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_4x2__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_4x2__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 <= 4);
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 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
40 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
41 if XNN_UNPREDICTABLE(mr < 2) {
42 a1 = a0;
43 c1 = c0;
44 }
45 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
46 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
47 if XNN_UNPREDICTABLE(mr <= 2) {
48 a2 = a1;
49 c2 = c1;
50 }
51 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 a3 = a2;
55 c3 = c2;
56 }
57
58 do {
59 float vacc00 = w[0];
60 float vacc01 = w[1];
61 w += 2;
62 float vacc10 = vacc00;
63 float vacc11 = vacc01;
64 float vacc20 = vacc00;
65 float vacc21 = vacc01;
66 float vacc30 = vacc00;
67 float vacc31 = vacc01;
68
69 size_t k = kc;
70 do {
71 const float va0 = *a0++;
72 const float va1 = *a1++;
73 const float va2 = *a2++;
74 const float va3 = *a3++;
75
76 const float vb0 = w[0];
77 const float vb1 = w[1];
78 w += 2;
79
80 vacc00 += va0 * vb0;
81 vacc01 += va0 * vb1;
82 vacc10 += va1 * vb0;
83 vacc11 += va1 * vb1;
84 vacc20 += va2 * vb0;
85 vacc21 += va2 * vb1;
86 vacc30 += va3 * vb0;
87 vacc31 += va3 * vb1;
88
89 k -= sizeof(float);
90 } while (k != 0);
91
92 vacc00 = math_max_f32(vacc00, 0.0f);
93 vacc01 = math_max_f32(vacc01, 0.0f);
94 vacc10 = math_max_f32(vacc10, 0.0f);
95 vacc11 = math_max_f32(vacc11, 0.0f);
96 vacc20 = math_max_f32(vacc20, 0.0f);
97 vacc21 = math_max_f32(vacc21, 0.0f);
98 vacc30 = math_max_f32(vacc30, 0.0f);
99 vacc31 = math_max_f32(vacc31, 0.0f);
100
101 if XNN_LIKELY(nc >= 2) {
102 c3[0] = vacc30;
103 c3[1] = vacc31;
104 c3 = (float*) ((uintptr_t) c3 + cn_stride);
105 c2[0] = vacc20;
106 c2[1] = vacc21;
107 c2 = (float*) ((uintptr_t) c2 + cn_stride);
108 c1[0] = vacc10;
109 c1[1] = vacc11;
110 c1 = (float*) ((uintptr_t) c1 + cn_stride);
111 c0[0] = vacc00;
112 c0[1] = vacc01;
113 c0 = (float*) ((uintptr_t) c0 + cn_stride);
114
115 a3 = (const void*) ((uintptr_t) a3 - kc);
116 a2 = (const void*) ((uintptr_t) a2 - kc);
117 a1 = (const void*) ((uintptr_t) a1 - kc);
118 a0 = (const void*) ((uintptr_t) a0 - kc);
119
120 nc -= 2;
121 } else {
122 if (nc & 1) {
123 c3[0] = vacc30;
124 c2[0] = vacc20;
125 c1[0] = vacc10;
126 c0[0] = vacc00;
127 }
128
129 nc = 0;
130 }
131 } while (nc != 0);
132 }
133