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_qs8_dwconv_minmax_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_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_dwconv_minmax_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_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const __mmask16 vblend_mask = _cvtu32_mask16(0xAAAA);
34   const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.multiplier));
35   const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.rounding));
36   const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
37   const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
38   const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
39   const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
40   const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_min));
41   const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_max));
42   const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 5, 1, 6, 2, 4, 0);
43 
44   do {
45     const int8_t* i0 = input[0];
46     assert(i0 != NULL);
47     if XNN_UNPREDICTABLE(i0 != zero) {
48       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
49     }
50     const int8_t* i1 = input[1];
51     assert(i1 != NULL);
52     if XNN_UNPREDICTABLE(i1 != zero) {
53       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
54     }
55     const int8_t* i2 = input[2];
56     assert(i2 != NULL);
57     if XNN_UNPREDICTABLE(i2 != zero) {
58       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
59     }
60     const int8_t* i3 = input[3];
61     assert(i3 != NULL);
62     if XNN_UNPREDICTABLE(i3 != zero) {
63       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
64     }
65     const int8_t* i4 = input[4];
66     assert(i4 != NULL);
67     if XNN_UNPREDICTABLE(i4 != zero) {
68       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
69     }
70     const int8_t* i5 = input[5];
71     assert(i5 != NULL);
72     if XNN_UNPREDICTABLE(i5 != zero) {
73       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
74     }
75     const int8_t* i6 = input[6];
76     assert(i6 != NULL);
77     if XNN_UNPREDICTABLE(i6 != zero) {
78       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
79     }
80     const int8_t* i7 = input[7];
81     assert(i7 != NULL);
82     if XNN_UNPREDICTABLE(i7 != zero) {
83       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
84     }
85     const int8_t* i8 = input[8];
86     assert(i8 != NULL);
87     if XNN_UNPREDICTABLE(i8 != zero) {
88       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
89     }
90     input = (const int8_t**) ((uintptr_t) input + input_stride);
91 
92     size_t c = channels;
93     const void* w = weights;
94     for (; c >= 32; c -= 32) {
95       __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
96       __m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((uintptr_t) w + 16 * sizeof(int32_t)));
97 
98 
99       const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
100       const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t))));
101       const __m512i vi0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i0 + 16)));
102       const __m512i vk0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t))));
103       i0 += 32;
104 
105       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
106       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV));
107 
108       const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
109       const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t))));
110       const __m512i vi1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i1 + 16)));
111       const __m512i vk1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t))));
112       i1 += 32;
113 
114       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
115       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV));
116 
117       const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
118       const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t))));
119       const __m512i vi2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i2 + 16)));
120       const __m512i vk2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t))));
121       i2 += 32;
122 
123       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
124       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV));
125 
126       const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
127       const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t))));
128       const __m512i vi3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i3 + 16)));
129       const __m512i vk3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t))));
130       i3 += 32;
131 
132       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
133       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV));
134 
135       const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
136       const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t))));
137       const __m512i vi4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i4 + 16)));
138       const __m512i vk4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t))));
139       i4 += 32;
140 
141       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
142       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV));
143 
144       const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
145       const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t))));
146       const __m512i vi5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i5 + 16)));
147       const __m512i vk5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t))));
148       i5 += 32;
149 
