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_2x8c2s4__neon_mull(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_2x8c2s4__neon_mull(
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 <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 c1 = c0;
47 }
48
49 kc = round_up_po2(kc, 8 * sizeof(int8_t));
50 do {
51 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc1x0123 = vacc0x0123;
54 int32x4_t vacc1x4567 = vacc0x4567;
55
56 size_t p = ks;
57 do {
58 const int8_t* restrict a0 = a[0];
59 if XNN_UNPREDICTABLE(a0 != zero) {
60 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61 }
62 const int8_t* restrict a1 = a[1];
63 if XNN_UNPREDICTABLE(a1 != zero) {
64 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
65 }
66 a += 2;
67
68 size_t k = kc;
69 do {
70 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
71 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
72
73 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81
82 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
83 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
84 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
85 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
86 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
87 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
88 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
89 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
90 va0x0 = vext_s8(va0x0, va0x0, 2);
91 va1x0 = vext_s8(va1x0, va1x0, 2);
92 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
93 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
94 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
95 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
96 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
97 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
98 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
99 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
100 va0x0 = vext_s8(va0x0, va0x0, 2);
101 va1x0 = vext_s8(va1x0, va1x0, 2);
102 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
103 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
104 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
105 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
106 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
107 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
108 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
109 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
110 va0x0 = vext_s8(va0x0, va0x0, 2);
111 va1x0 = vext_s8(va1x0, va1x0, 2);
112 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
113 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
114 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
115 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
116 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
117 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
118 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
119 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
120
121 k -= 8 * sizeof(int8_t);
122 } while (k != 0);
123
124 p -= 2 * sizeof(void*);
125 } while (p != 0);
126
127 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
128 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
129 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
130
131 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
132 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
133 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
134 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
135
136 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
137 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
138 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
139 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
140
141 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
142 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
143 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
144 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
145
146 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
147 #if XNN_ARCH_ARM64
148 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
149 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
150
151 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
152 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
153
154 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
155 #else
156 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
157 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
158
159 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
160 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
161
162 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
163 #endif
164
165 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
166 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
167
168 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
169 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
170
171 if (nc >= 8) {
172 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
173 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
174
175 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
176 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
177
178 a = (const int8_t**restrict) ((uintptr_t) a - ks);
179
180 nc -= 8;
181 } else {
182 if (nc & 4) {
183 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
184 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
185 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
186 }
187 if (nc & 2) {
188 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
189 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
190 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
191 }
192 if (nc & 1) {
193 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
194 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
195 }
196
197 nc = 0;
198 }
199 } while (nc != 0);
200 }
201