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_3x4__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_3x4__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 <= 3);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
42 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
43 if XNN_UNPREDICTABLE(mr <= 2) {
44 a2 = a1;
45 c2 = c1;
46 }
47
48 do {
49 int32_t vacc0x0 = ((const int32_t*) w)[0];
50 int32_t vacc0x1 = ((const int32_t*) w)[1];
51 int32_t vacc0x2 = ((const int32_t*) w)[2];
52 int32_t vacc0x3 = ((const int32_t*) w)[3];
53 int32_t vacc1x0 = vacc0x0;
54 int32_t vacc1x1 = vacc0x1;
55 int32_t vacc1x2 = vacc0x2;
56 int32_t vacc1x3 = vacc0x3;
57 int32_t vacc2x0 = vacc0x0;
58 int32_t vacc2x1 = vacc0x1;
59 int32_t vacc2x2 = vacc0x2;
60 int32_t vacc2x3 = vacc0x3;
61 w = (const void*) ((const int32_t*) w + 4);
62
63 size_t k = kc;
64 do {
65 const int32_t va0 = (int32_t) *a0++;
66 const int32_t va1 = (int32_t) *a1++;
67 const int32_t va2 = (int32_t) *a2++;
68
69 const int32_t vb0 = (int32_t) ((const int8_t*) w)[0];
70 const int32_t vb1 = (int32_t) ((const int8_t*) w)[1];
71 const int32_t vb2 = (int32_t) ((const int8_t*) w)[2];
72 const int32_t vb3 = (int32_t) ((const int8_t*) w)[3];
73 w = (const void*) ((const int8_t*) w + 4);
74
75 vacc0x0 += va0 * vb0;
76 vacc0x1 += va0 * vb1;
77 vacc0x2 += va0 * vb2;
78 vacc0x3 += va0 * vb3;
79 vacc1x0 += va1 * vb0;
80 vacc1x1 += va1 * vb1;
81 vacc1x2 += va1 * vb2;
82 vacc1x3 += va1 * vb3;
83 vacc2x0 += va2 * vb0;
84 vacc2x1 += va2 * vb1;
85 vacc2x2 += va2 * vb2;
86 vacc2x3 += va2 * vb3;
87
88 k -= sizeof(int8_t);
89 } while (k != 0);
90
91 float vfpacc0x0 = (float) vacc0x0;
92 float vfpacc0x1 = (float) vacc0x1;
93 float vfpacc0x2 = (float) vacc0x2;
94 float vfpacc0x3 = (float) vacc0x3;
95 float vfpacc1x0 = (float) vacc1x0;
96 float vfpacc1x1 = (float) vacc1x1;
97 float vfpacc1x2 = (float) vacc1x2;
98 float vfpacc1x3 = (float) vacc1x3;
99 float vfpacc2x0 = (float) vacc2x0;
100 float vfpacc2x1 = (float) vacc2x1;
101 float vfpacc2x2 = (float) vacc2x2;
102 float vfpacc2x3 = (float) vacc2x3;
103
104 const float vscale0 = ((const float*) w)[0];
105 vfpacc0x0 *= vscale0;
106 vfpacc1x0 *= vscale0;
107 vfpacc2x0 *= vscale0;
108 const float vscale1 = ((const float*) w)[1];
109 vfpacc0x1 *= vscale1;
110 vfpacc1x1 *= vscale1;
111 vfpacc2x1 *= vscale1;
112 const float vscale2 = ((const float*) w)[2];
113 vfpacc0x2 *= vscale2;
114 vfpacc1x2 *= vscale2;
115 vfpacc2x2 *= vscale2;
116 const float vscale3 = ((const float*) w)[3];
117 vfpacc0x3 *= vscale3;
118 vfpacc1x3 *= vscale3;
119 vfpacc2x3 *= vscale3;
120 w = (const void*) ((const float*) w + 4);
121
122 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
123 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
124 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
125 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
126 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
127 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
128 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
129 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
130 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
131 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
132 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
133 vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
134 vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
135
136 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
137 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
