26 …const __m128i vkernel_zero_point = _mm_load_si128((const __m128i*) params->sse2.kernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
27   const __m128i vzero = _mm_setzero_si128();  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
72       __m128i vacc_lo = _mm_loadu_si128((const __m128i*) w);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
73       __m128i vacc_hi = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
75       const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
76       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
77       const __m128i vk0 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
78       const __m128i vxk0 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
79       const __m128i vprod0_odd  = _mm_mullo_epi16(vxi0, vxk0);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
80       const __m128i vprod0_even = _mm_mulhi_epi16(vxi0, vxk0);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
84       const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
85       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
86       const __m128i vk1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
87       const __m128i vxk1 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
88       const __m128i vprod1_odd  = _mm_mullo_epi16(vxi1, vxk1);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
89       const __m128i vprod1_even = _mm_mulhi_epi16(vxi1, vxk1);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
93       const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
94       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
95       const __m128i vk2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
96       const __m128i vxk2 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
97       const __m128i vprod2_odd  = _mm_mullo_epi16(vxi2, vxk2);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
98       const __m128i vprod2_even = _mm_mulhi_epi16(vxi2, vxk2);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
102       const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
103       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
104       const __m128i vk3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
105       const __m128i vxk3 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
106       const __m128i vprod3_odd  = _mm_mullo_epi16(vxi3, vxk3);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
107       const __m128i vprod3_even = _mm_mulhi_epi16(vxi3, vxk3);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
111       const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
112       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
113       const __m128i vk4 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
114       const __m128i vxk4 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
115       const __m128i vprod4_odd  = _mm_mullo_epi16(vxi4, vxk4);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
116       const __m128i vprod4_even = _mm_mulhi_epi16(vxi4, vxk4);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
120       const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
121       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
122       const __m128i vk5 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
123       const __m128i vxk5 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
124       const __m128i vprod5_odd  = _mm_mullo_epi16(vxi5, vxk5);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
125       const __m128i vprod5_even = _mm_mulhi_epi16(vxi5, vxk5);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
129       const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
130       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
131       const __m128i vk6 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
132       const __m128i vxk6 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
133       const __m128i vprod6_odd  = _mm_mullo_epi16(vxi6, vxk6);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
134       const __m128i vprod6_even = _mm_mulhi_epi16(vxi6, vxk6);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
138       const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
139       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
140       const __m128i vk7 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
141       const __m128i vxk7 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
142       const __m128i vprod7_odd  = _mm_mullo_epi16(vxi7, vxk7);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
143       const __m128i vprod7_even = _mm_mulhi_epi16(vxi7, vxk7);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
147       const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); i8 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
148       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
149       const __m128i vk8 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
150       const __m128i vxk8 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
151       const __m128i vprod8_odd  = _mm_mullo_epi16(vxi8, vxk8);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
152       const __m128i vprod8_even = _mm_mulhi_epi16(vxi8, vxk8);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
158       const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
159       const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
161       const __m128i vnmask_lo0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
162       const __m128i vnmask_hi0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
164 …const __m128i vabsacc_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vnmask_lo0123), vnmask_lo0123);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
165 …const __m128i vabsacc_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vnmask_hi0123), vnmask_hi0123);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
167       const __m128i vabsacc_lo1032 = _mm_shuffle_epi32(vabsacc_lo0123, _MM_SHUFFLE(2, 3, 0, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
168       const __m128i vabsacc_hi1032 = _mm_shuffle_epi32(vabsacc_hi0123, _MM_SHUFFLE(2, 3, 0, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
170       const __m128i vabsprod_lo02 = _mm_mul_epu32(vabsacc_lo0123, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
171       const __m128i vabsprod_hi02 = _mm_mul_epu32(vabsacc_hi0123, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
173       const __m128i vnmask_lo02 = _mm_shuffle_epi32(vnmask_lo0123, _MM_SHUFFLE(2, 2, 0, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
174       const __m128i vnmask_hi02 = _mm_shuffle_epi32(vnmask_hi0123, _MM_SHUFFLE(2, 2, 0, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
176 …  const __m128i vprod_lo02 = _mm_sub_epi64(_mm_xor_si128(vabsprod_lo02, vnmask_lo02), vnmask_lo02);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
177 …  const __m128i vprod_hi02 = _mm_sub_epi64(_mm_xor_si128(vabsprod_hi02, vnmask_hi02), vnmask_hi02);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
