xref: /aosp_15_r20/external/XNNPACK/src/f32-gemm/gen/2x4-scalar.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_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_default_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_gemm_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_default_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   do {
47     float vacc00 = w[0];
48     float vacc01 = w[1];
49     float vacc02 = w[2];
50     float vacc03 = w[3];
51     w += 4;
52     float vacc10 = vacc00;
53     float vacc11 = vacc01;
54     float vacc12 = vacc02;
55     float vacc13 = vacc03;
56 
57     size_t k = kc;
58     do {
59       const float va0 = *a0++;
60       const float va1 = *a1++;
61 
62       const float vb0 = w[0];
63       const float vb1 = w[1];
64       const float vb2 = w[2];
65       const float vb3 = w[3];
66       w += 4;
67 
68       vacc00 = math_muladd_f32(va0, vb0, vacc00);
69       vacc01 = math_muladd_f32(va0, vb1, vacc01);
70       vacc02 = math_muladd_f32(va0, vb2, vacc02);
71       vacc03 = math_muladd_f32(va0, vb3, vacc03);
72       vacc10 = math_muladd_f32(va1, vb0, vacc10);
73       vacc11 = math_muladd_f32(va1, vb1, vacc11);
74       vacc12 = math_muladd_f32(va1, vb2, vacc12);
75       vacc13 = math_muladd_f32(va1, vb3, vacc13);
76 
77       k -= sizeof(float);
78     } while (k != 0);
79 
80 
81     if XNN_LIKELY(nc >= 4) {
82       c1[0] = vacc10;
83       c1[1] = vacc11;
84       c1[2] = vacc12;
85       c1[3] = vacc13;
86       c1 = (float*) ((uintptr_t) c1 + cn_stride);
87       c0[0] = vacc00;
88       c0[1] = vacc01;
89       c0[2] = vacc02;
90       c0[3] = vacc03;
91       c0 = (float*) ((uintptr_t) c0 + cn_stride);
92 
93       a1 = (const void*) ((uintptr_t) a1 - kc);
94       a0 = (const void*) ((uintptr_t) a0 - kc);
95 
96       nc -= 4;
97     } else {
98       if (nc & 2) {
99         c1[0] = vacc10;
100         c1[1] = vacc11;
101         vacc10 = vacc12;
102         c1 += 2;
103         c0[0] = vacc00;
104         c0[1] = vacc01;
105         vacc00 = vacc02;
106         c0 += 2;
107       }
108       if (nc & 1) {
109         c1[0] = vacc10;
110         c0[0] = vacc00;
111       }
112 
113       nc = 0;
114     }
115   } while (nc != 0);
116 }
117