1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-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_1x8c4s2__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_1x8c4s2__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 <= 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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
48 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
49 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
50 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
51
52 size_t p = ks;
53 do {
54 const int8_t* restrict a0 = a[0];
55 if XNN_UNPREDICTABLE(a0 != zero) {
56 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
57 }
58 a += 1;
59
60 size_t k = kc;
61 do {
62 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
63
64 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72
73 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
74 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
75 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
76 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
77 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
78 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
79 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
80 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
81 va0x0 = vext_s8(va0x0, va0x0, 4);
82 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
83 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
84 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
85 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
86 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
87 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
88 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
89 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
90
91 k -= 8 * sizeof(int8_t);
92 } while (k != 0);
93
94 p -= 1 * sizeof(void*);
95 } while (p != 0);
96
97 #if XNN_ARCH_ARM64
98 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
99 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
100 #else
101 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
102 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
103 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
104 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
105 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
106 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
107 #endif
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
116 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
117 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
118
119 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
120 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
121
122 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
123 #if XNN_ARCH_ARM64
124 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
125
126 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
127
128 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
129 #else
130 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
131
132 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
133
134 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
135 #endif
136
137 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
138 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
139
140 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
141 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
142
143 if (nc >= 8) {
144 vst1_s8(c0 + 0, vout0x01234567);
145
146 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
147
148 a = (const int8_t**restrict) ((uintptr_t) a - ks);
149
150 nc -= 8;
151 } else {
152 if (nc & 4) {
153 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
154 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
155 }
156 if (nc & 2) {
157 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
158 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
159 }
160 if (nc & 1) {
161 vst1_lane_s8(c0, vout0x01234567, 0);
162 }
163
164 nc = 0;
165 }
166 } while (nc != 0);
167 }
168