1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.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/intrinsics-polyfill.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul32(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)])24 void xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul32(
25 size_t channels,
26 size_t output_width,
27 const int8_t** input,
28 const void* weights,
29 int8_t* output,
30 size_t input_stride,
31 size_t output_increment,
32 size_t input_offset,
33 const int8_t* zero,
34 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(channels != 0);
37 assert(output_width != 0);
38
39 do {
40 const int8_t* i0 = input[0];
41 assert(i0 != NULL);
42 if XNN_UNPREDICTABLE(i0 != zero) {
43 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44 }
45 const int8_t* i1 = input[1];
46 assert(i1 != NULL);
47 if XNN_UNPREDICTABLE(i1 != zero) {
48 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
49 }
50 const int8_t* i2 = input[2];
51 assert(i2 != NULL);
52 if XNN_UNPREDICTABLE(i2 != zero) {
53 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
54 }
55 const int8_t* i3 = input[3];
56 assert(i3 != NULL);
57 if XNN_UNPREDICTABLE(i3 != zero) {
58 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
59 }
60 const int8_t* i4 = input[4];
61 assert(i4 != NULL);
62 if XNN_UNPREDICTABLE(i4 != zero) {
63 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
64 }
65 const int8_t* i5 = input[5];
66 assert(i5 != NULL);
67 if XNN_UNPREDICTABLE(i5 != zero) {
68 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
69 }
70 const int8_t* i6 = input[6];
71 assert(i6 != NULL);
72 if XNN_UNPREDICTABLE(i6 != zero) {
73 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
74 }
75 const int8_t* i7 = input[7];
76 assert(i7 != NULL);
77 if XNN_UNPREDICTABLE(i7 != zero) {
78 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
79 }
80 const int8_t* i8 = input[8];
81 assert(i8 != NULL);
82 if XNN_UNPREDICTABLE(i8 != zero) {
83 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
84 }
85 input = (const int8_t**) ((uintptr_t) input + input_stride);
86
87 size_t c = channels;
88 const void* w = weights;
89 for (; c >= 16; c -= 16) {
90 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
91 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
92 __m128i vacc89AB = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 8));
93 __m128i vaccCDEF = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 12));
94
95
96 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
97 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 0 * sizeof(int8_t)))));
98 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 4)));
99 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 4 * sizeof(int8_t)))));
100 const __m128i vi0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 8)));
101 const __m128i vk0x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 8 * sizeof(int8_t)))));
102 const __m128i vi0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0 + 12)));
103 const __m128i vk0xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 12 * sizeof(int8_t)))));
104 i0 += 16;
105
106 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
107 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
108 vacc89AB = _mm_macc_epi32(vi0x89AB, vk0x89AB, vacc89AB);
109 vaccCDEF = _mm_macc_epi32(vi0xCDEF, vk0xCDEF, vaccCDEF);
110
111 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
112 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 16 * sizeof(int8_t)))));
113 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 4)));
114 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 20 * sizeof(int8_t)))));
115 const __m128i vi1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 8)));
116 const __m128i vk1x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 24 * sizeof(int8_t)))));
117 const __m128i vi1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1 + 12)));
118 const __m128i vk1xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 28 * sizeof(int8_t)))));
119 i1 += 16;
120
121 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
122 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
123 vacc89AB = _mm_macc_epi32(vi1x89AB, vk1x89AB, vacc89AB);
124 vaccCDEF = _mm_macc_epi32(vi1xCDEF, vk1xCDEF, vaccCDEF);
125
126 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
127 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 32 * sizeof(int8_t)))));
128 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 4)));
129 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 36 * sizeof(int8_t)))));
130 const __m128i vi2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 8)));
131 const __m128i vk2x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 40 * sizeof(int8_t)))));
132 const __m128i vi2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2 + 12)));
133 const __m128i vk2xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 44 * sizeof(int8_t)))));
134 i2 += 16;
135
136 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
137 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
138 vacc89AB = _mm_macc_epi32(vi2x89AB, vk2x89AB, vacc89AB);
139 vaccCDEF = _mm_macc_epi32(vi2xCDEF, vk2xCDEF, vaccCDEF);
140
141 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
142 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 48 * sizeof(int8_t)))));
143 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 4)));
144 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 52 * sizeof(int8_t)))));
145 const __m128i vi3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 8)));
146 const __m128i vk3x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 56 * sizeof(int8_t)))));
147 const __m128i vi3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3 + 12)));
148 const __m128i vk3xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 60 * sizeof(int8_t)))));
149 i3 += 16;
150
151 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
152 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
153 vacc89AB = _mm_macc_epi32(vi3x89AB, vk3x89AB, vacc89AB);
154 vaccCDEF = _mm_macc_epi32(vi3xCDEF, vk3xCDEF, vaccCDEF);
155
156 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
157 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 64 * sizeof(int8_t)))));
158 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 4)));
159 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 68 * sizeof(int8_t)))));
160 const __m128i vi4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 8)));
161 const __m128i vk4x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 72 * sizeof(int8_t)))));
162 const __m128i vi4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4 + 12)));
163 const __m128i vk4xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 76 * sizeof(int8_t)))));
164 i4 += 16;
165
166 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
167 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
168 vacc89AB = _mm_macc_epi32(vi4x89AB, vk4x89AB, vacc89AB);
169 vaccCDEF = _mm_macc_epi32(vi4xCDEF, vk4xCDEF, vaccCDEF);
170
171 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
172 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 80 * sizeof(int8_t)))));
173 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 4)));
174 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 84 * sizeof(int8_t)))));
175 const __m128i vi5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 8)));
176 const __m128i vk5x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 88 * sizeof(int8_t)))));
177 const __m128i vi5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5 + 12)));
178 const __m128i vk5xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 92 * sizeof(int8_t)))));
179 i5 += 16;
180
181 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
182 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
183 vacc89AB = _mm_macc_epi32(vi5x89AB, vk5x89AB, vacc89AB);
