burn_tch/ops/
int_tensor.rs

1use std::ops::Range;
2
3use burn_backend::{
4    BoolDType, Distribution, ExecutionError, FloatDType, IntDType, Scalar, Shape, TensorData,
5    TensorMetadata,
6    ops::{FloatTensorOps, IntTensorOps},
7    tensor::IntTensor,
8};
9
10use crate::{IntoKind, LibTorch, LibTorchDevice, TchShape, TchTensor, element::TchElement};
11
12use super::TchOps;
13
14impl<E: TchElement> IntTensorOps<Self> for LibTorch<E> {
15    fn int_from_data(data: TensorData, device: &LibTorchDevice) -> TchTensor {
16        match data.dtype {
17            burn_backend::DType::I64 => TchTensor::from_data::<i64>(data, (*device).into()),
18            burn_backend::DType::I32 => TchTensor::from_data::<i32>(data, (*device).into()),
19            burn_backend::DType::I16 => TchTensor::from_data::<i16>(data, (*device).into()),
20            burn_backend::DType::I8 => TchTensor::from_data::<i8>(data, (*device).into()),
21            burn_backend::DType::U8 => TchTensor::from_data::<u8>(data, (*device).into()),
22            _ => unimplemented!("Unsupported dtype for `int_from_data`: {:?}", data.dtype),
23        }
24    }
25
26    fn int_repeat_dim(tensor: TchTensor, dim: usize, times: usize) -> TchTensor {
27        TchOps::repeat_dim(tensor, dim, times)
28    }
29
30    async fn int_into_data(tensor: TchTensor) -> Result<TensorData, ExecutionError> {
31        let shape = tensor.shape();
32        let tensor = Self::int_reshape(tensor.clone(), Shape::new([shape.num_elements()]));
33        let values: Result<Vec<i64>, tch::TchError> = tensor.tensor.shallow_clone().try_into();
34        Ok(TensorData::new(values.unwrap(), shape))
35    }
36
37    fn int_to_device(tensor: TchTensor, device: &LibTorchDevice) -> TchTensor {
38        TchOps::to_device(tensor, device)
39    }
40
41    fn int_reshape(tensor: TchTensor, shape: Shape) -> TchTensor {
42        TchOps::reshape(tensor, shape)
43    }
44
45    fn int_device(tensor: &TchTensor) -> LibTorchDevice {
46        tensor.tensor.device().into()
47    }
48
49    fn int_empty(shape: Shape, device: &LibTorchDevice, dtype: IntDType) -> TchTensor {
50        let tensor = tch::Tensor::empty(
51            TchShape::from(shape).dims,
52            (dtype.into_kind(), (*device).into()),
53        );
54
55        TchTensor::new(tensor)
56    }
57
58    fn int_slice(tensor: TchTensor, slices: &[burn_backend::Slice]) -> TchTensor {
59        TchOps::slice_with_steps(tensor, slices)
60    }
61
62    fn int_slice_assign(
63        tensor: TchTensor,
64        slices: &[burn_backend::Slice],
65        value: TchTensor,
66    ) -> TchTensor {
67        TchOps::slice_assign(tensor, slices, value)
68    }
69
70    fn int_cat(tensors: Vec<TchTensor>, dim: usize) -> TchTensor {
71        TchOps::cat(tensors, dim)
72    }
73
74    fn int_matmul(lhs: IntTensor<Self>, rhs: IntTensor<Self>) -> IntTensor<Self> {
75        let int_dtype = lhs.dtype();
76        let lhs = Self::int_into_float(lhs, FloatDType::F32);
77        let rhs = Self::int_into_float(rhs, FloatDType::F32);
78        let out = lhs.tensor.f_matmul(&rhs.tensor).unwrap();
79        Self::float_into_int(TchTensor::new(out), int_dtype.into())
80    }
81
82    fn int_equal(lhs: TchTensor, rhs: TchTensor, _out_dtype: BoolDType) -> TchTensor {
83        TchOps::equal(lhs, rhs)
84    }
85
86    fn int_equal_elem(lhs: TchTensor, rhs: Scalar, _out_dtype: BoolDType) -> TchTensor {
87        TchOps::equal_elem(lhs, rhs.elem::<i64>())
88    }
89
90    fn int_greater(lhs: TchTensor, rhs: TchTensor, _out_dtype: BoolDType) -> TchTensor {
91        TchOps::greater(lhs, rhs)
92    }
