1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neon-mull.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_1x16c8__neon_mull(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_1x16c8__neon_mull(
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, 8 * 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 size_t k = kc;
62
63 // Handle 8 bytes at a time using MUL.
64 while (k != 0) {
65 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
66
67 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
69 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
70 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
72 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
73 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
75 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
76 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
78 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
79 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
81 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
82 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
84 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
85 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
87 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
88 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
90 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
91 const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
93 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
94 const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
96 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
97 const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98 const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
99 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
100 const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
102 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
103 const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
105 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
106 const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107 const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
108 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
109 const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
111 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
112 const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
114 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
115
116 k -= 8 * sizeof(int8_t);
117 }
118
119 #if XNN_ARCH_ARM64
120 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
121 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
122 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
123 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
124 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
125 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
126 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
127 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
128
129 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
130 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
131 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
132 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
133 #else
134 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
135 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
136 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
137 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
138 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
139 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
140 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
141 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
142 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
143 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
144 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
145 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
146 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
147 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
148 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
149 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
150 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
151 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
152 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
153 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
154 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
155 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
156 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
157 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
158 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
159 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
160 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
161 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
162 #endif
163
164 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
165 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
166 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
167
168 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
169 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
170 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
171 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
172
173 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
174 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
175 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
176 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
177
178 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
179 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
180 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
181 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
182
183 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
184 #if XNN_ARCH_ARM64
185 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
186 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
187
188 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
189 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
190
191 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
192 #else
193 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
194 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
195
196 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
197 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
198
199 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
200 #endif
201
202 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
203 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
204
205 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
206 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
207
208 if (nc >= 16) {
209 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
210
211 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
212
213 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
214
215 nc -= 16;
216 } else {
217 // Final case where not all of the 16 columns fit in the destination.
218 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
219 if (nc & 8) {
220 vst1_s8(c0, vout0x01234567); c0 += 8;
221 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
222 }
223 if (nc & 4) {
224 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
225 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
226 }
227 if (nc & 2) {
228 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
229 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
230 }
231 if (nc & 1) {
232 vst1_lane_s8(c0, vout0x01234567, 0);
233 }
234
235 nc = 0;
236 }
237 } while (nc != 0);
238 }
239