ruTENSOR 0.1.29

Ruda tensor contractions, reductions, permutations and elementwise 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

use std::collections::BTreeMap;
use crate::{Error, Mode, OperationDescriptor, Result};
use crate::descriptor::product;
use crate::kernel::{Config, Launch, convert_operand};
use crate::operation::Kind;
use ruda_core::{device::Device, tensor::{Metadata, Shape}};
use ruda_kernel::{dsl::Runtime, tensor::RudaTensor};

/// A reusable index map and execution description, independent of buffer addresses.
#[derive(Clone, Debug)]
pub struct Plan {
    operation: OperationDescriptor,
    modes: Vec<Mode>,
    reduction_extents: Vec<usize>,
    reduction_count: usize,
    /// Logical mode index of every physical axis of each operand.
    axis_modes: Vec<Vec<usize>>,
}

pub(crate) fn merge_extent(extents: &mut BTreeMap<Mode, usize>, mode: Mode, value: usize) -> Result<()> {
    if let Some(&previous) = extents.get(&mode) {
        if previous != value && previous != 1 && value != 1 {
            return Err(Error::IncompatibleExtent { mode, left: previous, right: value });
        }
        // Broadcasting 0 with 1 produces an empty dimension, not a size-one dimension.
        extents.insert(mode, if previous == 1 { value } else { previous });
    } else {
        extents.insert(mode, value);
    }
    Ok(())
}

impl Plan {
    /// Resolve free, broadcast, diagonal and reduction axes without executing a kernel.
    pub fn new(operation: OperationDescriptor) -> Result<Self> {
        let mut extents = BTreeMap::new();
        for input in &operation.inputs {
            for (&mode, &extent) in input.modes.iter().zip(&input.tensor.extents) {
                merge_extent(&mut extents, mode, extent)?;
            }
        }
        for (&mode, &extent) in operation.output_modes.iter().zip(&operation.output.extents) {
            // Explicit output descriptors may introduce broadcast-only axes.
            merge_extent(&mut extents, mode, extent)?;
            if extents[&mode] != extent {
                return Err(Error::InvalidOperation(format!("output extent for mode {mode} would discard values")));
            }
        }
        let mut reduction_modes = Vec::new();
        for input in &operation.inputs[..operation.terms] {
            for &mode in &input.modes {
                if !operation.output_modes.contains(&mode) && !reduction_modes.contains(&mode) {
                    reduction_modes.push(mode);
                }
            }
        }
        if operation.addend && operation.inputs[operation.terms].modes.iter()
            .any(|mode| !operation.output_modes.contains(mode)) {
            return Err(Error::InvalidOperation("C may only contain output modes".into()));
        }
        if !reduction_modes.is_empty() && matches!(operation.kind, Kind::Permutation | Kind::Elementwise(_, _)) {
            return Err(Error::InvalidOperation("elementwise and permutation operations cannot discard modes".into()));
        }
        if operation.kind == Kind::Permutation {
            let input = &operation.inputs[0];
            if input.modes.len() != operation.output_modes.len()
                || input.modes.iter().any(|mode| !operation.output_modes.contains(mode))
                || input.modes.iter().enumerate().any(|(i, mode)| input.modes[..i].contains(mode)) {
                return Err(Error::InvalidOperation("a permutation must reorder each input mode exactly once".into()));
            }
            for (i, mode) in input.modes.iter().enumerate() {
                if input.tensor.extents[i] != extents[mode] {
                    return Err(Error::InvalidOperation("a permutation cannot expand an input axis".into()));
                }
            }
        }
        let reduction_extents: Vec<_> = reduction_modes.iter().map(|m| extents[m]).collect();
        let reduction_count = product(&reduction_extents)?;
        let mut modes = operation.output_modes.clone();
        modes.extend(reduction_modes);
        let axis_modes = operation.inputs.iter().map(|input| input.modes.iter()
            .map(|m| modes.iter().position(|x| x == m).expect("validated mode"))
            .collect()).collect();
        Ok(Self { operation, modes, reduction_extents, reduction_count, axis_modes })
    }

    pub fn operation(&self) -> &OperationDescriptor { &self.operation }
    pub fn reduction_extents(&self) -> &[usize] { &self.reduction_extents }
    pub fn reduction_elements(&self) -> usize { self.reduction_count }

