1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gavgpool/multipass-sse.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/gavgpool.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2(size_t rows,size_t channels,const int8_t * input,size_t input_stride,const int8_t * zero,int32_t * buffer,int8_t * output,const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2(
19     size_t rows,
20     size_t channels,
21     const int8_t* input,
22     size_t input_stride,
23     const int8_t* zero,
24     int32_t* buffer,
25     int8_t* output,
26     const union xnn_qs8_avgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
27 {
28   assert(rows > 7);
29   assert(channels != 0);
30 
31   const int8_t* i0 = input;
32   const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
33   const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
34   const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
35   const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
36   const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
37   const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride);
38   const size_t input_increment = 7 * input_stride - round_up_po2(channels, 16);
39 
40   const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
41   int32_t* b = buffer;
42   size_t c = channels;
43   for (; c != 0; c = doz(c, 16)) {
44     const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0));
45     const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
46     i0 += 16;
47     const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1));
48     const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
49     i1 += 16;
50     const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2));
51     const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
52     i2 += 16;
53     const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3));
54     const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
55     i3 += 16;
56     const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4));
57     const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
58     i4 += 16;
59     const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5));
60     const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
61     i5 += 16;
62     const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i6));
63     const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
64     i6 += 16;
65 
66 
67     __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
68     __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
69     __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
70     __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF);
71 
72     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
73     vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF);
74     vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
75     vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF);
76     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
77     vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi6x89ABCDEF);
78 
79     // Add up all accumulators to vacc0x0123456789ABCDEF
80     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
81     vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vacc1x89ABCDEF);
82 
83     const __m128i vacc0123 = _mm_add_epi32(vbias, _mm_cvtepi16_epi32(vacc0x01234567));
84     const __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x01234567, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567)));
85     const __m128i vacc89AB = _mm_add_epi32(vbias, _mm_cvtepi16_epi32(vacc0x89ABCDEF));
86     const __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vacc0x89ABCDEF, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x89ABCDEF)));
87 
88     _mm_store_si128((__m128i*) b, vacc0123);
89     _mm_store_si128((__m128i*) (b + 4), vacc4567);
90     _mm_store_si128((__m128i*) (b + 8), vacc89AB);
91     _mm_store_si128((__m128i*) (b + 12), vaccCDEF);
92     b += 16;
93   }
94 
95   for (rows -= 7; rows > 7; rows -= 7) {
96     i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
97     i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
98     i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
99     i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
100     i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
101     i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
102     i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
103 
104     int32_t* b = buffer;
105     size_t c = channels;
106     for (; c != 0; c = doz(c, 16)) {
107       const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0));
108       const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
109       i0 += 16;
110       const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1));
111       const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
112       i1 += 16;
113       const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2));
114       const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
115       i2 += 16;
116       const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3));
117       const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
118       i3 += 16;
119       const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4));
120       const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
121       i4 += 16;
122       const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5));
123       const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
124       i5 += 16;
125       const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i6));
126       const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
127       i6 += 16;
128 
129 
130       __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
131       __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
132       __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
133       __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF);
134 
135       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
136       vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF);
137       vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
138       vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF);
139       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
