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_gemminc_minmax_ukernel_2x4__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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_gemminc_minmax_ukernel_2x4__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 float*restrict acc,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 2);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37 assert(acc != NULL);
38
39 const float* a0 = a;
40 float* c0 = c;
41 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
42 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 const float vmin = params->scalar.min;
49 const float vmax = params->scalar.max;
50 do {
51 float vacc00 = acc[0];
52 float vacc01 = acc[1];
53 float vacc02 = acc[2];
54 float vacc03 = acc[3];
55 float vacc10 = acc[4];
56 float vacc11 = acc[5];
57 float vacc12 = acc[6];
58 float vacc13 = acc[7];
59 acc += 8;
60
61 size_t k = kc;
62 do {
63 const float va0 = *a0++;
64 const float va1 = *a1++;
65
66 const float vb0 = w[0];
67 const float vb1 = w[1];
68 const float vb2 = w[2];
69 const float vb3 = w[3];
70 w += 4;
71
72 vacc00 = math_muladd_f32(va0, vb0, vacc00);
73 vacc01 = math_muladd_f32(va0, vb1, vacc01);
74 vacc02 = math_muladd_f32(va0, vb2, vacc02);
75 vacc03 = math_muladd_f32(va0, vb3, vacc03);
76 vacc10 = math_muladd_f32(va1, vb0, vacc10);
77 vacc11 = math_muladd_f32(va1, vb1, vacc11);
78 vacc12 = math_muladd_f32(va1, vb2, vacc12);
79 vacc13 = math_muladd_f32(va1, vb3, vacc13);
80
81 k -= sizeof(float);
82 } while (k != 0);
83
84 vacc00 = math_max_f32(vacc00, vmin);
85 vacc01 = math_max_f32(vacc01, vmin);
86 vacc02 = math_max_f32(vacc02, vmin);
87 vacc03 = math_max_f32(vacc03, vmin);
88 vacc10 = math_max_f32(vacc10, vmin);
89 vacc11 = math_max_f32(vacc11, vmin);
90 vacc12 = math_max_f32(vacc12, vmin);
91 vacc13 = math_max_f32(vacc13, vmin);
92
93 vacc00 = math_min_f32(vacc00, vmax);
94 vacc01 = math_min_f32(vacc01, vmax);
95 vacc02 = math_min_f32(vacc02, vmax);
96 vacc03 = math_min_f32(vacc03, vmax);
97 vacc10 = math_min_f32(vacc10, vmax);
98 vacc11 = math_min_f32(vacc11, vmax);
99 vacc12 = math_min_f32(vacc12, vmax);
100 vacc13 = math_min_f32(vacc13, vmax);
101
102 if XNN_LIKELY(nc >= 4) {
103 c1[0] = vacc10;
104 c1[1] = vacc11;
105 c1[2] = vacc12;
106 c1[3] = vacc13;
107 c1 = (float*) ((uintptr_t) c1 + cn_stride);
108 c0[0] = vacc00;
109 c0[1] = vacc01;
110 c0[2] = vacc02;
111 c0[3] = vacc03;
112 c0 = (float*) ((uintptr_t) c0 + cn_stride);
113
114 a1 = (const void*) ((uintptr_t) a1 - kc);
115 a0 = (const void*) ((uintptr_t) a0 - kc);
116
117 nc -= 4;
118 } else {
119 if (nc & 2) {
120 c1[0] = vacc10;
121 c1[1] = vacc11;
122 vacc10 = vacc12;
123 c1 += 2;
124 c0[0] = vacc00;
125 c0[1] = vacc01;
126 vacc00 = vacc02;
127 c0 += 2;
128 }
129 if (nc & 1) {
130 c1[0] = vacc10;
131 c0[0] = vacc00;
132 }
133
134 nc = 0;
135 }
136 } while (nc != 0);
137 }
138