rublas 0.2.6

Linear algebra and matrix multiplication for Ruda.
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
231
232
233
234
235
236
237
238

use ruda_kernel::dsl as kernel_dsl;
use crate::kernel_ir::launch::{
    ConcreteInputsFactory, ConcreteOutputFactory, InputArg, MatmulArgs, OutputArg, TensorArgs,
    TensorMapArgs,
};
use crate::kernel_ir::routines::{BlueprintStrategy, Routine};
use crate::kernel_ir::{
    definition::MatmulProblem,
    definition::{AvailableVectorSizes, MatmulElems, TilingBlueprint},
    definition::{MatmulAvailabilityError, MatmulSetupError},
    launch::launch_kernel_concrete,
};
use ruda_kernel::dsl::ir::features::TypeUsage;
use ruda_kernel::library::tensor::MatrixBatchLayout;
use ruda_kernel::library::tensor::matrix_batch_layout;
use ruda_kernel::dsl::Runtime;
use ruda_kernel::dsl::client::ComputeClient;
use ruda_kernel::dsl::frontend::TensorBinding;
use ruda_kernel::tiling::InputBinding;

/// Launch a matrix multiplication kernel.
///
/// Cmma will be used if available and enabled,
/// otherwise it will fall back on a non-cmma implementation
#[allow(clippy::result_large_err)]
pub fn prepare_ref<R: Runtime, A: Routine<()>>(
    client: &ComputeClient<R>, lhs: InputBinding<R>, rhs: InputBinding<R>,
    out: TensorBinding<R>, strategy: &BlueprintStrategy<(), A>, dtypes: &MatmulElems,
) -> Result<ruda_kernel::dsl::prelude::PreparedKernel<R>, MatmulSetupError>
where InputArg<TensorArgs>: ConcreteInputsFactory<A>, OutputArg<TensorArgs>: ConcreteOutputFactory<A> {
    use crate::kernel_ir::{components::batch::BatchMatmulFamily, definition::ruda_mapping_launch};
    if lhs.scheme().is_some() || rhs.scheme().is_some()
        || matrix_batch_layout(&lhs.data().strides, None)==MatrixBatchLayout::HighlyPermuted
        || matrix_batch_layout(&rhs.data().strides, None)==MatrixBatchLayout::HighlyPermuted {
        return Err(MatmulSetupError::InvalidConfig(Box::new("prepared matmul requires directly bindable unquantized layouts")));
    }
    let address=lhs.required_address_type().max(rhs.required_address_type())
        .max(out.required_address_type(dtypes.acc_global.size()));
    let problem=MatmulProblem::from_shapes_and_strides(lhs.shape().into(),rhs.shape().into(),out.shape.clone(),
        lhs.data().strides.clone(),rhs.data().strides.clone(),out.strides.clone(),
        dtypes.as_global_elems(),address,None,None)?;
    for dtype in [dtypes.lhs_global,dtypes.rhs_global,dtypes.acc_global] {
        if !client.properties().features.type_usage(dtype).contains(TypeUsage::Conversion) {
            return Err(MatmulSetupError::Unavailable(MatmulAvailabilityError::TypesUnavailable {
                lhs:dtypes.lhs_global,rhs:dtypes.rhs_global,output:dtypes.acc_global}));
        }
    }
    let vectors=AvailableVectorSizes::from_type_sizes(client,lhs.data_elem_size(),rhs.data_elem_size(),dtypes.acc_global.size())
        .filter_lhs_with_tensor(&problem.lhs_strides,&problem.lhs_shape,problem.lhs_layout)
        .filter_rhs_with_tensor(&problem.rhs_strides,&problem.rhs_shape,problem.rhs_layout)
        .filter_out_with_tensor(&problem.out_strides,&problem.out_shape).pick_max()?;
    let settings=A::device_settings(client,vectors);
    let expanded=A::expand_blueprint(&problem,&settings,strategy)?;
    let info=A::prepare(&problem,&settings,expanded)?;
    let input=<InputArg<TensorArgs> as ConcreteInputsFactory<A>>::create(lhs,rhs,&info.blueprint,&problem,&info.vector_sizes,&info.dtypes);
    let output=<OutputArg<TensorArgs> as ConcreteOutputFactory<A>>::create(out,&info.blueprint,&problem,&info.vector_sizes,&info.dtypes);
    A::BatchMatmul::prepare_kernel::<TensorArgs,R>(client,info.ruda_dim,info.ruda_count_plan.resolve(),info.address_type,
        input,output,(),ruda_mapping_launch(&info.ruda_count_plan),info.blueprint,&info.dtypes,&info.vector_sizes)
        .ok_or_else(||MatmulSetupError::InvalidConfig(Box::new("matmul family does not support graph preparation")))
}

