xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/1x16c2-minmax-rndnu-neon-mull-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 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/math.h>
16 
xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 1);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41 
42   do {
43     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47 
48     size_t k = kc;
49 
50 
51     while (k >= 8 * sizeof(int8_t)) {
52       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
53 
54       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
55       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 
71       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
72 
73       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
74       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
75       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
76       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
77       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
78       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
79       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
80       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
81       const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
82 
83       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
84       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
85       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
86       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
87       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
88       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
89       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
90       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
91       const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
92 
93       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
94       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
95       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
96       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
97       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
98       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
99       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
100       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
101       const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
102 
103       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
104       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
105       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
106       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
107       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
108       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
109       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
110       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
111 
112       k -= 8 * sizeof(int8_t);
113     }
114 
115     if XNN_UNLIKELY(k != 0) {
116       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
117 
118       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 
123       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
124       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
125       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
126       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
127       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
128       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
129       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
130       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
131       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
132 
133       if (k > 2 * sizeof(int8_t)) {
134         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 
139         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
140         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
141         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
142         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
143         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
144         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
145         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
146         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
147         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
148 
149         if (k > 4 * sizeof(int8_t)) {
150           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 
155           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
156           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
157           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
158           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
159           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
160           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
161           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
162           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
163           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
164         }
165       }
166     }
167 
168     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
169     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
170     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
171 
172     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
173     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
174     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
175     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
176 
177     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
178     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
179     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
180     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
181 
182     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
183     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
184     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
185     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
186 
187     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
188 #if XNN_ARCH_ARM64
189     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
190     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
191 
192     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
193     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
194 
195     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
196 #else
197     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
198     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
199 
200     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
201     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
202 
203     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
204 #endif
205 
206     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
207     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
208 
209     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
210     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
211 
212     if (nc >= 16) {
213       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
214 
215       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
216 
217       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
218 
219       nc -= 16;
220     } else {
221       // Final case where not all of the 16 columns fit in the destination.
222       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
223       if (nc & 8) {
224         vst1_s8(c0, vout0x01234567); c0 += 8;
225         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
226       }
227       if (nc & 4) {
228         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
229         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
230       }
231       if (nc & 2) {
232         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
233         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
234       }
235       if (nc & 1) {
236         vst1_lane_s8(c0, vout0x01234567, 0);
237       }
238 
239       nc = 0;
240     }
241   } while (nc != 0);
242 }
243