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_3x4__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)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic(
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 <= 3);
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 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
44 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
45 if XNN_UNPREDICTABLE(mr <= 2) {
46 a2 = a1;
47 c2 = c1;
48 }
49
50 const int32_t vb_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
51 do {
52 int32_t vacc0x0 = ((const int32_t*) w)[0];
53 int32_t vacc0x1 = ((const int32_t*) w)[1];
54 int32_t vacc0x2 = ((const int32_t*) w)[2];
55 int32_t vacc0x3 = ((const int32_t*) w)[3];
56 int32_t vacc1x0 = vacc0x0;
57 int32_t vacc1x1 = vacc0x1;
58 int32_t vacc1x2 = vacc0x2;
59 int32_t vacc1x3 = vacc0x3;
60 int32_t vacc2x0 = vacc0x0;
61 int32_t vacc2x1 = vacc0x1;
62 int32_t vacc2x2 = vacc0x2;
63 int32_t vacc2x3 = vacc0x3;
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t k = kc;
67 do {
68 const int32_t va0 = (int32_t) (uint32_t) *a0++;
69 const int32_t va1 = (int32_t) (uint32_t) *a1++;
70 const int32_t va2 = (int32_t) (uint32_t) *a2++;
71
72 const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
73 const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
74 const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
75 const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
76 w = (const void*) ((const uint8_t*) w + 4);
77
78 vacc0x0 += va0 * vb0;
79 vacc0x1 += va0 * vb1;
80 vacc0x2 += va0 * vb2;
81 vacc0x3 += va0 * vb3;
82 vacc1x0 += va1 * vb0;
83 vacc1x1 += va1 * vb1;
84 vacc1x2 += va1 * vb2;
85 vacc1x3 += va1 * vb3;
86 vacc2x0 += va2 * vb0;
87 vacc2x1 += va2 * vb1;
88 vacc2x2 += va2 * vb2;
89 vacc2x3 += va2 * vb3;
90
91 k -= sizeof(uint8_t);
92 } while (k != 0);
93
94 float vfpacc0x0 = (float) vacc0x0;
95 float vfpacc0x1 = (float) vacc0x1;
96 float vfpacc0x2 = (float) vacc0x2;
97 float vfpacc0x3 = (float) vacc0x3;
98 float vfpacc1x0 = (float) vacc1x0;
99 float vfpacc1x1 = (float) vacc1x1;
100 float vfpacc1x2 = (float) vacc1x2;
101 float vfpacc1x3 = (float) vacc1x3;
102 float vfpacc2x0 = (float) vacc2x0;
103 float vfpacc2x1 = (float) vacc2x1;
104 float vfpacc2x2 = (float) vacc2x2;
105 float vfpacc2x3 = (float) vacc2x3;
106
107 const float vscale = params->fp32_scalar_fmagic.scale;
108 vfpacc0x0 *= vscale;
109 vfpacc0x1 *= vscale;
110 vfpacc0x2 *= vscale;
111 vfpacc0x3 *= vscale;
112 vfpacc1x0 *= vscale;
113 vfpacc1x1 *= vscale;
114 vfpacc1x2 *= vscale;
115 vfpacc1x3 *= vscale;
116 vfpacc2x0 *= vscale;
117 vfpacc2x1 *= vscale;
118 vfpacc2x2 *= vscale;
119 vfpacc2x3 *= vscale;
120
121 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
122 vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
123 vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
124 vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
125 vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
126 vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
127 vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
128 vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
129 vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
130 vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
131 vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
132 vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
133 vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
134
135 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
136 vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
