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_c24(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_c24(
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 >= 24; channels -= 24) {
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 const __m128i vxi0xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 16)));
65 i0 += 24;
66 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
67 const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
68 const __m128i vxi1xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16)));
69 i1 += 24;
70
71 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
72 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
73 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
74 const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
75 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN);
76 const __m128i vxi2xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 16)));
77 i2 += 24;
78
79 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
80 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
81 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
82 const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
83 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi2xGHIJKLMN);
84 const __m128i vxi3xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 16)));
85 i3 += 24;
86 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
87 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
88 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
89 const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
90 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi3xGHIJKLMN);
91 const __m128i vxi4xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 16)));
92 i4 += 24;
93 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
94 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
95 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
96 const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
97 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi4xGHIJKLMN);
98 const __m128i vxi5xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 16)));
99 i5 += 24;
100 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
101 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
102 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
103 const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
104 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi5xGHIJKLMN);
105 const __m128i vxi6xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 16)));
106 i6 += 24;
107
108 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
109 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
110 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi6xGHIJKLMN);
111
112 const __m128i vzero = _mm_setzero_si128();
113 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
114 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
115 __m128i vacc89AB = _mm_cvtepu16_epi32(vacc89ABCDEF);
116 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
117 __m128i vaccGHIJ = _mm_cvtepu16_epi32(vaccGHIJKLMN);
118 __m128i vaccKLMN = _mm_unpackhi_epi16(vaccGHIJKLMN, vzero);
119
120 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
121 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
122 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
123 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
124 vaccGHIJ = _mm_add_epi32(vaccGHIJ, vinit_bias);
125 vaccKLMN = _mm_add_epi32(vaccKLMN, vinit_bias);
126
127 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
128 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
129 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
130 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
131 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
132 __m128 vfpaccKLMN = _mm_cvtepi32_ps(vaccKLMN);
133
134 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
135 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
136 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
137 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
138 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale);
139 vfpaccKLMN = _mm_mul_ps(vfpaccKLMN, 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 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point);
146 vfpaccKLMN = _mm_min_ps(vfpaccKLMN, voutput_max_less_zero_point);
147
148 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
149 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
150 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
151 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
152 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ);
153 vaccKLMN = _mm_cvtps_epi32(vfpaccKLMN);
154
155 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
156 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
157 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
158
159 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
160 __m128i voutGHIJKLMNGHIJKLMN = _mm_packus_epi16(voutGHIJKLMN, voutGHIJKLMN);
161
162 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
163 voutGHIJKLMNGHIJKLMN = _mm_max_epu8(voutGHIJKLMNGHIJKLMN, voutput_min);
164
165 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
166 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
167 output += 24;
168 }
169 if XNN_UNLIKELY(channels != 0) {
170 do {
171 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
172 i0 += 8;
173 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
174 i1 += 8;
175
176 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
177 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
178 i2 += 8;
179
180 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
181 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
182 i3 += 8;
183 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
184 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
185 i4 += 8;
186 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
187 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
188 i5 += 8;
189 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
190 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
191 i6 += 8;
192
193 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
194
195 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
196 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, _mm_setzero_si128());
197
198 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
199 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
200
201 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
202 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
203
204 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
205 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
206
207 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
208 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
209
210 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
211 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
212
213 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
214
215 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
216 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
217
218 if XNN_LIKELY(channels >= 8) {
219 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
220 output += 8;
221 channels -= 8;
222 } else {
223 if (channels & 4) {
224 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
225 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
226 output += 4;
227 }
228 if (channels & 2) {
229 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
230 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
231 output += 2;
232 }
233 if (channels & 1) {
234 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
235 output += 1;
236 }
237 channels = 0;
238 }
239 } while (channels != 0);
240 }
241 }
242