1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-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
17
xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__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)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mull_dup(
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 kc = round_up_po2(kc, 4 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42
43 do {
44 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
45 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
46 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
47 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48
49 size_t k = kc;
50
51
52 while (k >= 8 * sizeof(int8_t)) {
53 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
54
55 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63
64 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
65
66 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
67 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
68 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
69 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
70 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
71 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
72 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
73 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
74 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
75
76 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
77 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
78 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
79 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
80 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
81 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
82 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
83 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
84
85 k -= 8 * sizeof(int8_t);
86 }
87
88 if XNN_UNLIKELY(k != 0) {
89 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
90
91 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95
96 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
97 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
98 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
99 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
100 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
101 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
102 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
103 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
104 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
105 }
106
107 #if XNN_ARCH_ARM64
108 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
109 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
110 #else
111 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
112 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
113 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
114 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
115 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
116 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
117 #endif
118
119 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
120 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
121 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
122
123 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
124 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
125
126 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
127 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
128
129 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
130 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
131
132 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
133 #if XNN_ARCH_ARM64
134 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
135
136 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
137
138 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
139 #else
140 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
141
142 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
143
144 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
145 #endif
146
147 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
148 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
149
150 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
151 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
152
153 if (nc >= 8) {
154 vst1_s8(c0 + 0, vout0x01234567);
155
156 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
157
158 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
159
160 nc -= 8;
161 } else {
162 // Final case where not all of the 8 columns fit in the destination.
163 if (nc & 4) {
164 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
165 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
166 }
167 if (nc & 2) {
168 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
169 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
170 }
171 if (nc & 1) {
172 vst1_lane_s8(c0, vout0x01234567, 0);
173 }
174
175 nc = 0;
176 }
177 } while (nc != 0);
178 }
179