137 vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
138 vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
139 vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
140 vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
141 vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
142 vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
143 vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
144 vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
145 vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
146 vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
147 vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
148
149 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
150 vfpacc0x0 += vmagic_bias;
151 vfpacc0x1 += vmagic_bias;
152 vfpacc0x2 += vmagic_bias;
153 vfpacc0x3 += vmagic_bias;
154 vfpacc1x0 += vmagic_bias;
155 vfpacc1x1 += vmagic_bias;
156 vfpacc1x2 += vmagic_bias;
157 vfpacc1x3 += vmagic_bias;
158 vfpacc2x0 += vmagic_bias;
159 vfpacc2x1 += vmagic_bias;
160 vfpacc2x2 += vmagic_bias;
161 vfpacc2x3 += vmagic_bias;
162
163 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
164 int32_t vout0x0 = (int32_t) fp32_to_bits(vfpacc0x0) - vmagic_bias_less_output_zero_point;
165 int32_t vout0x1 = (int32_t) fp32_to_bits(vfpacc0x1) - vmagic_bias_less_output_zero_point;
166 int32_t vout0x2 = (int32_t) fp32_to_bits(vfpacc0x2) - vmagic_bias_less_output_zero_point;
167 int32_t vout0x3 = (int32_t) fp32_to_bits(vfpacc0x3) - vmagic_bias_less_output_zero_point;
168 int32_t vout1x0 = (int32_t) fp32_to_bits(vfpacc1x0) - vmagic_bias_less_output_zero_point;
169 int32_t vout1x1 = (int32_t) fp32_to_bits(vfpacc1x1) - vmagic_bias_less_output_zero_point;
170 int32_t vout1x2 = (int32_t) fp32_to_bits(vfpacc1x2) - vmagic_bias_less_output_zero_point;
171 int32_t vout1x3 = (int32_t) fp32_to_bits(vfpacc1x3) - vmagic_bias_less_output_zero_point;
172 int32_t vout2x0 = (int32_t) fp32_to_bits(vfpacc2x0) - vmagic_bias_less_output_zero_point;
173 int32_t vout2x1 = (int32_t) fp32_to_bits(vfpacc2x1) - vmagic_bias_less_output_zero_point;
174 int32_t vout2x2 = (int32_t) fp32_to_bits(vfpacc2x2) - vmagic_bias_less_output_zero_point;
175 int32_t vout2x3 = (int32_t) fp32_to_bits(vfpacc2x3) - vmagic_bias_less_output_zero_point;
176
177 if XNN_LIKELY(nc >= 4) {
178 c0[0] = (uint8_t) vout0x0;
179 c0[1] = (uint8_t) vout0x1;
180 c0[2] = (uint8_t) vout0x2;
181 c0[3] = (uint8_t) vout0x3;
182 c1[0] = (uint8_t) vout1x0;
183 c1[1] = (uint8_t) vout1x1;
184 c1[2] = (uint8_t) vout1x2;
185 c1[3] = (uint8_t) vout1x3;
186 c2[0] = (uint8_t) vout2x0;
187 c2[1] = (uint8_t) vout2x1;
188 c2[2] = (uint8_t) vout2x2;
189 c2[3] = (uint8_t) vout2x3;
190
191 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
192 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
193 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
194
195 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
196 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
197 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
198
199 nc -= 4;
200 } else {
201 if (nc & 2) {
202 c0[0] = (uint8_t) vout0x0;
203 c0[1] = (uint8_t) vout0x1;
204 vout0x0 = vout0x2;
205 c0 += 2;
206 c1[0] = (uint8_t) vout1x0;
207 c1[1] = (uint8_t) vout1x1;
208 vout1x0 = vout1x2;
209 c1 += 2;
210 c2[0] = (uint8_t) vout2x0;
211 c2[1] = (uint8_t) vout2x1;
212 vout2x0 = vout2x2;
213 c2 += 2;
214 }
215 if (nc & 1) {
216 c0[0] = (uint8_t) vout0x0;
217 c1[0] = (uint8_t) vout1x0;
218 c2[0] = (uint8_t) vout2x0;
219 }
220
221 nc = 0;
222 }
223 } while (nc != 0);
224 }
225