• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #include <assert.h>
7 
8 #include <xnnpack/dwconv.h>
9 #include <xnnpack/math.h>
10 
11 
xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar(size_t m,size_t n,const float * input,const float * weights,float * output,size_t input_tuple_stride,size_t output_tuple_stride,size_t input_width_stride,size_t output_width_stride,const union xnn_f32_spchw_params params[restrict static1])12 void xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar(
13     size_t m,
14     size_t n,
15     const float* input,
16     const float* weights,
17     float* output,
18     size_t input_tuple_stride,
19     size_t output_tuple_stride,
20     size_t input_width_stride,
21     size_t output_width_stride,
22     const union xnn_f32_spchw_params params[restrict static 1])
23 {
24   assert(n != 0);
25 
26   const float params_max = params->scalar.max;
27   const float params_min = params->scalar.min;
28 
29   const size_t input_width_increment = input_width_stride * 2 - (1 + 2 * ((n - 1) / 2)) * input_tuple_stride;
30   const size_t output_width_increment = output_width_stride - (n - 1) / 2 * output_tuple_stride;
31 
32   // No vertical padding.
33   const float* i0 = input;
34   const float* i1 = (const float*) ((uintptr_t) i0 + input_width_stride);
35   const float* i2 = (const float*) ((uintptr_t) i1 + input_width_stride);
36   const float* i3 = (const float*) ((uintptr_t) i2 + input_width_stride);
37   const float* i4 = (const float*) ((uintptr_t) i3 + input_width_stride);
38 
39   float* output0 = output;
40 
41   // this almost certainly will use too many scalar registers
42   // hope the compiler is good at spilling...
43   const float vw0 = weights[0];
44   const float vw1 = weights[1];
45   const float vw2 = weights[2];
46   const float vw3 = weights[3];
47   const float vw4 = weights[4];
48   const float vw5 = weights[5];
49   const float vw6 = weights[6];
50   const float vw7 = weights[7];
51   const float vw8 = weights[8];
52   const float vw9 = weights[9];
53   const float vw10 = weights[10];
54   const float vw11 = weights[11];
55   const float vw12 = weights[12];
56   const float vw13 = weights[13];
57   const float vw14 = weights[14];
58   const float vw15 = weights[15];
59   const float vw16 = weights[16];
60   const float vw17 = weights[17];
61   const float vw18 = weights[18];
62   const float vw19 = weights[19];
63   const float vw20 = weights[20];
64   const float vw21 = weights[21];
65   const float vw22 = weights[22];
66   const float vw23 = weights[23];
67   const float vw24 = weights[24];
68   const float vw25 = weights[25];
69 
70   do {
71     float vi0x0 = 0.0f;
72     float vi1x0 = 0.0f;
73     float vi2x0 = 0.0f;
74     float vi3x0 = 0.0f;
75     float vi4x0 = 0.0f;
76     float vi0x1 = 0.0f;
77     float vi1x1 = 0.0f;
78     float vi2x1 = 0.0f;
79     float vi3x1 = 0.0f;
80     float vi4x1 = 0.0f;
81     float vi0x2 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride);
82     float vi1x2 = *i1; i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride);
83     float vi2x2 = *i2; i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride);
84     float vi3x2 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride);
85     float vi4x2 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride);
86 
87 
88     size_t k = n;
89     for (; k > 2; k -= 2) {
90       const float vi0x3 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride);
91       const float vi1x3 = *i1; i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride);
92       const float vi2x3 = *i2; i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride);
93       const float vi3x3 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride);
94       const float vi4x3 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride);
95 
96       const float vi0x4 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride);
97       const float vi1x4 = *i1; i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride);
98       const float vi2x4 = *i2; i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride);
99       const float vi3x4 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride);
100       const float vi4x4 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride);
101 
102       const float vrow0_accum = vw1  * vi0x0 + vw2  * vi0x1 + vw3  * vi0x2 + vw4  * vi0x3 + vw5  * vi0x4;
103       vi0x0 = vi0x2;
104       vi0x1 = vi0x3;
105       vi0x2 = vi0x4;
106       const float vrow1_accum = vw6  * vi1x0 + vw7  * vi1x1 + vw8  * vi1x2 + vw9  * vi1x3 + vw10 * vi1x4;
107       vi1x0 = vi1x2;
108       vi1x1 = vi1x3;
109       vi1x2 = vi1x4;
110       const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3 + vw15 * vi2x4;
111       vi2x0 = vi2x2;
112       vi2x1 = vi2x3;
113       vi2x2 = vi2x4;
114       const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3 + vw20 * vi3x4;
115       vi3x0 = vi3x2;
116       vi3x1 = vi3x3;
117       vi3x2 = vi3x4;
118       const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3 + vw25 * vi4x4;
119       vi4x0 = vi4x2;
120       vi4x1 = vi4x3;
121       vi4x2 = vi4x4;
122 
123       float voutput = (vw0 + vrow0_accum) + (vrow1_accum + vrow2_accum) + (vrow3_accum + vrow4_accum);
124 
125       voutput = math_max_f32(voutput, params_min);
126       voutput = math_min_f32(voutput, params_max);
127 
128       *output0 = voutput; output0 = (float*) ((uintptr_t) output0 + output_tuple_stride);
129     }
130     if XNN_UNPREDICTABLE(k == 2) {
131       const float vi0x3 = *i0;
132       const float vi1x3 = *i1;
133       const float vi2x3 = *i2;
134       const float vi3x3 = *i3;
135       const float vi4x3 = *i4;
136 
137       const float vrow0_accum = vw1  * vi0x0 + vw2  * vi0x1 + vw3  * vi0x2 + vw4  * vi0x3;
138       const float vrow1_accum = vw6  * vi1x0 + vw7  * vi1x1 + vw8  * vi1x2 + vw9  * vi1x3;
139       const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3;
140       const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3;
141       const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3;
142 
143       float voutput = (vw0 + vrow0_accum) + (vrow1_accum + vrow2_accum) + (vrow3_accum + vrow4_accum);
144 
145       voutput = math_max_f32(voutput, params_min);
146       voutput = math_min_f32(voutput, params_max);
147 
148       *output0 = voutput;
149     }
150     else {
151       const float vrow0_accum = vw1  * vi0x0 + vw2  * vi0x1 + vw3  * vi0x2;
152       const float vrow1_accum = vw6  * vi1x0 + vw7  * vi1x1 + vw8  * vi1x2;
153       const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2;
154       const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2;
155       const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2;
156 
157       float voutput = (vw0 + vrow0_accum) + (vrow1_accum + vrow2_accum) + (vrow3_accum + vrow4_accum);
158 
159       voutput = math_max_f32(voutput, params_min);
160       voutput = math_min_f32(voutput, params_max);
161 
162       *output0 = voutput;
163     }
164 
165     i0 = (const float*) ((uintptr_t) i0 + input_width_increment);
166     i1 = (const float*) ((uintptr_t) i1 + input_width_increment);
167     i2 = (const float*) ((uintptr_t) i2 + input_width_increment);
168     i3 = (const float*) ((uintptr_t) i3 + input_width_increment);
169     i4 = (const float*) ((uintptr_t) i4 + input_width_increment);
170     output0 = (float*) ((uintptr_t) output0 + output_width_increment);
171     m -= 1;
172   } while (m > 0);
173 }
174