1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_1x16c2s4__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_1x16c2s4__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 const int8_t* a0 = a;
40 int8_t* c0 = c;
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 do {
44 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
45 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
46 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
47 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
48
49 size_t k = kc;
50 do {
51 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
52
53 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
54 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
55 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
56 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
57 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
58 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
60 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
61 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
62 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
63 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
64 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69
70 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
71 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
72 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
73 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
74 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
75 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
76 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
77 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
78 va0x0 = vext_s8(va0x0, va0x0, 2);
79 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
80 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
81 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
82 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
83 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
84 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
85 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
86 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
87 va0x0 = vext_s8(va0x0, va0x0, 2);
88 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
89 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
90 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
91 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
92 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
93 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
94 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
95 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
96 va0x0 = vext_s8(va0x0, va0x0, 2);
97 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
98 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
99 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
100 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
101 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
102 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
103 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
104 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
105
106 k -= 8 * sizeof(int8_t);
107 } while (k != 0);
108
109 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
110 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
111 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
112
113 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
114 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
115 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
116 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
117
118 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
119 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
120 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
121 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
122
123 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
124 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
125 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
126 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
127
128 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
129 #if XNN_ARCH_ARM64
130 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
131 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
132
133 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
134 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
135
136 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
137 #else
138 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
139 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
140
141 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
142 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
143
144 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
145 #endif
146
147 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
148 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
149
150 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
151 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
152
153 if (nc >= 16) {
154 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
155
156 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
157
158 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
159
160 nc -= 16;
161 } else {
162 // Final case where not all of the 16 columns fit in the destination.
163 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
164 if (nc & 8) {
165 vst1_s8(c0, vout0x01234567); c0 += 8;
166 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
167 }
168 if (nc & 4) {
169 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
170 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
171 }
172 if (nc & 2) {
173 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
174 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
175 }
176 if (nc & 1) {
177 vst1_lane_s8(c0, vout0x01234567, 0);
178 }
179
180 nc = 0;
181 }
182 } while (nc != 0);
183 }
184