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
15
xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic(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)])16 void xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 const int8_t* restrict a,
21 size_t a_stride,
22 const void* restrict w,
23 int8_t* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mr != 0);
29 assert(mr <= 2);
30 assert(nc != 0);
31 assert(kc != 0);
32
33 const int8_t* a0 = a;
34 int8_t* c0 = c;
35 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
36 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
37 if XNN_UNPREDICTABLE(mr != 2) {
38 a1 = a0;
39 c1 = c0;
40 }
41
42 do {
43 int32_t vacc0x0 = ((const int32_t*) w)[0];
44 int32_t vacc0x1 = ((const int32_t*) w)[1];
45 int32_t vacc0x2 = ((const int32_t*) w)[2];
46 int32_t vacc0x3 = ((const int32_t*) w)[3];
47 int32_t vacc1x0 = vacc0x0;
48 int32_t vacc1x1 = vacc0x1;
49 int32_t vacc1x2 = vacc0x2;
50 int32_t vacc1x3 = vacc0x3;
51 w = (const void*) ((const int32_t*) w + 4);
52
53 size_t k = kc;
54 do {
55 const int32_t va0 = (int32_t) *a0++;
56 const int32_t va1 = (int32_t) *a1++;
57
58 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
59 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
60 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
61 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
62 w = (const void*) ((const int8_t*) w + 4);
63
64 vacc0x0 += va0 * vb0;
65 vacc0x1 += va0 * vb1;
66 vacc0x2 += va0 * vb2;
67 vacc0x3 += va0 * vb3;
68 vacc1x0 += va1 * vb0;
69 vacc1x1 += va1 * vb1;
70 vacc1x2 += va1 * vb2;
71 vacc1x3 += va1 * vb3;
72
73 k -= sizeof(int8_t);
74 } while (k != 0);
75
76 float vfpacc0x0 = (float) vacc0x0;
77 float vfpacc0x1 = (float) vacc0x1;
78 float vfpacc0x2 = (float) vacc0x2;
79 float vfpacc0x3 = (float) vacc0x3;
80 float vfpacc1x0 = (float) vacc1x0;
81 float vfpacc1x1 = (float) vacc1x1;
82 float vfpacc1x2 = (float) vacc1x2;
83 float vfpacc1x3 = (float) vacc1x3;
84
85 const float vscale0 = ((const float*) w)[0];
86 vfpacc0x0 *= vscale0;
87 vfpacc1x0 *= vscale0;
88 const float vscale1 = ((const float*) w)[1];
89 vfpacc0x1 *= vscale1;
90 vfpacc1x1 *= vscale1;
91 const float vscale2 = ((const float*) w)[2];
92 vfpacc0x2 *= vscale2;
93 vfpacc1x2 *= vscale2;
94 const float vscale3 = ((const float*) w)[3];
95 vfpacc0x3 *= vscale3;
96 vfpacc1x3 *= vscale3;
97 w = (const void*) ((const float*) w + 4);
98
99 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
100 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
101 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
102 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
103 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
104 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
105 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
106 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
107 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
108
109 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
110 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
111 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
112 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
113 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
114 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
115 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
116 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
117 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
118
119 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
120 vfpacc0x0 += vmagic_bias;
121 vfpacc0x1 += vmagic_bias;
122 vfpacc0x2 += vmagic_bias;
123 vfpacc0x3 += vmagic_bias;
124 vfpacc1x0 += vmagic_bias;
125 vfpacc1x1 += vmagic_bias;
126 vfpacc1x2 += vmagic_bias;
127 vfpacc1x3 += vmagic_bias;
128
129 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
130 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
131 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
132 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
133 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
134 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
135 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
136 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
137 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
138
139 if XNN_LIKELY(nc >= 4) {
140 c0[0] = (int8_t) vout0x0;
141 c0[1] = (int8_t) vout0x1;
142 c0[2] = (int8_t) vout0x2;
143 c0[3] = (int8_t) vout0x3;
144 c1[0] = (int8_t) vout1x0;
145 c1[1] = (int8_t) vout1x1;
146 c1[2] = (int8_t) vout1x2;
147 c1[3] = (int8_t) vout1x3;
148
149 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
150 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
151
152 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
153 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
154
155 nc -= 4;
156 } else {
157 if (nc & 2) {
158 c0[0] = (int8_t) vout0x0;
159 c0[1] = (int8_t) vout0x1;
160 vout0x0 = vout0x2;
161 c0 += 2;
162 c1[0] = (int8_t) vout1x0;
163 c1[1] = (int8_t) vout1x1;
164 vout1x0 = vout1x2;
165 c1 += 2;
166 }
167 if (nc & 1) {
168 c0[0] = (int8_t) vout0x0;
169 c1[0] = (int8_t) vout1x0;
170 }
171
172 nc = 0;
173 }
174 } while (nc != 0);
175 }
176