1 // Copyright 2022 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <immintrin.h>
9
10 #include <xnnpack/avgpool.h>
11 #include <xnnpack/intrinsics-polyfill.h>
12
13
xnn_f16_avgpool_minmax_ukernel_9x__f16c_c8(size_t output_pixels,size_t kernel_elements,size_t channels,const void ** input,size_t input_offset,const void * zero,void * output,size_t input_increment,size_t output_increment,const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])14 void xnn_f16_avgpool_minmax_ukernel_9x__f16c_c8(
15 size_t output_pixels,
16 size_t kernel_elements,
17 size_t channels,
18 const void** input,
19 size_t input_offset,
20 const void* zero,
21 void* output,
22 size_t input_increment,
23 size_t output_increment,
24 const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(output_pixels != 0);
27 assert(kernel_elements != 0);
28 assert(kernel_elements <= 9);
29 assert(channels != 0);
30
31 const __m256 vscale = _mm256_load_ps(params->avx.scale);
32 const __m256 vmin = _mm256_load_ps(params->avx.min);
33 const __m256 vmax = _mm256_load_ps(params->avx.max);
34
35 uint16_t* o = (uint16_t*) output;
36 do {
37 const uint16_t* i0 = (const uint16_t*) input[0];
38 assert(i0 != NULL);
39 const uint16_t* i1 = (const uint16_t*) input[1];
40 const uint16_t* i2 = (const uint16_t*) input[2];
41 const uint16_t* i3 = (const uint16_t*) input[3];
42 const uint16_t* i4 = (const uint16_t*) input[4];
43 const uint16_t* i5 = (const uint16_t*) input[5];
44 const uint16_t* i6 = (const uint16_t*) input[6];
45 const uint16_t* i7 = (const uint16_t*) input[7];
46 const uint16_t* i8 = (const uint16_t*) input[8];
47 input = (const void**) ((uintptr_t) input + input_increment);
48 if (kernel_elements < 2) {
49 i1 = (const uint16_t*) zero;
50 }
51 assert(i1 != NULL);
52 if (kernel_elements <= 2) {
53 i2 = (const uint16_t*) zero;
54 }
55 assert(i2 != NULL);
56 if (kernel_elements < 4) {
57 i3 = (const uint16_t*) zero;
58 }
59 assert(i3 != NULL);
60 if (kernel_elements <= 4) {
61 i4 = (const uint16_t*) zero;
62 }
63 assert(i4 != NULL);
64 if (kernel_elements < 6) {
65 i5 = (const uint16_t*) zero;
66 }
67 assert(i5 != NULL);
68 if (kernel_elements <= 6) {
69 i6 = (const uint16_t*) zero;
70 }
71 assert(i6 != NULL);
72 if (kernel_elements < 8) {
73 i7 = (const uint16_t*) zero;
74 }
75 assert(i7 != NULL);
76 if (kernel_elements <= 8) {
77 i8 = (const uint16_t*) zero;
78 }
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i0 != zero) {
81 i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
82 }
83 if XNN_UNPREDICTABLE(i1 != zero) {
84 i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
85 }
86 if XNN_UNPREDICTABLE(i2 != zero) {
87 i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
88 }
89 if XNN_UNPREDICTABLE(i3 != zero) {
90 i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
91 }
92 if XNN_UNPREDICTABLE(i4 != zero) {
93 i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
94 }
95 if XNN_UNPREDICTABLE(i5 != zero) {
96 i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
97 }
98 if XNN_UNPREDICTABLE(i6 != zero) {
99 i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
100 }
101 if XNN_UNPREDICTABLE(i7 != zero) {
102 i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
103 }
104 if XNN_UNPREDICTABLE(i8 != zero) {
105 i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
106 }
107
108 size_t c = channels;
109 while (c >= 8) {
110 const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
111 i0 += 8;
112 const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
113 i1 += 8;
114 const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
115 i2 += 8;
116 const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
117 i3 += 8;
118 const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
119 i4 += 8;
120 const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
121 i5 += 8;
122 const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
123 i6 += 8;
124 const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
125 i7 += 8;
126 const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
127 i8 += 8;
128
129 const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
130 const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
131 const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
132 const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
133 const __m256 vsum018 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vi8), _MM_FROUND_NO_EXC));
134 const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
135 const __m256 vsum01678 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum018, vsum67), _MM_FROUND_NO_EXC));
136 const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum01678), _MM_FROUND_NO_EXC));
137
138 __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vscale), _MM_FROUND_NO_EXC));
139 vout = _mm256_max_ps(vout, vmin);
140 vout = _mm256_min_ps(vout, vmax);
141
142 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vout, _MM_FROUND_NO_EXC));
143 o += 8;
144
145 c -= 8;
146 }
147 if (c != 0) {
148 const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
149 const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
150 const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
151 const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
152 const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
153 const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
154 const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
155 const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
156 const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
157
158 const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
159 const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
160 const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
161 const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
162 const __m256 vsum018 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vi8), _MM_FROUND_NO_EXC));
163 const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
164 const __m256 vsum01678 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum018, vsum67), _MM_FROUND_NO_EXC));
165 const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum01678), _MM_FROUND_NO_EXC));
166
167 __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vscale), _MM_FROUND_NO_EXC));
168 vout = _mm256_max_ps(vout, vmin);
169 vout = _mm256_min_ps(vout, vmax);
170
171 __m128i vh = _mm256_cvtps_ph(vout, _MM_FROUND_NO_EXC);
172 if (c & 4) {
173 _mm_storel_epi64((__m128i*) o, vh);
174 vh = _mm_unpackhi_epi64(vh, vh);
175 o += 4;
176 }
177 if (c & 2) {
178 _mm_storeu_si32(o, vh);
179 vh = _mm_srli_epi64(vh, 32);
180 o += 2;
181 }
182 if (c & 1) {
183 *o = (uint16_t) _mm_extract_epi16(vh, 0);
184 o += 1;
185 }
186 }
187 o = (uint16_t*) ((uintptr_t) o + output_increment);
188 } while (--output_pixels != 0);
189 }
190