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
16
xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c16(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c16(
18 size_t rows,
19 size_t channels,
20 const uint8_t* input,
21 size_t input_stride,
22 const uint8_t* zero,
23 uint8_t* output,
24 const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(rows != 0);
27 assert(rows <= 7);
28 assert(channels != 0);
29
30 const uint8_t* i0 = input;
31 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
32 if XNN_UNPREDICTABLE(rows < 2) {
33 i1 = zero;
34 }
35 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
36 if XNN_UNPREDICTABLE(rows <= 2) {
37 i2 = zero;
38 }
39 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
40 if XNN_UNPREDICTABLE(rows < 4) {
41 i3 = zero;
42 }
43 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
44 if XNN_UNPREDICTABLE(rows <= 4) {
45 i4 = zero;
46 }
47 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
48 if XNN_UNPREDICTABLE(rows < 6) {
49 i5 = zero;
50 }
51 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
52 if XNN_UNPREDICTABLE(rows <= 6) {
53 i6 = zero;
54 }
55
56 const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias);
57 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
58 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
59 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
60 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
61 const __m128i vzero = _mm_setzero_si128();
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_unpacklo_epi8(vi0x01234567, vzero);
69 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
70 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
71 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
72 i1 += 16;
73
74 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
75 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
76 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
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_unpacklo_epi8(vi2x01234567, vzero);
82 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
83 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
84 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
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_unpacklo_epi8(vi3x01234567, vzero);
90 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
91 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
92 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
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_unpacklo_epi8(vi4x01234567, vzero);
98 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
99 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
100 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
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_unpacklo_epi8(vi5x01234567, vzero);
106 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
107 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
108 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
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_unpacklo_epi8(vi6x01234567, vzero);
114 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
115 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
116
117 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
118 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
119
120 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
121 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
122 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
123 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
124
125 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
126 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
127 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
128 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
129
130 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
131 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
132 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
133 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
134
135 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
136 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
137 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
138 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
139
140 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
141 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
142 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
143 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
144
145 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
146 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
147 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
148 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
149
150 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
151 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
152
153
154 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
155
156 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
157
158 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
159 output += 16;
160 }
161 if XNN_UNLIKELY(channels != 0) {
162 do {
163
164 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
165 i0 += 8;
166
167 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
168 i1 += 8;
169
170 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
171 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
172 i2 += 8;
173
174 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
175 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
176 i3 += 8;
177
178 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
179 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
180 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
181 i4 += 8;
182
183 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
184 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
185 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
186 i5 += 8;
187
188 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
189 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
190 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
191 i6 += 8;
192
193 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
194 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
195
196 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
197 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
198
199 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
200
201 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
202 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
203
204 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
205 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
206
207 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
208 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
209
210 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
211 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
212
213 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
214 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
215
216 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
217 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
218
219 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
220
221 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
222 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
223
224 if XNN_LIKELY(channels >= 8) {
225 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
226 output += 8;
227 channels -= 8;
228 } else {
229 if (channels & 4) {
230 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
231 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
232 output += 4;
233 }
234 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
235 if (channels & 2) {
236 *((uint16_t*) output) = (uint16_t) vout0123;
237 vout0123 >>= 16;
238 output += 2;
239 }
240 if (channels & 1) {
241 *output = (uint8_t) vout0123;
242 output += 1;
243 }
244 channels = 0;
245 }
246 } while (channels != 0);
247 }
248 }
249