• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <math.h>
12 
13 #include <xnnpack/math.h>
14 #include <xnnpack/gemm.h>
15 
16 
xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf(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)])17 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const uint8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     uint8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(mr != 0);
30   assert(mr <= 3);
31   assert(nc != 0);
32   assert(kc != 0);
33 
34   const uint8_t* a0 = a;
35   uint8_t* c0 = c;
36   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
37   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
38   if XNN_UNPREDICTABLE(mr < 2) {
39     a1 = a0;
40     c1 = c0;
41   }
42   const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
43   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
44   if XNN_UNPREDICTABLE(mr <= 2) {
45     a2 = a1;
46     c2 = c1;
47   }
48 
49   const int32_t vb_zero_point = params->fp32_scalar_lrintf.kernel_zero_point;
50   do {
51     int32_t vacc0x0 = ((const int32_t*) w)[0];
52     int32_t vacc0x1 = ((const int32_t*) w)[1];
53     int32_t vacc0x2 = ((const int32_t*) w)[2];
54     int32_t vacc0x3 = ((const int32_t*) w)[3];
55     int32_t vacc1x0 = vacc0x0;
56     int32_t vacc1x1 = vacc0x1;
57     int32_t vacc1x2 = vacc0x2;
58     int32_t vacc1x3 = vacc0x3;
59     int32_t vacc2x0 = vacc0x0;
60     int32_t vacc2x1 = vacc0x1;
61     int32_t vacc2x2 = vacc0x2;
62     int32_t vacc2x3 = vacc0x3;
63     w = (const void*) ((const int32_t*) w + 4);
64 
65     size_t k = kc;
66     do {
67       const int32_t va0 = (int32_t) (uint32_t) *a0++;
68       const int32_t va1 = (int32_t) (uint32_t) *a1++;
69       const int32_t va2 = (int32_t) (uint32_t) *a2++;
70 
71       const int32_t vb0 = (int32_t) (uint32_t) ((const uint8_t*) w)[0] - vb_zero_point;
72       const int32_t vb1 = (int32_t) (uint32_t) ((const uint8_t*) w)[1] - vb_zero_point;
73       const int32_t vb2 = (int32_t) (uint32_t) ((const uint8_t*) w)[2] - vb_zero_point;
74       const int32_t vb3 = (int32_t) (uint32_t) ((const uint8_t*) w)[3] - vb_zero_point;
75       w = (const void*) ((const uint8_t*) w + 4);
76 
77       vacc0x0 += va0 * vb0;
78       vacc0x1 += va0 * vb1;
79       vacc0x2 += va0 * vb2;
80       vacc0x3 += va0 * vb3;
81       vacc1x0 += va1 * vb0;
82       vacc1x1 += va1 * vb1;
83       vacc1x2 += va1 * vb2;
84       vacc1x3 += va1 * vb3;
85       vacc2x0 += va2 * vb0;
86       vacc2x1 += va2 * vb1;
87       vacc2x2 += va2 * vb2;
88       vacc2x3 += va2 * vb3;
89 
90       k -= sizeof(uint8_t);
91     } while (k != 0);
92 
93     float vfpacc0x0 = (float) vacc0x0;
94     float vfpacc0x1 = (float) vacc0x1;
95     float vfpacc0x2 = (float) vacc0x2;
96     float vfpacc0x3 = (float) vacc0x3;
97     float vfpacc1x0 = (float) vacc1x0;
98     float vfpacc1x1 = (float) vacc1x1;
99     float vfpacc1x2 = (float) vacc1x2;
100     float vfpacc1x3 = (float) vacc1x3;
101     float vfpacc2x0 = (float) vacc2x0;
102     float vfpacc2x1 = (float) vacc2x1;
103     float vfpacc2x2 = (float) vacc2x2;
104     float vfpacc2x3 = (float) vacc2x3;
105 
106     const float vscale = params->fp32_scalar_lrintf.scale;
107     vfpacc0x0 *= vscale;
108     vfpacc0x1 *= vscale;
109     vfpacc0x2 *= vscale;
110     vfpacc0x3 *= vscale;
111     vfpacc1x0 *= vscale;
112     vfpacc1x1 *= vscale;
113     vfpacc1x2 *= vscale;
114     vfpacc1x3 *= vscale;
115     vfpacc2x0 *= vscale;
116     vfpacc2x1 *= vscale;
117     vfpacc2x2 *= vscale;
118     vfpacc2x3 *= vscale;
119 
120     const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
121     vfpacc0x0 = math_max_f32(vfpacc0x0, voutput_min_less_zero_point);
122     vfpacc0x1 = math_max_f32(vfpacc0x1, voutput_min_less_zero_point);
123     vfpacc0x2 = math_max_f32(vfpacc0x2, voutput_min_less_zero_point);
124     vfpacc0x3 = math_max_f32(vfpacc0x3, voutput_min_less_zero_point);
125     vfpacc1x0 = math_max_f32(vfpacc1x0, voutput_min_less_zero_point);
126     vfpacc1x1 = math_max_f32(vfpacc1x1, voutput_min_less_zero_point);
127     vfpacc1x2 = math_max_f32(vfpacc1x2, voutput_min_less_zero_point);
128     vfpacc1x3 = math_max_f32(vfpacc1x3, voutput_min_less_zero_point);
129     vfpacc2x0 = math_max_f32(vfpacc2x0, voutput_min_less_zero_point);
