1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2022 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker
6*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
7*4bdc9457SAndroid Build Coastguard Worker
8*4bdc9457SAndroid Build Coastguard Worker #include <immintrin.h>
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/avgpool.h>
11*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/intrinsics-polyfill.h>
12*4bdc9457SAndroid Build Coastguard Worker
13*4bdc9457SAndroid Build Coastguard Worker
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*4bdc9457SAndroid Build Coastguard Worker void xnn_f16_avgpool_minmax_ukernel_9x__f16c_c8(
15*4bdc9457SAndroid Build Coastguard Worker size_t output_pixels,
16*4bdc9457SAndroid Build Coastguard Worker size_t kernel_elements,
17*4bdc9457SAndroid Build Coastguard Worker size_t channels,
18*4bdc9457SAndroid Build Coastguard Worker const void** input,
19*4bdc9457SAndroid Build Coastguard Worker size_t input_offset,
20*4bdc9457SAndroid Build Coastguard Worker const void* zero,
21*4bdc9457SAndroid Build Coastguard Worker void* output,
22*4bdc9457SAndroid Build Coastguard Worker size_t input_increment,
23*4bdc9457SAndroid Build Coastguard Worker size_t output_increment,
24*4bdc9457SAndroid Build Coastguard Worker const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25*4bdc9457SAndroid Build Coastguard Worker {
26*4bdc9457SAndroid Build Coastguard Worker assert(output_pixels != 0);
27*4bdc9457SAndroid Build Coastguard Worker assert(kernel_elements != 0);
28*4bdc9457SAndroid Build Coastguard Worker assert(kernel_elements <= 9);
29*4bdc9457SAndroid Build Coastguard Worker assert(channels != 0);
30*4bdc9457SAndroid Build Coastguard Worker
31*4bdc9457SAndroid Build Coastguard Worker const __m256 vscale = _mm256_load_ps(params->avx.scale);
32*4bdc9457SAndroid Build Coastguard Worker const __m256 vmin = _mm256_load_ps(params->avx.min);
33*4bdc9457SAndroid Build Coastguard Worker const __m256 vmax = _mm256_load_ps(params->avx.max);
34*4bdc9457SAndroid Build Coastguard Worker
35*4bdc9457SAndroid Build Coastguard Worker uint16_t* o = (uint16_t*) output;
36*4bdc9457SAndroid Build Coastguard Worker do {
37*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i0 = (const uint16_t*) input[0];
38*4bdc9457SAndroid Build Coastguard Worker assert(i0 != NULL);
39*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i1 = (const uint16_t*) input[1];
40*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i2 = (const uint16_t*) input[2];
41*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i3 = (const uint16_t*) input[3];
42*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i4 = (const uint16_t*) input[4];
43*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i5 = (const uint16_t*) input[5];
44*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i6 = (const uint16_t*) input[6];
45*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i7 = (const uint16_t*) input[7];
46*4bdc9457SAndroid Build Coastguard Worker const uint16_t* i8 = (const uint16_t*) input[8];
47*4bdc9457SAndroid Build Coastguard Worker input = (const void**) ((uintptr_t) input + input_increment);
48*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements < 2) {
49*4bdc9457SAndroid Build Coastguard Worker i1 = (const uint16_t*) zero;
50*4bdc9457SAndroid Build Coastguard Worker }
51*4bdc9457SAndroid Build Coastguard Worker assert(i1 != NULL);
52*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements <= 2) {
53*4bdc9457SAndroid Build Coastguard Worker i2 = (const uint16_t*) zero;
54*4bdc9457SAndroid Build Coastguard Worker }
55*4bdc9457SAndroid Build Coastguard Worker assert(i2 != NULL);
56*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements < 4) {
57*4bdc9457SAndroid Build Coastguard Worker i3 = (const uint16_t*) zero;
58*4bdc9457SAndroid Build Coastguard Worker }
59*4bdc9457SAndroid Build Coastguard Worker assert(i3 != NULL);
60*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements <= 4) {
61*4bdc9457SAndroid Build Coastguard Worker i4 = (const uint16_t*) zero;
62*4bdc9457SAndroid Build Coastguard Worker }
63*4bdc9457SAndroid Build Coastguard Worker assert(i4 != NULL);
64*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements < 6) {
65*4bdc9457SAndroid Build Coastguard Worker i5 = (const uint16_t*) zero;
66*4bdc9457SAndroid Build Coastguard Worker }
67*4bdc9457SAndroid Build Coastguard Worker assert(i5 != NULL);
68*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements <= 6) {
69*4bdc9457SAndroid Build Coastguard Worker i6 = (const uint16_t*) zero;
