1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal(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_1x16c16__neon_mlal(
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, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 do {
44 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
45 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
46 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
47 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
48 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
49 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
50 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55 int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60
61 // KC loop of 16
62 size_t k = kc;
63 while (k != 0) {
64 const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
65
66 const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
67 const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
68 const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
69 const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
70 const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
71 const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
72 const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
73 const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
74 const int8x16_t vb8 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
75 const int8x16_t vb9 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
76 const int8x16_t vb10 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
77 const int8x16_t vb11 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
78 const int8x16_t vb12 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
79 const int8x16_t vb13 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
80 const int8x16_t vb14 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
81 const int8x16_t vb15 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
82
83 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
84 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
85 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
86 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
87 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
88 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
89 int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
90 vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
91 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
92 int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
93 vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
94 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
95 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
96 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
97 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
98 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
99 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
100 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
101 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
102 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
103 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
104 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
105 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
106 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
107 int16x8_t vprod0x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va0));
108 vprod0x8 = vmlal_s8(vprod0x8, vget_high_s8(vb8), vget_high_s8(va0));
109 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
110 int16x8_t vprod0x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va0));
111 vprod0x9 = vmlal_s8(vprod0x9, vget_high_s8(vb9), vget_high_s8(va0));
112 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
113 int16x8_t vprod0x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va0));
114 vprod0x10 = vmlal_s8(vprod0x10, vget_high_s8(vb10), vget_high_s8(va0));
115 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
116 int16x8_t vprod0x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va0));
117 vprod0x11 = vmlal_s8(vprod0x11, vget_high_s8(vb11), vget_high_s8(va0));
118 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
119 int16x8_t vprod0x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va0));
120 vprod0x12 = vmlal_s8(vprod0x12, vget_high_s8(vb12), vget_high_s8(va0));
121 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
122 int16x8_t vprod0x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va0));
123 vprod0x13 = vmlal_s8(vprod0x13, vget_high_s8(vb13), vget_high_s8(va0));
124 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
125 int16x8_t vprod0x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va0));
126 vprod0x14 = vmlal_s8(vprod0x14, vget_high_s8(vb14), vget_high_s8(va0));
127 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
128 int16x8_t vprod0x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va0));
129 vprod0x15 = vmlal_s8(vprod0x15, vget_high_s8(vb15), vget_high_s8(va0));
130 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
131
132 k -= 16 * sizeof(int8_t);
133 }
134
135 #if XNN_ARCH_ARM64
136 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
137 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
138 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
139 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
140 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
141 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
142 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
143 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
144 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
145 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
146 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
147 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
148 #else
149 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
150 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
151 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
152 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
153 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
154 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
155 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
156 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
157 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
158 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
159 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
160 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
161 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
162 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
163 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
164 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
165 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
166 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
167 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
168 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
169 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
170 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
171 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
172 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
173 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
174 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
175 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
176 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
177 #endif
178
179 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
180 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
181 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
182
183 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
184 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
185 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
186 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
187
188 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
189 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
190 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
191 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
192
193 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
194 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
195 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
196 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
197
198 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
199 #if XNN_ARCH_ARM64
200 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
201 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
202 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
203 #else
204 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
205 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
206
207 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
208 #endif
209 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
210 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
211
212 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
213
214 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
215
216 if (nc >= 16) {
217 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
218
219 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
220
221 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
222
223 nc -= 16;
224 } else {
225 // Final case where not all of the 16 columns fit in the destination.
226 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
227 if (nc & 8) {
228 vst1_s8(c0, vout0x01234567); c0 += 8;
229 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
230 }
231 if (nc & 4) {
232 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
233 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
234 }
235 if (nc & 2) {
236 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
237 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
238 }
239 if (nc & 1) {
240 vst1_lane_s8(c0, vout0x01234567, 0);
241 }
242
243 nc = 0;
244 }
245 } while (nc != 0);
246 }
247