hodu_cuda_kernels 0.2.4

hodu cuda kernels
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217

use hodu_cuda_kernels::{compat::*, kernel::Kernels, kernels::*};

fn device() -> Arc<cudarc::driver::CudaContext> {
    cudarc::driver::CudaContext::new(0).unwrap()
}

fn kernels() -> Kernels {
    Kernels::new()
}

#[test]
fn concat_f32_dim0() {
    let kernels = kernels();

    let device = device();
    let stream = device.default_stream();

    // Two 2x2 matrices concatenated along dim 0 to form 4x2
    // Input 1: [[1, 2], [3, 4]]
    // Input 2: [[5, 6], [7, 8]]
    // Output: [[1, 2], [3, 4], [5, 6], [7, 8]]
    let input = vec![1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let input_dev = stream.memcpy_stod(&input).unwrap();
    let mut output: cudarc::driver::CudaSlice<f32> = unsafe { stream.alloc(8).unwrap() };

    let output_shape = vec![4, 2];
    let concat_dim = 0;
    let num_inputs = 2;

    // Input shapes for 2 inputs: [2, 2, 2, 2] (flattened)
    let input_shapes = vec![2, 2, 2, 2];
    // Input strides: [2, 1, 2, 1]
    let input_strides = vec![2, 1, 2, 1];
    // Tensor-level offsets (usually 0): [0, 0]
    let input_offsets = vec![0, 0];
    // Buffer offsets (where each tensor starts in physical buffer): [0, 4]
    let input_buffer_offsets = vec![0, 4];

    // Build metadata
    let num_els = 8;
    let num_dims = 2;
    let mut metadata = Vec::new();
    metadata.push(num_els);
    metadata.push(num_dims);
    metadata.extend(&output_shape);
    metadata.push(concat_dim);
    metadata.push(num_inputs);
    metadata.extend(&input_shapes);
    metadata.extend(&input_strides);
    metadata.extend(&input_offsets);
    metadata.extend(&input_buffer_offsets);

    call_ops_concat(concat::F32, &kernels, &device, &input_dev, &mut output, &metadata).unwrap();

    let mut results = vec![0.0f32; 8];
    stream.memcpy_dtoh(&output, &mut results).unwrap();
    assert_eq!(results, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]);
}

#[test]
fn concat_f32_dim1() {
    let kernels = kernels();

    let device = device();
    let stream = device.default_stream();

    // Two 2x2 matrices concatenated along dim 1 to form 2x4
    // Input 1: [[1, 2], [3, 4]]
    // Input 2: [[5, 6], [7, 8]]
    // Output: [[1, 2, 5, 6], [3, 4, 7, 8]]
    let input = vec![1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let input_dev = stream.memcpy_stod(&input).unwrap();
    let mut output: cudarc::driver::CudaSlice<f32> = unsafe { stream.alloc(8).unwrap() };

    let output_shape = vec![2, 4];
    let concat_dim = 1;
    let num_inputs = 2;

    let input_shapes = vec![2, 2, 2, 2];
    let input_strides = vec![2, 1, 2, 1];
    let input_offsets = vec![0, 0];
    let input_buffer_offsets = vec![0, 4];

    let num_els = 8;
    let num_dims = 2;
    let mut metadata = Vec::new();
    metadata.push(num_els);
    metadata.push(num_dims);
    metadata.extend(&output_shape);
    metadata.push(concat_dim);
    metadata.push(num_inputs);
    metadata.extend(&input_shapes);
    metadata.extend(&input_strides);
    metadata.extend(&input_offsets);
    metadata.extend(&input_buffer_offsets);

    call_ops_concat(concat::F32, &kernels, &device, &input_dev, &mut output, &metadata).unwrap();

    let mut results = vec![0.0f32; 8];
    stream.memcpy_dtoh(&output, &mut results).unwrap();
    assert_eq!(results, vec![1.0, 2.0, 5.0, 6.0, 3.0, 4.0, 7.0, 8.0]);
}

#[test]
fn split_f32_dim0() {
    let kernels = kernels();

    let device = device();
    let stream = device.default_stream();