184 vaccCDEF = _mm_macc_epi32(vi5xCDEF, vk5xCDEF, vaccCDEF);
185
186 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
187 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 96 * sizeof(int8_t)))));
188 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 4)));
189 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 100 * sizeof(int8_t)))));
190 const __m128i vi6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 8)));
191 const __m128i vk6x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 104 * sizeof(int8_t)))));
192 const __m128i vi6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6 + 12)));
193 const __m128i vk6xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 108 * sizeof(int8_t)))));
194 i6 += 16;
195
196 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
197 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
198 vacc89AB = _mm_macc_epi32(vi6x89AB, vk6x89AB, vacc89AB);
199 vaccCDEF = _mm_macc_epi32(vi6xCDEF, vk6xCDEF, vaccCDEF);
200
201 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
202 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 112 * sizeof(int8_t)))));
203 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 4)));
204 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 116 * sizeof(int8_t)))));
205 const __m128i vi7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 8)));
206 const __m128i vk7x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 120 * sizeof(int8_t)))));
207 const __m128i vi7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7 + 12)));
208 const __m128i vk7xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 124 * sizeof(int8_t)))));
209 i7 += 16;
210
211 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
212 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
213 vacc89AB = _mm_macc_epi32(vi7x89AB, vk7x89AB, vacc89AB);
214 vaccCDEF = _mm_macc_epi32(vi7xCDEF, vk7xCDEF, vaccCDEF);
215
216 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
217 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 128 * sizeof(int8_t)))));
218 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 4)));
219 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 132 * sizeof(int8_t)))));
220 const __m128i vi8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 8)));
221 const __m128i vk8x89AB = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 136 * sizeof(int8_t)))));
222 const __m128i vi8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8 + 12)));
223 const __m128i vk8xCDEF = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) ((uintptr_t) w + 16 * sizeof(int32_t) + 140 * sizeof(int8_t)))));
224 i8 += 16;
225
226 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
227 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
228 vacc89AB = _mm_macc_epi32(vi8x89AB, vk8x89AB, vacc89AB);
229 vaccCDEF = _mm_macc_epi32(vi8xCDEF, vk8xCDEF, vaccCDEF);
230
231 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t));
232
233 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
234 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
235 __m128 vscaled89AB = _mm_cvtepi32_ps(vacc89AB);
236 __m128 vscaledCDEF = _mm_cvtepi32_ps(vaccCDEF);
237
238 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
239 const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4);
240 const __m128 vscale89AB = _mm_loadu_ps((const float*) w + 8);
241 const __m128 vscaleCDEF = _mm_loadu_ps((const float*) w + 12);
242 w = (const void*) ((const float*) w + 16);
243 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
244 vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567);
245 vscaled89AB = _mm_mul_ps(vscaled89AB, vscale89AB);
246 vscaledCDEF = _mm_mul_ps(vscaledCDEF, vscaleCDEF);
247
248 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
249 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
250 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
251 vscaled89AB = _mm_min_ps(vscaled89AB, voutput_max_less_zero_point);
252 vscaledCDEF = _mm_min_ps(vscaledCDEF, voutput_max_less_zero_point);
253
254 vacc0123 = _mm_cvtps_epi32(vscaled0123);
255 vacc4567 = _mm_cvtps_epi32(vscaled4567);
256 vacc89AB = _mm_cvtps_epi32(vscaled89AB);
257 vaccCDEF = _mm_cvtps_epi32(vscaledCDEF);
258
259 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
260 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
261 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
262
263 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
264 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
265 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
266
267 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
268 output += 16;
269 }
270 if XNN_UNLIKELY(c != 0) {
271 const int8_t* k = (const int8_t*) ((const int32_t*) w + 16);
272 do {
273 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
274
275 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i0)));
276 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) k)));
277 i0 += 4;
278
279 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
280 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i1)));
281 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 16))));
282 i1 += 4;
283
284 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
285 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i2)));
286 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 32))));
287 i2 += 4;
288
289 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
290 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i3)));
291 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 48))));
292 i3 += 4;
293
294 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
295 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i4)));
296 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 64))));
297 i4 += 4;
298
299 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
300 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i5)));
301 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 80))));
302 i5 += 4;
303
304 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
305 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i6)));
306 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 96))));
307 i6 += 4;
308
309 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
310 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i7)));
311 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 112))));
312 i7 += 4;
313
314 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
315 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(i8)));
316 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128(*((const int*) (k + 128))));
317 i8 += 4;
318
319 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
320
321 k += 4;
322
323 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
324 const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 16 * sizeof(int32_t) + 144 * sizeof(int8_t)));
325 vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123);
326 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
327 vacc0123 = _mm_cvtps_epi32(vscaled0123);
328
329 w = (const void*) ((const int32_t*) w + 4);
330
331 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
332 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
333
334 vout0123 = _mm_packs_epi16(vout0123, vout0123);
335 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
336
337 if XNN_LIKELY(c >= 4) {
338 _mm_storeu_si32(output, vout0123);
339 output += 4;
340 c -= 4;
341 } else {
342 if (c & 2) {
343 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123, 0));
344 vout0123 = _mm_srli_epi32(vout0123, 16);
345 output += 2;
346 }
347 if (c & 1) {
348 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
349 output += 1;
350 }
351 c = 0;
352 }
353 } while (c != 0);
354 }
355
356 output = (int8_t*) ((uintptr_t) output + output_increment);
357 } while (--output_width != 0);
358 }
359