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 <fp16.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic(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)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34
35 const uint8_t* a0 = a;
36 uint8_t* c0 = c;
37 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
38 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
39 if XNN_UNPREDICTABLE(mr != 2) {
40 a1 = a0;
41 c1 = c0;
42 }
43
44 const int32_t vb_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
45 do {
46 int32_t vacc0x0 = ((const int32_t*) w)[0];
47 int32_t vacc0x1 = ((const int32_t*) w)[1];
48 int32_t vacc0x2 = ((const int32_t*) w)[2];
49 int32_t vacc0x3 = ((const int32_t*) w)[3];
50 int32_t vacc1x0 = vacc0x0;
51 int32_t vacc1x1 = vacc0x1;
52 int32_t vacc1x2 = vacc0x2;
53 int32_t vacc1x3 = vacc0x3;
54 w = (const void*) ((const int32_t*) w + 4);
55
56 size_t k = kc;
57 do {
58 const int32_t va0 = (int32_t) (uint32_t) *a0++;
59 const int32_t va1 = (int32_t) (uint32_t) *a1++;
60
61 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
62 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
63 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
64 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
65 w = (const void*) ((const uint8_t*) w + 4);
66
67 vacc0x0 += va0 * vb0;
68 vacc0x1 += va0 * vb1;
69 vacc0x2 += va0 * vb2;
70 vacc0x3 += va0 * vb3;
71 vacc1x0 += va1 * vb0;
72 vacc1x1 += va1 * vb1;
73 vacc1x2 += va1 * vb2;
74 vacc1x3 += va1 * vb3;
75
76 k -= sizeof(uint8_t);
77 } while (k != 0);
78
79 float vfpacc0x0 = (float) vacc0x0;
80 float vfpacc0x1 = (float) vacc0x1;
81 float vfpacc0x2 = (float) vacc0x2;
82 float vfpacc0x3 = (float) vacc0x3;
83 float vfpacc1x0 = (float) vacc1x0;
84 float vfpacc1x1 = (float) vacc1x1;
85 float vfpacc1x2 = (float) vacc1x2;
86 float vfpacc1x3 = (float) vacc1x3;
87
88 const float vscale = params->fp32_scalar_imagic.scale;
89 vfpacc0x0 *= vscale;
90 vfpacc0x1 *= vscale;
91 vfpacc0x2 *= vscale;
92 vfpacc0x3 *= vscale;
93 vfpacc1x0 *= vscale;
94 vfpacc1x1 *= vscale;
95 vfpacc1x2 *= vscale;
96 vfpacc1x3 *= vscale;
97
98 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
99 vfpacc0x0 += vmagic_bias;
100 vfpacc0x1 += vmagic_bias;
101 vfpacc0x2 += vmagic_bias;
102 vfpacc0x3 += vmagic_bias;
103 vfpacc1x0 += vmagic_bias;
104 vfpacc1x1 += vmagic_bias;
105 vfpacc1x2 += vmagic_bias;
106 vfpacc1x3 += vmagic_bias;
107
108 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0);
109 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1);
110 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2);
111 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3);
112 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0);
113 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1);
114 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2);
115 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3);
116
117 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
118 vout0x0 = math_max_s32(vout0x0, vmagic_min);
119 vout0x1 = math_max_s32(vout0x1, vmagic_min);
120 vout0x2 = math_max_s32(vout0x2, vmagic_min);
121 vout0x3 = math_max_s32(vout0x3, vmagic_min);
122 vout1x0 = math_max_s32(vout1x0, vmagic_min);
123 vout1x1 = math_max_s32(vout1x1, vmagic_min);
124 vout1x2 = math_max_s32(vout1x2, vmagic_min);
125 vout1x3 = math_max_s32(vout1x3, vmagic_min);
126
127 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
128 vout0x0 = math_min_s32(vout0x0, vmagic_max);
129 vout0x1 = math_min_s32(vout0x1, vmagic_max);
130 vout0x2 = math_min_s32(vout0x2, vmagic_max);
131 vout0x3 = math_min_s32(vout0x3, vmagic_max);
132 vout1x0 = math_min_s32(vout1x0, vmagic_max);
133 vout1x1 = math_min_s32(vout1x1, vmagic_max);
134 vout1x2 = math_min_s32(vout1x2, vmagic_max);
135 vout1x3 = math_min_s32(vout1x3, vmagic_max);
136
137 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
138 vout0x0 -= vmagic_bias_less_zero_point;
139 vout0x1 -= vmagic_bias_less_zero_point;
140 vout0x2 -= vmagic_bias_less_zero_point;
141 vout0x3 -= vmagic_bias_less_zero_point;
142 vout1x0 -= vmagic_bias_less_zero_point;
143 vout1x1 -= vmagic_bias_less_zero_point;
144 vout1x2 -= vmagic_bias_less_zero_point;
145 vout1x3 -= vmagic_bias_less_zero_point;
146
147 if XNN_LIKELY(nc >= 4) {
148 c0[0] = (uint8_t) vout0x0;
149 c0[1] = (uint8_t) vout0x1;
150 c0[2] = (uint8_t) vout0x2;
151 c0[3] = (uint8_t) vout0x3;
152 c1[0] = (uint8_t) vout1x0;
153 c1[1] = (uint8_t) vout1x1;
154 c1[2] = (uint8_t) vout1x2;
155 c1[3] = (uint8_t) vout1x3;
156
157 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
158 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
159
160 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
161 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
162
163 nc -= 4;
164 } else {
165 if (nc & 2) {
166 c0[0] = (uint8_t) vout0x0;
167 c0[1] = (uint8_t) vout0x1;
168 vout0x0 = vout0x2;
169 c0 += 2;
170 c1[0] = (uint8_t) vout1x0;
171 c1[1] = (uint8_t) vout1x1;
172 vout1x0 = vout1x2;
173 c1 += 2;
174 }
175 if (nc & 1) {
176 c0[0] = (uint8_t) vout0x0;
177 c1[0] = (uint8_t) vout1x0;
178 }
179
180 nc = 0;
181 }
182 } while (nc != 0);
183 }
184