1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-shuffle.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_1x8c2s4__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_1x8c2s4__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 int8_t* c0 = c;
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 do {
47 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
48 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
49
50 size_t p = ks;
51 do {
52 const int8_t* restrict a0 = a[0];
53 if XNN_UNPREDICTABLE(a0 != zero) {
54 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55 }
56 a += 1;
57
58 size_t k = kc;
59 while (k >= 16 * sizeof(int8_t)) {
60 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
61 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
62
63 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
64 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71
72 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
73 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
74 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
75 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
76 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
77 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
78 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
79 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
80 va0x0 = vext_s8(va0x0, va0x0, 2);
81 va0x1 = vext_s8(va0x1, va0x1, 2);
82 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
83 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
84 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
85 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
86 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
87 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
88 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
89 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
90 va0x0 = vext_s8(va0x0, va0x0, 2);
91 va0x1 = vext_s8(va0x1, va0x1, 2);
92 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
93 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
94 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
95 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
96 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
97 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
98 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
100 va0x0 = vext_s8(va0x0, va0x0, 2);
101 va0x1 = vext_s8(va0x1, va0x1, 2);
102 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
103 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
104 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
105 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
106 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
107 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
108 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
109 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
110
111 k -= 16 * sizeof(int8_t);
112 }
113 if (k != 0) {
114 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
115
116 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
118 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
119 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
120 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
121 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
122 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
123 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
124
125 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
126 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
127 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
128 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
129 va0x0 = vext_s8(va0x0, va0x0, 2);
130 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
131 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
132 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
133 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
134 va0x0 = vext_s8(va0x0, va0x0, 2);
135 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
136 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
137 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
138 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
139 va0x0 = vext_s8(va0x0, va0x0, 2);
140 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
141 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
142 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
144
145 }
146
147 p -= 1 * sizeof(void*);
148 } while (p != 0);
149
150 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
151 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
152 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
153
154 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
155 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
156
157 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
158 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
159
160 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
161 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
162
163 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
164 #if XNN_ARCH_ARM64
165 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
166
167 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
168
169 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
170 #else
171 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
172
173 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
174
175 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
176 #endif
177
178 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
179 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
180
181 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
182 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
183
184 if (nc >= 8) {
185 vst1_s8(c0 + 0, vout0x01234567);
186
187 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
188
189 a = (const int8_t**restrict) ((uintptr_t) a - ks);
190
191 nc -= 8;
192 } else {
193 if (nc & 4) {
194 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
195 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
196 }
197 if (nc & 2) {
198 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
199 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
200 }
201 if (nc & 1) {
202 vst1_lane_s8(c0, vout0x01234567, 0);
203 }
204
205 nc = 0;
206 }
207 } while (nc != 0);
208 }
209