1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
13 #include <xnnpack/math.h>
14
15
xnn_f32_gemm_relu_ukernel_1x4__scalar(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_gemm_relu_ukernel_1x4__scalar(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 const float* restrict a,
21 size_t a_stride,
22 const float* restrict w,
23 float* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mr != 0);
29 assert(mr <= 1);
30 assert(nc != 0);
31 assert(kc != 0);
32 assert(kc % sizeof(float) == 0);
33 assert(a != NULL);
34 assert(w != NULL);
35 assert(c != NULL);
36
37 const float* a0 = a;
38 float* c0 = c;
39
40 do {
41 float vacc00 = w[0];
42 float vacc01 = w[1];
43 float vacc02 = w[2];
44 float vacc03 = w[3];
45 w += 4;
46
47 size_t k = kc;
48 do {
49 const float va0 = *a0++;
50
51 const float vb0 = w[0];
52 const float vb1 = w[1];
53 const float vb2 = w[2];
54 const float vb3 = w[3];
55 w += 4;
56
57 vacc00 += va0 * vb0;
58 vacc01 += va0 * vb1;
59 vacc02 += va0 * vb2;
60 vacc03 += va0 * vb3;
61
62 k -= sizeof(float);
63 } while (k != 0);
64
65 vacc00 = math_max_f32(vacc00, 0.0f);
66 vacc01 = math_max_f32(vacc01, 0.0f);
67 vacc02 = math_max_f32(vacc02, 0.0f);
68 vacc03 = math_max_f32(vacc03, 0.0f);
69
70 if XNN_LIKELY(nc >= 4) {
71 c0[0] = vacc00;
72 c0[1] = vacc01;
73 c0[2] = vacc02;
74 c0[3] = vacc03;
75 c0 = (float*) ((uintptr_t) c0 + cn_stride);
76
77 a0 = (const void*) ((uintptr_t) a0 - kc);
78
79 nc -= 4;
80 } else {
81 if (nc & 2) {
82 c0[0] = vacc00;
83 c0[1] = vacc01;
84 vacc00 = vacc02;
85 c0 += 2;
86 }
87 if (nc & 1) {
88 c0[0] = vacc00;
89 }
90
91 nc = 0;
92 }
93 } while (nc != 0);
94 }
95