1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/unipass-sse2.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 #include <xnnpack/unaligned.h>
16
17
xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c16(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c16(
19 size_t rows,
20 size_t channels,
21 const int8_t* input,
22 size_t input_stride,
23 const int8_t* zero,
24 int8_t* output,
25 const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(rows <= 7);
29 assert(channels != 0);
30
31 const int8_t* i0 = input;
32 const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
33 if XNN_UNPREDICTABLE(rows < 2) {
34 i1 = zero;
35 }
36 const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
37 if XNN_UNPREDICTABLE(rows <= 2) {
38 i2 = zero;
39 }
40 const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
41 if XNN_UNPREDICTABLE(rows < 4) {
42 i3 = zero;
43 }
44 const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
45 if XNN_UNPREDICTABLE(rows <= 4) {
46 i4 = zero;
47 }
48 const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
49 if XNN_UNPREDICTABLE(rows < 6) {
50 i5 = zero;
51 }
52 const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride);
53 if XNN_UNPREDICTABLE(rows <= 6) {
54 i6 = zero;
55 }
56
57 const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias);
58 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
59 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
60 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
61 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
62 for (; channels >= 16; channels -= 16) {
63
64 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
65 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
66 i0 += 16;
67
68 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
69 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
70 const __m128i vxi0x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x89ABCDEF, vi0x89ABCDEF), 8);
71 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
72 i1 += 16;
73
74 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
75 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
76 const __m128i vxi1x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x89ABCDEF, vi1x89ABCDEF), 8);
77 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
78 i2 += 16;
79
80 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
81 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
82 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
83 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
84 const __m128i vxi2x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x89ABCDEF, vi2x89ABCDEF), 8);
85 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
86 i3 += 16;
87
88 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
89 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
90 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
91 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
92 const __m128i vxi3x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x89ABCDEF, vi3x89ABCDEF), 8);
93 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
94 i4 += 16;
95
96 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
97 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
98 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
99 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
100 const __m128i vxi4x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x89ABCDEF, vi4x89ABCDEF), 8);
101 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
102 i5 += 16;
103
104 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
105 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
106 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
107 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
108 const __m128i vxi5x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x89ABCDEF, vi5x89ABCDEF), 8);
109 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
110 i6 += 16;
111
112 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
113 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
114 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
115 const __m128i vxi6x89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x89ABCDEF, vi6x89ABCDEF), 8);
116
117 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
118 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
119
120 const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567);
121 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567);
122 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567);
123 const __m128i vsgnacc89ABCDEF = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc89ABCDEF);
124 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vsgnacc89ABCDEF);
125 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vsgnacc89ABCDEF);
126
127 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
128 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
129 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
130 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
131
132 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
133 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
134 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
135 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
136
137 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
138 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
139 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
140 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
141
142 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
143 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
144 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
145 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
146
147 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
148 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
149 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
150 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
151
152 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
153 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
154
155 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
156 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
157
158 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
159
160
161 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
162 output += 16;
163 }
164 if XNN_UNLIKELY(channels != 0) {
165 do {
166
167 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
168 i0 += 8;
169
170 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
171 i1 += 8;
172
173 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
174 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
175 i2 += 8;
176
177 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
178 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
179 i3 += 8;
180
181 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
182 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
183 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
184 i4 += 8;
185
186 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
187 const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8);
188 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
189 i5 += 8;
190
191 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
192 const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8);
193 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
194 i6 += 8;
195
196 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
197 const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8);
198
199 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
200 const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8);
201
202 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
203
204 const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567);
205 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567);
206 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567);
207
208 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
209 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
210
211 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
212 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
213
214 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
215 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
216
217 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
218 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
219
220 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
221 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
222
223 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
224 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
225
226 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
227
228 if XNN_LIKELY(channels >= 8) {
229 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
230 output += 8;
231 channels -= 8;
232 } else {
233 if (channels & 4) {
234 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
235 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
236 output += 4;
237 }
238 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
239 if (channels & 2) {
240 unaligned_store_u16(output, (uint16_t) vout0123);
241 vout0123 >>= 16;
242 output += 2;
243 }
244 if (channels & 1) {
245 *output = (int8_t) vout0123;
246 output += 1;
247 }
248 channels = 0;
249 }
250 } while (channels != 0);
251 }
252 }
253