• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c4-neondot.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 <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 6);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   kc = round_up_po2(kc, 4 * sizeof(int8_t));
41   const int8_t* a0 = a;
42   int8_t* c0 = c;
43   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     a1 = a0;
47     c1 = c0;
48   }
49   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     a2 = a1;
53     c2 = c1;
54   }
55   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
57   if XNN_UNPREDICTABLE(mr < 4) {
58     a3 = a2;
59     c3 = c2;
60   }
61   const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
63   if XNN_UNPREDICTABLE(mr <= 4) {
64     a4 = a3;
65     c4 = c3;
66   }
67   const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
69   if XNN_UNPREDICTABLE(mr != 6) {
70     a5 = a4;
71     c5 = c4;
72   }
73 
74   // Loop over groups of 16 columns.
75   do {
76     // Initialize accumulators with bias. 16 bias values are loaded from the
77     // weight matrix, at the start of the group of 16 columns.
78     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
79     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
80     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
81     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
82     int32x4_t vacc1x0123 = vacc0x0123;
83     int32x4_t vacc1x4567 = vacc0x4567;
84     int32x4_t vacc1x89AB = vacc0x89AB;
85     int32x4_t vacc1xCDEF = vacc0xCDEF;
86     int32x4_t vacc2x0123 = vacc0x0123;
87     int32x4_t vacc2x4567 = vacc0x4567;
88     int32x4_t vacc2x89AB = vacc0x89AB;
89     int32x4_t vacc2xCDEF = vacc0xCDEF;
90     int32x4_t vacc3x0123 = vacc0x0123;
91     int32x4_t vacc3x4567 = vacc0x4567;
92     int32x4_t vacc3x89AB = vacc0x89AB;
93     int32x4_t vacc3xCDEF = vacc0xCDEF;
94     int32x4_t vacc4x0123 = vacc0x0123;
95     int32x4_t vacc4x4567 = vacc0x4567;
96     int32x4_t vacc4x89AB = vacc0x89AB;
97     int32x4_t vacc4xCDEF = vacc0xCDEF;
98     int32x4_t vacc5x0123 = vacc0x0123;
99     int32x4_t vacc5x4567 = vacc0x4567;
100     int32x4_t vacc5x89AB = vacc0x89AB;
101     int32x4_t vacc5xCDEF = vacc0xCDEF;
102 
103     // Inner accumulation loop along the 16 columns.
104     size_t k = kc;
105     // 2x partial unrolled loop to load 8 bytes at a time.
106     while (k >= 8 * sizeof(int8_t)) {
107       // Load a 6x8 block of activations.
108       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
109       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
110       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
111       const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
112       const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
113       const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
114 
115       // Load a 8x16 block of weights.
116       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
117       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
118       const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
119       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
120       const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
121       const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
122       const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
123       const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
124 
125       // Multiply-accumulate: 6x8 * 8x16 --> 6x16.
126       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
127       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
128       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
129       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
130       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
131       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
132       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
133       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
134       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
135       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
136       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
137       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
138       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
139       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
140       vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
141       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
142       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
143       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
144       vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
145       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
146       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
147       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
148       vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
149       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
150       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
151       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
152       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
153       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
154       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
155       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
156       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
157       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
158       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
159       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
160       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
161       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
162       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
163       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
164       vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
165       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
166       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
167       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
168       vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
169       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
170       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
171       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
172       vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
173       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
174 
175       k -= 8 * sizeof(int8_t);
176     }
177     // Handle up to 4 final positions of `k`
178     if XNN_UNLIKELY(k != 0) {
179       // Load a 6x4 block of activations.
180       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
181       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
182       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
183       const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
184       const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
185       const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
186 
187       // Load a 4x16 block of weights.