93
94    fn int_greater_elem(lhs: TchTensor, rhs: Scalar, _out_dtype: BoolDType) -> TchTensor {
95        TchOps::greater_elem(lhs, rhs.elem::<i64>())
96    }
97
98    fn int_greater_equal(lhs: TchTensor, rhs: TchTensor, _out_dtype: BoolDType) -> TchTensor {
99        TchOps::greater_equal(lhs, rhs)
100    }
101
102    fn int_greater_equal_elem(lhs: TchTensor, rhs: Scalar, _out_dtype: BoolDType) -> TchTensor {
103        TchOps::greater_equal_elem(lhs, rhs.elem::<i64>())
104    }
105
106    fn int_lower(lhs: TchTensor, rhs: TchTensor, _out_dtype: BoolDType) -> TchTensor {
107        TchOps::lower(lhs, rhs)
108    }
109
110    fn int_lower_elem(lhs: TchTensor, rhs: Scalar, _out_dtype: BoolDType) -> TchTensor {
111        TchOps::lower_elem(lhs, rhs.elem::<i64>())
112    }
113
114    fn int_lower_equal(lhs: TchTensor, rhs: TchTensor, _out_dtype: BoolDType) -> TchTensor {
115        TchOps::lower_equal(lhs, rhs)
116    }
117
118    fn int_lower_equal_elem(lhs: TchTensor, rhs: Scalar, _out_dtype: BoolDType) -> TchTensor {
119        TchOps::lower_equal_elem(lhs, rhs.elem::<i64>())
120    }
121
122    fn int_add(lhs: TchTensor, rhs: TchTensor) -> TchTensor {
123        TchOps::add(lhs, rhs)
124    }
125
126    fn int_add_scalar(lhs: TchTensor, rhs: Scalar) -> TchTensor {
127        lhs.unary_ops(
128            |mut tensor| tensor.f_add_scalar_(rhs.elem::<i64>()).unwrap(),
129            |tensor| tensor.f_add_scalar(rhs.elem::<i64>()).unwrap(),
130        )
131    }
132
133    fn int_sub(lhs: TchTensor, rhs: TchTensor) -> TchTensor {
134        TchOps::sub(lhs, rhs)
135    }
136
137    fn int_sub_scalar(lhs: TchTensor, rhs: Scalar) -> TchTensor {
138        lhs.unary_ops(
139            |mut tensor| tensor.f_sub_scalar_(rhs.elem::<i64>()).unwrap(),
140            |tensor| tensor.f_sub_scalar(rhs.elem::<i64>()).unwrap(),
141        )
142    }
143
144    fn int_mul(lhs: TchTensor, rhs: TchTensor) -> TchTensor {
145        TchOps::mul(lhs, rhs)
146    }
147
148    fn int_mul_scalar(lhs: TchTensor, rhs: Scalar) -> TchTensor {
149        lhs.unary_ops(
150            |mut tensor| tensor.f_mul_scalar_(rhs.elem::<i64>()).unwrap(),
151            |tensor| tensor.f_mul_scalar(rhs.elem::<i64>()).unwrap(),
152        )
153    }
154
155    fn int_div(lhs: TchTensor, rhs: TchTensor) -> TchTensor {
156        let dtype = lhs.tensor.kind();
157        let copy = false;
158        let non_blocking = true;
159        let lhs: TchTensor =
160            TchTensor::new(lhs.tensor.to_dtype(tch::Kind::Float, non_blocking, copy));
161        let rhs: TchTensor =
162            TchTensor::new(rhs.tensor.to_dtype(tch::Kind::Float, non_blocking, copy));
163
164        let out = TchOps::div(lhs, rhs);
165
166        TchTensor::new(out.tensor.to_dtype(dtype, non_blocking, copy))
167    }
168
169    fn int_div_scalar(lhs: TchTensor, rhs: Scalar) -> TchTensor {
170        let dtype = lhs.tensor.kind();
171        let copy = false;
172        let non_blocking = true;
173        let lhs: TchTensor =
174            TchTensor::new(lhs.tensor.to_dtype(tch::Kind::Float, non_blocking, copy));
175
176        let out: TchTensor = lhs.unary_ops(
177            |mut tensor| tensor.f_div_scalar_(rhs.elem::<i64>()).unwrap(),
178            |tensor| tensor.f_div_scalar(rhs.elem::<i64>()).unwrap(),
179        );
180
181        TchTensor::new(out.tensor.to_dtype(dtype, non_blocking, copy))
182    }
183
184    fn int_remainder(lhs: TchTensor, rhs: TchTensor) -> TchTensor {
185        let dtype = lhs.tensor.kind();
186        let copy = false;
187        let non_blocking = true;
188        let lhs: TchTensor =