150       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
151       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV));
152 
153       const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
154       const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t))));
155       const __m512i vi6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i6 + 16)));
156       const __m512i vk6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t))));
157       i6 += 32;
158 
159       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
160       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV));
161 
162       const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
163       const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t))));
164       const __m512i vi7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i7 + 16)));
165       const __m512i vk7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t))));
166       i7 += 32;
167 
168       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
169       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV));
170 
171       const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
172       const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t))));
173       const __m512i vi8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i8 + 16)));
174       const __m512i vk8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t))));
175       i8 += 32;
176 
177       vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
178       vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV));
179 
180       w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t));
181 
182       const __m512i vacc13579BDF = _mm512_shuffle_epi32(vacc0123456789ABCDEF, _MM_SHUFFLE(3, 3, 1, 1));
183       const __m512i vaccHJLNPRTV = _mm512_shuffle_epi32(vaccGHIJKLMNOPQRSTUV, _MM_SHUFFLE(3, 3, 1, 1));
184 
185       const __m512i vprod02468ACE = _mm512_add_epi64(_mm512_mul_epi32(vacc0123456789ABCDEF, vmultiplier), vrounding);
186       const __m512i vprod13579BDF = _mm512_add_epi64(_mm512_mul_epi32(vacc13579BDF, vmultiplier), vrounding);
187       const __m512i vprodGIKMOQSU = _mm512_add_epi64(_mm512_mul_epi32(vaccGHIJKLMNOPQRSTUV, vmultiplier), vrounding);
188       const __m512i vprodHJLNPRTV = _mm512_add_epi64(_mm512_mul_epi32(vaccHJLNPRTV, vmultiplier), vrounding);
189 
190       const __m512i vq31prod02468ACE = _mm512_srli_epi64(vprod02468ACE, 31);
191       const __m512i vq31prod13579BDF = _mm512_add_epi64(vprod13579BDF, vprod13579BDF);
192       const __m512i vq31prodGIKMOQSU = _mm512_srli_epi64(vprodGIKMOQSU, 31);
193       const __m512i vq31prodHJLNPRTV = _mm512_add_epi64(vprodHJLNPRTV, vprodHJLNPRTV);
194 
195       const __m512i vq31prod0123456789ABCDEF = _mm512_mask_blend_epi32(vblend_mask, vq31prod02468ACE, vq31prod13579BDF);
196       const __m512i vq31prodGHIJKLMNOPQRSTUV = _mm512_mask_blend_epi32(vblend_mask, vq31prodGIKMOQSU, vq31prodHJLNPRTV);
197 
198       const __m512i vrem0123456789ABCDEF =
199         _mm512_add_epi32(_mm512_and_epi32(vq31prod0123456789ABCDEF, vremainder_mask), _mm512_srai_epi32(vq31prod0123456789ABCDEF, 31));
200       const __m512i vremGHIJKLMNOPQRSTUV =
201         _mm512_add_epi32(_mm512_and_epi32(vq31prodGHIJKLMNOPQRSTUV, vremainder_mask), _mm512_srai_epi32(vq31prodGHIJKLMNOPQRSTUV, 31));
202 
203       vacc0123456789ABCDEF = _mm512_sra_epi32(vq31prod0123456789ABCDEF, vshift);
204       vaccGHIJKLMNOPQRSTUV = _mm512_sra_epi32(vq31prodGHIJKLMNOPQRSTUV, vshift);
205 
206       const __m512i vminus_one = _mm512_set1_epi32(-1);
207       vacc0123456789ABCDEF = _mm512_mask_sub_epi32(vacc0123456789ABCDEF, _mm512_cmpgt_epi32_mask(vrem0123456789ABCDEF, vremainder_threshold), vacc0123456789ABCDEF, vminus_one);
208       vaccGHIJKLMNOPQRSTUV = _mm512_mask_sub_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_cmpgt_epi32_mask(vremGHIJKLMNOPQRSTUV, vremainder_threshold), vaccGHIJKLMNOPQRSTUV, vminus_one);
209 
210       __m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point);
211       __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vaccGHIJKLMNOPQRSTUV), _mm512_extracti32x8_epi32(vaccGHIJKLMNOPQRSTUV, 1)), _mm512_castsi512_si256(voutput_zero_point));
212 
213       vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_min_epi16(_mm512_max_epi16(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV, voutput_min), voutput_max);
214       voutGHIJOPQRKLMNSTUV = _mm256_min_epi16(_mm256_max_epi16(voutGHIJOPQRKLMNSTUV, _mm512_castsi512_si256(voutput_min)), _mm512_castsi512_si256(voutput_max));
215 
216       const __m256i vout0123GHIJ4567KLMN = _mm512_castsi512_si256(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV);
217       const __m256i vout89ABOPQRCDEFSTUV = _mm512_extracti32x8_epi32(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV, 1);
218       const __m256i vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV = _mm256_packs_epi16(vout0123GHIJ4567KLMN, vout89ABOPQRCDEFSTUV);
219       __m256i vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_permutevar8x32_epi32(vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV, vpermute_mask);
220       const __m128i voutGHIJOPQR = _mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV);
221       const __m128i voutKLMNSTUV = _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1);
222       __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(voutGHIJOPQR, voutKLMNSTUV), _MM_SHUFFLE(3, 1, 2, 0));
223 
224       _mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV);
225       _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
226       output += 32;
227     }
228     if XNN_UNLIKELY(c != 0) {
229       // Prepare mask for valid 8-bit elements (depends on nc).
