ug 0.3.0

Micro compiler for tensor operations.
Documentation 
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
218
219
220
221
222
223
224
225
226
227
228
229

use ug::{Device, Slice};
use ug::{LazyBuffer as LB, Result};

#[test]
fn schedule_interpret() -> Result<()> {
    let cpu = ug::CpuDevice;
    let lhs = LB::cst(40., (5, 2), &cpu)?;
    let rhs = LB::cst(2., (5, 2), &cpu)?;
    let lb = lhs.binary(ug::lang::BinaryOp::Add, rhs)?;
    let schedule = ug::Schedule::create_one(&lb)?;
    let items = schedule.items();
    assert_eq!(items.len(), 1);
    for item in items.iter() {
        let kernel = match item {
            ug::ScheduleItem::Kernel(k) => k.kernel()?,
            _ => ug::bail!("unexpected item"),
        };
        let ssa = kernel.lower(&Default::default())?;
        println!("{ssa:?}");
        let mut buffer = ug::interpreter::Buffer::F32(vec![0f32; 10]);
        let args = vec![];
        ug::interpreter::eval_ssa::<1>(&ssa, vec![&mut buffer], &args, 0)?;
        let buffer = buffer.as_f32()?;
        assert_eq!(buffer, [42., 42., 42., 42., 42., 42., 42., 42., 42., 42.]);
    }
    Ok(())
}

#[test]
fn schedule_cpu() -> Result<()> {
    let cpu = ug::CpuDevice;
    let lhs = LB::cst(40., (5, 2), &cpu)?;
    let rhs = LB::cst(2., (5, 2), &cpu)?;
    let lb = lhs.binary(ug::lang::BinaryOp::Add, rhs)?;
    let schedule = ug::Schedule::create_one(&lb)?;
    let items = schedule.items();
    assert_eq!(items.len(), 1);
    let kernel = match &items[0] {
        ug::ScheduleItem::Kernel(k) => k.kernel()?,
        _ => ug::bail!("unexpected item"),
    };
    let ssa = kernel.lower(&Default::default())?;
    let func = cpu.compile(&ssa, None)?;
    let mut buffer = unsafe { cpu.allocate_uninit(ug::DType::F32, 10)? };
    cpu.run(&func, &mut [&mut buffer])?;
    let buffer = buffer.to_vec::<f32>()?;
    assert_eq!(buffer, [42., 42., 42., 42., 42., 42., 42., 42., 42., 42.]);
    Ok(())
}

#[test]
fn schedule_cpu_compile() -> Result<()> {
    ug::common_tests::lb::add(&ug::CpuDevice)
}

#[test]
fn schedule_mm() -> Result<()> {
    ug::common_tests::lb::mm(&ug::CpuDevice)
}

#[test]
fn schedule_cat() -> Result<()> {
    ug::common_tests::lb::cat(&ug::CpuDevice)
}

#[test]
fn lb_copy() -> Result<()> {
    let cpu = ug::CpuDevice;
    let shape = (2, 3);
    let buf = [0f32, 1f32, 2f32, 3f32, 4f32, 5f32].as_slice();
    let lhs = LB::copy(buf, shape, &cpu)?;
    let two = LB::cst(2., shape, &cpu)?;
    let half = LB::cst(0.5, shape, &cpu)?;
    let lb = lhs.binary(ug::lang::BinaryOp::Add, two)?;
    let lb = lb.binary(ug::lang::BinaryOp::Mul, half)?;
    let schedule = ug::Schedule::create_one(&lb)?;
    let schedule = schedule.compile()?;
    schedule.run()?;
    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [1.0, 1.5, 2.0, 2.5, 3.0, 3.5]);

    {
        let mut data = lhs.data().try_borrow_mut()?;
        let data = data.as_mut().unwrap();
        if let ug::CpuStorage::F32(vs) = data {
            vs[1] = 0.;
            vs[2] = -8.;
            vs[5] = -2.;
        }
    }

    schedule.run()?;
    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [1.0, 1.0, -3.0, 2.5, 3.0, 0.0]);

    Ok(())
}

#[test]
fn schedule_broadcast() -> Result<()> {
    let cpu = ug::CpuDevice;
    let lhs = LB::cst(40., (5, 2), &cpu)?;
    let rhs = LB::cst(2., (), &cpu)?;
    let rhs = rhs.broadcast((5, 2))?;
    let lb = lhs.binary(ug::lang::BinaryOp::Add, rhs)?;
    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [42., 42., 42., 42., 42., 42., 42., 42., 42., 42.]);
    Ok(())
}

