ruPRIM 0.1.1

Parallel primitives, reductions, scans, and indexing 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
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
267
268

use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::dsl::Runtime;
use ruda_kernel::tensor::layout::address_type;
use ruda_kernel::tensor::layout::shape_divmod;
use ruda_kernel::tensor::allocation::empty_device_dtype;
use ruda_kernel::tensor::RudaTensor;
use ruda_core::tensor::Slice;
use ruda_core::tensor::TensorMetadata;
use ruda_core::tensor::Metadata;
use ruda_core::tensor::SliceOps;
use ruda_kernel::dsl::calculate_ruda_count_elemwise;
use ruda_kernel::dsl::intrinsic;
use ruda_kernel::dsl::prelude::*;
use ruda_kernel::library::FastDivmod;
use ruda_kernel::library::tensor::layout::linear::LinearView;
use std::ops::Range;

/// Slice a jit tensor with a set of ranges
pub fn slice<R: Runtime>(tensor: RudaTensor<R>, indices: &[Range<usize>]) -> RudaTensor<R> {
    let mut dims = tensor.shape();
    for (dim, range) in dims.iter_mut().zip(indices) {
        *dim = range.end.saturating_sub(range.start);
    }
    if dims.contains(&0) {
        return empty_device_dtype(tensor.client, tensor.device, dims, tensor.dtype);
    }
    let mut offset_start = 0u64;
    let mut offset_end = 0u64;

    for i in 0..indices.len() {
        offset_start += (tensor.meta.strides()[i] * indices[i].start) as u64;
        offset_end += (tensor.meta.strides()[i] * (tensor.meta.shape()[i] - indices[i].end)) as u64;
    }

    let offset_start = offset_start * tensor.dtype.size() as u64;
    let offset_end = offset_end * tensor.dtype.size() as u64;

    let memory_offset_alignment = tensor.client.properties().memory.alignment;

    if offset_start.is_multiple_of(memory_offset_alignment)
        && offset_end.is_multiple_of(memory_offset_alignment)
    {
        RudaTensor::new(
            tensor.client.clone(),
            tensor
                .handle
                .clone()
                .offset_start(offset_start)
                .offset_end(offset_end),
            Metadata::new(dims, tensor.meta.strides.clone()),
            tensor.device.clone(),
            tensor.dtype,
        )
    } else {
        let output = empty_device_dtype(
            tensor.client.clone(),
            tensor.device.clone(),
            dims,
            tensor.dtype,
        );
        slice_on_output(tensor, output, indices)
    }
}

#[ruda(launch_unchecked, address_type = "dynamic")]
fn slice_kernel<E: Numeric>(
    input: &Tensor<E>,
    output: &mut LinearView<E, ReadWrite>,
    out_shape: Sequence<FastDivmod<usize>>,
    indices: Sequence<usize>,
    #[define(E)] _dtype: StorageType,
) {
    if !output.is_in_bounds(ABSOLUTE_POS) {
        terminate!();
    }

    let rank = comptime![out_shape.len()];
    let mut offset_output = ABSOLUTE_POS;
    let mut offset_input = 0;

    #[unroll]
    for i in 0..rank {
        // Iterate in reverse to use divmod
        let dim = rank - i - 1;

        let range_start = indices[dim];
        let (rem, offset_local) = out_shape[dim].div_mod(offset_output);
        offset_output = rem;

        let offset_local = offset_local + range_start;

        offset_input += offset_local * input.stride(dim);
    }

    output[ABSOLUTE_POS] = input[offset_input];
}

pub fn slice_on_output<R: Runtime>(
    tensor: RudaTensor<R>,
    output: RudaTensor<R>,
    indices: &[Range<usize>],
) -> RudaTensor<R> {
    if output.meta.num_elements() == 0 {
        return output;
    }
    let ndims = tensor.meta.num_dims();
    let mut indices_sequence = SequenceArg::<R, usize>::new();

    for i in 0..ndims {
        let start = indices.get(i).map(|index| index.start).unwrap_or(0);
        indices_sequence.push(start);
    }

    let working_units = output.meta.num_elements();
    let ruda_dim = RudaDim::new(tensor.client.properties(), working_units);
    let ruda_count = calculate_ruda_count_elemwise(&tensor.client, working_units, ruda_dim);
    let dtype = tensor.dtype;

