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_up8x3__sse41_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__sse41_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 vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
60 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
61 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
62 i0 += 8;
63
64
65 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
66
67 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
68 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
69
70 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
71 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
72 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
73 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
74 i1 += 8;
75
76
77 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
78
79 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
80 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
81
82 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
83 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
84 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
85 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
86 i2 += 8;
87
88
89 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
90
91 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
92 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
93
94 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t));
95
96 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
97 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
98
99 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
100 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
101 w = (const void*) ((const float*) w + 8);
102 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
103 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
104
105 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
106 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
107 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
108
109 vacc0123 = _mm_cvtps_epi32(vscaled0123);
110 vacc4567 = _mm_cvtps_epi32(vscaled4567);
111
112 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
113 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
114
115
116 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
117
118 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
119 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
120
121 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
122 output += 8;
123 }
124 if XNN_UNLIKELY(c != 0) {
125 {
126 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
127 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
128
129
130 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
131 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
132 const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
133 const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
134
135
136 __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
137
138 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
139 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
140
141 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
142 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
143 const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
144 const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
145
146
147 vprod01234567 = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
148
149 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
151
152 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
153 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
154 const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
155 const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
156
157
158 vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
159
160 vacc0123 = _mm_add_epi32(vacc0123, _mm_cvtepi16_epi32(vprod01234567));
161 vacc4567 = _mm_add_epi32(vacc4567, _mm_srai_epi32(_mm_unpackhi_epi16(vprod01234567, vprod01234567), 16));
162
163
164 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
165 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
166
167 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
168 const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t) + 4 * sizeof(float)));
169 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
170 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
171
172 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
173 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
174 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
175
176 vacc0123 = _mm_cvtps_epi32(vscaled0123);
177 vacc4567 = _mm_cvtps_epi32(vscaled4567);
178
179
180 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
181 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
182
183
184 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
185
186 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
187
188 if (c & 4) {
189 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
190 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
191 output += 4;
192 }
193 if (c & 2) {
194 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
195 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
196 output += 2;
197 }
198 if (c & 1) {
199 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
200 output += 1;
201 }
202 }
203 }
204
205 output = (int8_t*) ((uintptr_t) output + output_increment);
206 } while (--output_width != 0);
207 }
208