1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-neondot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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__neondot(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__neondot(
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 // Loop over groups of 8 columns.
44 do {
45 // Initialize accumulators with bias. 8 bias values are loaded from the
46 // weight matrix, at the start of the group of 8 columns.
47 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
48 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49
50 // Inner accumulation loop along the 8 columns.
51 size_t k = kc;
52 // 2x partial unrolled loop to load 8 bytes at a time.
53 while (k >= 8 * sizeof(int8_t)) {
54 // Load a 1x8 block of activations.
55 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
56
57 // Load a 8x8 block of weights.
58 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
59 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
60 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
61 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
62
63 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
64 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
65 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
66 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
67 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
68
69 k -= 8 * sizeof(int8_t);
70 }
71 // Handle up to 4 final positions of `k`
72 if XNN_UNLIKELY(k != 0) {
73 // Load a 1x4 block of activations.
74 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
75
76 // Load a 4x8 block of weights.
77 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
78 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
79
80 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
81 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
82 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
83 }
84
85 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
86 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
87 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
88
89 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
90 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
91
92 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
93 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
94
95 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
96 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
97
98 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
99 #if XNN_ARCH_ARM64
100 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
101
102 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
103 #else
104 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
105
106 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
107 #endif
108 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
109 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
110
111 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
112
113 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
114
115 if (nc >= 8) {
116 // Main case where there the 8 columns fit in the destination.
117 vst1_s8(c0 + 0, vout0x01234567);
118
119 // Advance to the next 8 columns.
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