    unsafe {
        slice_kernel::launch_unchecked(
            &output.client,
            ruda_count,
            ruda_dim,
            address_type!(tensor, output),
            tensor.into_tensor_arg(),
            output.clone().into_linear_view(),
            shape_divmod(&output),
            indices_sequence,
            dtype.into(),
        )
    };

    output
}

/// Kernel for slicing with steps
#[ruda(launch_unchecked, address_type = "dynamic")]
fn slice_with_steps_kernel<E: Numeric>(
    input: &Tensor<E>,
    output: &mut LinearView<E, ReadWrite>,
    out_shape: Sequence<FastDivmod<usize>>,
    starts: Sequence<usize>,
    ends: Sequence<usize>,
    steps: Sequence<usize>,
    reversed: Sequence<usize>,
    #[define(E)] _dtype: StorageType,
) {
    if !output.is_in_bounds(ABSOLUTE_POS) {
        terminate!();
    }

    let rank = comptime![out_shape.len()];
    let mut output_offset = ABSOLUTE_POS;
    let mut input_offset = 0;

    // Calculate the input offset based on output position and slice info
    #[unroll]
    for i in 0..rank {
        // Iterate in reverse to use divmod
        let dim = rank - i - 1;
        let start = starts[dim];
        let end = ends[dim];
        let step = steps[dim];

        let (rem, output_idx) = out_shape[dim].div_mod(output_offset);
        output_offset = rem;

        let input_idx = if reversed[dim] == 0 {
            // Forward stepping
            start + output_idx * step
        } else {
            // Backward stepping - start from end-1
            let end_minus_1 = end - 1;
            end_minus_1 - output_idx * step
        };

        input_offset += input_idx * input.stride(dim);
    }

    output[ABSOLUTE_POS] = input[input_offset];
}

/// Slice a tensor with steps
pub fn slice_with_steps<R: Runtime>(tensor: RudaTensor<R>, slices: &[Slice]) -> RudaTensor<R> {
    // Check if all steps are 1 - if so, use the optimized regular slice
    let all_steps_one = slices.iter().all(|info| info.step == 1);

    if all_steps_one {
        // Convert Slice to Range for step=1
        let simple_ranges: Vec<Range<usize>> = slices
            .iter()
            .enumerate()
            .map(|(i, slice)| slice.to_range(tensor.meta.shape()[i]))
            .collect();
        return slice(tensor, &simple_ranges);
    }

    // Calculate output shape
    let shape_output = tensor.shape().slice(slices).unwrap();

    // Create output tensor
    let output = empty_device_dtype(
        tensor.client.clone(),
        tensor.device.clone(),
        shape_output.clone(),
        tensor.dtype,
    );

    if shape_output.num_elements() == 0 {
        return output;
    }

    let mut starts = SequenceArg::<R, usize>::new();
    let mut ends = SequenceArg::<R, usize>::new();
    let mut steps = SequenceArg::<R, usize>::new();
    let mut reversed = SequenceArg::<R, usize>::new();

    for (dim, slice) in slices.iter().enumerate() {
        let range = slice.to_range(tensor.meta.shape()[dim]);
        starts.push(range.start);
        ends.push(range.end);
        steps.push(slice.step.unsigned_abs());
        reversed.push(usize::from(slice.is_reversed()));
    }

    // Pad with default values if needed to match tensor dimensions
    for dim in slices.len()..tensor.meta.num_dims() {
        starts.push(0);
        ends.push(tensor.meta.shape[dim]);
        steps.push(1);
        reversed.push(0);
    }

    // Launch kernel
    let working_units = shape_output.num_elements();
    let ruda_dim = RudaDim::new(tensor.client.properties(), working_units);
    let ruda_count = calculate_ruda_count_elemwise(&tensor.client, working_units, ruda_dim);
    let dtype = tensor.dtype;

    let step_address = AddressType::from_len(
        slices.iter().map(|slice| slice.step.unsigned_abs()).max().unwrap_or(1),
    );
    unsafe {
        slice_with_steps_kernel::launch_unchecked(
            &output.client,
            ruda_count,
            ruda_dim,
            address_type!(tensor, output).max(step_address),
            tensor.into_tensor_arg(),
            output.clone().into_linear_view(),
            shape_divmod(&output),
            starts,
            ends,
            steps,
            reversed,
            dtype.into(),
        );
    }

    output
}

/// This is annoying and we need to find a way to do this automatically at some point
#[allow(unused)]
#[ruda]
fn unwrap(value: u32) -> comptime_type!(u32) {
    intrinsic!(|_| value.constant().unwrap().as_u32())
}