    // Input: 4x2 matrix [[1, 2], [3, 4], [5, 6], [7, 8]]
    // Split along dim 0, taking first 2 rows: [[1, 2], [3, 4]]
    let input = vec![1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let input_dev = stream.memcpy_stod(&input).unwrap();
    let mut output: cudarc::driver::CudaSlice<f32> = unsafe { stream.alloc(4).unwrap() };

    let input_shape = vec![4, 2];
    let strides = vec![2, 1];
    let split_dim = 0;
    let output_size_on_dim = 2; // Take 2 rows
    let split_offset = 0; // Start from beginning

    // Build metadata
    let num_els = 4; // output size
    let num_dims = 2;
    let mut metadata = Vec::new();
    metadata.push(num_els);
    metadata.push(num_dims);
    metadata.extend(&input_shape);
    metadata.extend(&strides);
    metadata.push(0); // input offset
    metadata.push(split_dim);
    metadata.push(output_size_on_dim);
    metadata.push(split_offset);

    call_ops_split(split::F32, &kernels, &device, &input_dev, &mut output, &metadata).unwrap();

    let mut results = vec![0.0f32; 4];
    stream.memcpy_dtoh(&output, &mut results).unwrap();
    assert_eq!(results, vec![1.0, 2.0, 3.0, 4.0]);
}

#[test]
fn split_f32_dim0_offset() {
    let kernels = kernels();

    let device = device();
    let stream = device.default_stream();

    // Input: 4x2 matrix [[1, 2], [3, 4], [5, 6], [7, 8]]
    // Split along dim 0, taking 2 rows starting from offset 2: [[5, 6], [7, 8]]
    let input = vec![1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let input_dev = stream.memcpy_stod(&input).unwrap();
    let mut output: cudarc::driver::CudaSlice<f32> = unsafe { stream.alloc(4).unwrap() };

    let input_shape = vec![4, 2];
    let strides = vec![2, 1];
    let split_dim = 0;
    let output_size_on_dim = 2;
    let split_offset = 2; // Start from row 2

    let num_els = 4;
    let num_dims = 2;
    let mut metadata = Vec::new();
    metadata.push(num_els);
    metadata.push(num_dims);
    metadata.extend(&input_shape);
    metadata.extend(&strides);
    metadata.push(0);
    metadata.push(split_dim);
    metadata.push(output_size_on_dim);
    metadata.push(split_offset);

    call_ops_split(split::F32, &kernels, &device, &input_dev, &mut output, &metadata).unwrap();

    let mut results = vec![0.0f32; 4];
    stream.memcpy_dtoh(&output, &mut results).unwrap();
    assert_eq!(results, vec![5.0, 6.0, 7.0, 8.0]);
}

#[test]
fn split_f32_dim1() {
    let kernels = kernels();

    let device = device();
    let stream = device.default_stream();

    // Input: 2x4 matrix [[1, 2, 3, 4], [5, 6, 7, 8]]
    // Split along dim 1, taking 2 columns: [[1, 2], [5, 6]]
    let input = vec![1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let input_dev = stream.memcpy_stod(&input).unwrap();
    let mut output: cudarc::driver::CudaSlice<f32> = unsafe { stream.alloc(4).unwrap() };

    let input_shape = vec![2, 4];
    let strides = vec![4, 1];
    let split_dim = 1;
    let output_size_on_dim = 2;
    let split_offset = 0;

    let num_els = 4;
    let num_dims = 2;
    let mut metadata = Vec::new();
    metadata.push(num_els);
    metadata.push(num_dims);
    metadata.extend(&input_shape);
    metadata.extend(&strides);
    metadata.push(0);
    metadata.push(split_dim);
    metadata.push(output_size_on_dim);
    metadata.push(split_offset);

    call_ops_split(split::F32, &kernels, &device, &input_dev, &mut output, &metadata).unwrap();

    let mut results = vec![0.0f32; 4];
    stream.memcpy_dtoh(&output, &mut results).unwrap();
    assert_eq!(results, vec![1.0, 2.0, 5.0, 6.0]);
}