189            TchTensor::new(lhs.tensor.to_dtype(tch::Kind::Float, non_blocking, copy));
190        let rhs: TchTensor =
191            TchTensor::new(rhs.tensor.to_dtype(tch::Kind::Float, non_blocking, copy));
192
193        let out = TchOps::remainder(lhs, rhs);
194
195        TchTensor::new(out.tensor.to_dtype(dtype, non_blocking, copy))
196    }
197
198    fn int_remainder_scalar(lhs: TchTensor, rhs: Scalar) -> TchTensor {
199        lhs.unary_ops(
200            |tensor| tensor.f_remainder(rhs.elem::<i64>()).unwrap(),
201            |tensor| tensor.f_remainder(rhs.elem::<i64>()).unwrap(),
202        )
203    }
204
205    fn int_zeros(shape: Shape, device: &LibTorchDevice, dtype: IntDType) -> TchTensor {
206        let shape = TchShape::from(shape);
207        let device: tch::Device = (*device).into();
208
209        TchTensor::new(tch::Tensor::zeros(shape.dims, (dtype.into_kind(), device)))
210    }
211
212    fn int_ones(shape: Shape, device: &LibTorchDevice, dtype: IntDType) -> TchTensor {
213        let shape = TchShape::from(shape);
214        let device: tch::Device = (*device).into();
215
216        TchTensor::new(tch::Tensor::ones(shape.dims, (dtype.into_kind(), device)))
217    }
218
219    fn int_full(
220        shape: Shape,
221        fill_value: Scalar,
222        device: &LibTorchDevice,
223        dtype: IntDType,
224    ) -> TchTensor {
225        let shape = TchShape::from(shape);
226        let device: tch::Device = (*device).into();
227
228        TchTensor::new(tch::Tensor::full(
229            shape.dims,
230            fill_value.elem::<i64>(),
231            (dtype.into_kind(), device),
232        ))
233    }
234
235    fn int_sum(tensor: TchTensor) -> TchTensor {
236        TchOps::sum(tensor)
237    }
238
239    fn int_sum_dim(tensor: TchTensor, dim: usize) -> TchTensor {
240        TchOps::sum_dim(tensor, dim)
241    }
242
243    fn int_prod(tensor: TchTensor) -> TchTensor {
244        TchOps::prod(tensor)
245    }
246
247    fn int_prod_dim(tensor: TchTensor, dim: usize) -> TchTensor {
248        TchOps::prod_dim(tensor, dim)
249    }
250
251    fn int_mean(tensor: TchTensor) -> TchTensor {
252        let dtype = tensor.tensor.kind();
253        let tensor: TchTensor =
254            TchTensor::new(tensor.tensor.to_dtype(tch::Kind::Float, true, false));
255        let output: TchTensor = TchTensor::new(TchOps::mean(tensor).tensor);
256
257        TchTensor::new(output.tensor.to_dtype(dtype, true, false))
258    }
259
260    fn int_mean_dim(tensor: TchTensor, dim: usize) -> TchTensor {
261        let dtype = tensor.tensor.kind();
262        let tensor: TchTensor =
263            TchTensor::new(tensor.tensor.to_dtype(tch::Kind::Float, true, false));
264
265        let output: TchTensor = TchTensor::new(TchOps::mean_dim(tensor, dim).tensor);
266
267        TchTensor::new(output.tensor.to_dtype(dtype, true, false))
268    }
269
270    fn int_cumsum(tensor: TchTensor, dim: usize) -> TchTensor {
271        TchOps::cumsum(tensor, dim)
272    }
273
274    fn int_cumprod(tensor: TchTensor, dim: usize) -> TchTensor {
275        TchOps::cumprod(tensor, dim)
276    }
277
278    fn int_cummin(tensor: TchTensor, dim: usize) -> TchTensor {
279        TchOps::cummin(tensor, dim)
280    }
281
282    fn int_cummax(tensor: TchTensor, dim: usize) -> TchTensor {
283        TchOps::cummax(tensor, dim)
284    }
285
286    fn int_gather(dim: usize, tensor: TchTensor, indices: TchTensor) -> TchTensor {
287        TchOps::gather(dim, tensor, indices)
288    }
289
290    fn int_scatter_add(
291        dim: usize,
292        tensor: TchTensor,
