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_1x16c4__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_1x16c4__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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
46 do {
47 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55
56 size_t p = ks;
57 do {
58 const int8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 a += 1;
63
64 size_t k = kc;
65
66
67 while (k >= 8 * sizeof(int8_t)) {
68 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
69
70 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86
87 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
88
89 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
90 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
91 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
92 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
93 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
94 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
95 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
96 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
97 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
98 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
99 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
100 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
101 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
102 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
103 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
104 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
105 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
106
107 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
108 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
109 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
110 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
111 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
112 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
113 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
114 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
115 const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
116 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
117 const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
118 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
119 const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
120 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
121 const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
122 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
123
124 k -= 8 * sizeof(int8_t);
125 }
126
127 if XNN_UNLIKELY(k != 0) {
128 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
129
130 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137 const int8x8_t vbEFc0 = 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 int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
149 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
150 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
151 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
152 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
153 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
154 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
155 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
156 }
157 p -= 1 * sizeof(void*);
158 } while (p != 0);
159
160 #if XNN_ARCH_ARM64
161 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
162 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
163 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
164 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
165 #else
166 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
167 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
168 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
169 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
170 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
171 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
172 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
173 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
174 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
175 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
176 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
177 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
178 #endif
179
180 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
181 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
182 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
183
184 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
185 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
186 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
187 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
188
189 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
190 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
191 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
192 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
193
194 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
195 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
196 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
197 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
198
199 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
200 #if XNN_ARCH_ARM64
201 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
202 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
203
204 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
205 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
206
207 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
208 #else
209 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
210 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
211
212 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
213 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
214
215 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
216 #endif
217
218 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
219 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
220
221 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
222 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
223
224 if (nc >= 16) {
225 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
226
227 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
228
229 a = (const int8_t**restrict) ((uintptr_t) a - ks);
230
231 nc -= 16;
232 } else {
233 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
234 if (nc & 8) {
235 vst1_s8(c0, vout0x01234567); c0 += 8;
236 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
237 }
238 if (nc & 4) {
239 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
240 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
241 }
242 if (nc & 2) {
243 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
244 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
245 }
246 if (nc & 1) {
247 vst1_lane_s8(c0, vout0x01234567, 0);
248 }
249
250 nc = 0;
251 }
252 } while (nc != 0);
253 }
254