#[test]
fn lb_custom() -> Result<()> {
    let cpu = ug::CpuDevice;
    let add_one = |vs: Vec<&mut ug::CpuStorage>| -> Result<()> {
        let [src, dst]: [&mut ug::CpuStorage; 2] = vs.try_into().unwrap();
        let dst = dst.data_mut::<f32>()?;
        let src = src.data::<f32>()?;
        dst.iter_mut().zip(src.iter()).for_each(|(d, s)| *d = *s + 1.0);
        Ok(())
    };
    let shape = (2, 3);
    let buf = [0f32, 1f32, 2f32, 3f32, 4f32, 5f32].as_slice();
    let buf_lb = LB::copy(buf, shape, &cpu)?;
    let two_lb = LB::cst(2., shape, &cpu)?;
    let lb = LB::custom(add_one, vec![buf_lb.clone()], shape, ug::DType::F32, &cpu)?;
    let lb = lb.binary(ug::lang::BinaryOp::Mul, two_lb)?;

    let schedule = ug::Schedule::create_one(&lb)?;
    let schedule = schedule.compile()?;
    schedule.run()?;

    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [2., 4., 6., 8., 10., 12.]);

    {
        let mut data = buf_lb.data().try_borrow_mut()?;
        let data = data.as_mut().unwrap();
        if let ug::CpuStorage::F32(vs) = data {
            vs[1] = 0.;
            vs[2] = -8.;
            vs[5] = -2.;
        }
    }

    schedule.run()?;
    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [2., 2., -14., 8., 10., -2.]);

    Ok(())
}

#[test]
fn lb_layout() -> Result<()> {
    let cpu = ug::CpuDevice;
    let shape = (2, 3);
    let buf = [0f32, 1f32, 2f32, 3f32, 4f32, 5f32].as_slice();
    let lb = LB::copy(buf, shape, &cpu)?;

    let lb = lb.transpose(0, 1)?;
    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [0.0, 3.0, 1.0, 4.0, 2.0, 5.0]);

    let lb = lb.transpose(0, 1)?;
    let lb = lb.merge_dims(0)?;
    assert_eq!(lb.dims(), [6]);

    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [0.0, 1.0, 2.0, 3.0, 4.0, 5.0]);

    Ok(())
}

#[test]
fn lb_custom_in_place() -> Result<()> {
    let cpu = ug::CpuDevice;
    let add_one = |vs: Vec<&mut ug::CpuStorage>| -> Result<()> {
        let [dst]: [&mut ug::CpuStorage; 1] = vs.try_into().unwrap();
        let dst = dst.data_mut::<f32>()?;
        dst.iter_mut().for_each(|d| *d += 1.0);
        Ok(())
    };
    let shape = (2, 3);
    let buf = [0f32, 1f32, 2f32, 3f32, 4f32, 5f32].as_slice();
    let lb = LB::copy(buf, shape, &cpu)?;
    let lb = lb.custom_ip(add_one, vec![])?;
    let lb = lb.custom_ip(add_one, vec![])?;
    let lb = lb.custom_ip(add_one, vec![])?;
    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [3., 4., 5., 6., 7., 8.]);

    {
        let mut data = lb.data().try_borrow_mut()?;
        let data = data.as_mut().unwrap();
        if let ug::CpuStorage::F32(vs) = data {
            vs[1] = 0.;
            vs[2] = -8.;
            vs[5] = -2.;
        }
    }

    let data = lb.data_vec::<f32>()?;
    assert_eq!(data, [6., 3., -5., 9., 10., 1.]);
    Ok(())
}

#[test]
fn lb_set() -> Result<()> {
    let cpu = ug::CpuDevice;
    let shape = (2, 3);
    let buf = [0f32, 1f32, 2f32, 3f32, 4f32, 5f32].as_slice();
    let dst_lb = LB::copy(buf, shape, &cpu)?;
    let src_lb = dst_lb.binary(ug::lang::BinaryOp::Mul, dst_lb.clone())?;
    let data = src_lb.data_vec::<f32>()?;
    assert_eq!(data, [0.0, 1.0, 4.0, 9.0, 16.0, 25.0]);
    let data = dst_lb.data_vec::<f32>()?;
    assert_eq!(data, [0., 1., 2., 3., 4., 5.]);
    let dst_lb = dst_lb.set(src_lb)?;
    // TODO: the current behavior is very error prone as if realize is not called, the
    // 'set' op is not executed and the data is not modified.
    let data = dst_lb.data_vec::<f32>()?;
    assert_eq!(data, [0.0, 1.0, 4.0, 9.0, 16.0, 25.0]);

    let buf = [3f32, 1f32, 2f32].as_slice();
    let vs = LB::copy(buf, (3,), &cpu)?;
    let dst_lb = dst_lb.set_l(vs, ug::Layout::from_shape(3))?;
    let data = dst_lb.data_vec::<f32>()?;
    assert_eq!(data, [3.0, 1.0, 2.0, 9.0, 16.0, 25.0]);
    Ok(())
}