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 #include <xnnpack/unaligned.h>
16
17
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)])18 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24(
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_sse4.init_bias);
58 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
59 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
60 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
61 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
62 for (; channels >= 24; channels -= 24) {
63 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
64 const __m128i vxi0x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
65 const __m128i vxi0xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 16)));
66 i0 += 24;
67 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
68 const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
69 const __m128i vxi1xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16)));
70 i1 += 24;
71
72 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
73 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
74 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
75 const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
76 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN);
77 const __m128i vxi2xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 16)));
78 i2 += 24;
79
80 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
81 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
82 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
83 const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
84 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi2xGHIJKLMN);
85 const __m128i vxi3xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 16)));
86 i3 += 24;
87 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
88 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
89 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
90 const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
91 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi3xGHIJKLMN);
92 const __m128i vxi4xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 16)));
93 i4 += 24;
94 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
95 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
96 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
97 const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
98 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi4xGHIJKLMN);
99 const __m128i vxi5xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 16)));
100 i5 += 24;
101 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
102 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
103 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
104 const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
105 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi5xGHIJKLMN);
106 const __m128i vxi6xGHIJKLMN = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 16)));
107 i6 += 24;
108
109 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
110 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
111 vaccGHIJKLMN = _mm_add_epi16(vaccGHIJKLMN, vxi6xGHIJKLMN);
112
113 const __m128i vzero = _mm_setzero_si128();
114 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
115 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
116 __m128i vacc89AB = _mm_cvtepu16_epi32(vacc89ABCDEF);
117 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
118 __m128i vaccGHIJ = _mm_cvtepu16_epi32(vaccGHIJKLMN);
119 __m128i vaccKLMN = _mm_unpackhi_epi16(vaccGHIJKLMN, vzero);
120
121 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
122 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
123 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
124 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
125 vaccGHIJ = _mm_add_epi32(vaccGHIJ, vinit_bias);
126 vaccKLMN = _mm_add_epi32(vaccKLMN, vinit_bias);
127
128 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
129 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
130 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
131 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
132 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ);
133 __m128 vfpaccKLMN = _mm_cvtepi32_ps(vaccKLMN);
134
135 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
136 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
137 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
138 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
139 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale);
140 vfpaccKLMN = _mm_mul_ps(vfpaccKLMN, vscale);
141
142 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
143 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
144 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
145 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
146 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point);
147 vfpaccKLMN = _mm_min_ps(vfpaccKLMN, voutput_max_less_zero_point);
148
149 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
150 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
151 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
152 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
153 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ);
154 vaccKLMN = _mm_cvtps_epi32(vfpaccKLMN);
155
156 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
157 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
158 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
159
160 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
161 __m128i voutGHIJKLMNGHIJKLMN = _mm_packus_epi16(voutGHIJKLMN, voutGHIJKLMN);
162
163 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
164 voutGHIJKLMNGHIJKLMN = _mm_max_epu8(voutGHIJKLMNGHIJKLMN, voutput_min);
165
166 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
167 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
168 output += 24;
169 }
170 if XNN_UNLIKELY(channels != 0) {
171 do {
172 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
173 i0 += 8;
174 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
175 i1 += 8;
176
177 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
178 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
179 i2 += 8;
180
181 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
182 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
183 i3 += 8;
184 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
185 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
186 i4 += 8;
187 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
188 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
189 i5 += 8;
190 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
191 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
192 i6 += 8;
193
194 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
195
196 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
197 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, _mm_setzero_si128());
198
199 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
200 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
201
202 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
203 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
204
205 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
206 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
207
208 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
209 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
210
211 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
212 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
213
214 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
215
216 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
217 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
218
219 if XNN_LIKELY(channels >= 8) {
220 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
221 output += 8;
222 channels -= 8;
223 } else {
224 if (channels & 4) {
225 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
226 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
227 output += 4;
228 }
229 if (channels & 2) {
230 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
231 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
232 output += 2;
233 }
234 if (channels & 1) {
235 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
236 output += 1;
237 }
238 channels = 0;
239 }
240 } while (channels != 0);
241 }
242 }
243