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_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 const uint8_t* restrict a,
21 size_t a_stride,
22 const void* restrict w,
23 uint8_t* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const union xnn_qu8_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 uint8_t* a0 = a;
34 uint8_t* c0 = c;
35 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
36 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
37 if XNN_UNPREDICTABLE(mr != 2) {
38 a1 = a0;
39 c1 = c0;
40 }
41
42 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
43 do {
44 int32_t vacc0x0 = ((const int32_t*) w)[0];
45 int32_t vacc0x1 = ((const int32_t*) w)[1];
46 int32_t vacc0x2 = ((const int32_t*) w)[2];
47 int32_t vacc0x3 = ((const int32_t*) w)[3];
48 int32_t vacc1x0 = vacc0x0;
49 int32_t vacc1x1 = vacc0x1;
50 int32_t vacc1x2 = vacc0x2;
51 int32_t vacc1x3 = vacc0x3;
52 w = (const void*) ((const int32_t*) w + 4);
53
54 size_t k = kc;
55 do {
56 const int32_t va0 = (int32_t) (uint32_t) *a0++;
57 const int32_t va1 = (int32_t) (uint32_t) *a1++;
58
59 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
60 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
61 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
62 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
63 w = (const void*) ((const uint8_t*) w + 4);
64
65 vacc0x0 += va0 * vb0;
66 vacc0x1 += va0 * vb1;
67 vacc0x2 += va0 * vb2;
68 vacc0x3 += va0 * vb3;
69 vacc1x0 += va1 * vb0;
70 vacc1x1 += va1 * vb1;
71 vacc1x2 += va1 * vb2;
72 vacc1x3 += va1 * vb3;
73
74 k -= sizeof(uint8_t);
75 } while (k != 0);
76
77 float vfpacc0x0 = (float) vacc0x0;
78 float vfpacc0x1 = (float) vacc0x1;
79 float vfpacc0x2 = (float) vacc0x2;
80 float vfpacc0x3 = (float) vacc0x3;
81 float vfpacc1x0 = (float) vacc1x0;
82 float vfpacc1x1 = (float) vacc1x1;
83 float vfpacc1x2 = (float) vacc1x2;
84 float vfpacc1x3 = (float) vacc1x3;
85
86 const float vscale = params->fp32_scalar_fmagic.scale;
87 vfpacc0x0 *= vscale;
88 vfpacc0x1 *= vscale;
89 vfpacc0x2 *= vscale;
90 vfpacc0x3 *= vscale;
91 vfpacc1x0 *= vscale;
92 vfpacc1x1 *= vscale;
93 vfpacc1x2 *= vscale;
94 vfpacc1x3 *= vscale;
95
96 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
97 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
98 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
99 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
100 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
101 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
102 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
103 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
104 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
105
106 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
107 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
108 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
109 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
110 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
111 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
112 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
113 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
114 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
115
116 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
117 vfpacc0x0 += vmagic_bias;
118 vfpacc0x1 += vmagic_bias;
119 vfpacc0x2 += vmagic_bias;
120 vfpacc0x3 += vmagic_bias;
121 vfpacc1x0 += vmagic_bias;
122 vfpacc1x1 += vmagic_bias;
123 vfpacc1x2 += vmagic_bias;
124 vfpacc1x3 += vmagic_bias;
125
126 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
127 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
128 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
129 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
130 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
131 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
132 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
133 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
134 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
135
136 if XNN_LIKELY(nc >= 4) {
137 c0[0] = (uint8_t) vout0x0;
138 c0[1] = (uint8_t) vout0x1;
139 c0[2] = (uint8_t) vout0x2;
140 c0[3] = (uint8_t) vout0x3;
141 c1[0] = (uint8_t) vout1x0;
142 c1[1] = (uint8_t) vout1x1;
143 c1[2] = (uint8_t) vout1x2;
144 c1[3] = (uint8_t) vout1x3;
145
146 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
147 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
148
149 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
150 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
151
152 nc -= 4;
153 } else {
154 if (nc & 2) {
155 c0[0] = (uint8_t) vout0x0;
156 c0[1] = (uint8_t) vout0x1;
157 vout0x0 = vout0x2;
158 c0 += 2;
159 c1[0] = (uint8_t) vout1x0;
160 c1[1] = (uint8_t) vout1x1;
161 vout1x0 = vout1x2;
162 c1 += 2;
163 }
164 if (nc & 1) {
165 c0[0] = (uint8_t) vout0x0;
166 c1[0] = (uint8_t) vout1x0;
167 }
168
169 nc = 0;
170 }
171 } while (nc != 0);
172 }
173