1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/unipass-sse4.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_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_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__sse41_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_sse4.init_bias);
57 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
58 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
59 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
60 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
61 for (; channels >= 16; channels -= 16) {
62 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
63 const __m128i vxi0x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
64 i0 += 16;
65 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
66 const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
67 i1 += 16;
68
69 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
70 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
71 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
72 const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
73 i2 += 16;
74
75 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
76 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
77 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
78 const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
79 i3 += 16;
80 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
81 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
82 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
83 const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
84 i4 += 16;
85 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
86 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
87 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
88 const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
89 i5 += 16;
90 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
91 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
92 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
93 const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
94 i6 += 16;
95
96 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
97 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
98
99 const __m128i vzero = _mm_setzero_si128();
100 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
101 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
102 __m128i vacc89AB = _mm_cvtepu16_epi32(vacc89ABCDEF);
103 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
104
105 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
106 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
107 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
108 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
109
110 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
111 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
112 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
113 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
114
115 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
116 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
117 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
118 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
119
120 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
121 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
122 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
123 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
124
125 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
126 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
127 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
128 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
129
130 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
131 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
132
133 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
134
135 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
136
137 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
138 output += 16;
139 }
140 if XNN_UNLIKELY(channels != 0) {
141 do {
142 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
143 i0 += 8;
144 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
145 i1 += 8;
146
147 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
148 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
149 i2 += 8;
150
151 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
152 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
153 i3 += 8;
154 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
155 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
156 i4 += 8;
157 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
158 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
159 i5 += 8;
160 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
161 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
162 i6 += 8;
163
164 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
165
166 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
167 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, _mm_setzero_si128());
168
169 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
170 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
171
172 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
173 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
174
175 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
176 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
177
178 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
179 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
180
181 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
182 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
183
184 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
185
186 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
187 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
188
189 if XNN_LIKELY(channels >= 8) {
190 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
191 output += 8;
192 channels -= 8;
193 } else {
194 if (channels & 4) {
195 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
196 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
197 output += 4;
198 }
199 if (channels & 2) {
200 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
201 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
202 output += 2;
203 }
204 if (channels & 1) {
205 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
206 output += 1;
207 }
208 channels = 0;
209 }
210 } while (channels != 0);
211 }
212 }
213