1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-neon-mull-dup.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
xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_dup(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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_dup(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41
42 do {
43 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45
46 size_t k = kc;
47
48
49 while (k >= 8 * sizeof(int8_t)) {
50 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
51
52 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
53 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
54 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
55 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60
61 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
62
63 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
64 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
65 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
66 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
67 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
68
69 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
70 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
71 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
72 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
73 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
74
75 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
76 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
77 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
78 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
79 const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
80
81 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
82 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
83 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
84 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
85
86 k -= 8 * sizeof(int8_t);
87 }
88
89 if XNN_UNLIKELY(k != 0) {
90 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
91
92 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94
95 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
96 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
97 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
98 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
100
101 if (k > 2 * sizeof(int8_t)) {
102 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104
105 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
106 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
107 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
108 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
109 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
110
111 if (k > 4 * sizeof(int8_t)) {
112 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114
115 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
116 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
117 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
118 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
119 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
120 }
121 }
122 }
123
124 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
125 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
126 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
127
128 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
129 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
130
131 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
132 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
133
134 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
135 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
136
137 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
138 #if XNN_ARCH_ARM64
139 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
140
141 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
142
143 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
144 #else
145 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
146
147 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
148
149 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
150 #endif
151
152 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
153 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
154
155 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
156 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
157
158 if (nc >= 8) {
159 vst1_s8(c0 + 0, vout0x01234567);
160
161 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
162
163 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
164
165 nc -= 8;
166 } else {
167 // Final case where not all of the 8 columns fit in the destination.
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