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_ld1r(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_ld1r(
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 int32x2_t va00 = vld1_dup_s32((const void*)a0);
64 const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
65 const int32x2_t va10 = vld1_dup_s32((const void*)a1);
66 const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
67
68 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76
77 const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
78 const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
79
80 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
81 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
82 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
83 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
84 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
85 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
86 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
87 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
88 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
89 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
90 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
91 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
92 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
93 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
94 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
95 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
96 const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
97 const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
98
99 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
100 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
101 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
102 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
103 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
104 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
105 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
106 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
107 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
108 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
109 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
110 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
111 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
112 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
113 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
114 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
115
116 k -= 8 * sizeof(int8_t);
117 }
118
119 if XNN_UNLIKELY(k != 0) {
120 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
121 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
122
123 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127
128 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
129 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
130 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
131 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
132 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
133 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
134 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
135 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
136 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
137 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
138 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
139 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
140 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
141 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
142 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
143 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
144 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
145 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
146 }
147
148 #if XNN_ARCH_ARM64
149 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
150 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
151 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
152 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
153 #else
154 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
155 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
156 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
157 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
158 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
159 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
160 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
161 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
162 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
163 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
164 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
165 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
166 #endif
167
168 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
169 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
170 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
171
172 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
173 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
174 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
175 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
176
177 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
178 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
179 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
180 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
181
182 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
183 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
184 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
185 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
186
187 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
188 #if XNN_ARCH_ARM64
189 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
190 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
191
192 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
193 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
194
195 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
196 #else
197 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
198 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
199
200 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
201 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
202
203 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
204 #endif
205
206 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
207 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
208
209 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
210 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
211
212 if (nc >= 8) {
213 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
214 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
215
216 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
217 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
218
219 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
220 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
221
222 nc -= 8;
223 } else {
224 // Final case where not all of the 8 columns fit in the destination.
225 if (nc & 4) {
226 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
227 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
228 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
229 }
230 if (nc & 2) {
231 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
232 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
233 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
234 }
235 if (nc & 1) {
236 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
237 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
238 }
239
240 nc = 0;
241 }
242 } while (nc != 0);
243 }
244