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_1x16c4__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_1x16c4__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 <= 1);
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
43 do {
44 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
45 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
46 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
47 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52
53 size_t k = kc;
54
55
56 while (k >= 8 * sizeof(int8_t)) {
57 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
58
59 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75
76 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
77
78 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
79 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
80 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
81 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
82 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
83 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
84 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
85 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
86 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
87 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
88 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
89 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
90 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
91 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
92 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
93 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
94 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
95
96 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
97 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
98 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
99 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
100 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
101 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
102 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
103 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
104 const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
105 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
106 const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
107 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
108 const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
109 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
110 const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
111 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
112
113 k -= 8 * sizeof(int8_t);
114 }
115
116 if XNN_UNLIKELY(k != 0) {
117 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
118
119 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int8x8_t vbEFc0 = 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 int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
138 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
139 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
140 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
141 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
142 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
143 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
144 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
145 }
146
147 #if XNN_ARCH_ARM64
148 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
149 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
150 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
151 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
152 #else
153 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
154 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
155 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
156 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
157 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
158 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
159 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
160 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
161 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
162 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
163 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
164 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
165 #endif
166
167 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
168 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
169 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
170
171 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
172 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
173 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
174 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
175
176 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
177 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
178 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
179 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
180
181 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
182 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
183 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
184 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
185
186 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
187 #if XNN_ARCH_ARM64
188 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
189 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
190
191 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
192 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
193
194 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
195 #else
196 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
197 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
198
199 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
200 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
201
202 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
203 #endif
204
205 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
206 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
207
208 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
209 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
210
211 if (nc >= 16) {
212 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
213
214 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
215
216 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
217
218 nc -= 16;
219 } else {
220 // Final case where not all of the 16 columns fit in the destination.
221 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
222 if (nc & 8) {
223 vst1_s8(c0, vout0x01234567); c0 += 8;
224 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
225 }
226 if (nc & 4) {
227 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
228 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
229 }
230 if (nc & 2) {
231 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
232 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
233 }
234 if (nc & 1) {
235 vst1_lane_s8(c0, vout0x01234567, 0);
236 }
237
238 nc = 0;
239 }
240 } while (nc != 0);
241 }
242