1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c16-neon-mlal.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_1x8c16__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_1x8c16__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 kc = round_up_po2(kc, 16 * sizeof(int8_t));
44 int8_t* c0 = c;
45
46 do {
47 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
48 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
49 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
50 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
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 a += 1;
63
64 // KC loop of 16 with up to 15 remainder
65 size_t k = kc;
66 while (k != 0) {
67 const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
68
69 const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
70 const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
71 const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
72 const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
73 const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
74 const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
75 const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
76 const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
77
78 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
79 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
80 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
81 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
82 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
83 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
84 int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
85 vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
86 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
87 int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
88 vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
89 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
90 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
91 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
92 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
93 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
94 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
95 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
96 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
97 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
98 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
99 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
100 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
101 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
102
103 k -= 16 * sizeof(int8_t);
104 }
105
106 p -= 1 * sizeof(void*);
107 } while (p != 0);
108
109 #if XNN_ARCH_ARM64
110 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
111 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
112 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
113 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
114 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
115 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
116 #else
117 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
118 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
119 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
120 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
121 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
122 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
123 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
124 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
125 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
126 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
127 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
128 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
129 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
130 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
131 #endif
132
133 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
134 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
135 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
136
137 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
138 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
139
140 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
141 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
142
143 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
144 vacc0x4567 = vrshlq_s32(vacc0x4567, 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 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
149 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
150 #else
151 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
152
153 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
154 #endif
155 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
156 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
157
158 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
159
160 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
161
162 if (nc >= 8) {
163 vst1_s8(c0 + 0, vout0x01234567);
164
165 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
166
167 a = (const int8_t**restrict) ((uintptr_t) a - ks);
168
169 nc -= 8;
170 } else {
171 if (nc & 4) {
172 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
173 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
174 }
175 if (nc & 2) {
176 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
177 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
178 }
179 if (nc & 1) {
180 vst1_lane_s8(c0, vout0x01234567, 0);
181 }
182
183 nc = 0;
184 }
185 } while (nc != 0);
186 }
187