tenferro-einsum 0.3.0

Subscripts, contraction planning, concrete/traced/eager einsum APIs, extension runtime, and AD rule for tenferro.
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
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

#![cfg(all(feature = "autodiff", feature = "webgpu"))]

use num_complex::Complex32;
use tenferro_ad::{EagerRuntime, EagerTensor};
use tenferro_einsum::{EagerEinsumExt, TraceContextEinsumExt};
use tenferro_gpu::{
    webgpu::download_webgpu_tensor, webgpu::upload_webgpu_tensor, webgpu::webgpu_available,
};
use tenferro_gpu::{webgpu::WebGpuBackend, webgpu::WebGpuRuntime};
use tenferro_ops::dim_expr::DimExpr;
use tenferro_runtime::program::ProgramInputSpec;
use tenferro_runtime::{CompiledGraph, DType, GraphCompiler, Runtime, Tensor, TraceContext};

fn matmul2_col_major(lhs: &[Complex32], rhs: &[Complex32]) -> [Complex32; 4] {
    let a00 = lhs[0];
    let a10 = lhs[1];
    let a01 = lhs[2];
    let a11 = lhs[3];
    let b00 = rhs[0];
    let b10 = rhs[1];
    let b01 = rhs[2];
    let b11 = rhs[3];
    [
        a00 * b00 + a01 * b10,
        a10 * b00 + a11 * b10,
        a00 * b01 + a01 * b11,
        a10 * b01 + a11 * b11,
    ]
}

fn assert_complex_close(actual: &[Complex32], expected: &[Complex32]) {
    for (actual, expected) in actual.iter().zip(expected) {
        assert!(
            (actual.re - expected.re).abs() <= 1e-4,
            "real mismatch: actual={actual:?} expected={expected:?}"
        );
        assert!(
            (actual.im - expected.im).abs() <= 1e-4,
            "imag mismatch: actual={actual:?} expected={expected:?}"
        );
    }
}

fn batched_matmul_f32_reference() -> Vec<f32> {
    vec![58.0, 139.0, 64.0, 154.0, 5800.0, 13900.0, 6400.0, 15400.0]
}

fn assert_f32_close(actual: &[f32], expected: &[f32]) {
    for (actual, expected) in actual.iter().zip(expected) {
        assert!(
            (actual - expected).abs() <= 1e-4,
            "f32 mismatch: actual={actual} expected={expected}"
        );
    }
}

fn compile_einsum(dtype: DType, shapes: &[&[usize]], subscripts: &str) -> CompiledGraph {
    let mut trace = TraceContext::new();
    let inputs = shapes
        .iter()
        .map(|shape| {
            trace
                .input(ProgramInputSpec::new(dtype, DimExpr::from_concrete(shape)))
                .unwrap()
        })
        .collect::<Vec<_>>();
    let output = trace.einsum(&inputs, subscripts).unwrap();
    let graph = trace.finish(&[output]).unwrap();
    GraphCompiler::new().compile_traced_graph(&graph).unwrap()
}

fn webgpu_runtime_with_einsum(backend: &WebGpuBackend) -> Runtime {
    let mut builder = Runtime::builder();
    builder
        .register_engine(tenferro_gpu::webgpu::webgpu_runtime_engine_registration(backend).unwrap())
        .unwrap();
    builder
        .install_extension_module(
            tenferro_einsum::extension_module::<WebGpuBackend>(
                tenferro_gpu::webgpu::webgpu_runtime_engine_id().unwrap(),
            )
            .unwrap(),
        )
        .unwrap();
    builder.build().unwrap()
}

