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 <smmintrin.h>
13
14 #include <xnnpack/gavgpool.h>
15
16
xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_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__sse41_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 vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0));
62 i0 += 8;
63 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1));
64 i1 += 8;
65 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2));
66 i2 += 8;
67 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3));
68 i3 += 8;
69 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4));
70 i4 += 8;
71 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5));
72 i5 += 8;
73 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i6));
74 i6 += 8;
75
76
77 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
78 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
79
80 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
81 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
82 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
83
84 // Add up all accumulators to vacc0x01234567
85 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
86
87 const __m128i vacc0123 = _mm_add_epi32(vbias, _mm_cvtepi16_epi32(vacc0x01234567));
88 const __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x01234567, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567)));
89
90 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123);
91 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567);
92
93 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1));
94 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1));
95
96 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier);
97 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier);
98 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier);
99 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier);
100
101 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift);
102 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift);
103 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift);
104 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift);
105
106 const __m128i vabsout0123 = _mm_blend_epi16(vabsout02, _mm_shuffle_epi32(vabsout13, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
107 const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
108
109 const __m128i vout0123 = _mm_sign_epi32(vabsout0123, vacc0123);
110 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567);
111
112 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
113 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vout0123, vout4567), voutput_zero_point);
114
115 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
116 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
117 vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
118
119 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
120
121 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
122 output += 8;
123
124 channels -= 8;
125 }
126 if XNN_UNLIKELY(channels != 0) {
127 {
128 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0));
129 i0 += 8;
130 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1));
131 i1 += 8;
132 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2));
133 i2 += 8;
134 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3));
135 i3 += 8;
136 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4));
137 i4 += 8;
138 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5));
139 i5 += 8;
140 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i6));
141 i6 += 8;
142
143
144 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
145 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
146
147 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
148 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
149 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
150
151 // Add up all accumulators to vacc0x01234567
152 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
153
154 const __m128i vacc0123 = _mm_add_epi32(vbias, _mm_cvtepi16_epi32(vacc0x01234567));
155 const __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x01234567, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567)));
156
157 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123);
158 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567);
159
160 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1));
161 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1));
162
163 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier);
164 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier);
165 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier);
166 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier);
167
168 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift);
169 const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift);
170 const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift);
171 const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift);
172
173 const __m128i vabsout0123 = _mm_blend_epi16(vabsout02, _mm_shuffle_epi32(vabsout13, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
174 const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
175
176 const __m128i vout0123 = _mm_sign_epi32(vabsout0123, vacc0123);
177 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567);
178
179 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
180 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vout0123, vout4567), voutput_zero_point);
181
182 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
183 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
184 vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
185
186 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
187
188 if (channels & 4) {
189 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
190 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
191 output += 4;
192 }
193 if (channels & 2) {
194 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
195 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
196 output += 2;
197 }
198 if (channels & 1) {
199 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
200 }
201 }
202 }
203 }
204