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_4x4__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_4x4__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 float vacc02 = w[2];
62 float vacc03 = w[3];
63 w += 4;
64 float vacc10 = vacc00;
65 float vacc11 = vacc01;
66 float vacc12 = vacc02;
67 float vacc13 = vacc03;
68 float vacc20 = vacc00;
69 float vacc21 = vacc01;
70 float vacc22 = vacc02;
71 float vacc23 = vacc03;
72 float vacc30 = vacc00;
73 float vacc31 = vacc01;
74 float vacc32 = vacc02;
75 float vacc33 = vacc03;
76
77 size_t k = kc;
78 do {
79 const float va0 = *a0++;
80 const float va1 = *a1++;
81 const float va2 = *a2++;
82 const float va3 = *a3++;
83
84 const float vb0 = w[0];
85 const float vb1 = w[1];
86 const float vb2 = w[2];
87 const float vb3 = w[3];
88 w += 4;
89
90 vacc00 = math_muladd_f32(va0, vb0, vacc00);
91 vacc01 = math_muladd_f32(va0, vb1, vacc01);
92 vacc02 = math_muladd_f32(va0, vb2, vacc02);
93 vacc03 = math_muladd_f32(va0, vb3, vacc03);
94 vacc10 = math_muladd_f32(va1, vb0, vacc10);
95 vacc11 = math_muladd_f32(va1, vb1, vacc11);
96 vacc12 = math_muladd_f32(va1, vb2, vacc12);
97 vacc13 = math_muladd_f32(va1, vb3, vacc13);
98 vacc20 = math_muladd_f32(va2, vb0, vacc20);
99 vacc21 = math_muladd_f32(va2, vb1, vacc21);
100 vacc22 = math_muladd_f32(va2, vb2, vacc22);
101 vacc23 = math_muladd_f32(va2, vb3, vacc23);
102 vacc30 = math_muladd_f32(va3, vb0, vacc30);
103 vacc31 = math_muladd_f32(va3, vb1, vacc31);
104 vacc32 = math_muladd_f32(va3, vb2, vacc32);
105 vacc33 = math_muladd_f32(va3, vb3, vacc33);
106
107 k -= sizeof(float);
108 } while (k != 0);
109
110 vacc00 = math_max_f32(vacc00, 0.0f);
111 vacc01 = math_max_f32(vacc01, 0.0f);
112 vacc02 = math_max_f32(vacc02, 0.0f);
113 vacc03 = math_max_f32(vacc03, 0.0f);
114 vacc10 = math_max_f32(vacc10, 0.0f);
115 vacc11 = math_max_f32(vacc11, 0.0f);
116 vacc12 = math_max_f32(vacc12, 0.0f);
117 vacc13 = math_max_f32(vacc13, 0.0f);
118 vacc20 = math_max_f32(vacc20, 0.0f);
119 vacc21 = math_max_f32(vacc21, 0.0f);
120 vacc22 = math_max_f32(vacc22, 0.0f);
121 vacc23 = math_max_f32(vacc23, 0.0f);
122 vacc30 = math_max_f32(vacc30, 0.0f);
123 vacc31 = math_max_f32(vacc31, 0.0f);
124 vacc32 = math_max_f32(vacc32, 0.0f);
125 vacc33 = math_max_f32(vacc33, 0.0f);
126
127 if XNN_LIKELY(nc >= 4) {
128 c3[0] = vacc30;
129 c3[1] = vacc31;
130 c3[2] = vacc32;
131 c3[3] = vacc33;
132 c3 = (float*) ((uintptr_t) c3 + cn_stride);
133 c2[0] = vacc20;
134 c2[1] = vacc21;
135 c2[2] = vacc22;
136 c2[3] = vacc23;
137 c2 = (float*) ((uintptr_t) c2 + cn_stride);
138 c1[0] = vacc10;
139 c1[1] = vacc11;
140 c1[2] = vacc12;
141 c1[3] = vacc13;
142 c1 = (float*) ((uintptr_t) c1 + cn_stride);
143 c0[0] = vacc00;
144 c0[1] = vacc01;
145 c0[2] = vacc02;
146 c0[3] = vacc03;
147 c0 = (float*) ((uintptr_t) c0 + cn_stride);
148
149 a3 = (const void*) ((uintptr_t) a3 - kc);
150 a2 = (const void*) ((uintptr_t) a2 - kc);
151 a1 = (const void*) ((uintptr_t) a1 - kc);
152 a0 = (const void*) ((uintptr_t) a0 - kc);
153
154 nc -= 4;
155 } else {
156 if (nc & 2) {
157 c3[0] = vacc30;
158 c3[1] = vacc31;
159 vacc30 = vacc32;
160 c3 += 2;
161 c2[0] = vacc20;
162 c2[1] = vacc21;
163 vacc20 = vacc22;
164 c2 += 2;
165 c1[0] = vacc10;
166 c1[1] = vacc11;
167 vacc10 = vacc12;
168 c1 += 2;
169 c0[0] = vacc00;
170 c0[1] = vacc01;
171 vacc00 = vacc02;
172 c0 += 2;
173 }
174 if (nc & 1) {
175 c3[0] = vacc30;
176 c2[0] = vacc20;
177 c1[0] = vacc10;
178 c0[0] = vacc00;
179 }
180
181 nc = 0;
182 }
183 } while (nc != 0);
184 }
185