179       const __m128i vq31prod_lo02 = _mm_srli_epi64(_mm_add_epi64(vprod_lo02, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
180       const __m128i vq31prod_hi02 = _mm_srli_epi64(_mm_add_epi64(vprod_hi02, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
182       const __m128i vabsprod_lo13 = _mm_mul_epu32(vabsacc_lo1032, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
183       const __m128i vabsprod_hi13 = _mm_mul_epu32(vabsacc_hi1032, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
185       const __m128i vnmask_lo13 = _mm_shuffle_epi32(vnmask_lo0123, _MM_SHUFFLE(3, 3, 1, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
186       const __m128i vnmask_hi13 = _mm_shuffle_epi32(vnmask_hi0123, _MM_SHUFFLE(3, 3, 1, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
188 …  const __m128i vprod_lo13 = _mm_sub_epi64(_mm_xor_si128(vabsprod_lo13, vnmask_lo13), vnmask_lo13);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
189 …  const __m128i vprod_hi13 = _mm_sub_epi64(_mm_xor_si128(vabsprod_hi13, vnmask_hi13), vnmask_hi13);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
191       const __m128i vq31prod_lo13 = _mm_srli_epi64(_mm_add_epi64(vprod_lo13, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
192       const __m128i vq31prod_hi13 = _mm_srli_epi64(_mm_add_epi64(vprod_hi13, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
194       const __m128i vq31prod_lo0213 = _mm_castps_si128(_mm_shuffle_ps(  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
196       const __m128i vq31prod_hi0213 = _mm_castps_si128(_mm_shuffle_ps(  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
199       const __m128i vq31prod_lo0123 = _mm_shuffle_epi32(vq31prod_lo0213, _MM_SHUFFLE(3, 1, 2, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
200       const __m128i vq31prod_hi0123 = _mm_shuffle_epi32(vq31prod_hi0213, _MM_SHUFFLE(3, 1, 2, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
202       const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
204       const __m128i vrem_lo0123 =  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
206       const __m128i vrem_hi0123 =  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
209 …const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_thresh…  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
210       const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
212 …const __m128i vout_lo = _mm_sub_epi32(_mm_sra_epi32(vq31prod_lo0123, vshift), _mm_cmpgt_epi32(vrem…  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
213 …const __m128i vout_hi = _mm_sub_epi32(_mm_sra_epi32(vq31prod_hi0123, vshift), _mm_cmpgt_epi32(vrem…  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
215 …const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
216       __m128i vout = _mm_adds_epi16(_mm_packs_epi32(vout_lo, vout_hi), voutput_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
218       vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_max));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
219       vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_min));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
221       _mm_storel_epi64((__m128i*) output, vout); output += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
224       __m128i vacc_lo = _mm_loadu_si128((const __m128i*) w);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
225       __m128i vacc_hi = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
227       const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
228       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
229       const __m128i vk0 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 32));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
230       const __m128i vxk0 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
231       const __m128i vprod0_odd  = _mm_mullo_epi16(vxi0, vxk0);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
232       const __m128i vprod0_even = _mm_mulhi_epi16(vxi0, vxk0);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
236       const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
237       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
238       const __m128i vk1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 40));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
239       const __m128i vxk1 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
240       const __m128i vprod1_odd  = _mm_mullo_epi16(vxi1, vxk1);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
241       const __m128i vprod1_even = _mm_mulhi_epi16(vxi1, vxk1);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
245       const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
246       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
247       const __m128i vk2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 48));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
248       const __m128i vxk2 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
249       const __m128i vprod2_odd  = _mm_mullo_epi16(vxi2, vxk2);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
250       const __m128i vprod2_even = _mm_mulhi_epi16(vxi2, vxk2);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
254       const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
255       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
256       const __m128i vk3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 56));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
257       const __m128i vxk3 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
258       const __m128i vprod3_odd  = _mm_mullo_epi16(vxi3, vxk3);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
259       const __m128i vprod3_even = _mm_mulhi_epi16(vxi3, vxk3);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
263       const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
264       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
265       const __m128i vk4 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 64));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
266       const __m128i vxk4 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
267       const __m128i vprod4_odd  = _mm_mullo_epi16(vxi4, vxk4);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
268       const __m128i vprod4_even = _mm_mulhi_epi16(vxi4, vxk4);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
272       const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
273       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
274       const __m128i vk5 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 72));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
275       const __m128i vxk5 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
276       const __m128i vprod5_odd  = _mm_mullo_epi16(vxi5, vxk5);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
