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_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28   assert(mr != 0);
29   assert(mr <= 2);
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 
46   const float vmin = params->scalar.min;
47   const float vmax = params->scalar.max;
48   do {
49     float vacc00 = w[0];
50     float vacc01 = w[1];
51     float vacc02 = w[2];
52     float vacc03 = w[3];
53     w += 4;
54     float vacc10 = vacc00;
55     float vacc11 = vacc01;
56     float vacc12 = vacc02;
57     float vacc13 = vacc03;
58 
59     size_t k = kc;
60     do {
61       const float va0 = *a0++;
62       const float va1 = *a1++;
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 += va0 * vb0;
71       vacc01 += va0 * vb1;
72       vacc02 += va0 * vb2;
73       vacc03 += va0 * vb3;
74       vacc10 += va1 * vb0;
75       vacc11 += va1 * vb1;
76       vacc12 += va1 * vb2;
77       vacc13 += va1 * vb3;
78 
79       k -= sizeof(float);
80     } while (k != 0);
81 
82     vacc00 = math_max_f32(vacc00, vmin);
83     vacc01 = math_max_f32(vacc01, vmin);
84     vacc02 = math_max_f32(vacc02, vmin);
85     vacc03 = math_max_f32(vacc03, vmin);
86     vacc10 = math_max_f32(vacc10, vmin);
87     vacc11 = math_max_f32(vacc11, vmin);
88     vacc12 = math_max_f32(vacc12, vmin);
89     vacc13 = math_max_f32(vacc13, vmin);
90 
91     vacc00 = math_min_f32(vacc00, vmax);
92     vacc01 = math_min_f32(vacc01, vmax);
93     vacc02 = math_min_f32(vacc02, vmax);
94     vacc03 = math_min_f32(vacc03, vmax);
95     vacc10 = math_min_f32(vacc10, vmax);
96     vacc11 = math_min_f32(vacc11, vmax);
97     vacc12 = math_min_f32(vacc12, vmax);
98     vacc13 = math_min_f32(vacc13, vmax);
99 
100     if XNN_LIKELY(nc >= 4) {
101       c1[0] = vacc10;
102       c1[1] = vacc11;
103       c1[2] = vacc12;
104       c1[3] = vacc13;
105       c1 = (float*) ((uintptr_t) c1 + cn_stride);
106       c0[0] = vacc00;
107       c0[1] = vacc01;
108       c0[2] = vacc02;
109       c0[3] = vacc03;
110       c0 = (float*) ((uintptr_t) c0 + cn_stride);
111 
112       a1 = (const void*) ((uintptr_t) a1 - kc);
113       a0 = (const void*) ((uintptr_t) a0 - kc);
114 
115       nc -= 4;
116     } else {
117       if (nc & 2) {
118         c1[0] = vacc10;
119         c1[1] = vacc11;
120         vacc10 = vacc12;
121         c1 += 2;
122         c0[0] = vacc00;
123         c0[1] = vacc01;
124         vacc00 = vacc02;
125         c0 += 2;
126       }
127       if (nc & 1) {
128         c1[0] = vacc10;
129         c0[0] = vacc00;
130       }
131 
132       nc = 0;
133     }
134   } while (nc != 0);
135 }
136