    /// Allocate D and enqueue the operation. All inputs remain unchanged.
    pub fn execute<R: Runtime>(&self, inputs: &[&RudaTensor<R>], scalars: &[f64]) -> Result<RudaTensor<R>> {
        self.validate_inputs(inputs, scalars)?;
        let reference = inputs[0];
        let descriptor = &self.operation.output;
        let handle = reference.client.empty(descriptor.storage_bytes().max(descriptor.dtype.size()));
        let output = RudaTensor::new(reference.client.clone(), handle,
            Metadata::new(Shape::from(descriptor.extents.clone()), descriptor.strides.clone()),
            reference.device.clone(), descriptor.dtype);
        self.submit(inputs, output, scalars)
    }

    /// Write into an exclusively owned D buffer and return it. D cannot alias any input.
    pub fn execute_into<R: Runtime>(
        &self, inputs: &[&RudaTensor<R>], output: RudaTensor<R>, scalars: &[f64],
    ) -> Result<RudaTensor<R>> {
        self.validate_inputs(inputs, scalars)?;
        if !self.operation.output.matches(&output) { return Err(Error::OutputMismatch); }
        self.operation.output.check_buffer(&output)?;
        if inputs[0].device.to_id() != output.device.to_id() { return Err(Error::DeviceMismatch); }
        if !output.can_mut() { return Err(Error::SharedOutput); }
        self.submit(inputs, output, scalars)
    }

    fn validate_inputs<R: Runtime>(&self, inputs: &[&RudaTensor<R>], scalars: &[f64]) -> Result<()> {
        if inputs.len() != self.operation.inputs.len() {
            return Err(Error::InputCount { expected: self.operation.inputs.len(), actual: inputs.len() });
        }
        if scalars.len() != self.operation.scalar_count() {
            return Err(Error::ScalarCount { expected: self.operation.scalar_count(), actual: scalars.len() });
        }
        for (index, (input, descriptor)) in inputs.iter().zip(&self.operation.inputs).enumerate() {
            if !descriptor.tensor.matches(input) { return Err(Error::TensorMismatch { input: index }); }
            descriptor.tensor.check_buffer(input)?;
            if inputs[0].device.to_id() != input.device.to_id() { return Err(Error::DeviceMismatch); }
        }
        Ok(())
    }

    fn submit<R: Runtime>(
        &self, inputs: &[&RudaTensor<R>], output: RudaTensor<R>, scalars: &[f64],
    ) -> Result<RudaTensor<R>> {
        let count = self.operation.output.num_elements();
        if count == 0 { return Ok(output); }
        let hardware = &output.client.properties().hardware;
        let mut maximum = (hardware.max_ruda_dim.0 as usize)
            .min(hardware.max_units_per_ruda as usize).min(128);
        if !self.reduction_extents.is_empty() {
            maximum = maximum.min(hardware.max_shared_memory_size / self.operation.compute.dtype().size());
        }
        if maximum == 0 {
            return Err(Error::UnsupportedDevice("no workgroup threads available".into()));
        }
        let desired = maximum.min(self.reduction_count.max(1));
        let threads = 1usize << desired.ilog2();
        count.checked_mul(threads).ok_or(Error::Overflow)?;
        let compute = self.operation.compute;
        let converted = inputs.iter().map(|input| convert_operand(input, compute.dtype()))
            .collect::<Result<Vec<_>>>()?;
        let stride_count = self.modes.len().checked_mul(inputs.len()).ok_or(Error::Overflow)?;
        let mut strides = vec![0usize; stride_count];
        for (operand, tensor) in converted.iter().enumerate() {
            for (axis, &mode) in self.axis_modes[operand].iter().enumerate() {
                if tensor.meta.shape()[axis] != 1 {
                    let position = operand * self.modes.len() + mode;
                    strides[position] = strides[position].checked_add(tensor.meta.strides()[axis])
                        .ok_or(Error::Overflow)?;
                }
            }
        }
        Launch {
            inputs: &converted, output: &output, strides,
            reduction_extents: &self.reduction_extents, scalars, count,
            reduction_count: self.reduction_count, threads, compute,
            config: Config {
                kind: self.operation.kind,
                unary: self.operation.inputs.iter().map(|x| x.unary).collect(),
                terms: self.operation.terms, addend: self.operation.addend,
            },
        }.submit()?;
        Ok(output)
    }
}