293        indices: TchTensor,
294        value: TchTensor,
295    ) -> TchTensor {
296        TchOps::scatter(dim, tensor, indices, value)
297    }
298
299    fn int_scatter_nd(
300        data: TchTensor,
301        indices: TchTensor,
302        values: TchTensor,
303        reduction: burn_backend::tensor::IndexingUpdateOp,
304    ) -> TchTensor {
305        TchOps::scatter_nd(data, indices, values, reduction)
306    }
307
308    fn int_gather_nd(data: TchTensor, indices: TchTensor) -> TchTensor {
309        TchOps::gather_nd(data, indices)
310    }
311
312    fn int_select(tensor: TchTensor, dim: usize, indices: TchTensor) -> TchTensor {
313        TchOps::index_select_dim(tensor, dim, indices)
314    }
315
316    fn int_select_add(
317        tensor: TchTensor,
318        dim: usize,
319        indices: TchTensor,
320        value: TchTensor,
321    ) -> TchTensor {
322        TchOps::select_assign(tensor, dim, indices, value)
323    }
324
325    fn int_mask_where(tensor: TchTensor, mask: TchTensor, source: TchTensor) -> TchTensor {
326        TchTensor::binary_ops_tensor(
327            tensor,
328            source,
329            |tensor, source| source.f_where_self(&mask.tensor, tensor).unwrap(),
330            |tensor, source| source.f_where_self(&mask.tensor, tensor).unwrap(),
331            |tensor, source| source.f_where_self(&mask.tensor, tensor).unwrap(),
332        )
333    }
334
335    fn int_mask_fill(tensor: TchTensor, mask: TchTensor, value: Scalar) -> TchTensor {
336        let value = value.elem::<i64>();
337        tensor.unary_ops(
338            |mut tensor| tensor.f_masked_fill_(&mask.tensor, value).unwrap(),
339            |tensor| tensor.f_masked_fill(&mask.tensor, value).unwrap(),
340        )
341    }
342
343    fn int_argmax(tensor: TchTensor, dim: usize) -> TchTensor {
344        TchOps::argmax(tensor, dim)
345    }
346
347    fn int_argtopk(_tensor: TchTensor, _dim: usize, _k: usize) -> TchTensor {
348        panic!("argtopk not implemented for torch")
349    }
350
351    fn int_topk(tensor: TchTensor, dim: usize, k: usize) -> TchTensor {
352        TchOps::topk(tensor, dim, k)
353    }
354
355    fn int_argmin(tensor: TchTensor, dim: usize) -> TchTensor {
356        TchOps::argmin(tensor, dim)
357    }
358
359    fn int_max_dim(tensor: TchTensor, dim: usize) -> TchTensor {
360        TchOps::max_dim(tensor, dim)
361    }
362
363    fn int_max_dim_with_indices(tensor: TchTensor, dim: usize) -> (TchTensor, TchTensor) {
364        TchOps::max_dim_with_indices(tensor, dim)
365    }
366
367    fn int_min_dim(tensor: TchTensor, dim: usize) -> TchTensor {
368        TchOps::min_dim(tensor, dim)
369    }
370
371    fn int_min_dim_with_indices(tensor: TchTensor, dim: usize) -> (TchTensor, TchTensor) {
372        TchOps::min_dim_with_indices(tensor, dim)
373    }
374
375    fn int_clamp_min(tensor: TchTensor, min: Scalar) -> TchTensor {
376        TchOps::clamp_min(tensor, min.elem::<i64>())
377    }
378
379    fn int_clamp_max(tensor: TchTensor, max: Scalar) -> TchTensor {
380        TchOps::clamp_max(tensor, max.elem::<i64>())
381    }
382
383    fn int_clamp(tensor: TchTensor, min: Scalar, max: Scalar) -> TchTensor {
384        TchOps::clamp(tensor, min.elem::<i64>(), max.elem::<i64>())
385    }
386
387    fn int_abs(tensor: TchTensor) -> TchTensor {
388        tensor.unary_ops(|mut tensor| tensor.abs_(), |tensor| tensor.abs())
389    }
390
391    fn int_into_float(tensor: TchTensor, out_dtype: FloatDType) -> TchTensor {
392        let tensor = tensor.tensor.to_kind(out_dtype.into_kind());
393        TchTensor::new(tensor)
