1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-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
xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41
42 do {
43 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47
48 size_t k = kc;
49
50
51 while (k >= 8 * sizeof(int8_t)) {
52 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
53 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
54 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
55 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
56
57 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73
74 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
75
76 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
77 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
78 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
79 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
80 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
81 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
82 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
83 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
84 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
85
86 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
87 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
88 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
89 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
90 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
91 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
92 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
93 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
94 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
95
96 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
97 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
98 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
100 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
101 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
102 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
103 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
104 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
105
106 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
107 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
108 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
109 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
110 const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
111 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
112 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
113 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
114
115 k -= 8 * sizeof(int8_t);
116 }
117
118 if XNN_UNLIKELY(k != 0) {
119 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
120
121 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125
126 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
127 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
128 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
129 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
130 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
131 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
132 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
133 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
134 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
135
136 if (k > 2 * sizeof(int8_t)) {
137 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141
142 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
143 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
144 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
145 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
146 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
147 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
148 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
149 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
150 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
151
152 if (k > 4 * sizeof(int8_t)) {
153 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157
158 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
159 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
160 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
161 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
162 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
163 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
164 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
165 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
166 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
167 }
168 }
169 }
170
171 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
172 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
173 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
174
175 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
176 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
177 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
178 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
179
180 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
181 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
182 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
183 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
184
185 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
186 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
187 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
188 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
189
190 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
191 #if XNN_ARCH_ARM64
192 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
193 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
194
195 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
196 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
197
198 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
199 #else
200 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
201 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
202
203 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
204 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
205
206 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
207 #endif
208
209 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
210 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
211
212 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
213 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
214
215 if (nc >= 16) {
216 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
217
218 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
219
220 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
221
222 nc -= 16;
223 } else {
224 // Final case where not all of the 16 columns fit in the destination.
225 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
226 if (nc & 8) {
227 vst1_s8(c0, vout0x01234567); c0 += 8;
228 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
229 }
230 if (nc & 4) {
231 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
232 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
233 }
234 if (nc & 2) {
235 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
236 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
237 }
238 if (nc & 1) {
239 vst1_lane_s8(c0, vout0x01234567, 0);
240 }
241
242 nc = 0;
243 }
244 } while (nc != 0);
245 }
246