#[test]
fn eager_tensor_einsum_runs_rank2_f32_matmul_on_webgpu_when_adapter_available() {
    if !webgpu_available() {
        return;
    }

    let runtime = WebGpuRuntime::new_default().unwrap();
    let ctx =
        EagerRuntime::with_webgpu_backend(WebGpuBackend::from_runtime(runtime.clone())).unwrap();
    let lhs =
        Tensor::from_vec_col_major(vec![2, 3], vec![1.0_f32, 4.0, 2.0, 5.0, 3.0, 6.0]).unwrap();
    let rhs =
        Tensor::from_vec_col_major(vec![3, 2], vec![7.0_f32, 9.0, 11.0, 8.0, 10.0, 12.0]).unwrap();
    let lhs =
        EagerTensor::from_tensor_in(upload_webgpu_tensor(&runtime, &lhs).unwrap(), ctx.clone())
            .unwrap();
    let rhs =
        EagerTensor::from_tensor_in(upload_webgpu_tensor(&runtime, &rhs).unwrap(), ctx).unwrap();

    let out = [&lhs, &rhs].einsum("ij,jk->ik").unwrap();
    let host = download_webgpu_tensor(&runtime, out.to_tensor().unwrap()).unwrap();

    assert_eq!(host.shape(), &[2, 2]);
    let actual = host.as_slice::<f32>().unwrap();
    let expected = [58.0_f32, 139.0, 64.0, 154.0];
    for (actual, expected) in actual.iter().zip(expected) {
        assert!((actual - expected).abs() <= 1e-4);
    }
}

#[test]
fn eager_tensor_einsum_runs_batched_f32_matmul_on_webgpu_when_adapter_available() {
    if !webgpu_available() {
        return;
    }

    let runtime = WebGpuRuntime::new_default().unwrap();
    let ctx =
        EagerRuntime::with_webgpu_backend(WebGpuBackend::from_runtime(runtime.clone())).unwrap();
    let lhs = Tensor::from_vec_col_major(
        vec![2, 3, 2],
        vec![
            1.0_f32, 4.0, 2.0, 5.0, 3.0, 6.0, 10.0, 40.0, 20.0, 50.0, 30.0, 60.0,
        ],
    )
    .unwrap();
    let rhs = Tensor::from_vec_col_major(
        vec![3, 2, 2],
        vec![
            7.0_f32, 9.0, 11.0, 8.0, 10.0, 12.0, 70.0, 90.0, 110.0, 80.0, 100.0, 120.0,
        ],
    )
    .unwrap();
    let lhs =
        EagerTensor::from_tensor_in(upload_webgpu_tensor(&runtime, &lhs).unwrap(), ctx.clone())
            .unwrap();
    let rhs =
        EagerTensor::from_tensor_in(upload_webgpu_tensor(&runtime, &rhs).unwrap(), ctx).unwrap();

    let out = [&lhs, &rhs].einsum("ikb,kjb->ijb").unwrap();
    let host = download_webgpu_tensor(&runtime, out.to_tensor().unwrap()).unwrap();

    assert_eq!(host.shape(), &[2, 2, 2]);
    assert_f32_close(
        host.as_slice::<f32>().unwrap(),
        &batched_matmul_f32_reference(),
    );
}

#[test]
fn eager_tensor_einsum_runs_rank2_c32_matmul_on_webgpu_when_adapter_available() {
    if !webgpu_available() {
        return;
    }

    let runtime = WebGpuRuntime::new_default().unwrap();
    let ctx =
        EagerRuntime::with_webgpu_backend(WebGpuBackend::from_runtime(runtime.clone())).unwrap();
    let lhs_data = vec![
        Complex32::new(1.0, 0.5),
        Complex32::new(2.0, -1.0),
        Complex32::new(3.0, 0.25),
        Complex32::new(4.0, 1.0),
    ];
    let rhs_data = vec![
        Complex32::new(5.0, -0.5),
        Complex32::new(6.0, 0.25),
        Complex32::new(7.0, 1.0),
        Complex32::new(8.0, -0.75),
    ];
    let lhs = Tensor::from_vec_col_major(vec![2, 2], lhs_data.clone()).unwrap();
    let rhs = Tensor::from_vec_col_major(vec![2, 2], rhs_data.clone()).unwrap();
    let lhs =
        EagerTensor::from_tensor_in(upload_webgpu_tensor(&runtime, &lhs).unwrap(), ctx.clone())
            .unwrap();
    let rhs =
        EagerTensor::from_tensor_in(upload_webgpu_tensor(&runtime, &rhs).unwrap(), ctx).unwrap();

    let out = [&lhs, &rhs].einsum("ij,jk->ik").unwrap();
    let host = download_webgpu_tensor(&runtime, out.to_tensor().unwrap()).unwrap();

    assert_eq!(host.shape(), &[2, 2]);
    let actual = host.as_slice::<Complex32>().unwrap();
    let expected = matmul2_col_major(&lhs_data, &rhs_data);
    assert_complex_close(actual, &expected);
}