138 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
139 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
140 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
141 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
142 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
143 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
144 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
145 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
146 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
147 vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
148 vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
149
150 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
151 vfpacc0x0 += vmagic_bias;
152 vfpacc0x1 += vmagic_bias;
153 vfpacc0x2 += vmagic_bias;
154 vfpacc0x3 += vmagic_bias;
155 vfpacc1x0 += vmagic_bias;
156 vfpacc1x1 += vmagic_bias;
157 vfpacc1x2 += vmagic_bias;
158 vfpacc1x3 += vmagic_bias;
159 vfpacc2x0 += vmagic_bias;
160 vfpacc2x1 += vmagic_bias;
161 vfpacc2x2 += vmagic_bias;
162 vfpacc2x3 += vmagic_bias;
163
164 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
165 int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0) - vmagic_bias_less_output_zero_point;
166 int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1) - vmagic_bias_less_output_zero_point;
167 int32_t vout0x2 = (int32_t) float_as_uint32(vfpacc0x2) - vmagic_bias_less_output_zero_point;
168 int32_t vout0x3 = (int32_t) float_as_uint32(vfpacc0x3) - vmagic_bias_less_output_zero_point;
169 int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0) - vmagic_bias_less_output_zero_point;
170 int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1) - vmagic_bias_less_output_zero_point;
171 int32_t vout1x2 = (int32_t) float_as_uint32(vfpacc1x2) - vmagic_bias_less_output_zero_point;
172 int32_t vout1x3 = (int32_t) float_as_uint32(vfpacc1x3) - vmagic_bias_less_output_zero_point;
173 int32_t vout2x0 = (int32_t) float_as_uint32(vfpacc2x0) - vmagic_bias_less_output_zero_point;
174 int32_t vout2x1 = (int32_t) float_as_uint32(vfpacc2x1) - vmagic_bias_less_output_zero_point;
175 int32_t vout2x2 = (int32_t) float_as_uint32(vfpacc2x2) - vmagic_bias_less_output_zero_point;
176 int32_t vout2x3 = (int32_t) float_as_uint32(vfpacc2x3) - vmagic_bias_less_output_zero_point;
177
178 if XNN_LIKELY(nc >= 4) {
179 c0[0] = (int8_t) vout0x0;
180 c0[1] = (int8_t) vout0x1;
181 c0[2] = (int8_t) vout0x2;
182 c0[3] = (int8_t) vout0x3;
183 c1[0] = (int8_t) vout1x0;
184 c1[1] = (int8_t) vout1x1;
185 c1[2] = (int8_t) vout1x2;
186 c1[3] = (int8_t) vout1x3;
187 c2[0] = (int8_t) vout2x0;
188 c2[1] = (int8_t) vout2x1;
189 c2[2] = (int8_t) vout2x2;
190 c2[3] = (int8_t) vout2x3;
191
192 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
193 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
194 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
195
196 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
197 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
198 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
199
200 nc -= 4;
201 } else {
202 if (nc & 2) {
203 c0[0] = (int8_t) vout0x0;
204 c0[1] = (int8_t) vout0x1;
205 vout0x0 = vout0x2;
206 c0 += 2;
207 c1[0] = (int8_t) vout1x0;
208 c1[1] = (int8_t) vout1x1;
209 vout1x0 = vout1x2;
210 c1 += 2;
211 c2[0] = (int8_t) vout2x0;
212 c2[1] = (int8_t) vout2x1;
213 vout2x0 = vout2x2;
214 c2 += 2;
215 }
216 if (nc & 1) {
217 c0[0] = (int8_t) vout0x0;
218 c1[0] = (int8_t) vout1x0;
219 c2[0] = (int8_t) vout2x0;
220 }
221
222 nc = 0;
223 }
224 } while (nc != 0);
225 }
226