1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c8-neon-mull-padal.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 
17 
xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal(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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   assert(mr != 0);
31   assert(mr <= 3);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 8);
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54 
55   do {
56     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc1x0 = vacc0x0;
65     int32x4_t vacc1x1 = vacc0x1;
66     int32x4_t vacc1x2 = vacc0x2;
67     int32x4_t vacc1x3 = vacc0x3;
68     int32x4_t vacc1x4 = vacc0x4;
69     int32x4_t vacc1x5 = vacc0x5;
70     int32x4_t vacc1x6 = vacc0x6;
71     int32x4_t vacc1x7 = vacc0x7;
72     int32x4_t vacc2x0 = vacc0x0;
73     int32x4_t vacc2x1 = vacc0x1;
74     int32x4_t vacc2x2 = vacc0x2;
75     int32x4_t vacc2x3 = vacc0x3;
76     int32x4_t vacc2x4 = vacc0x4;
77     int32x4_t vacc2x5 = vacc0x5;
78     int32x4_t vacc2x6 = vacc0x6;
79     int32x4_t vacc2x7 = vacc0x7;
80 
81     size_t k = kc;
82 
83     // Handle 8 bytes at a time using MUL.
84     while (k > 0) {
85       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
86       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
87       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
88 
89       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
91       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
92       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
93       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
94       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
95       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
96       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
98       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
99       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
100       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
101       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
102       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
103       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
105       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
106       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
107       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
108       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
109       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
110       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
112       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
113       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
114       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
115       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
116       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
117       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
119       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
120       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
121       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
122       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
123       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
124       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
126       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
127       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
128       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
129       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
130       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
131       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
133       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
134       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
135       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
136       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
137       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
138       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
140       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
141       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
142       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
143       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
144       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
145 
146       k -= 8 * sizeof(int8_t);
147     }
148 
149 #if XNN_ARCH_ARM64
150     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
151     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
152     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
153     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
154     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
155     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
156     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
157     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
158     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
159     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
160     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
161     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
162     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
163     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
164     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
165     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
166     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
167     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
168 #else
169     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
170     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
171     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
172     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
173     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
174     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
175     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
176     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
177     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
178     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
179     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
180     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
181     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
182     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
183     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
184     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
185     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
186     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
187     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
188     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
189     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
190     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
191     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
192     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
193     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
194     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
195     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
196     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
197     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
198     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
199     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
200     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
201     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
202     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
203     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
204     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
205     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
206     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
207     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
208     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
209     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
210     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
211 #endif
212 
213     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
214     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
215     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
216     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
217     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
218     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
219     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
220 
221     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
222     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
223     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
224     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
225     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
226     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
227     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
228     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
229 
230     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
231     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
232     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
233     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
234     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
235     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
236 
237     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
238 #if XNN_ARCH_ARM64
239     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
240     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
241     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
242     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
243     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
244 #else
245     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
246     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
247     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
248 
249     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
250     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
251 #endif
252     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
253     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
254 
255     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
256     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
257 
258     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
259     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
260 
261     if (nc >= 8) {
262       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
263       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
264       vst1_s8(c2 + 0, vout2x01234567);
265 
266       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
267       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
268       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
269 
270       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
271       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
272       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
273 
274       nc -= 8;
275     } else {
276       if (nc & 4) {
277         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
278         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
279         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
280         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
281         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
282       }
283       if (nc & 2) {
284         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
285         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
286         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
287         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
288         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
289       }
290       if (nc & 1) {
291         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
292         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
293         vst1_lane_s8(c2, vout2x01234567, 0);
294       }
295 
296       nc = 0;
297     }
298   } while (nc != 0);
299 }
300