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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qc8_gemm_minmax_fp32_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43
44 // Loop over groups of 8 columns.
45 do {
46 // Initialize accumulators with bias. 8 bias values are loaded from the
47 // weight matrix, at the start of the group of 8 columns.
48 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50
51 // Inner accumulation loop along the 8 columns.
52 size_t k = kc;
53 // 2x partial unrolled loop to load 8 bytes at a time.
54 while (k >= 8 * sizeof(int8_t)) {
55 // Load a 1x8 block of activations.
56 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
57
58 // Load a 8x8 block of weights.
59 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
60 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
61 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
62 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
63
64 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
65 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
66 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
67 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
68 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
69
70 k -= 8 * sizeof(int8_t);
71 }
72 // Handle up to 4 final positions of `k`
73 if XNN_UNLIKELY(k != 0) {
74 // Load a 1x4 block of activations.
75 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
76
77 // Load a 4x8 block of weights.
78 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
79 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
80
81 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
82 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
83 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
84 }
85
86 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
87 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
88
89 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
90 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
91 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
92 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
93
94 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
95 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
96
97 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
98 #if XNN_ARCH_ARM64
99 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
100
101 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
102 #else
103 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
104
105 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
106 #endif
107 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
108 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
109
110 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
111
112 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
113
114 if (nc >= 8) {
115 // Main case where there the 8 columns fit in the destination.
116 vst1_s8(c0 + 0, vout0x01234567);
117
118 // Advance to the next 8 columns.
119 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
120
121 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
122
123 nc -= 8;
124 } else {
125 // Final case where not all of the 8 columns fit in the destination.
126 if (nc & 4) {
127 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
128 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
129 }
130 if (nc & 2) {
131 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
132 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
133 }
134 if (nc & 1) {
135 vst1_lane_s8(c0, vout0x01234567, 0);
136 }
137
138 nc = 0;
139 }
140 } while (nc != 0);
141 }
142