sim-lib-compute-wgpu 0.1.0

Evidence-based wgpu tensor compute site discovery for SIM.
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
230

use super::*;

#[test]
fn portable_reductions_match_cpu_and_handle_zero_extent() {
    let mut cx = test_cx();
    let executor = WgpuTensorExecutor::new(
        WgpuDiscovery::from_probes(vec![adapter("alpha", "Vulkan", true)], Vec::new()).adapters[0]
            .clone(),
    );
    let source = tensor(&mut cx, vec![2, 3], &["3", "-1", "5", "7", "0.5", "2"]);
    let dtype = Symbol::qualified("numbers", "f32");

    for symbol in [
        sum_op_symbol(),
        min_op_symbol(),
        max_op_symbol(),
        norm_op_symbol(),
    ] {
        let wgpu = execute_wgpu(
            &mut cx,
            &executor,
            symbol.clone(),
            vec![source.clone()],
            Vec::new(),
            dtype.clone(),
        );
        let cpu = execute_cpu(
            &mut cx,
            symbol,
            vec![source.clone()],
            Vec::new(),
            dtype.clone(),
        );
        assert_same_f32_cells(&wgpu, &cpu);
    }

    let empty = tensor(&mut cx, vec![0, 3], &[]);
    let zero_sum = execute_wgpu(
        &mut cx,
        &executor,
        sum_op_symbol(),
        vec![empty.clone()],
        Vec::new(),
        dtype.clone(),
    );
    let zero_norm = execute_wgpu(
        &mut cx,
        &executor,
        norm_op_symbol(),
        vec![empty],
        Vec::new(),
        dtype,
    );
    assert_eq!(f32_cells(&zero_sum), vec![0.0]);
    assert_eq!(f32_cells(&zero_norm), vec![0.0]);
}

#[test]
fn portable_transpose_dot_and_matmul_match_cpu() {
    let mut cx = test_cx();
    let executor = WgpuTensorExecutor::new(
        WgpuDiscovery::from_probes(vec![adapter("alpha", "Vulkan", true)], Vec::new()).adapters[0]
            .clone(),
    );
    let dtype = Symbol::qualified("numbers", "f32");
    let matrix = tensor(&mut cx, vec![2, 3], &["1", "2", "3", "4", "5", "6"]);
    let transposed = execute_wgpu(
        &mut cx,
        &executor,
        transpose_exec_op_symbol(),
        vec![matrix.clone()],
        vec![3, 2],
        dtype.clone(),
    );
    assert_same_f32_cells(
        &transposed,
        &execute_cpu(
            &mut cx,
            transpose_exec_op_symbol(),
            vec![matrix.clone()],
            vec![3, 2],
            dtype.clone(),
        ),
    );

    let left_vec = tensor(&mut cx, vec![3], &["1", "2", "3"]);
    let right_vec = tensor(&mut cx, vec![3], &["4", "5", "6"]);
    let dot = execute_wgpu(
        &mut cx,
        &executor,
        dot_op_symbol(),
        vec![left_vec.clone(), right_vec.clone()],
        Vec::new(),
        dtype.clone(),
    );
    assert_same_f32_cells(
        &dot,
        &execute_cpu(
            &mut cx,
            dot_op_symbol(),
            vec![left_vec.clone(), right_vec.clone()],
            Vec::new(),
            dtype.clone(),
        ),
    );

    let right_matrix = tensor(&mut cx, vec![3, 2], &["7", "8", "9", "10", "11", "12"]);
    let matmul = execute_wgpu(
        &mut cx,
        &executor,
        matmul_exec_op_symbol(),
        vec![matrix.clone(), right_matrix.clone()],
        vec![2, 2],
        dtype.clone(),
    );
    assert_same_f32_cells(
        &matmul,
        &execute_cpu(
            &mut cx,
            matmul_exec_op_symbol(),
            vec![matrix, right_matrix],
            vec![2, 2],
            dtype.clone(),
        ),
    );

    let matrix_vector_left = tensor(&mut cx, vec![2, 3], &["1", "2", "3", "4", "5", "6"]);
    let matrix_vector = execute_wgpu(
        &mut cx,
        &executor,
        matmul_exec_op_symbol(),
        vec![matrix_vector_left.clone(), left_vec.clone()],
        vec![2],
        dtype.clone(),
    );
    assert_same_f32_cells(
        &matrix_vector,
        &execute_cpu(
            &mut cx,
            matmul_exec_op_symbol(),
            vec![matrix_vector_left, left_vec],
            vec![2],
            dtype,
        ),
    );
}

#[test]
fn portable_linalg_handles_aliases_segments_oom_and_half_accumulate() {
    let mut cx = test_cx();
    let mut small_binding = limits(128);
    small_binding.max_storage_buffer_binding_size = 8;
    let executor = WgpuTensorExecutor::new(
        WgpuDiscovery::from_probes(
            vec![adapter_with_limits("alpha", "Vulkan", true, small_binding)],
            Vec::new(),
        )
        .adapters[0]
            .clone(),
    );
    let dtype = Symbol::qualified("numbers", "f32");
    let alias = tensor(&mut cx, vec![3], &["1", "2", "3"]);
    let alias_dot = execute_wgpu(
        &mut cx,
        &executor,
        dot_op_symbol(),
        vec![alias.clone(), alias],
        Vec::new(),
        dtype.clone(),
    );
    assert_eq!(f32_cells(&alias_dot), vec![14.0]);

    let segment_left = tensor(&mut cx, vec![2, 2], &["1", "2", "3", "4"]);
    let segment_right = tensor(&mut cx, vec![2, 3], &["5", "6", "7", "8", "9", "10"]);
    let segmented = execute_wgpu(
        &mut cx,
        &executor,
        matmul_exec_op_symbol(),
        vec![segment_left, segment_right],
        vec![2, 3],
        dtype,
    );
    assert_eq!(resident_storage(&segmented).segments().len(), 3);
    assert_eq!(
        f32_cells(&segmented),
        vec![21.0, 24.0, 27.0, 47.0, 54.0, 61.0]
    );

    let f16_dtype = Symbol::qualified("numbers", "f16");
    let half_left = tensor(&mut cx, vec![2], &["1.5", "2.5"]);
    let half_right = tensor(&mut cx, vec![2], &["2", "4"]);
    let half = execute_wgpu(
        &mut cx,
        &executor,
        dot_op_symbol(),
        vec![half_left, half_right],
        Vec::new(),
        f16_dtype,
    );
    assert_eq!(
        parse_f16_literal_cell(&half.cells().unwrap()[0])
            .unwrap()
            .to_f32(),
        half::f16::from_f32(13.0).to_f32()
    );

    let tiny = WgpuTensorExecutor::new(
        WgpuDiscovery::from_probes(
            vec![adapter_with_limits("tiny", "Vulkan", true, limits(4))],
            Vec::new(),
        )
        .adapters[0]
            .clone(),
    );
    let op = TensorOp::without_attributes(&mut cx, add_op_symbol()).unwrap();
    let tiny_left = tensor(&mut cx, vec![2], &["1", "2"]);
    let tiny_right = tensor(&mut cx, vec![2], &["3", "4"]);
    let result = tiny.execute(
        &mut cx,
        TensorRequest::new(
            op,
            vec![tiny_left, tiny_right],
            TensorMeta::new(vec![2], Symbol::qualified("numbers", "f32")),
        ),
    );
    match result {
        Ok(_) => panic!("tiny wgpu profile must reject the submission"),
        Err(error) => assert!(error.to_string().contains("byte limit")),
    }
}