130     vfpacc2x1 = math_max_f32(vfpacc2x1, voutput_min_less_zero_point);
131     vfpacc2x2 = math_max_f32(vfpacc2x2, voutput_min_less_zero_point);
132     vfpacc2x3 = math_max_f32(vfpacc2x3, voutput_min_less_zero_point);
133 
134     const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
135     vfpacc0x0 = math_min_f32(vfpacc0x0, voutput_max_less_zero_point);
136     vfpacc0x1 = math_min_f32(vfpacc0x1, voutput_max_less_zero_point);
137     vfpacc0x2 = math_min_f32(vfpacc0x2, voutput_max_less_zero_point);
138     vfpacc0x3 = math_min_f32(vfpacc0x3, voutput_max_less_zero_point);
139     vfpacc1x0 = math_min_f32(vfpacc1x0, voutput_max_less_zero_point);
140     vfpacc1x1 = math_min_f32(vfpacc1x1, voutput_max_less_zero_point);
141     vfpacc1x2 = math_min_f32(vfpacc1x2, voutput_max_less_zero_point);
142     vfpacc1x3 = math_min_f32(vfpacc1x3, voutput_max_less_zero_point);
143     vfpacc2x0 = math_min_f32(vfpacc2x0, voutput_max_less_zero_point);
144     vfpacc2x1 = math_min_f32(vfpacc2x1, voutput_max_less_zero_point);
145     vfpacc2x2 = math_min_f32(vfpacc2x2, voutput_max_less_zero_point);
146     vfpacc2x3 = math_min_f32(vfpacc2x3, voutput_max_less_zero_point);
147 
148     const int32_t vrndacc0x0 = (int32_t) lrintf(vfpacc0x0);
149     const int32_t vrndacc0x1 = (int32_t) lrintf(vfpacc0x1);
150     const int32_t vrndacc0x2 = (int32_t) lrintf(vfpacc0x2);
151     const int32_t vrndacc0x3 = (int32_t) lrintf(vfpacc0x3);
152     const int32_t vrndacc1x0 = (int32_t) lrintf(vfpacc1x0);
153     const int32_t vrndacc1x1 = (int32_t) lrintf(vfpacc1x1);
154     const int32_t vrndacc1x2 = (int32_t) lrintf(vfpacc1x2);
155     const int32_t vrndacc1x3 = (int32_t) lrintf(vfpacc1x3);
156     const int32_t vrndacc2x0 = (int32_t) lrintf(vfpacc2x0);
157     const int32_t vrndacc2x1 = (int32_t) lrintf(vfpacc2x1);
158     const int32_t vrndacc2x2 = (int32_t) lrintf(vfpacc2x2);
159     const int32_t vrndacc2x3 = (int32_t) lrintf(vfpacc2x3);
160 
161     const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
162     int32_t vout0x0 = vrndacc0x0 + voutput_zero_point;
163     int32_t vout0x1 = vrndacc0x1 + voutput_zero_point;
164     int32_t vout0x2 = vrndacc0x2 + voutput_zero_point;
165     int32_t vout0x3 = vrndacc0x3 + voutput_zero_point;
166     int32_t vout1x0 = vrndacc1x0 + voutput_zero_point;
167     int32_t vout1x1 = vrndacc1x1 + voutput_zero_point;
168     int32_t vout1x2 = vrndacc1x2 + voutput_zero_point;
169     int32_t vout1x3 = vrndacc1x3 + voutput_zero_point;
170     int32_t vout2x0 = vrndacc2x0 + voutput_zero_point;
171     int32_t vout2x1 = vrndacc2x1 + voutput_zero_point;
172     int32_t vout2x2 = vrndacc2x2 + voutput_zero_point;
173     int32_t vout2x3 = vrndacc2x3 + voutput_zero_point;
174 
175     if XNN_LIKELY(nc >= 4) {
176       c0[0] = (uint8_t) vout0x0;
177       c0[1] = (uint8_t) vout0x1;
178       c0[2] = (uint8_t) vout0x2;
179       c0[3] = (uint8_t) vout0x3;
180       c1[0] = (uint8_t) vout1x0;
181       c1[1] = (uint8_t) vout1x1;
182       c1[2] = (uint8_t) vout1x2;
183       c1[3] = (uint8_t) vout1x3;
184       c2[0] = (uint8_t) vout2x0;
185       c2[1] = (uint8_t) vout2x1;
186       c2[2] = (uint8_t) vout2x2;
187       c2[3] = (uint8_t) vout2x3;
188 
189       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
190       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
191       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
192 
193       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
194       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
195       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
196 
197       nc -= 4;
198     } else {
199       if (nc & 2) {
200         c0[0] = (uint8_t) vout0x0;
201         c0[1] = (uint8_t) vout0x1;
202         vout0x0 = vout0x2;
203         c0 += 2;
204         c1[0] = (uint8_t) vout1x0;
205         c1[1] = (uint8_t) vout1x1;
206         vout1x0 = vout1x2;
207         c1 += 2;
208         c2[0] = (uint8_t) vout2x0;
209         c2[1] = (uint8_t) vout2x1;
210         vout2x0 = vout2x2;
211         c2 += 2;
212       }
213       if (nc & 1) {
214         c0[0] = (uint8_t) vout0x0;
215         c1[0] = (uint8_t) vout1x0;
216         c2[0] = (uint8_t) vout2x0;
217       }
218 
219       nc = 0;
220     }
221   } while (nc != 0);
222 }
223