1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gavgpool/unipass-sse.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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 <emmintrin.h>
13 
14 #include <xnnpack/gavgpool.h>
15 
16 
xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2(size_t rows,size_t channels,const int8_t * input,size_t input_stride,const int8_t * zero,int8_t * output,const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2(
18     size_t rows,
19     size_t channels,
20     const int8_t* input,
21     size_t input_stride,
22     const int8_t* zero,
23     int8_t* output,
24     const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
25 {
26   assert(rows != 0);
27   assert(rows <= 7);
28   assert(channels != 0);
29 
30   const int8_t* i0 = input;
31   const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
32   if XNN_UNPREDICTABLE(rows < 2) {
33     i1 = zero;
34   }
35   const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
36   if XNN_UNPREDICTABLE(rows <= 2) {
37     i2 = zero;
38   }
39   const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
40   if XNN_UNPREDICTABLE(rows < 4) {
41     i3 = zero;
42   }
43   const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
44   if XNN_UNPREDICTABLE(rows <= 4) {
45     i4 = zero;
46   }
47   const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
48   if XNN_UNPREDICTABLE(rows < 6) {
49     i5 = zero;
50   }
51   const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride);
52   if XNN_UNPREDICTABLE(rows <= 6) {
53     i6 = zero;
54   }
55 
56   const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
57   const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
58   const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
59   const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->sse2.shift);
60   while (channels >= 8) {
61     const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
62     i0 += 8;
63     const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
64     i1 += 8;
65     const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
66     i2 += 8;
67     const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
68     i3 += 8;
69     const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
70     i4 += 8;
71     const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
72     i5 += 8;
73     const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
74     i6 += 8;
75 
76     const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234567));
77     const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234567));
78     const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi2x01234567));
79     const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi3x01234567));
80     const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4x01234567));
81     const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi5x01234567));
82     const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi6x01234567));
83 
84     __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
85     __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
86 
87     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
88     vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
89     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
90 
91     // Add up all accumulators to vacc0x01234567
92     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
93 
94     const __m128i vsgnacc0x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567);
95     const __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vacc0x01234567, vsgnacc0x01234567));
96     const __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x01234567, vsgnacc0x01234567));
97 
98     const __m128i vsgnacc0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123);
99     const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567);
100 
101     const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123);
102     const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567);
103 
104     const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1));
105     const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1));
106 
107     const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier);
108     const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier);
109     const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier);
110     const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier);
111 
112     const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift);
113     const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift);
114     const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift);
115     const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift);
116 
117     const __m128i vabsout0213 = _mm_castps_si128(
118         _mm_shuffle_ps(_mm_castsi128_ps(vabsout02), _mm_castsi128_ps(vabsout13), _MM_SHUFFLE(2, 0, 2, 0)));
119     const __m128i vabsout4657 = _mm_castps_si128(
120         _mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0)));
121 
122     const __m128i vabsout0123 = _mm_shuffle_epi32(vabsout0213, _MM_SHUFFLE(3, 1, 2, 0));
123     const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0));
124 
125     const __m128i vout0123 = _mm_sub_epi32(_mm_xor_si128(vabsout0123, vsgnacc0123), vsgnacc0123);
126     const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567);
127 
128     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
129     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vout0123, vout4567), voutput_zero_point);
130 
131     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
132     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
133     vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
134 
135     __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
136 
137     _mm_storel_epi64((__m128i*) output, vout0123456701234567);
138     output += 8;
139 
140     channels -= 8;
141   }
142   if XNN_UNLIKELY(channels != 0) {
143     {
144       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
145       i0 += 8;
146       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
147       i1 += 8;
148       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
149       i2 += 8;
150       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
151       i3 += 8;
152       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
153       i4 += 8;
154       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
155       i5 += 8;
156       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
157       i6 += 8;
158 
159       const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234567));
160       const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234567));
161       const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi2x01234567));
162       const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi3x01234567));
163       const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4x01234567));
164       const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi5x01234567));
165       const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi6x01234567));
166 
167       __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
168       __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
169 
170       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
171       vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
172       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
173 
174       // Add up all accumulators to vacc0x01234567
175       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
176 
177       const __m128i vsgnacc0x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567);
178       const __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vacc0x01234567, vsgnacc0x01234567));
179       const __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x01234567, vsgnacc0x01234567));
180 
181       const __m128i vsgnacc0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123);
182       const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567);
183 
184       const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123);
185       const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567);
186 
187       const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1));
188       const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1));
189 
190       const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier);
191       const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier);
192       const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier);
193       const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier);
194 
195       const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift);
196       const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift);
197       const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift);
198       const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift);
199 
200       const __m128i vabsout0213 = _mm_castps_si128(
201           _mm_shuffle_ps(_mm_castsi128_ps(vabsout02), _mm_castsi128_ps(vabsout13), _MM_SHUFFLE(2, 0, 2, 0)));
202       const __m128i vabsout4657 = _mm_castps_si128(
203           _mm_shuffle_ps(_mm_castsi128_ps(vabsout46), _mm_castsi128_ps(vabsout57), _MM_SHUFFLE(2, 0, 2, 0)));
204 
205       const __m128i vabsout0123 = _mm_shuffle_epi32(vabsout0213, _MM_SHUFFLE(3, 1, 2, 0));
206       const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0));
207 
208       const __m128i vout0123 = _mm_sub_epi32(_mm_xor_si128(vabsout0123, vsgnacc0123), vsgnacc0123);
209       const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567);
210 
211       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
212       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vout0123, vout4567), voutput_zero_point);
213 
214       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
215       const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
216       vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
217 
218       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
219 
220       if (channels & 4) {
221         *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
222         vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
223         output += 4;
224       }
225       if (channels & 2) {
226         *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
227         vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
228         output += 2;
229       }
230       if (channels & 1) {
231         *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
232       }
233     }
234   }
235 }
236