277       const __m128i vprod5_even = _mm_mulhi_epi16(vxi5, vxk5);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
281       const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
282       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
283       const __m128i vk6 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 80));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
284       const __m128i vxk6 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
285       const __m128i vprod6_odd  = _mm_mullo_epi16(vxi6, vxk6);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
286       const __m128i vprod6_even = _mm_mulhi_epi16(vxi6, vxk6);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
290       const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
291       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
292       const __m128i vk7 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 88));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
293       const __m128i vxk7 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
294       const __m128i vprod7_odd  = _mm_mullo_epi16(vxi7, vxk7);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
295       const __m128i vprod7_even = _mm_mulhi_epi16(vxi7, vxk7);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
299       const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); i8 += 8;  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
300       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
301       const __m128i vk8 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 96));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
302       const __m128i vxk8 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8, vzero), vkernel_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
303       const __m128i vprod8_odd  = _mm_mullo_epi16(vxi8, vxk8);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
304       const __m128i vprod8_even = _mm_mulhi_epi16(vxi8, vxk8);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
308       const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
309       const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
311       const __m128i vnmask_lo0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
312       const __m128i vnmask_hi0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
314 …const __m128i vabsacc_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vnmask_lo0123), vnmask_lo0123);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
315 …const __m128i vabsacc_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vnmask_hi0123), vnmask_hi0123);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
317       const __m128i vabsacc_lo1032 = _mm_shuffle_epi32(vabsacc_lo0123, _MM_SHUFFLE(2, 3, 0, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
318       const __m128i vabsacc_hi1032 = _mm_shuffle_epi32(vabsacc_hi0123, _MM_SHUFFLE(2, 3, 0, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
320       const __m128i vabsprod_lo02 = _mm_mul_epu32(vabsacc_lo0123, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
321       const __m128i vabsprod_hi02 = _mm_mul_epu32(vabsacc_hi0123, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
323       const __m128i vnmask_lo02 = _mm_shuffle_epi32(vnmask_lo0123, _MM_SHUFFLE(2, 2, 0, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
324       const __m128i vnmask_hi02 = _mm_shuffle_epi32(vnmask_hi0123, _MM_SHUFFLE(2, 2, 0, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
326 …  const __m128i vprod_lo02 = _mm_sub_epi64(_mm_xor_si128(vabsprod_lo02, vnmask_lo02), vnmask_lo02);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
327 …  const __m128i vprod_hi02 = _mm_sub_epi64(_mm_xor_si128(vabsprod_hi02, vnmask_hi02), vnmask_hi02);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
329       const __m128i vq31prod_lo02 = _mm_srli_epi64(_mm_add_epi64(vprod_lo02, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
330       const __m128i vq31prod_hi02 = _mm_srli_epi64(_mm_add_epi64(vprod_hi02, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
332       const __m128i vabsprod_lo13 = _mm_mul_epu32(vabsacc_lo1032, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
333       const __m128i vabsprod_hi13 = _mm_mul_epu32(vabsacc_hi1032, vmultiplier);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
335       const __m128i vnmask_lo13 = _mm_shuffle_epi32(vnmask_lo0123, _MM_SHUFFLE(3, 3, 1, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
336       const __m128i vnmask_hi13 = _mm_shuffle_epi32(vnmask_hi0123, _MM_SHUFFLE(3, 3, 1, 1));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
338 …  const __m128i vprod_lo13 = _mm_sub_epi64(_mm_xor_si128(vabsprod_lo13, vnmask_lo13), vnmask_lo13);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
339 …  const __m128i vprod_hi13 = _mm_sub_epi64(_mm_xor_si128(vabsprod_hi13, vnmask_hi13), vnmask_hi13);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
341       const __m128i vq31prod_lo13 = _mm_srli_epi64(_mm_add_epi64(vprod_lo13, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
342       const __m128i vq31prod_hi13 = _mm_srli_epi64(_mm_add_epi64(vprod_hi13, vrounding), 31);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
344       const __m128i vq31prod_lo0213 = _mm_castps_si128(_mm_shuffle_ps(  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
346       const __m128i vq31prod_hi0213 = _mm_castps_si128(_mm_shuffle_ps(  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
349       const __m128i vq31prod_lo0123 = _mm_shuffle_epi32(vq31prod_lo0213, _MM_SHUFFLE(3, 1, 2, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
350       const __m128i vq31prod_hi0123 = _mm_shuffle_epi32(vq31prod_hi0213, _MM_SHUFFLE(3, 1, 2, 0));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
352       const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
354       const __m128i vrem_lo0123 =  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
356       const __m128i vrem_hi0123 =  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
359 …const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_thresh…  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
360       const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
362 …const __m128i vout_lo = _mm_sub_epi32(_mm_sra_epi32(vq31prod_lo0123, vshift), _mm_cmpgt_epi32(vrem…  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
363 …const __m128i vout_hi = _mm_sub_epi32(_mm_sra_epi32(vq31prod_hi0123, vshift), _mm_cmpgt_epi32(vrem…  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
365 …const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
366       __m128i vout = _mm_adds_epi16(_mm_packs_epi32(vout_lo, vout_hi), voutput_zero_point);  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
368       vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_max));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
369       vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->sse2.output_min));  in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()