230       const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1)));
231       const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t));
232       do {
233         __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w);
234 
235 
236         const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0));
237         const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) k));
238         i0 += 16;
239 
240         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF));
241 
242         const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1));
243         const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 32)));
244         i1 += 16;
245 
246         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF));
247 
248         const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2));
249         const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 64)));
250         i2 += 16;
251 
252         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF));
253 
254         const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3));
255         const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 96)));
256         i3 += 16;
257 
258         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF));
259 
260         const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4));
261         const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 128)));
262         i4 += 16;
263 
264         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF));
265 
266         const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5));
267         const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 160)));
268         i5 += 16;
269 
270         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF));
271 
272         const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6));
273         const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 192)));
274         i6 += 16;
275 
276         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF));
277 
278         const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7));
279         const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 224)));
280         i7 += 16;
281 
282         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF));
283 
284         const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8));
285         const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 256)));
286         i8 += 16;
287 
288         vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF));
289 
290         w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
291         k += 16;
292 
293         const __m512i vacc13579BDF = _mm512_shuffle_epi32(vacc0123456789ABCDEF, _MM_SHUFFLE(3, 3, 1, 1));
294 
295         const __m512i vprod02468ACE = _mm512_add_epi64(_mm512_mul_epi32(vacc0123456789ABCDEF, vmultiplier), vrounding);
296         const __m512i vprod13579BDF = _mm512_add_epi64(_mm512_mul_epi32(vacc13579BDF, vmultiplier), vrounding);
297 
298         const __m512i vq31prod02468ACE = _mm512_srli_epi64(vprod02468ACE, 31);
299         const __m512i vq31prod13579BDF = _mm512_add_epi64(vprod13579BDF, vprod13579BDF);
300 
301         const __m512i vq31prod0123456789ABCDEF = _mm512_mask_blend_epi32(vblend_mask, vq31prod02468ACE, vq31prod13579BDF);
302 
303         const __m512i vrem0123456789ABCDEF = _mm512_add_epi32(_mm512_and_epi32(vq31prod0123456789ABCDEF, vremainder_mask), _mm512_srai_epi32(vq31prod0123456789ABCDEF, 31));
304 
305         vacc0123456789ABCDEF = _mm512_sra_epi32(vq31prod0123456789ABCDEF, vshift);
306         vacc0123456789ABCDEF = _mm512_mask_sub_epi32(vacc0123456789ABCDEF, _mm512_cmpgt_epi32_mask(vrem0123456789ABCDEF, vremainder_threshold), vacc0123456789ABCDEF, _mm512_set1_epi32(-1));
307 
308         __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point));
309 
310         vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, _mm512_castsi512_si256(voutput_min)), _mm512_castsi512_si256(voutput_max));
311 
312         const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF);
313         const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1);
314         __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0));
315 
316         if XNN_LIKELY(c >= 16) {
317           _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
318           output += 16;
319           c -= 16;
320         } else {
321           _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF);
322           output = (int8_t*) ((uintptr_t) output + c);
323           c = 0;
324         }
325       } while (c != 0);
326     }
327 
328     output = (int8_t*) ((uintptr_t) output + output_increment);
329   } while (--output_width != 0);
330 }
331