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_c8(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_c8(
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 >= 8; channels -= 8) {
64
65 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
66 i0 += 8;
67
68 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
69 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
70 i1 += 8;
71
72 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
73 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
74 i2 += 8;
75
76 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
77 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
78 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
79 i3 += 8;
80
81 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
82 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
83 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
84 i4 += 8;
85
86 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
87 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
88 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
89 i5 += 8;
90
91 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
92 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
93 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
94 i6 += 8;
95
96 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
97 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
98
99 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
100
101 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
102 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
103
104 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
105 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
106
107 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
108 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
109
110 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
111 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
112
113 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
114 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
115
116 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
117 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
118
119 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
120
121
122 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
123
124 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
125
126 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
127 output += 8;
128 }
129 if XNN_UNLIKELY(channels != 0) {
130 {
131
132 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
133 i0 += 8;
134
135 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
136 i1 += 8;
137
138 const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero);
139 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
140 i2 += 8;
141
142 const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero);
143 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
144 i3 += 8;
145
146 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
147 const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero);
148 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
149 i4 += 8;
150
151 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
152 const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero);
153 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
154 i5 += 8;
155
156 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
157 const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero);
158 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
159 i6 += 8;
160
161 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
162 const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero);
163
164 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
165 const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero);
166
167 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
168
169 __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero);
170 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
171
172 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
173 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
174
175 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
176 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
177
178 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
179 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
180
181 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
182 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
183
184 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
185 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
186
187 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
188
189 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
190 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
191
192 if (channels & 4) {
193 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
194 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
195 output += 4;
196 }
197 uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
198 if (channels & 2) {
199 unaligned_store_u16(output, (uint16_t) vout0123);
200 vout0123 >>= 16;
201 output += 2;
202 }
203 if (channels & 1) {
204 *output = (uint8_t) vout0123;
205 }
206 }
207 }
208 }
209