;
111
112 const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 32)));
113 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
114
115 const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
116 i4 += 8;
117
118 const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 40)));
119 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
120
121 const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
122 i5 += 8;
123
124 const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 48)));
125 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
126
127 const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
128 i6 += 8;
129
130 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56)));
131 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
132
133 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
134 i7 += 8;
135
136 const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 64)));
137 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
138
139 const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
140 i8 += 8;
141
142 const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 72)));
143 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
144
145 w += 80;
146
147 // Add up all accumulators to vacc01234567p0
148 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
149
150 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
151 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
152
153 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
154 o += 8;
155 }
156 if XNN_UNLIKELY(c != 0) {
157 assert(c >= 1);
158 assert(c <= 7);
159
160 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
161
162 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
163
164 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 8)));
165 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
166
167 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
168
169 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
170 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
171
172 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
173
174 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 24)));
175 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
176
177 const __m256 vi3x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
178
179 const __m256 vk3x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
180 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi3x01234567, vk3x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
181
182 const __m256 vi4x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
183
184 const __m256 vk4x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 40)));
185 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
186
187 const __m256 vi5x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
188
189 const __m256 vk5x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 48)));
190 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
191
192 const __m256 vi6x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
193
194 const __m256 vk6x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 56)));
195 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
196
197 const __m256 vi7x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
198
199 const __m256 vk7x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
200 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi7x01234567, vk7x01234567, vacc01234567p1), _MM_FROUND_NO_EXC));
201
202 const __m256 vi8x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
203
204 const __m256 vk8x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 72)));
205 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
206
207 // Add up all accumulators to vacc01234567p0
208 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
209
210 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
211 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
212
213 __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC);
214 if (c & 4) {
215 _mm_storel_epi64((__m128i*) o, vh01234567);
216 vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
217 o += 4;
218 }
219 if (c & 2) {
220 _mm_storeu_si32(o, vh01234567);
221 vh01234567 = _mm_srli_epi64(vh01234567, 32);
222 o += 2;
223 }
224 if (c & 1) {
225 *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
226 o += 1;
227 }
228 }
229
230 o = (uint16_t*) ((uintptr_t) o + output_increment);
231 } while (--output_width != 0);
232 }
233