70*4bdc9457SAndroid Build Coastguard Worker }
71*4bdc9457SAndroid Build Coastguard Worker assert(i6 != NULL);
72*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements < 8) {
73*4bdc9457SAndroid Build Coastguard Worker i7 = (const uint16_t*) zero;
74*4bdc9457SAndroid Build Coastguard Worker }
75*4bdc9457SAndroid Build Coastguard Worker assert(i7 != NULL);
76*4bdc9457SAndroid Build Coastguard Worker if (kernel_elements <= 8) {
77*4bdc9457SAndroid Build Coastguard Worker i8 = (const uint16_t*) zero;
78*4bdc9457SAndroid Build Coastguard Worker }
79*4bdc9457SAndroid Build Coastguard Worker assert(i8 != NULL);
80*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i0 != zero) {
81*4bdc9457SAndroid Build Coastguard Worker i0 = (const uint16_t*) ((uintptr_t) i0 + input_offset);
82*4bdc9457SAndroid Build Coastguard Worker }
83*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i1 != zero) {
84*4bdc9457SAndroid Build Coastguard Worker i1 = (const uint16_t*) ((uintptr_t) i1 + input_offset);
85*4bdc9457SAndroid Build Coastguard Worker }
86*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i2 != zero) {
87*4bdc9457SAndroid Build Coastguard Worker i2 = (const uint16_t*) ((uintptr_t) i2 + input_offset);
88*4bdc9457SAndroid Build Coastguard Worker }
89*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i3 != zero) {
90*4bdc9457SAndroid Build Coastguard Worker i3 = (const uint16_t*) ((uintptr_t) i3 + input_offset);
91*4bdc9457SAndroid Build Coastguard Worker }
92*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i4 != zero) {
93*4bdc9457SAndroid Build Coastguard Worker i4 = (const uint16_t*) ((uintptr_t) i4 + input_offset);
94*4bdc9457SAndroid Build Coastguard Worker }
95*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i5 != zero) {
96*4bdc9457SAndroid Build Coastguard Worker i5 = (const uint16_t*) ((uintptr_t) i5 + input_offset);
97*4bdc9457SAndroid Build Coastguard Worker }
98*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i6 != zero) {
99*4bdc9457SAndroid Build Coastguard Worker i6 = (const uint16_t*) ((uintptr_t) i6 + input_offset);
100*4bdc9457SAndroid Build Coastguard Worker }
101*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i7 != zero) {
102*4bdc9457SAndroid Build Coastguard Worker i7 = (const uint16_t*) ((uintptr_t) i7 + input_offset);
103*4bdc9457SAndroid Build Coastguard Worker }
104*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(i8 != zero) {
105*4bdc9457SAndroid Build Coastguard Worker i8 = (const uint16_t*) ((uintptr_t) i8 + input_offset);
106*4bdc9457SAndroid Build Coastguard Worker }
107*4bdc9457SAndroid Build Coastguard Worker
108*4bdc9457SAndroid Build Coastguard Worker size_t c = channels;
109*4bdc9457SAndroid Build Coastguard Worker while (c >= 8) {
110*4bdc9457SAndroid Build Coastguard Worker const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
111*4bdc9457SAndroid Build Coastguard Worker i0 += 8;
112*4bdc9457SAndroid Build Coastguard Worker const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
113*4bdc9457SAndroid Build Coastguard Worker i1 += 8;
114*4bdc9457SAndroid Build Coastguard Worker const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
115*4bdc9457SAndroid Build Coastguard Worker i2 += 8;
116*4bdc9457SAndroid Build Coastguard Worker const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
117*4bdc9457SAndroid Build Coastguard Worker i3 += 8;
118*4bdc9457SAndroid Build Coastguard Worker const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
119*4bdc9457SAndroid Build Coastguard Worker i4 += 8;
120*4bdc9457SAndroid Build Coastguard Worker const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
121*4bdc9457SAndroid Build Coastguard Worker i5 += 8;
122*4bdc9457SAndroid Build Coastguard Worker const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
123*4bdc9457SAndroid Build Coastguard Worker i6 += 8;
124*4bdc9457SAndroid Build Coastguard Worker const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
125*4bdc9457SAndroid Build Coastguard Worker i7 += 8;
126*4bdc9457SAndroid Build Coastguard Worker const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
127*4bdc9457SAndroid Build Coastguard Worker i8 += 8;
128*4bdc9457SAndroid Build Coastguard Worker
129*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
130*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
131*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
132*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
133*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum018 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vi8), _MM_FROUND_NO_EXC));
