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 
16 
xnn_qs8_dwconv_minmax_ukernel_up8x9__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_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16(
18     size_t channels,
19     size_t output_width,
20     const int8_t** input,
21     const void* weights,
22     int8_t* output,
23     size_t input_stride,
24     size_t output_increment,
25     size_t input_offset,
26     const int8_t* zero,
27     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
28 {
29   assert(channels != 0);
30   assert(output_width != 0);
31 
32   do {
33     const int8_t* i0 = input[0];
34     assert(i0 != NULL);
35     if XNN_UNPREDICTABLE(i0 != zero) {
36       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
37     }
38     const int8_t* i1 = input[1];
39     assert(i1 != NULL);
40     if XNN_UNPREDICTABLE(i1 != zero) {
41       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
42     }
43     const int8_t* i2 = input[2];
44     assert(i2 != NULL);
45     if XNN_UNPREDICTABLE(i2 != zero) {
46       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
47     }
48     const int8_t* i3 = input[3];
49     assert(i3 != NULL);
50     if XNN_UNPREDICTABLE(i3 != zero) {
51       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
52     }
53     const int8_t* i4 = input[4];
54     assert(i4 != NULL);
55     if XNN_UNPREDICTABLE(i4 != zero) {
56       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
57     }
58     const int8_t* i5 = input[5];
59     assert(i5 != NULL);
60     if XNN_UNPREDICTABLE(i5 != zero) {
61       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
62     }
63     const int8_t* i6 = input[6];
64     assert(i6 != NULL);
65     if XNN_UNPREDICTABLE(i6 != zero) {
66       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
67     }
68     const int8_t* i7 = input[7];
69     assert(i7 != NULL);
70     if XNN_UNPREDICTABLE(i7 != zero) {
71       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
72     }
73     const int8_t* i8 = input[8];
74     assert(i8 != NULL);
75     if XNN_UNPREDICTABLE(i8 != zero) {
76       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
77     }
78     input = (const int8_t**) ((uintptr_t) input + input_stride);
79 
80     size_t c = channels;
81     const void* w = weights;
82     for (; c >= 8; c -= 8) {
83       __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
84       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 4 * sizeof(int32_t)));
85 
86 
87       const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
88       const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
89       const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
90       const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
91       i0 += 8;
92 
93 
94       const __m128i vp0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
95       const __m128i vp0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
96 
97       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp0x01234567lo, vp0x01234567hi));
98       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi));
99 
100       const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
101       const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
102       const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
103       const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
104       i1 += 8;
105 
106 
107       const __m128i vp1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
108       const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
109 
110       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi));
111       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi));
112 
113       const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
114       const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
115       const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
116       const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
117       i2 += 8;
118 
119 
120       const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
121       const __m128i vp2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
122 
123       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi));
124       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi));
125 
126       const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
127       const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
128       const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
129       const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
130       i3 += 8;
131 
132 
133       const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
134       const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
135 
136       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi));
137       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi));
138 
139       const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
140       const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
141       const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
142       const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
143       i4 += 8;
144 
145 
146       const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
147       const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
148 
149       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi));
150       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi));
151 
152       const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
153       const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
154       const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
155       const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
156       i5 += 8;
157 
158 
159       const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
160       const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
161 
162       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp5x01234567lo, vp5x01234567hi));
163       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi));
164 
165       const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
166       const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
167       const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
168       const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
169       i6 += 8;
170 
171 
172       const __m128i vp6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
173       const __m128i vp6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
174 
175       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp6x01234567lo, vp6x01234567hi));
176       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi));
177 
178       const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
179       const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
180       const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
181       const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
182       i7 += 8;
183 
184 
185       const __m128i vp7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
186       const __m128i vp7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
187 
188       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp7x01234567lo, vp7x01234567hi));
189       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi));
190 
191       const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
192       const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
193       const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
194       const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
195       i8 += 8;
196 
197 
198       const __m128i vp8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
199       const __m128i vp8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
200 
201       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp8x01234567lo, vp8x01234567hi));
202       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi));
203 
204       w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
205 
206       const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
207       const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
208 
209       const __m128i vacc13 = _mm_shuffle_epi32(vacc0123, _MM_SHUFFLE(3, 3, 1, 1));
210       const __m128i vprod02 = _mm_add_epi64(_mm_mul_epi32(vacc0123, vmultiplier), vrounding);
211       const __m128i vprod13 = _mm_add_epi64(_mm_mul_epi32(vacc13, vmultiplier), vrounding);
212       const __m128i vacc57 = _mm_shuffle_epi32(vacc4567, _MM_SHUFFLE(3, 3, 1, 1));
213       const __m128i vprod46 = _mm_add_epi64(_mm_mul_epi32(vacc4567, vmultiplier), vrounding);
214       const __m128i vprod57 = _mm_add_epi64(_mm_mul_epi32(vacc57, vmultiplier), vrounding);
215 
216       const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31);
217       const __m128i vq31prod13 = _mm_add_epi64(vprod13, vprod13);
218       const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31);
219       const __m128i vq31prod57 = _mm_add_epi64(vprod57, vprod57);
220 
221       const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC);
222       const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC);
223 
224       const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
225       const __m128i vrem0123 =
226         _mm_add_epi32(_mm_and_si128(vq31prod0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0123));
227       const __m128i vrem4567 =
228         _mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod4567));
229 
230       const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
231       const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
232       vacc0123 =
