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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20 #include <xnnpack/unaligned.h>
21
22
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__xop_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)])23 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__xop_mul16_add16(
24 size_t channels,
25 size_t output_width,
26 const int8_t** input,
27 const void* weights,
28 int8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const int8_t* zero,
33 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
38 do {
39 const int8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const int8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const int8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 input = (const int8_t**) ((uintptr_t) input + input_stride);
55
56 size_t c = channels;
57 const void* w = weights;
58 for (; c >= 16; c -= 16) {
59 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
60 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
61 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
62 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
63
64
65 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
66 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
67 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)));
68 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
69 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
70 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(vi0x89ABCDEF);
71 const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)));
72 const __m128i vxk0x89ABCDEF = _mm_cvtepi8_epi16(vk0x89ABCDEF);
73 i0 += 16;
74
75
76 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
77 __m128i vprod89ABCDEF = _mm_mullo_epi16(vxi0x89ABCDEF, vxk0x89ABCDEF);
78
79
80 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
81 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
82 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)));
83 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
84 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
85 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(vi1x89ABCDEF);
86 const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)));
87 const __m128i vxk1x89ABCDEF = _mm_cvtepi8_epi16(vk1x89ABCDEF);
88 i1 += 16;
89
90
91 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
92 vprod89ABCDEF = _mm_macc_epi16(vxi1x89ABCDEF, vxk1x89ABCDEF, vprod89ABCDEF);
93
94 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
95 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
96 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
97 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
98
99 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
100 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
101 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)));
102 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
103 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
104 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(vi2x89ABCDEF);
105 const __m128i vk2x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)));
106 const __m128i vxk2x89ABCDEF = _mm_cvtepi8_epi16(vk2x89ABCDEF);
107 i2 += 16;
108
109
110 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
111 vprod89ABCDEF = _mm_mullo_epi16(vxi2x89ABCDEF, vxk2x89ABCDEF);
112
113 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
114 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
115 vacc89AB = _mm_add_epi32(vacc89AB, _mm_cvtepi16_epi32(vprod89ABCDEF));
116 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_srai_epi32(_mm_unpackhi_epi16(vprod89ABCDEF, vprod89ABCDEF), 16));
117
118 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t));
119
120 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
121 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
122 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
123 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
124
125 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
126 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
127 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
128 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
129 w = (const void*) ((const float*) w + 16);
130 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
131 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
132 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
133 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
134
135 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
136 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
137 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
138 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
139 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
140
141 vacc0123 = _mm_cvtps_epi32(vscaled0123);
142 vacc4567 = _mm_cvtps_epi32(vscaled4567);
143 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
144 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
145
146 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
147 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
148 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
149
150
151 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
152
153 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
154 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
155
156 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
157 output += 16;
158 }
159 if XNN_UNLIKELY(c != 0) {
160 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
161 do {
162 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
163 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
164
165
166 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
167 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
168 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) k);
169 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
170 i0 += 8;
171
172
173 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
174
175
176 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
177 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
178 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) (k + 16));
179 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
180 i1 += 8;
181
182
183 vprod01234567 = _mm_macc_epi16(vxi1x01234567, vxk1x01234567, vprod01234567);
184
185 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
186 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
187
188 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
189 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
190 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) (k + 32));
191 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
192 i2 += 8;
193
194
195 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
196
197 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
198 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
199
200 k += 8;
201
202 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
203 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
204
205 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)));
206 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t) + 4 * sizeof(float)));
207 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
208 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
209
210 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
211 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
212 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
213
214 vacc0123 = _mm_cvtps_epi32(vscaled0123);
215 vacc4567 = _mm_cvtps_epi32(vscaled4567);
216
217 w = (const void*) ((const int32_t*) w + 8);
218
219 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
220 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
221
222
223 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
224
225 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
226
227 if XNN_LIKELY(c >= 8) {
228 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
229 output += 8;
230 c -= 8;
231 } else {
232 if (c & 4) {
233 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
234 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
235 output += 4;
236 }
237 if (c & 2) {
238 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
239 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
240 output += 2;
241 }
242 if (c & 1) {
243 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
244 output += 1;
245 }
246 c = 0;
247 }
248 } while (c != 0);
249 }
250
251 output = (int8_t*) ((uintptr_t) output + output_increment);
252 } while (--output_width != 0);
253 }
254