xref: /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/up32x9-minmax-fp32-avx512skx-mul32.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-dwconv/unipass-avx512skx-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 #include <immintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 
17 
xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_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)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32(
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_DISABLE_MSAN
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->fp32_avx512.output_max_less_zero_point);
34   const __m512i voutput_zero_point = _mm512_load_si512(params->fp32_avx512.output_zero_point);
35   const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->fp32_avx512.output_min);
36   const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 5, 1, 6, 2, 4, 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     const int8_t* i3 = input[3];
55     assert(i3 != NULL);
56     if XNN_UNPREDICTABLE(i3 != zero) {
57       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58     }
59     const int8_t* i4 = input[4];
60     assert(i4 != NULL);
61     if XNN_UNPREDICTABLE(i4 != zero) {
62       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63     }
64     const int8_t* i5 = input[5];
65     assert(i5 != NULL);
66     if XNN_UNPREDICTABLE(i5 != zero) {
67       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68     }
69     const int8_t* i6 = input[6];
70     assert(i6 != NULL);
71     if XNN_UNPREDICTABLE(i6 != zero) {
72       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73     }
74     const int8_t* i7 = input[7];
75     assert(i7 != NULL);
76     if XNN_UNPREDICTABLE(i7 != zero) {
77       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78     }
79     const int8_t* i8 = input[8];
80     assert(i8 != NULL);
81     if XNN_UNPREDICTABLE(i8 != zero) {
82       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83     }
84     input = (const int8_t**) ((uintptr_t) input + input_stride);
85 
86     size_t c = channels;
87     const void* w = weights;
88     for (; c >= 32; c -= 32) {
89       __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
90       __m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((uintptr_t) w + 16 * sizeof(int32_t)));
91 
92 
93       const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
94       const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
95       const __m512i vi0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i0 + 16)));
96       const __m512i vk0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
97       i0 += 32;
98 
99       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
100       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV));
101 
102       const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
103       const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
104       const __m512i vi1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i1 + 16)));
105       const __m512i vk1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
106       i1 += 32;
107 
108       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
109       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV));
110 
111       const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
112       const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
113       const __m512i vi2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i2 + 16)));
114       const __m512i vk2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
115       i2 += 32;
116 
117       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
118       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV));
119 
120       const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
121       const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
122       const __m512i vi3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i3 + 16)));
123       const __m512i vk3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
124       i3 += 32;
125 
126       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
127       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV));
128 
129       const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
130       const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
131       const __m512i vi4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i4 + 16)));
132       const __m512i vk4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
133       i4 += 32;
134 
135       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
136       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV));
137 
138       const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
139       const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
140       const __m512i vi5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i5 + 16)));
141       const __m512i vk5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
142       i5 += 32;
143 
144       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
145       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV));
146 
147       const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
148       const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
149       const __m512i vi6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i6 + 16)));
150       const __m512i vk6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
151       i6 += 32;
152 
153       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
154       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV));
155 
156       const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
157       const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
158       const __m512i vi7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i7 + 16)));
159       const __m512i vk7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
160       i7 += 32;
161 
162       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
163       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV));
164 
165       const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
166       const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
167       const __m512i vi8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i8 + 16)));
168       const __m512i vk8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
169       i8 += 32;
170 
171       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
172       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV));
173 
174       w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t));
175 
176       __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
177       __m512 vscaledGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vaccGHIJKLMNOPQRSTUV);
178 
179       const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps(w);
180       const __m512 vscaleGHIJKLMNOPQRSTUV = _mm512_loadu_ps((const void*) ((uintptr_t) w + 16 * sizeof(float)));
181       w = (const void*) ((uintptr_t) w + 32 * sizeof(float));
182       vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale0123456789ABCDEF);
183       vscaledGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaledGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
184 
185       vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
186       vscaledGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaledGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
187 
188       vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
189       vaccGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaledGHIJKLMNOPQRSTUV);
190 
191       __m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point);
192       __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vaccGHIJKLMNOPQRSTUV), _mm512_extracti32x8_epi32(vaccGHIJKLMNOPQRSTUV, 1)), _mm512_castsi512_si256(voutput_zero_point));
193 
194       const __m256i vout0123GHIJ4567KLMN = _mm512_castsi512_si256(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV);
195       const __m256i vout89ABOPQRCDEFSTUV = _mm512_extracti32x8_epi32(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV, 1);
196       const __m256i vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV = _mm256_packs_epi16(vout0123GHIJ4567KLMN, vout89ABOPQRCDEFSTUV);
197       __m256i vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_permutevar8x32_epi32(vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV, vpermute_mask);
198       const __m128i voutGHIJOPQR = _mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV);
199       const __m128i voutKLMNSTUV = _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1);
200       __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(voutGHIJOPQR, voutKLMNSTUV), _MM_SHUFFLE(3, 1, 2, 0));
201 
202       vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_max_epi8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_min);
203       voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, _mm256_castsi256_si128(voutput_min));
204 
205       _mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV);
206       _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
207       output += 32;
208     }
209     if XNN_UNLIKELY(c != 0) {
210       // Prepare mask for valid 8-bit elements (depends on nc).
211       const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
212       const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
213       do {
214         __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
215 
216 
217         const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
218         const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) k));
219         i0 += 16;
220 
221         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
222 
223         const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
224         const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 32)));
225         i1 += 16;
226 
227         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
228 
229         const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
230         const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 64)));
231         i2 += 16;
232 
233         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
234 
235         const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
236         const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 96)));
237         i3 += 16;
238 
239         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
240 
241         const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
242         const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 128)));
243         i4 += 16;
244 
245         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
246 
247         const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
248         const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 160)));
249         i5 += 16;
250 
251         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
252 
253         const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
254         const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 192)));
255         i6 += 16;
256 
257         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
258 
259         const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
260         const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 224)));
261         i7 += 16;
262 
263         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
264 
265         const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
266         const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 256)));
267         i8 += 16;
268 
269         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
270 
271         k += 16;
272 
273         __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF);
274         const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps((const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t)));
275         vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale0123456789ABCDEF);
276         vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point);
277         vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF);
278 
279         w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
280 
281         __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point));
282 
283         const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
284         const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
285         __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
286         vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_min));
287 
288         if XNN_LIKELY(c >= 16) {
289           _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
290           output += 16;
291           c -= 16;
292         } else {
293           _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
294           output = (int8_t*) ((uintptr_t) output + c);
295           c = 0;
296         }
297       } while (c != 0);
298     }
299 
300     output = (int8_t*) ((uintptr_t) output + output_increment);
301   } while (--output_width != 0);
302 }
303