#[allow(clippy::result_large_err)]
pub fn launch_ref<R: Runtime, A: Routine<()>>(
    client: &ComputeClient<R>,
    lhs: InputBinding<R>,
    rhs: InputBinding<R>,
    out: TensorBinding<R>,
    blueprint_strategy: &BlueprintStrategy<(), A>,
    dtypes: &mut MatmulElems,
) -> Result<(), MatmulSetupError> {
    let lhs = if matrix_batch_layout(&lhs.data().strides, lhs.scheme())
        == MatrixBatchLayout::HighlyPermuted
    {
        lhs.into_contiguous(client)?
    } else {
        lhs
    };

    let rhs = if matrix_batch_layout(&rhs.data().strides, rhs.scheme())
        == MatrixBatchLayout::HighlyPermuted
    {
        rhs.into_contiguous(client)?
    } else {
        rhs
    };

    let vector_sizes = AvailableVectorSizes::from_type_sizes(
        client,
        lhs.data_elem_size(),
        rhs.data_elem_size(),
        dtypes.acc_global.size(),
    );
    launch_inner_ref::<R, TensorArgs, A>(
        client,
        lhs,
        rhs,
        out,
        blueprint_strategy,
        vector_sizes,
        dtypes,
    )
}

/// Launch a matrix multiplication kernel, with TMA restrictions enabled.
/// TMA doesn't support permuted batches, so checks are slightly different.
///
/// Cmma will be used if available and enabled,
/// otherwise it will fall back on a non-cmma implementation
#[allow(clippy::result_large_err)]
pub fn launch_ref_tma<R: Runtime, A: Routine<(), Blueprint = TilingBlueprint>>(
    client: &ComputeClient<R>,
    lhs: InputBinding<R>,
    rhs: InputBinding<R>,
    out: TensorBinding<R>,
    blueprint_strategy: &BlueprintStrategy<(), A>,
    dtypes: &mut MatmulElems,
) -> Result<(), MatmulSetupError> {
    let lhs = match matrix_batch_layout(&lhs.data().strides, lhs.scheme()) {
        MatrixBatchLayout::Contiguous
        | MatrixBatchLayout::MildlyPermuted {
            transposed: _,
            batch_swap: false,
        } => lhs,
        MatrixBatchLayout::MildlyPermuted {
            transposed: _,
            batch_swap: true,
        }
        | MatrixBatchLayout::HighlyPermuted => lhs.into_contiguous(client)?,
    };

    let rhs = match matrix_batch_layout(&rhs.data().strides, rhs.scheme()) {
        MatrixBatchLayout::Contiguous
        | MatrixBatchLayout::MildlyPermuted {
            transposed: _,
            batch_swap: false,
        } => rhs,
        MatrixBatchLayout::MildlyPermuted {
            transposed: _,
            batch_swap: true,
        }
        | MatrixBatchLayout::HighlyPermuted => rhs.into_contiguous(client)?,
    };

    let vector_sizes = AvailableVectorSizes::from_type_size_tma(client, dtypes.acc_global.size());
    launch_inner_ref::<R, TensorMapArgs, A>(
        client,
        lhs,
        rhs,
        out,
        blueprint_strategy,
        vector_sizes,
        dtypes,
    )
}

#[allow(clippy::result_large_err, clippy::too_many_arguments)]
fn launch_inner_ref<R: Runtime, MA: MatmulArgs<Config = ()>, A: Routine<()>>(
    client: &ComputeClient<R>,
    lhs: InputBinding<R>,
    rhs: InputBinding<R>,
    out: TensorBinding<R>,
    blueprint_strategy: &BlueprintStrategy<(), A>,
    vector_sizes: AvailableVectorSizes,
    dtypes: &mut MatmulElems,
) -> Result<(), MatmulSetupError>
where
    InputArg<MA>: ConcreteInputsFactory<A>,
    OutputArg<MA>: ConcreteOutputFactory<A>,
{
    let address_type = lhs
        .required_address_type()
        .max(rhs.required_address_type())
        .max(out.required_address_type(dtypes.acc_global.size()));

    let problem = MatmulProblem::from_shapes_and_strides(
        lhs.shape().into(),
        rhs.shape().into(),
        out.shape.clone(),
        lhs.data().strides.clone(),
        rhs.data().strides.clone(),
        out.strides.clone(),
        dtypes.as_global_elems(),
        address_type,
        lhs.scheme(),
        rhs.scheme(),
    )?;

    if !client
        .properties()
        .features
        .type_usage(dtypes.lhs_global)
        .contains(TypeUsage::Conversion)
        || !client
            .properties()
            .features
            .type_usage(dtypes.rhs_global)
            .contains(TypeUsage::Conversion)
        || !client
            .properties()
            .features
            .type_usage(dtypes.acc_global)
            .contains(TypeUsage::Conversion)
    {
        return Err(MatmulSetupError::Unavailable(
            MatmulAvailabilityError::TypesUnavailable {
                lhs: dtypes.lhs_global,
                rhs: dtypes.rhs_global,
                output: dtypes.acc_global,
            },
        ));
    }

    let mut vector_sizes = vector_sizes
        .filter_lhs_with_tensor(&problem.lhs_strides, &problem.lhs_shape, problem.lhs_layout)
        .filter_rhs_with_tensor(&problem.rhs_strides, &problem.rhs_shape, problem.rhs_layout)
        .filter_out_with_tensor(&problem.out_strides, &problem.out_shape)
        .pick_max()?;

    // The large vector size resulting from dequantizing ends up slower due to restrictions on
    // algorithms. Use this as a quick and dirty fix.
    if lhs.scale().is_some() {
        vector_sizes.lhs = 1;
    }
    if rhs.scale().is_some() {
        vector_sizes.rhs = 1;
    }

    launch_kernel_concrete::<MA, R, A>(
        client,
        lhs,
        rhs,
        out,
        problem,
        vector_sizes,
        blueprint_strategy,
        dtypes,
    )
}