1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c16-neon-mlal.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_1x16c16__neon_mlal(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_1x16c16__neon_mlal(
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, 16 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46 do {
47 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
48 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
49 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
50 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55 int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63
64 size_t p = ks;
65 do {
66 const int8_t* restrict a0 = a[0];
67 if XNN_UNPREDICTABLE(a0 != zero) {
68 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
69 }
70 a += 1;
71
72 // KC loop of 16 with up to 15 remainder
73 size_t k = kc;
74 while (k != 0) {
75 const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
76
77 const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
78 const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
79 const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
80 const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
81 const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
82 const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
83 const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
84 const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
85 const int8x16_t vb8 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
86 const int8x16_t vb9 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
87 const int8x16_t vb10 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
88 const int8x16_t vb11 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
89 const int8x16_t vb12 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
90 const int8x16_t vb13 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
91 const int8x16_t vb14 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
92 const int8x16_t vb15 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
93
94 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
95 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
96 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
97 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
98 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
99 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
100 int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
101 vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
102 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
103 int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
104 vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
105 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
106 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
107 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
108 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
109 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
110 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
111 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
112 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
113 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
114 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
115 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
116 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
117 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
118 int16x8_t vprod0x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va0));
119 vprod0x8 = vmlal_s8(vprod0x8, vget_high_s8(vb8), vget_high_s8(va0));
120 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
121 int16x8_t vprod0x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va0));
122 vprod0x9 = vmlal_s8(vprod0x9, vget_high_s8(vb9), vget_high_s8(va0));
123 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
124 int16x8_t vprod0x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va0));
125 vprod0x10 = vmlal_s8(vprod0x10, vget_high_s8(vb10), vget_high_s8(va0));
126 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
127 int16x8_t vprod0x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va0));
128 vprod0x11 = vmlal_s8(vprod0x11, vget_high_s8(vb11), vget_high_s8(va0));
129 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
130 int16x8_t vprod0x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va0));
131 vprod0x12 = vmlal_s8(vprod0x12, vget_high_s8(vb12), vget_high_s8(va0));
132 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
133 int16x8_t vprod0x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va0));
134 vprod0x13 = vmlal_s8(vprod0x13, vget_high_s8(vb13), vget_high_s8(va0));
135 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
136 int16x8_t vprod0x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va0));
137 vprod0x14 = vmlal_s8(vprod0x14, vget_high_s8(vb14), vget_high_s8(va0));
138 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
139 int16x8_t vprod0x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va0));
140 vprod0x15 = vmlal_s8(vprod0x15, vget_high_s8(vb15), vget_high_s8(va0));
141 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
142
143 k -= 16 * sizeof(int8_t);
144 }
145
146 p -= 1 * sizeof(void*);
147 } while (p != 0);
148
149 #if XNN_ARCH_ARM64
150 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
151 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
152 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
153 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
154 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
155 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
156 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
157 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
158 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
159 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
160 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
161 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
162 #else
163 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
164 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
165 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
166 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
167 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
168 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
169 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
170 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
171 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
172 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
173 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
174 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
175 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
176 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
177 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
178 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
179 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
180 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
181 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
182 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
183 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
184 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
185 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
186 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
187 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
188 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
189 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
190 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
191 #endif
192
193 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
194 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
195 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
196
197 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
198 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
199 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
200 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
201
202 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
203 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
204 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
205 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
206
207 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
208 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
209 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
210 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
211
212 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
213 #if XNN_ARCH_ARM64
214 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
215 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
216 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
217 #else
218 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
219 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
220
221 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
222 #endif
223 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
224 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
225
226 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
227
228 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
229
230 if (nc >= 16) {
231 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
232
233 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
234
235 a = (const int8_t**restrict) ((uintptr_t) a - ks);
236
237 nc -= 16;
238 } else {
239 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
240 if (nc & 8) {
241 vst1_s8(c0, vout0x01234567); c0 += 8;
242 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
243 }
244 if (nc & 4) {
245 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
246 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
247 }
248 if (nc & 2) {
249 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
250 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
251 }
252 if (nc & 1) {
253 vst1_lane_s8(c0, vout0x01234567, 0);
254 }
255
256 nc = 0;
257 }
258 } while (nc != 0);
259 }
260