233         _mm_sub_epi32(_mm_sra_epi32(vq31prod0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
234       vacc4567 =
235         _mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
236 
237       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
238       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
239 
240       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
241       const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
242       vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
243 
244       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
245 
246       _mm_storel_epi64((__m128i*) output, vout0123456701234567);
247       output += 8;
248     }
249     if XNN_UNLIKELY(c != 0) {
250       {
251         __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
252         __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 4 * sizeof(int32_t)));
253 
254 
255         const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0);
256         const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(vi0x01234567);
257         const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t)));
258         const __m128i vxk0x01234567 = _mm_cvtepi8_epi16(vk0x01234567);
259 
260 
261         const __m128i vp0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567);
262         const __m128i vp0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567);
263 
264         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp0x01234567lo, vp0x01234567hi));
265         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi));
266 
267         const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1);
268         const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(vi1x01234567);
269         const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t)));
270         const __m128i vxk1x01234567 = _mm_cvtepi8_epi16(vk1x01234567);
271 
272 
273         const __m128i vp1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567);
274         const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567);
275 
276         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi));
277         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi));
278 
279         const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2);
280         const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(vi2x01234567);
281         const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t)));
282         const __m128i vxk2x01234567 = _mm_cvtepi8_epi16(vk2x01234567);
283 
284 
285         const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567);
286         const __m128i vp2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567);
287 
288         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi));
289         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi));
290 
291         const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3);
292         const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567);
293         const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)));
294         const __m128i vxk3x01234567 = _mm_cvtepi8_epi16(vk3x01234567);
295 
296 
297         const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567);
298         const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567);
299 
300         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi));
301         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi));
302 
303         const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4);
304         const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567);
305         const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t)));
306         const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567);
307 
308 
309         const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567);
310         const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567);
311 
312         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi));
313         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi));
314 
315         const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5);
316         const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567);
317         const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t)));
318         const __m128i vxk5x01234567 = _mm_cvtepi8_epi16(vk5x01234567);
319 
320 
321         const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567);
322         const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567);
323 
324         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp5x01234567lo, vp5x01234567hi));
325         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi));
326 
327         const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6);
328         const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(vi6x01234567);
329         const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t)));
330         const __m128i vxk6x01234567 = _mm_cvtepi8_epi16(vk6x01234567);
331 
332 
333         const __m128i vp6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567);
334         const __m128i vp6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567);
335 
336         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp6x01234567lo, vp6x01234567hi));
337         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi));
338 
339         const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7);
340         const __m128i vxi7x01234567 = _mm_cvtepi8_epi16(vi7x01234567);
341         const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t)));
342         const __m128i vxk7x01234567 = _mm_cvtepi8_epi16(vk7x01234567);
343 
344 
345         const __m128i vp7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567);
346         const __m128i vp7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567);
347 
348         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp7x01234567lo, vp7x01234567hi));
349         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi));
350 
351         const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8);
352         const __m128i vxi8x01234567 = _mm_cvtepi8_epi16(vi8x01234567);
353         const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t)));
354         const __m128i vxk8x01234567 = _mm_cvtepi8_epi16(vk8x01234567);
355 
356 
357         const __m128i vp8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567);
358         const __m128i vp8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567);
359 
360         vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp8x01234567lo, vp8x01234567hi));
361         vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi));
362 
363 
364         const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
365         const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
366 
367         const __m128i vacc13 = _mm_shuffle_epi32(vacc0123, _MM_SHUFFLE(3, 3, 1, 1));
368         const __m128i vacc57 = _mm_shuffle_epi32(vacc4567, _MM_SHUFFLE(3, 3, 1, 1));
369 
370         const __m128i vprod02 = _mm_add_epi64(_mm_mul_epi32(vacc0123, vmultiplier), vrounding);
371         const __m128i vprod46 = _mm_add_epi64(_mm_mul_epi32(vacc4567, vmultiplier), vrounding);
372 
373         const __m128i vprod13 = _mm_add_epi64(_mm_mul_epi32(vacc13, vmultiplier), vrounding);
374         const __m128i vprod57 = _mm_add_epi64(_mm_mul_epi32(vacc57, vmultiplier), vrounding);
375 
376         const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31);
377         const __m128i vq31prod13 = _mm_add_epi64(vprod13, vprod13);
378         const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31);
379         const __m128i vq31prod57 = _mm_add_epi64(vprod57, vprod57);
380 
381         const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC);
382         const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC);
383 
384         const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
385         const __m128i vrem0123 =
386           _mm_add_epi32(_mm_and_si128(vq31prod0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0123));
387         const __m128i vrem4567 =
388           _mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod4567));
389 
390         const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
391         const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
392         vacc0123 =
393           _mm_sub_epi32(_mm_sra_epi32(vq31prod0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
394         vacc4567 =
395           _mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
396 
397         const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
398         __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
399 
400         const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
401         const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
402         vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
403 
404         __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
405 
406         if (c & 4) {
407           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
408           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
409           output += 4;
410         }
411         if (c & 2) {
412           *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
413           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
414           output += 2;
415         }
416         if (c & 1) {
417           *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
418           output += 1;
419         }
420       }
421     }
422 
423     output = (int8_t*) ((uintptr_t) output + output_increment);
424   } while (--output_width != 0);
425 }
426