1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
13 #include <xnnpack/math.h>
14
15
xnn_f32_igemm_minmax_ukernel_2x4__scalar(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_f32_igemm_minmax_ukernel_2x4__scalar(
17 size_t mr,
18 size_t nc,
19 size_t kc,
20 size_t ks,
21 const float**restrict a,
22 const float*restrict w,
23 float*restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 size_t a_offset,
27 const float* zero,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(ks != 0);
36 assert(ks % (2 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(float) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 float* c0 = c;
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 c1 = c0;
46 }
47
48 const float vmin = params->scalar.min;
49 const float vmax = params->scalar.max;
50 do {
51 float vacc00 = w[0];
52 float vacc01 = w[1];
53 float vacc02 = w[2];
54 float vacc03 = w[3];
55 float vacc10 = vacc00;
56 float vacc11 = vacc01;
57 float vacc12 = vacc02;
58 float vacc13 = vacc03;
59 w += 4;
60
61 size_t p = ks;
62 do {
63 const float* restrict a0 = a[0];
64 assert(a0 != NULL);
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const float*) ((uintptr_t) a0 + a_offset);
67 }
68 const float* restrict a1 = a[1];
69 assert(a1 != NULL);
70 if XNN_UNPREDICTABLE(a1 != zero) {
71 a1 = (const float*) ((uintptr_t) a1 + a_offset);
72 }
73 a += 2;
74
75 size_t k = kc;
76 do {
77 const float va0 = *a0++;
78 const float va1 = *a1++;
79
80 const float vb0 = w[0];
81 const float vb1 = w[1];
82 const float vb2 = w[2];
83 const float vb3 = w[3];
84 w += 4;
85
86 vacc00 = math_muladd_f32(va0, vb0, vacc00);
87 vacc01 = math_muladd_f32(va0, vb1, vacc01);
88 vacc02 = math_muladd_f32(va0, vb2, vacc02);
89 vacc03 = math_muladd_f32(va0, vb3, vacc03);
90 vacc10 = math_muladd_f32(va1, vb0, vacc10);
91 vacc11 = math_muladd_f32(va1, vb1, vacc11);
92 vacc12 = math_muladd_f32(va1, vb2, vacc12);
93 vacc13 = math_muladd_f32(va1, vb3, vacc13);
94
95 k -= sizeof(float);
96 } while (k != 0);
97 p -= 2 * sizeof(void*);
98 } while (p != 0);
99
100 vacc00 = math_max_f32(vacc00, vmin);
101 vacc01 = math_max_f32(vacc01, vmin);
102 vacc02 = math_max_f32(vacc02, vmin);
103 vacc03 = math_max_f32(vacc03, vmin);
104 vacc10 = math_max_f32(vacc10, vmin);
105 vacc11 = math_max_f32(vacc11, vmin);
106 vacc12 = math_max_f32(vacc12, vmin);
107 vacc13 = math_max_f32(vacc13, vmin);
108
109 vacc00 = math_min_f32(vacc00, vmax);
110 vacc01 = math_min_f32(vacc01, vmax);
111 vacc02 = math_min_f32(vacc02, vmax);
112 vacc03 = math_min_f32(vacc03, vmax);
113 vacc10 = math_min_f32(vacc10, vmax);
114 vacc11 = math_min_f32(vacc11, vmax);
115 vacc12 = math_min_f32(vacc12, vmax);
116 vacc13 = math_min_f32(vacc13, vmax);
117
118 if XNN_LIKELY(nc >= 4) {
119 c1[0] = vacc10;
120 c1[1] = vacc11;
121 c1[2] = vacc12;
122 c1[3] = vacc13;
123 c1 = (float*) ((uintptr_t) c1 + cn_stride);
124 c0[0] = vacc00;
125 c0[1] = vacc01;
126 c0[2] = vacc02;
127 c0[3] = vacc03;
128 c0 = (float*) ((uintptr_t) c0 + cn_stride);
129
130 a = (const float**restrict) ((uintptr_t) a - ks);
131 nc -= 4;
132 } else {
133 if (nc & 2) {
134 c1[0] = vacc10;
135 c1[1] = vacc11;
136 vacc10 = vacc12;
137 c1 += 2;
138 c0[0] = vacc00;
139 c0[1] = vacc01;
140 vacc00 = vacc02;
141 c0 += 2;
142 }
143 if (nc & 1) {
144 c1[0] = vacc10;
145 c0[0] = vacc00;
146 }
147
148 nc = 0;
149 }
150 } while (nc != 0);
151 }
152