1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_2x8c4__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)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
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, 4 * sizeof(int8_t));
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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54 int32x4_t vacc1x01 = vacc0x01;
55 int32x4_t vacc1x23 = vacc0x23;
56 int32x4_t vacc1x45 = vacc0x45;
57 int32x4_t vacc1x67 = vacc0x67;
58
59 size_t k = kc;
60
61
62 while (k >= 8 * sizeof(int8_t)) {
63 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
64 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
65
66 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74
75 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
76 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
77
78 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
79 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
80 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
81 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
82 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
83 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
84 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
85 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
86 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
87 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
88 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
89 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
90 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
91 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
92 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
93 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
94 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
95 const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
96
97 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
98 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
99 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
101 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
102 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
103 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
104 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
105 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
106 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
107 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
108 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
109 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
110 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
111 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
112 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
113
114 k -= 8 * sizeof(int8_t);
115 }
116
117 if XNN_UNLIKELY(k != 0) {
118 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
119 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
120
121 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125
126 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
127 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
128 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
129 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
130 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
131 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
132 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
133 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
134 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
135 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
136 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
137 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
138 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
139 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
140 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
141 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
142 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
143 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
144 }
145
146 #if XNN_ARCH_ARM64
147 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
148 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
149 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
150 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
151 #else
152 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
153 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
154 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
155 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
156 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
157 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
158 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
159 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
160 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
161 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
162 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
163 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
164 #endif
165
166 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
167 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
168 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
169
170 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
171 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
172 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
173 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
174
175 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
176 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
177 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
178 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
179
180 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
181 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
182 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
183 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
184
185 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
186 #if XNN_ARCH_ARM64
187 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
188 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
189
190 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
191 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
192
193 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
194 #else
195 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
196 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
197
198 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
199 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
200
201 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
202 #endif
203
204 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
205 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
206
207 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
208 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
209
210 if (nc >= 8) {
211 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
212 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
213
214 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
215 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
216
217 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
218 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
219
220 nc -= 8;
221 } else {
222 // Final case where not all of the 8 columns fit in the destination.
223 if (nc & 4) {
224 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
225 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
226 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
227 }
228 if (nc & 2) {
229 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
230 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
231 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
232 }
233 if (nc & 1) {
234 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
235 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
236 }
237
238 nc = 0;
239 }
240 } while (nc != 0);
241 }
242