134*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
135*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum01678 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum018, vsum67), _MM_FROUND_NO_EXC));
136*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum01678), _MM_FROUND_NO_EXC));
137*4bdc9457SAndroid Build Coastguard Worker
138*4bdc9457SAndroid Build Coastguard Worker __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vscale), _MM_FROUND_NO_EXC));
139*4bdc9457SAndroid Build Coastguard Worker vout = _mm256_max_ps(vout, vmin);
140*4bdc9457SAndroid Build Coastguard Worker vout = _mm256_min_ps(vout, vmax);
141*4bdc9457SAndroid Build Coastguard Worker
142*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vout, _MM_FROUND_NO_EXC));
143*4bdc9457SAndroid Build Coastguard Worker o += 8;
144*4bdc9457SAndroid Build Coastguard Worker
145*4bdc9457SAndroid Build Coastguard Worker c -= 8;
146*4bdc9457SAndroid Build Coastguard Worker }
147*4bdc9457SAndroid Build Coastguard Worker if (c != 0) {
148*4bdc9457SAndroid Build Coastguard Worker const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
149*4bdc9457SAndroid Build Coastguard Worker const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
150*4bdc9457SAndroid Build Coastguard Worker const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
151*4bdc9457SAndroid Build Coastguard Worker const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
152*4bdc9457SAndroid Build Coastguard Worker const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i4));
153*4bdc9457SAndroid Build Coastguard Worker const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i5));
154*4bdc9457SAndroid Build Coastguard Worker const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6));
155*4bdc9457SAndroid Build Coastguard Worker const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i7));
156*4bdc9457SAndroid Build Coastguard Worker const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i8));
157*4bdc9457SAndroid Build Coastguard Worker
158*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC));
159*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC));
160*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum45 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi4, vi5), _MM_FROUND_NO_EXC));
161*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC));
162*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum018 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum01, vi8), _MM_FROUND_NO_EXC));
163*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum2345 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum23, vsum45), _MM_FROUND_NO_EXC));
164*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum01678 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum018, vsum67), _MM_FROUND_NO_EXC));
165*4bdc9457SAndroid Build Coastguard Worker const __m256 vsum = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vsum2345, vsum01678), _MM_FROUND_NO_EXC));
166*4bdc9457SAndroid Build Coastguard Worker
167*4bdc9457SAndroid Build Coastguard Worker __m256 vout = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vsum, vscale), _MM_FROUND_NO_EXC));
168*4bdc9457SAndroid Build Coastguard Worker vout = _mm256_max_ps(vout, vmin);
169*4bdc9457SAndroid Build Coastguard Worker vout = _mm256_min_ps(vout, vmax);
170*4bdc9457SAndroid Build Coastguard Worker
171*4bdc9457SAndroid Build Coastguard Worker __m128i vh = _mm256_cvtps_ph(vout, _MM_FROUND_NO_EXC);
172*4bdc9457SAndroid Build Coastguard Worker if (c & 4) {
173*4bdc9457SAndroid Build Coastguard Worker _mm_storel_epi64((__m128i*) o, vh);
174*4bdc9457SAndroid Build Coastguard Worker vh = _mm_unpackhi_epi64(vh, vh);
175*4bdc9457SAndroid Build Coastguard Worker o += 4;
176*4bdc9457SAndroid Build Coastguard Worker }
177*4bdc9457SAndroid Build Coastguard Worker if (c & 2) {
178*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si32(o, vh);
179*4bdc9457SAndroid Build Coastguard Worker vh = _mm_srli_epi64(vh, 32);
180*4bdc9457SAndroid Build Coastguard Worker o += 2;
181*4bdc9457SAndroid Build Coastguard Worker }
182*4bdc9457SAndroid Build Coastguard Worker if (c & 1) {
183*4bdc9457SAndroid Build Coastguard Worker *o = (uint16_t) _mm_extract_epi16(vh, 0);
184*4bdc9457SAndroid Build Coastguard Worker o += 1;
185*4bdc9457SAndroid Build Coastguard Worker }
186*4bdc9457SAndroid Build Coastguard Worker }
187*4bdc9457SAndroid Build Coastguard Worker o = (uint16_t*) ((uintptr_t) o + output_increment);
188*4bdc9457SAndroid Build Coastguard Worker } while (--output_pixels != 0);
189*4bdc9457SAndroid Build Coastguard Worker }
190