1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-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
17
xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55 int32x4_t vacc1x01 = vacc0x01;
56 int32x4_t vacc1x23 = vacc0x23;
57 int32x4_t vacc1x45 = vacc0x45;
58 int32x4_t vacc1x67 = vacc0x67;
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const int8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73
74
75 while (k >= 8 * sizeof(int8_t)) {
76 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
77 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
78
79 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87
88 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
89 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
90
91 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
92 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
93 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
94 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
95 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
96 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
97 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
98 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
99 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
100 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
101 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
102 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
103 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
104 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
105 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
106 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
107 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
108 const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
109
110 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
111 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
112 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
113 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
114 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
115 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
116 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
117 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
118 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
119 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
120 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
121 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
122 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
123 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
124 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
125 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
126
127 k -= 8 * sizeof(int8_t);
128 }
129
130 if XNN_UNLIKELY(k != 0) {
131 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
132 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
133
134 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138
139 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
140 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
141 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
142 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
143 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
144 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
145 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
146 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
147 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
148 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
149 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
150 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
151 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
152 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
153 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
154 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
155 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
156 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
157 }
158 p -= 2 * sizeof(void*);
159 } while (p != 0);
160
161 #if XNN_ARCH_ARM64
162 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
163 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
164 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
165 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
166 #else
167 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
168 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
169 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
170 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
171 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
172 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
173 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
174 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
175 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
176 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
177 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
178 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
179 #endif
180
181 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
182 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
183 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
184
185 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
186 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
187 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
188 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
189
190 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
191 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
192 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
193 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
194
195 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
196 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
197 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
198 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
199
200 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
201 #if XNN_ARCH_ARM64
202 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
203 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
204
205 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
206 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
207
208 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
209 #else
210 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
211 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
212
213 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
214 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
215
216 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
217 #endif
218
219 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
220 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
221
222 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
223 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
224
225 if (nc >= 8) {
226 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
227 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
228
229 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
230 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
231
232 a = (const int8_t**restrict) ((uintptr_t) a - ks);
233
234 nc -= 8;
235 } else {
236 if (nc & 4) {
237 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
238 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
239 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
240 }
241 if (nc & 2) {
242 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
243 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
244 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
245 }
246 if (nc & 1) {
247 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
248 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
249 }
250
251 nc = 0;
252 }
253 } while (nc != 0);
254 }
255