188       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
189       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
190       const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
191       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
192 
193       // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
194       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
195       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
196       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
197       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
198       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
199       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
200       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
201       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
202       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
203       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
204       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
205       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
206       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
207       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
208       vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
209       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
210       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
211       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
212       vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
213       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
214       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
215       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
216       vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
217       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
218     }
219 
220     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
221     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
222     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
223     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
224     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
225     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
226     float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
227     float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
228     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
229     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
230     float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
231     float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
232     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
233     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
234     float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
235     float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
236     float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
237     float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
238     float32x4_t vfpacc4x89AB = vcvtq_f32_s32(vacc4x89AB);
239     float32x4_t vfpacc4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
240     float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
241     float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
242     float32x4_t vfpacc5x89AB = vcvtq_f32_s32(vacc5x89AB);
243     float32x4_t vfpacc5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
244 
245     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
246     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
247     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
248     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
249     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
250     vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
251     vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
252     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
253     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
254     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
255     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
256     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
257     vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
258     vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
259     const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
260     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
261     vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
262     vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
263     vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale89AB);
264     vfpacc4x89AB = vmulq_f32(vfpacc4x89AB, vscale89AB);
265     vfpacc5x89AB = vmulq_f32(vfpacc5x89AB, vscale89AB);
266     const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
267     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
268     vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
269     vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
270     vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscaleCDEF);
271     vfpacc4xCDEF = vmulq_f32(vfpacc4xCDEF, vscaleCDEF);
272     vfpacc5xCDEF = vmulq_f32(vfpacc5xCDEF, vscaleCDEF);
273 
274     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
275     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
276     vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
277     vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
278     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
279     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
280     vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
281     vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
282     vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
283     vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
284     vacc2x89AB = vcvtnq_s32_f32(vfpacc2x89AB);
285     vacc2xCDEF = vcvtnq_s32_f32(vfpacc2xCDEF);
286     vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
287     vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
288     vacc3x89AB = vcvtnq_s32_f32(vfpacc3x89AB);
289     vacc3xCDEF = vcvtnq_s32_f32(vfpacc3xCDEF);
290     vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
291     vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
292     vacc4x89AB = vcvtnq_s32_f32(vfpacc4x89AB);
293     vacc4xCDEF = vcvtnq_s32_f32(vfpacc4xCDEF);
294     vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
295     vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
296     vacc5x89AB = vcvtnq_s32_f32(vfpacc5x89AB);
297     vacc5xCDEF = vcvtnq_s32_f32(vfpacc5xCDEF);
298 
299     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
300 #if XNN_ARCH_ARM64
301     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
302     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
303     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
304     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
305     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
306     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
307     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
308     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
309     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
310     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
311     const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
312     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF), voutput_zero_point);
313 
314     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
315     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
316     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
317     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
318     int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
319     int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
320 #else
321     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
322     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
323     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
324     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
325     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
326     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
327     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
328     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
329     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
330     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
331     const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
332     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF)), voutput_zero_point);
333 
334     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
335     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
336     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
337     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
338     int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
339     int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
340 #endif
341     const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
342     const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
343 
344     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
345     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
346     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
347     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
348     vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
349     vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
350 
351     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
352     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
353     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
354     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
355     vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
356     vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
357 
358     if (nc >= 16) {
359       // Main case where there the 16 columns fit in the destination.
360       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
361       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
362       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
363       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
364       vst1q_s8(c4 + 0, vout4x0123456789ABCDEF);
365       vst1q_s8(c5 + 0, vout5x0123456789ABCDEF);
366 
367       // Advance to the next 16 columns.
368       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
369       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
370       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
371       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
372       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
373       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
374 
375       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
376       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
377       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
378       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
379       a4 = (const int8_t*) ((uintptr_t) a4 - kc);
380       a5 = (const int8_t*) ((uintptr_t) a5 - kc);
381 
382       nc -= 16;
383     } else {
384       // Final case where not all of the 16 columns fit in the destination.
385       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
386       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
387       int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
388       if (nc & 8) {
389         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
390         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
391         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
392         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
393         vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
394         vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
395         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
396         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
397         vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
398       }
399       if (nc & 4) {
400         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
401         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
402         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
403         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
404         vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
405         vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
406         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
407         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
408         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
409       }
410       if (nc & 2) {
411         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
412         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
413         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
414         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
415         vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
416         vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
417         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
418         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
419         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
420       }
421       if (nc & 1) {
422         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
423         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
424         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
425         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
426         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
427         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
428       }
429 
430       nc = 0;
431     }
432   } while (nc != 0);
433 }
434