• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-igemm/scalar.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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/igemm.h>
13 #include <xnnpack/math.h>
14 
15 
xnn_f32_igemm_ukernel_1x4__scalar(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_igemm_ukernel_1x4__scalar(
17     size_t mr,
18     size_t nc,
19     size_t kc,
20     size_t ks,
21     const float**restrict a,
22     const float*restrict w,
23     float*restrict c,
24     size_t cm_stride,
25     size_t cn_stride,
26     size_t a_offset,
27     const float* zero,
28     const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(float) == 0);
35   assert(ks != 0);
36   assert(ks % (1 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(float) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   float* c0 = c;
43 
44   do {
45     float vacc00 = w[0];
46     float vacc01 = w[1];
47     float vacc02 = w[2];
48     float vacc03 = w[3];
49     w += 4;
50 
51     size_t p = ks;
52     do {
53       const float* restrict a0 = a[0];
54       assert(a0 != NULL);
55       if XNN_UNPREDICTABLE(a0 != zero) {
56         a0 = (const float*) ((uintptr_t) a0 + a_offset);
57       }
58       a += 1;
59 
60       size_t k = kc;
61       do {
62         const float va0 = *a0++;
63 
64         const float vb0 = w[0];
65         const float vb1 = w[1];
66         const float vb2 = w[2];
67         const float vb3 = w[3];
68         w += 4;
69 
70         vacc00 = math_muladd_f32(va0, vb0, vacc00);
71         vacc01 = math_muladd_f32(va0, vb1, vacc01);
72         vacc02 = math_muladd_f32(va0, vb2, vacc02);
73         vacc03 = math_muladd_f32(va0, vb3, vacc03);
74 
75         k -= sizeof(float);
76       } while (k != 0);
77       p -= 1 * sizeof(void*);
78     } while (p != 0);
79 
80 
81     if XNN_LIKELY(nc >= 4) {
82       c0[0] = vacc00;
83       c0[1] = vacc01;
84       c0[2] = vacc02;
85       c0[3] = vacc03;
86       c0 = (float*) ((uintptr_t) c0 + cn_stride);
87 
88       a = (const float**restrict) ((uintptr_t) a - ks);
89       nc -= 4;
90     } else {
91       if (nc & 2) {
92         c0[0] = vacc00;
93         c0[1] = vacc01;
94         vacc00 = vacc02;
95         c0 += 2;
96       }
97       if (nc & 1) {
98         c0[0] = vacc00;
99       }
100 
101       nc = 0;
102     }
103   } while (nc != 0);
104 }
105