1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/scalar.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 <xnnpack/math.h>
13 #include <xnnpack/gemm.h>
14 #include <xnnpack/unaligned.h>
15
16
xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic(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)])17 void xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic(
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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 2);
31 assert(nc != 0);
32 assert(kc != 0);
33
34 const int8_t* a0 = a;
35 int8_t* c0 = c;
36 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
37 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
38 if XNN_UNPREDICTABLE(mr != 2) {
39 a1 = a0;
40 c1 = c0;
41 }
42
43 do {
44 int32_t vacc0x0 = unaligned_indexed_load_s32(w, 0);
45 int32_t vacc0x1 = unaligned_indexed_load_s32(w, 1);
46 int32_t vacc1x0 = vacc0x0;
47 int32_t vacc1x1 = vacc0x1;
48 w = (const void*) ((const int32_t*) w + 2);
49
50 size_t k = kc;
51 do {
52 const int32_t va0 = (int32_t) *a0++;
53 const int32_t va1 = (int32_t) *a1++;
54
55 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
56 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
57 w = (const void*) ((const int8_t*) w + 2);
58
59 vacc0x0 += va0 * vb0;
60 vacc0x1 += va0 * vb1;
61 vacc1x0 += va1 * vb0;
62 vacc1x1 += va1 * vb1;
63
64 k -= sizeof(int8_t);
65 } while (k != 0);
66
67 float vfpacc0x0 = (float) vacc0x0;
68 float vfpacc0x1 = (float) vacc0x1;
69 float vfpacc1x0 = (float) vacc1x0;
70 float vfpacc1x1 = (float) vacc1x1;
71
72 const float vscale0 = unaligned_indexed_load_f32(w, 0);
73 vfpacc0x0 *= vscale0;
74 vfpacc1x0 *= vscale0;
75 const float vscale1 = unaligned_indexed_load_f32(w, 1);
76 vfpacc0x1 *= vscale1;
77 vfpacc1x1 *= vscale1;
78 w = (const void*) ((const float*) w + 2);
79
80 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
81 vfpacc0x0 += vmagic_bias;
82 vfpacc0x1 += vmagic_bias;
83 vfpacc1x0 += vmagic_bias;
84 vfpacc1x1 += vmagic_bias;
85
86 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0);
87 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1);
88 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0);
89 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1);
90
91 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
92 vout0x0 = math_max_s32(vout0x0, vmagic_min);
93 vout0x1 = math_max_s32(vout0x1, vmagic_min);
94 vout1x0 = math_max_s32(vout1x0, vmagic_min);
95 vout1x1 = math_max_s32(vout1x1, vmagic_min);
96
97 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
98 vout0x0 = math_min_s32(vout0x0, vmagic_max);
99 vout0x1 = math_min_s32(vout0x1, vmagic_max);
100 vout1x0 = math_min_s32(vout1x0, vmagic_max);
101 vout1x1 = math_min_s32(vout1x1, vmagic_max);
102
103 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
104 vout0x0 -= vmagic_bias_less_zero_point;
105 vout0x1 -= vmagic_bias_less_zero_point;
106 vout1x0 -= vmagic_bias_less_zero_point;
107 vout1x1 -= vmagic_bias_less_zero_point;
108
109 if XNN_LIKELY(nc >= 2) {
110 c0[0] = (int8_t) vout0x0;
111 c0[1] = (int8_t) vout0x1;
112 c1[0] = (int8_t) vout1x0;
113 c1[1] = (int8_t) vout1x1;
114
115 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
116 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
117
118 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
119 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
120
121 nc -= 2;
122 } else {
123 if (nc & 1) {
124 c0[0] = (int8_t) vout0x0;
125 c1[0] = (int8_t) vout1x0;
126 }
127
128 nc = 0;
129 }
130 } while (nc != 0);
131 }
132