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 <emmintrin.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__sse2_mul16(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_up8x3__sse2_mul16(
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 >= 8; c -= 8) {
54 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
55 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
56
57
58 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
59 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
60 i0 += 8;
61
62 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
63 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
64
65 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
66 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
67
68 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
69 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
70
71 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
72 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
73 i1 += 8;
74
75 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
76 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
77
78 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
79 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
80
81 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
82 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
83
84 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
85 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
86 i2 += 8;
87
88 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
89 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
90
91 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
92 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
93
94 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
95 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
96
97 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t));
98
99 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
100 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
101
102 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
103 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
104 w = (const void*) ((const float*) w + 8);
105 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
106 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
107
108 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
109 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
110 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
111
112 vacc0123 = _mm_cvtps_epi32(vscaled0123);
113 vacc4567 = _mm_cvtps_epi32(vscaled4567);
114
115 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
116 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
117
118 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
119 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
120
121 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
122
123
124 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
125 output += 8;
126 }
127 if XNN_UNLIKELY(c != 0) {
128 {
129 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
130 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
131
132
133 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
134 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
135
136 const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8);
137 const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8);
138
139 const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
140 const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
141
142 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi));
143 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi));
144
145 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
146 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
147
148 const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8);
149 const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8);
150
151 const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
152 const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
153
154 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi));
155 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi));
156
157 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
158 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
159
160 const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8);
161 const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8);
162
163 const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
164 const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
165
166 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi));
167 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi));
168
169
170 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
171 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
172
173 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
174 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t) + 4 * sizeof(float)));
175 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
176 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
177
178 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
179 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
180 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
181
182 vacc0123 = _mm_cvtps_epi32(vscaled0123);
183 vacc4567 = _mm_cvtps_epi32(vscaled4567);
184
185
186 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
187 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
188
189 vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
190
191 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
192
193
194 if (c & 4) {
195 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
196 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
197 output += 4;
198 }
199 if (c & 2) {
200 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
201 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
202 output += 2;
203 }
204 if (c & 1) {
205 *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567);
206 output += 1;
207 }
208 }
209 }
210
211 output = (int8_t*) ((uintptr_t) output + output_increment);
212 } while (--output_width != 0);
213 }
214