140       vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi6x89ABCDEF);
141 
142       // Add up all accumulators to vacc0x0123456789ABCDEF
143       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
144       vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vacc1x89ABCDEF);
145 
146       const __m128i vacc0123 = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x01234567), _mm_load_si128((const __m128i*) (b + 0)));
147       const __m128i vacc4567 = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x01234567, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567)), _mm_load_si128((const __m128i*) (b + 4)));
148       const __m128i vacc89AB = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x89ABCDEF), _mm_load_si128((const __m128i*) (b + 8)));
149       const __m128i vaccCDEF = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x89ABCDEF, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x89ABCDEF)), _mm_load_si128((const __m128i*) (b + 12)));
150 
151       _mm_store_si128((__m128i*) b, vacc0123);
152       _mm_store_si128((__m128i*) (b + 4), vacc4567);
153       _mm_store_si128((__m128i*) (b + 8), vacc89AB);
154       _mm_store_si128((__m128i*) (b + 12), vaccCDEF);
155       b += 16;
156     }
157   }
158 
159   i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
160   i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
161   if XNN_UNPREDICTABLE(rows < 2) {
162     i1 = zero;
163   }
164   i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
165   if XNN_UNPREDICTABLE(rows <= 2) {
166     i2 = zero;
167   }
168   i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
169   if XNN_UNPREDICTABLE(rows < 4) {
170     i3 = zero;
171   }
172   i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
173   if XNN_UNPREDICTABLE(rows <= 4) {
174     i4 = zero;
175   }
176   i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
177   if XNN_UNPREDICTABLE(rows < 6) {
178     i5 = zero;
179   }
180   i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
181   if XNN_UNPREDICTABLE(rows <= 6) {
182     i6 = zero;
183   }
184 
185   const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
186   const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
187   const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->sse2.shift);
188   while (channels >= 16) {
189     const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0));
190     const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
191     i0 += 16;
192     const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1));
193     const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
194     i1 += 16;
195     const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2));
196     const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
197     i2 += 16;
198     const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3));
199     const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
200     i3 += 16;
201     const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4));
202     const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
203     i4 += 16;
204     const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5));
205     const __m128i vxi5x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
206     i5 += 16;
207     const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i6));
208     const __m128i vxi6x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
209     i6 += 16;
210 
211 
212     __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
213     __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
214     __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
215     __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF);
216 
217     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
218     vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF);
219     vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
220     vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF);
221     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
222     vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi6x89ABCDEF);
223 
224     // Add up all accumulators to vacc0x0123456789ABCDEF
225     vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
226     vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vacc1x89ABCDEF);
227 
228     const __m128i vacc0123 = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x01234567), _mm_load_si128((const __m128i*) (buffer + 0)));
229     const __m128i vacc4567 = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x01234567, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567)), _mm_load_si128((const __m128i*) (buffer + 4)));
230     const __m128i vacc89AB = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x89ABCDEF), _mm_load_si128((const __m128i*) (buffer + 8)));
231     const __m128i vaccCDEF = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x89ABCDEF, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x89ABCDEF)), _mm_load_si128((const __m128i*) (buffer + 12)));
232     buffer += 16;
233 
234     const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123);
235     const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567);
236     const __m128i vabsacc89AB = _mm_abs_epi32(vacc89AB);
237     const __m128i vabsaccCDEF = _mm_abs_epi32(vaccCDEF);
238 
239     const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1));
240     const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1));
241     const __m128i vabsacc9B = _mm_shuffle_epi32(vabsacc89AB, _MM_SHUFFLE(3, 3, 1, 1));
242     const __m128i vabsaccDF = _mm_shuffle_epi32(vabsaccCDEF, _MM_SHUFFLE(3, 3, 1, 1));
243 
244     const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier);
245     const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier);
246     const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier);
247     const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier);
248     const __m128i vabsprod8A = _mm_mul_epu32(vabsacc89AB, vmultiplier);
249     const __m128i vabsprod9B = _mm_mul_epu32(vabsacc9B, vmultiplier);
250     const __m128i vabsprodCE = _mm_mul_epu32(vabsaccCDEF, vmultiplier);
251     const __m128i vabsprodDF = _mm_mul_epu32(vabsaccDF, vmultiplier);
252 
253     const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift);
