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_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)])18 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c16(
19 size_t rows,
20 size_t channels,
21 const uint8_t* input,
22 size_t input_stride,
23 const uint8_t* zero,
24 uint8_t* output,
25 const union xnn_qu8_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 uint8_t* i0 = input;
32 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
33 if XNN_UNPREDICTABLE(rows < 2) {
34 i1 = zero;
35 }
36 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
37 if XNN_UNPREDICTABLE(rows <= 2) {
38 i2 = zero;
39 }
40 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
41 if XNN_UNPREDICTABLE(rows < 4) {
42 i3 = zero;
43 }
44 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
45 if XNN_UNPREDICTABLE(rows <= 4) {
46 i4 = zero;
47 }
48 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
49 if XNN_UNPREDICTABLE(rows < 6) {
50 i5 = zero;
51 }
52 const uint8_t* i6 = (const uint8_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 const __m128i vzero = _mm_setzero_si128();
63 for (; channels >= 16; channels -= 16) {
64
65 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
66 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
67 i0 += 16;
68
69 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
70 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
71 const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, vzero);
72 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
73 i1 += 16;
74
75 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
76 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
77 const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, vzero);
78 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
79 i2 += 16;
80
81 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
82 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
83 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
84 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
85 const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, vzero);
86 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
87 i3 += 16;
88
89 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
90 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
91 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
92 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
93 const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, vzero);
94 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8));
95 i4 += 16;
96
97 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
98 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
99 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
100 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
101 const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, vzero);
102 const __m128i vi5x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i5 + 8));
103 i5 += 16;
104
105 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
106 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
107 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
108 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
109 const __m128i vxi5x89ABCDEF = _mm_unpacklo_epi8(vi5x89ABCDEF, vzero);
110 const __m128i vi6x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i6 + 8));
111 i6 += 16;
112
113 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
114 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
115 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
116 const __m128i vxi6x89ABCDEF = _mm_unpacklo_epi8(vi6x89ABCDEF, vzero);
117
118 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
119 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
120
121 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
122 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
123 __m128i vacc89AB = _mm_unpacklo_epi16(vacc89ABCDEF, vzero);
124 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
125
126 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
127 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
128 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
129 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
130
131 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
132 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
133 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
134 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
135
136 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
137 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
138 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
139 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
140
141 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
142 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
143 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
144 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
145
146 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
147 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
148 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
149 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
150
151 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
152 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
153
154
155 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
156
157 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
158
159 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
160 output += 16;
161 }
162 if XNN_UNLIKELY(channels != 0) {
163 do {
164
165 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
166 i0 += 8;
167
168 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
169 i1 += 8;
170
171 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
172 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
173 i2 += 8;
174
175 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
176 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
177 i3 += 8;
178
179 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
180 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
181 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
182 i4 += 8;
183
184 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
185 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
186 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
187 i5 += 8;
188
189 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
190 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
191 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
192 i6 += 8;
193
194 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
195 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
196
197 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
198 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
199
200 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
201
202 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
203 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
204
205 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
206 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
207
208 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
209 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
210
211 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
212 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
213
214 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
215 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
216
217 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
218 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
219
220 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
221
222 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
223 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
224
225 if XNN_LIKELY(channels >= 8) {
226 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
227 output += 8;
228 channels -= 8;
229 } else {
230 if (channels & 4) {
231 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
232 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
233 output += 4;
234 }
235 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
236 if (channels & 2) {
237 unaligned_store_u16(output, (uint16_t) vout0123);
238 vout0123 >>= 16;
239 output += 2;
240 }
241 if (channels & 1) {
242 *output = (uint8_t) vout0123;
243 output += 1;
244 }
245 channels = 0;
246 }
247 } while (channels != 0);
248 }
249 }
250