#[test]
fn traced_einsum_runs_rank2_f32_matmul_on_webgpu_when_adapter_available() {
    if !webgpu_available() {
        return;
    }

    let program = compile_einsum(DType::F32, &[&[2, 3], &[3, 2]], "ij,jk->ik");
    let gpu_runtime = WebGpuRuntime::new_default().unwrap();
    let lhs_host =
        Tensor::from_vec_col_major(vec![2, 3], vec![1.0_f32, 4.0, 2.0, 5.0, 3.0, 6.0]).unwrap();
    let rhs_host =
        Tensor::from_vec_col_major(vec![3, 2], vec![7.0_f32, 9.0, 11.0, 8.0, 10.0, 12.0]).unwrap();
    let lhs_gpu = upload_webgpu_tensor(&gpu_runtime, &lhs_host).unwrap();
    let rhs_gpu = upload_webgpu_tensor(&gpu_runtime, &rhs_host).unwrap();
    let backend = WebGpuBackend::from_runtime(gpu_runtime.clone());
    let exec_runtime = webgpu_runtime_with_einsum(&backend);

    let out = exec_runtime
        .run_compiled(&program, &[&lhs_gpu, &rhs_gpu])
        .unwrap()
        .remove(0);
    backend.synchronize().unwrap();
    let host = download_webgpu_tensor(backend.runtime(), &out).unwrap();

    assert_eq!(host.shape(), &[2, 2]);
    let actual = host.as_slice::<f32>().unwrap();
    let expected = [58.0_f32, 139.0, 64.0, 154.0];
    for (actual, expected) in actual.iter().zip(expected) {
        assert!((actual - expected).abs() <= 1e-4);
    }
}

#[test]
fn traced_einsum_runs_batched_f32_matmul_on_webgpu_when_adapter_available() {
    if !webgpu_available() {
        return;
    }

    let program = compile_einsum(DType::F32, &[&[2, 3, 2], &[3, 2, 2]], "ikb,kjb->ijb");
    let gpu_runtime = WebGpuRuntime::new_default().unwrap();
    let lhs_host = Tensor::from_vec_col_major(
        vec![2, 3, 2],
        vec![
            1.0_f32, 4.0, 2.0, 5.0, 3.0, 6.0, 10.0, 40.0, 20.0, 50.0, 30.0, 60.0,
        ],
    )
    .unwrap();
    let rhs_host = Tensor::from_vec_col_major(
        vec![3, 2, 2],
        vec![
            7.0_f32, 9.0, 11.0, 8.0, 10.0, 12.0, 70.0, 90.0, 110.0, 80.0, 100.0, 120.0,
        ],
    )
    .unwrap();
    let lhs_gpu = upload_webgpu_tensor(&gpu_runtime, &lhs_host).unwrap();
    let rhs_gpu = upload_webgpu_tensor(&gpu_runtime, &rhs_host).unwrap();
    let backend = WebGpuBackend::from_runtime(gpu_runtime.clone());
    let exec_runtime = webgpu_runtime_with_einsum(&backend);

    let out = exec_runtime
        .run_compiled(&program, &[&lhs_gpu, &rhs_gpu])
        .unwrap()
        .remove(0);
    backend.synchronize().unwrap();
    let host = download_webgpu_tensor(backend.runtime(), &out).unwrap();

    assert_eq!(host.shape(), &[2, 2, 2]);
    assert_f32_close(
        host.as_slice::<f32>().unwrap(),
        &batched_matmul_f32_reference(),
    );
}