254     const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift);
255     const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift);
256     const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift);
257     const __m128i vabsout8A = _mm_srl_epi64(_mm_add_epi64(vabsprod8A, vrounding), vshift);
258     const __m128i vabsout9B = _mm_srl_epi64(_mm_add_epi64(vabsprod9B, vrounding), vshift);
259     const __m128i vabsoutCE = _mm_srl_epi64(_mm_add_epi64(vabsprodCE, vrounding), vshift);
260     const __m128i vabsoutDF = _mm_srl_epi64(_mm_add_epi64(vabsprodDF, vrounding), vshift);
261 
262     const __m128i vabsout0123 = _mm_blend_epi16(vabsout02, _mm_shuffle_epi32(vabsout13, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
263     const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
264     const __m128i vabsout89AB = _mm_blend_epi16(vabsout8A, _mm_shuffle_epi32(vabsout9B, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
265     const __m128i vabsoutCDEF = _mm_blend_epi16(vabsoutCE, _mm_shuffle_epi32(vabsoutDF, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
266 
267     const __m128i vout0123 = _mm_sign_epi32(vabsout0123, vacc0123);
268     const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567);
269     const __m128i vout89AB = _mm_sign_epi32(vabsout89AB, vacc89AB);
270     const __m128i voutCDEF = _mm_sign_epi32(vabsoutCDEF, vaccCDEF);
271 
272     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
273     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vout0123, vout4567), voutput_zero_point);
274     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vout89AB, voutCDEF), voutput_zero_point);
275 
276     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
277     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
278     vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
279     vout89ABCDEF = _mm_min_epi16(_mm_max_epi16(vout89ABCDEF, voutput_min), voutput_max);
280 
281     __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
282 
283     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
284     output += 16;
285 
286     channels -= 16;
287   }
288   if XNN_UNLIKELY(channels != 0) {
289     do {
290       const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0));
291       i0 += 8;
292       const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1));
293       i1 += 8;
294       const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2));
295       i2 += 8;
296       const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3));
297       i3 += 8;
298       const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4));
299       i4 += 8;
300       const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5));
301       i5 += 8;
302       const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i6));
303       i6 += 8;
304 
305 
306       __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
307       __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567);
308 
309       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567);
310       vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567);
311       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567);
312 
313       // Add up all accumulators to vacc0x01234567
314       vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567);
315 
316       const __m128i vacc0123 = _mm_add_epi32(_mm_cvtepi16_epi32(vacc0x01234567), _mm_load_si128((const __m128i*) buffer));
317       const __m128i vacc4567 = _mm_add_epi32(_mm_unpackhi_epi16(vacc0x01234567, _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567)), _mm_load_si128((const __m128i*) (buffer + 4)));
318       buffer += 8;
319 
320       const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123);
321       const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567);
322 
323       const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1));
324       const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1));
325 
326       const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier);
327       const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier);
328       const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier);
329       const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier);
330 
331       const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift);
332       const __m128i vabsout13 = _mm_srl_epi64(_mm_add_epi64(vabsprod13, vrounding), vshift);
333       const __m128i vabsout46 = _mm_srl_epi64(_mm_add_epi64(vabsprod46, vrounding), vshift);
334       const __m128i vabsout57 = _mm_srl_epi64(_mm_add_epi64(vabsprod57, vrounding), vshift);
335 
336       const __m128i vabsout0123 = _mm_blend_epi16(vabsout02, _mm_shuffle_epi32(vabsout13, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
337       const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2, 2, 0, 0)), 0xCC);
338 
339       const __m128i vout0123 = _mm_sign_epi32(vabsout0123, vacc0123);
340       const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567);
341 
342       const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
343       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vout0123, vout4567), voutput_zero_point);
344 
345       const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
346       const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
347       vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
348 
349       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
350 
351       if XNN_LIKELY(channels >= 8) {
352         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
353         output += 8;
354         channels -= 8;
355       } else {
356         if (channels & 4) {
357           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
358           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
359           output += 4;
360         }
361         if (channels & 2) {
362           *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
363           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
364           output += 2;
365         }
366         if (channels & 1) {
367           *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
368           output += 1;
369         }
370         channels = 0;
371       }
372     } while (channels != 0);
373   }
374 }
375