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_up32x3__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_up32x3__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 >= 32; c -= 32) {
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 __m256 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 16)));
61 __m256 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 24)));
62
63
64 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
65 const __m256 vi0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 8)));
66 const __m256 vi0xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 16)));
67 const __m256 vi0xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 24)));
68 i0 += 32;
69
70 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 32)));
71 const __m256 vk0x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 40)));
72 const __m256 vk0xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 48)));
73 const __m256 vk0xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 56)));
74 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
75 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
76 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0xGHIJKLMN, vk0xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
77 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0xOPQRSTUV, vk0xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
78
79 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
80 const __m256 vi1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 8)));
81 const __m256 vi1xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 16)));
82 const __m256 vi1xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 24)));
83 i1 += 32;
84
85 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 64)));
86 const __m256 vk1x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 72)));
87 const __m256 vk1xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 80)));
88 const __m256 vk1xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 88)));
89 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
90 __m256 vacc89ABCDEFp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x89ABCDEF, vk1x89ABCDEF), _MM_FROUND_NO_EXC));
91 __m256 vaccGHIJKLMNp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1xGHIJKLMN, vk1xGHIJKLMN), _MM_FROUND_NO_EXC));
92 __m256 vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1xOPQRSTUV, vk1xOPQRSTUV), _MM_FROUND_NO_EXC));
93
94 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
95 const __m256 vi2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 8)));
96 const __m256 vi2xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 16)));
97 const __m256 vi2xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 24)));
98 i2 += 32;
99
100 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 96)));
101 const __m256 vk2x89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 104)));
102 const __m256 vk2xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 112)));
103 const __m256 vk2xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 120)));
104 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
105 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0), _MM_FROUND_NO_EXC));
106 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2xGHIJKLMN, vk2xGHIJKLMN, vaccGHIJKLMNp0), _MM_FROUND_NO_EXC));
107 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2xOPQRSTUV, vk2xOPQRSTUV, vaccOPQRSTUVp0), _MM_FROUND_NO_EXC));
108
109 w += 128;
110
111 // Add up all accumulators to vacc0123456789ABCDEFGHIJKLMNOPQRSTUVp0
112 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
113 vacc89ABCDEFp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc89ABCDEFp0, vacc89ABCDEFp1), _MM_FROUND_NO_EXC));
114 vaccGHIJKLMNp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vaccGHIJKLMNp0, vaccGHIJKLMNp1), _MM_FROUND_NO_EXC));
115 vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vaccOPQRSTUVp0, vaccOPQRSTUVp1), _MM_FROUND_NO_EXC));
116
117 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
118 __m256 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEFp0, vmin);
119 __m256 vaccGHIJKLMN = _mm256_max_ps(vaccGHIJKLMNp0, vmin);
120 __m256 vaccOPQRSTUV = _mm256_max_ps(vaccOPQRSTUVp0, vmin);
121 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
122 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vmax);
123 vaccGHIJKLMN = _mm256_min_ps(vaccGHIJKLMN, vmax);
124 vaccOPQRSTUV = _mm256_min_ps(vaccOPQRSTUV, vmax);
125
126 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
127 _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vacc89ABCDEF, _MM_FROUND_NO_EXC));
128 _mm_storeu_si128((__m128i*) (o + 16), _mm256_cvtps_ph(vaccGHIJKLMN, _MM_FROUND_NO_EXC));
129 _mm_storeu_si128((__m128i*) (o + 24), _mm256_cvtps_ph(vaccOPQRSTUV, _MM_FROUND_NO_EXC));
130 o += 32;
131 }
132 for (; c >= 8; c -= 8) {
133 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
134
135 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
136 i0 += 8;
137
138 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
139 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
140
141 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
142 i1 += 8;
143
144 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
145 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
146
147 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
148 i2 += 8;
149
150 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 96)));
151 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
152
153 w += 8;
154
155 // Add up all accumulators to vacc01234567p0
156 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
157
158 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
159 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
160
161 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC));
162 o += 8;
163 }
164 if XNN_UNLIKELY(c != 0) {
165 assert(c >= 1);
166 assert(c <= 7);
167
168 __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
169
170 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
171
172 const __m256 vk0x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 32)));
173 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
174
175 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
176
177 const __m256 vk1x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 64)));
178 __m256 vacc01234567p1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi1x01234567, vk1x01234567), _MM_FROUND_NO_EXC));
179
180 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
181
182 const __m256 vk2x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 96)));
183 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0), _MM_FROUND_NO_EXC));
184
185 // Add up all accumulators to vacc01234567p0
186 vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p0, vacc01234567p1), _MM_FROUND_NO_EXC));
187
188 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
189 vacc01234567 = _mm256_min_ps(vacc01234567, vmax);
190
191 __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_NO_EXC);
192 if (c & 4) {
193 _mm_storel_epi64((__m128i*) o, vh01234567);
194 vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
195 o += 4;
196 }
197 if (c & 2) {
198 _mm_storeu_si32(o, vh01234567);
199 vh01234567 = _mm_srli_epi64(vh01234567, 32);
200 o += 2;
201 }
202 if (c & 1) {
203 *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
204 o += 1;
205 }
206 }
207
208 o = (uint16_t*) ((uintptr_t) o + output_increment);
209 } while (--output_width != 0);
210 }
211