1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_fp32_ukernel_2x4__scalar_fmagic(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4__scalar_fmagic(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 size_t ks,
21 const int8_t**restrict a,
22 const void*restrict w,
23 int8_t*restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 size_t a_offset,
27 const int8_t* zero,
28 const union xnn_qc8_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 assert(ks != 0);
35 assert(ks % (2 * sizeof(void*)) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 int8_t* c0 = c;
41 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr != 2) {
43 c1 = c0;
44 }
45
46 do {
47 int32_t vacc0x0 = ((const int32_t*) w)[0];
48 int32_t vacc0x1 = ((const int32_t*) w)[1];
49 int32_t vacc0x2 = ((const int32_t*) w)[2];
50 int32_t vacc0x3 = ((const int32_t*) w)[3];
51 int32_t vacc1x0 = vacc0x0;
52 int32_t vacc1x1 = vacc0x1;
53 int32_t vacc1x2 = vacc0x2;
54 int32_t vacc1x3 = vacc0x3;
55 w = (const void*) ((const int32_t*) w + 4);
56
57 size_t p = ks;
58 do {
59 const int8_t* restrict a0 = a[0];
60 assert(a0 != NULL);
61 if XNN_UNPREDICTABLE(a0 != zero) {
62 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
63 }
64 const int8_t* restrict a1 = a[1];
65 assert(a1 != NULL);
66 if XNN_UNPREDICTABLE(a1 != zero) {
67 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
68 }
69 a += 2;
70
71 size_t k = kc;
72 do {
73 const int32_t va0 = (int32_t) *a0++;
74 const int32_t va1 = (int32_t) *a1++;
75
76 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
77 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
78 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
79 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
80 w = (const void*) ((const int8_t*) w + 4);
81
82 vacc0x0 += va0 * vb0;
83 vacc0x1 += va0 * vb1;
84 vacc0x2 += va0 * vb2;
85 vacc0x3 += va0 * vb3;
86 vacc1x0 += va1 * vb0;
87 vacc1x1 += va1 * vb1;
88 vacc1x2 += va1 * vb2;
89 vacc1x3 += va1 * vb3;
90
91 k -= sizeof(int8_t);
92 } while (k != 0);
93 p -= 2 * sizeof(void*);
94 } while (p != 0);
95
96 float vfpacc0x0 = (float) vacc0x0;
97 float vfpacc0x1 = (float) vacc0x1;
98 float vfpacc0x2 = (float) vacc0x2;
99 float vfpacc0x3 = (float) vacc0x3;
100 float vfpacc1x0 = (float) vacc1x0;
101 float vfpacc1x1 = (float) vacc1x1;
102 float vfpacc1x2 = (float) vacc1x2;
103 float vfpacc1x3 = (float) vacc1x3;
104
105 const float vscale0 = ((const float*) w)[0];
106 vfpacc0x0 *= vscale0;
107 vfpacc1x0 *= vscale0;
108 const float vscale1 = ((const float*) w)[1];
109 vfpacc0x1 *= vscale1;
110 vfpacc1x1 *= vscale1;
111 const float vscale2 = ((const float*) w)[2];
112 vfpacc0x2 *= vscale2;
113 vfpacc1x2 *= vscale2;
114 const float vscale3 = ((const float*) w)[3];
115 vfpacc0x3 *= vscale3;
116 vfpacc1x3 *= vscale3;
117 w = (const void*) ((const float*) w + 4);
118
119 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
120 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
121 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
122 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
123 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
124 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
125 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
126 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
127 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
128
129 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
130 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
131 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
132 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
133 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
134 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
135 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
136 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
137 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
138
139 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
140 vfpacc0x0 += vmagic_bias;
141 vfpacc0x1 += vmagic_bias;
142 vfpacc0x2 += vmagic_bias;
143 vfpacc0x3 += vmagic_bias;
144 vfpacc1x0 += vmagic_bias;
145 vfpacc1x1 += vmagic_bias;
146 vfpacc1x2 += vmagic_bias;
147 vfpacc1x3 += vmagic_bias;
148
149 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
150 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
151 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
152 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
153 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
154 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
155 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
156 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
157 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
158
159 if XNN_LIKELY(nc >= 4) {
160 c1[0] = (int8_t) vout1x0;
161 c1[1] = (int8_t) vout1x1;
162 c1[2] = (int8_t) vout1x2;
163 c1[3] = (int8_t) vout1x3;
164 c0[0] = (int8_t) vout0x0;
165 c0[1] = (int8_t) vout0x1;
166 c0[2] = (int8_t) vout0x2;
167 c0[3] = (int8_t) vout0x3;
168
169 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
170 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
171
172 a = (const int8_t**restrict) ((uintptr_t) a - ks);
173 nc -= 4;
174 } else {
175 if (nc & 2) {
176 c1[0] = (int8_t) vout1x0;
177 c1[1] = (int8_t) vout1x1;
178 vout1x0 = vout1x2;
179 c1 += 2;
180 c0[0] = (int8_t) vout0x0;
181 c0[1] = (int8_t) vout0x1;
182 vout0x0 = vout0x2;
183 c0 += 2;
184 }
185 if (nc & 1) {
186 c1[0] = (int8_t) vout1x0;
187 c0[0] = (int8_t) vout0x0;
188 }
189
190 nc = 0;
191 }
192 } while (nc != 0);
193 }
194