1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-padal-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 
17 
xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mull_padal_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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mull_padal_dup(
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 <= 2);
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, 2);
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 
49   do {
50     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52     int32x4_t vacc1x0123 = vacc0x0123;
53     int32x4_t vacc1x4567 = vacc0x4567;
54 
55     size_t k = kc;
56 
57 
58     while (k >= 8 * sizeof(int8_t)) {
59       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
60       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
61 
62       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 
71       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
72       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
73       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
74       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
75       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
76       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
77       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
78       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
79       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
80       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
81       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
82       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
83       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
84       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
85       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
86       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
87       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
88       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
89       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
90       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
91       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
92       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
93       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
94       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
95       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
96       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
97       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
98       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
99       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
100       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
101       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
102       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
103 
104       k -= 8 * sizeof(int8_t);
105     }
106 
107     if XNN_UNLIKELY(k != 0) {
108       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
109       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
110 
111       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 
114       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
115       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
116       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
117       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
118       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
119       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
120       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
121       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
122 
123       if (k > 2 * sizeof(int8_t)) {
124         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 
127         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
128         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
129         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
130         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
131         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
132         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
133         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
134         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
135 
136         if (k > 4 * sizeof(int8_t)) {
137           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 
140           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
141           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
142           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
143           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
144           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
145           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
146           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
147           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
148         }
149       }
150     }
151     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
152     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
153     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
154     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
155     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
156 
157     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
158     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
159     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
160     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
161     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
162     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
163 
164     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
165     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
166     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
167     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
168 
169     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
170 #if XNN_ARCH_ARM64
171     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
172     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
173 
174     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
175 #else
176     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
177     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
178 
179     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
180 #endif
181     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
182     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
183 
184     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
185 
186     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
187 
188     if (nc >= 8) {
189       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
190       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
191 
192       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
193       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
194 
195       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
196       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
197 
198       nc -= 8;
199     } else {
200       if (nc & 4) {
201         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
202         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
203         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
204       }
205       if (nc & 2) {
206         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
207         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
208         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
209       }
210       if (nc & 1) {
211         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
212         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
213       }
214 
215       nc = 0;
216     }
217   } while (nc != 0);
218 }
219