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_1x8c2s4__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_1x8c2s4__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
47 size_t k = kc;
48 do {
49 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
50
51 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
52 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
53 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
54 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
55 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
56 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
57 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
58 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59
60 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
61 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
62 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
63 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
64 va0x0 = vext_s8(va0x0, va0x0, 2);
65 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
66 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
67 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
68 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
69 va0x0 = vext_s8(va0x0, va0x0, 2);
70 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
71 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
72 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
73 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
74 va0x0 = vext_s8(va0x0, va0x0, 2);
75 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
76 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
77 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
78 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
79
80 k -= 8 * sizeof(int8_t);
81 } while (k != 0);
82
83 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
84 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
85 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
86
87 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
88 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
89
90 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
91 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
92
93 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
94 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
95
96 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
97 #if XNN_ARCH_ARM64
98 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
99
100 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
101
102 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
103 #else
104 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
105
106 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
107
108 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
109 #endif
110
111 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
112 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
113
114 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
115 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
116
117 if (nc >= 8) {
118 vst1_s8(c0 + 0, vout0x01234567);
119
120 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
121
122 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
123
124 nc -= 8;
125 } else {
126 // Final case where not all of the 8 columns fit in the destination.
127 if (nc & 4) {
128 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
129 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
130 }
131 if (nc & 2) {
132 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
133 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
134 }
135 if (nc & 1) {
136 vst1_lane_s8(c0, vout0x01234567, 0);
137 }
138
139 nc = 0;
140 }
141 } while (nc != 0);
142 }
143