1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul16.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/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx_mul16_add16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx_mul16_add16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 do {
34 const int8_t* i0 = input[0];
35 assert(i0 != NULL);
36 if XNN_UNPREDICTABLE(i0 != zero) {
37 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
38 }
39 const int8_t* i1 = input[1];
40 assert(i1 != NULL);
41 if XNN_UNPREDICTABLE(i1 != zero) {
42 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
43 }
44 const int8_t* i2 = input[2];
45 assert(i2 != NULL);
46 if XNN_UNPREDICTABLE(i2 != zero) {
47 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
48 }
49 input = (const int8_t**) ((uintptr_t) input + input_stride);
50
51 size_t c = channels;
52 const void* w = weights;
53 for (; c >= 16; c -= 16) {
54 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
55 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
56 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
57 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
58
59
60 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
61 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
62 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
63 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
64 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
65 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
66 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
67 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
68 i0 += 16;
69
70
71 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
72 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
73
74
75 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
76 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
77 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
78 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
79 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
80 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
81 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
82 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
83 i1 += 16;
84
85
86 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
87 vprod89ABCDEF = _mm_add_epi16(vprod89ABCDEF, _mm_mullo_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF));
88
89 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
90 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
91 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
92 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
93
94 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
95 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
96 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
97 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
98 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
99 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
100 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
101 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
102 i2 += 16;
103
104
105 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
106 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
107
108 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
109 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
110 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
111 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
112
113 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t));
114
115 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
116 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
117 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
118 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
119
120 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
121 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
122 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
123 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
124 w = (const void*) ((const float*) w + 16);
125 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
126 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
127 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
128 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
129
130 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
131 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
132 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
133 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
134 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
135
136 vacc0123 = _mm_cvtps_epi32(vscaled0123);
137 vacc4567 = _mm_cvtps_epi32(vscaled4567);
138 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
139 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
140
141 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
142 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
143 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
144
145
146 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
147
148 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
149 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
150
151 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
152 output += 16;
153 }
154 if XNN_UNLIKELY(c != 0) {
155 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
156 do {
157 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
158 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
159
160
161 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
162 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
163 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
164 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
165 i0 += 8;
166
167
168 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
169
170
171 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
172 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
173 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
174 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
175 i1 += 8;
176
177
178 vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567));
179
180 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
181 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
182
183 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
184 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
185 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
186 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
187 i2 += 8;
188
189
190 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
191
192 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
193 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
194
195 k += 8;
196
197 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
198 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
199
200 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
201 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t) + 4 * sizeof(float)));
202 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
203 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
204
205 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
206 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
207 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
208
209 vacc0123 = _mm_cvtps_epi32(vscaled0123);
210 vacc4567 = _mm_cvtps_epi32(vscaled4567);
211
212 w = (const void*) ((const int32_t*) w + 8);
213
214 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
215 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
216
217
218 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
219
220 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
221
222 if XNN_LIKELY(c >= 8) {
223 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
224 output += 8;
225 c -= 8;
226 } else {
227 if (c & 4) {
228 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
229 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
230 output += 4;
231 }
232 if (c & 2) {
233 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
234 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
235 output += 2;
236 }
237 if (c & 1) {
238 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
239 output += 1;
240 }
241 c = 0;
242 }
243 } while (c != 0);
244 }
245
246 output = (int8_t*) ((uintptr_t) output + output_increment);
247 } while (--output_width != 0);
248 }
249