394    }
395
396    fn int_swap_dims(tensor: IntTensor<Self>, dim1: usize, dim2: usize) -> IntTensor<Self> {
397        TchOps::swap_dims(tensor, dim1, dim2)
398    }
399
400    fn int_random(
401        shape: Shape,
402        distribution: Distribution,
403        device: &LibTorchDevice,
404        dtype: IntDType,
405    ) -> TchTensor {
406        match distribution {
407            Distribution::Default => TchTensor::new(tch::Tensor::randint_low(
408                0,
409                255,
410                shape.iter().map(|i| *i as i64).collect::<Vec<_>>(),
411                (dtype.into_kind(), (*device).into()),
412            )),
413            Distribution::Bernoulli(prob) => {
414                let mut tensor = TchTensor::empty(shape, *device, dtype.into());
415                tensor
416                    .mut_ops(|tensor| tensor.f_bernoulli_float_(prob).unwrap())
417                    .unwrap()
418            }
419            Distribution::Uniform(from, to) => TchTensor::new(tch::Tensor::randint_low(
420                from as i64,
421                to as i64,
422                shape.iter().map(|i| *i as i64).collect::<Vec<_>>(),
423                (dtype.into_kind(), (*device).into()),
424            )),
425            Distribution::Normal(mean, std) => {
426                let mut tensor = TchTensor::empty(shape, *device, dtype.into());
427                tensor.mut_ops(|tensor| tensor.normal_(mean, std)).unwrap()
428            }
429        }
430    }
431
432    fn int_arange(range: Range<i64>, device: &LibTorchDevice, dtype: IntDType) -> TchTensor {
433        let device: tch::Device = (*device).into();
434        let mut tensor = tch::Tensor::arange(range.end - range.start, (dtype.into_kind(), device));
435
436        if range.start != 0 {
437            tensor = tensor.f_add_scalar_(range.start).unwrap();
438        }
439
440        TchTensor::new(tensor)
441    }
442
443    fn int_permute(tensor: IntTensor<Self>, axes: &[usize]) -> IntTensor<Self> {
444        TchOps::permute(tensor, axes)
445    }
446
447    fn int_flip(tensor: IntTensor<Self>, axes: &[usize]) -> IntTensor<Self> {
448        TchOps::flip(tensor, axes)
449    }
450
451    fn int_sign(tensor: IntTensor<Self>) -> IntTensor<Self> {
452        TchOps::sign(tensor)
453    }
454
455    fn int_expand(tensor: IntTensor<Self>, shape: Shape) -> IntTensor<Self> {
456        TchOps::expand(tensor, shape)
457    }
458
459    fn int_sort(tensor: IntTensor<Self>, dim: usize, descending: bool) -> IntTensor<Self> {
460        TchOps::sort(tensor, dim, descending)
461    }
462
463    fn int_argsort(tensor: IntTensor<Self>, dim: usize, descending: bool) -> IntTensor<Self> {
464        TchOps::argsort(tensor, dim, descending)
465    }
466
467    fn bitwise_and(lhs: IntTensor<Self>, rhs: IntTensor<Self>) -> IntTensor<Self> {
468        TchOps::bitwise_and(lhs, rhs)
469    }
470
471    fn bitwise_or(lhs: IntTensor<Self>, rhs: IntTensor<Self>) -> IntTensor<Self> {
472        TchOps::bitwise_or(lhs, rhs)
473    }
474
475    fn bitwise_xor(lhs: IntTensor<Self>, rhs: IntTensor<Self>) -> IntTensor<Self> {
476        TchOps::bitwise_xor(lhs, rhs)
477    }
478
479    fn bitwise_not(tensor: IntTensor<Self>) -> IntTensor<Self> {
480        TchOps::bitwise_not(tensor)
481    }
482
483    fn bitwise_and_scalar(lhs: IntTensor<Self>, rhs: Scalar) -> IntTensor<Self> {
484        TchOps::bitwise_and_scalar(lhs, rhs.elem::<i64>())
485    }
486
487    fn bitwise_or_scalar(lhs: IntTensor<Self>, rhs: Scalar) -> IntTensor<Self> {
488        TchOps::bitwise_or_scalar(lhs, rhs.elem::<i64>())
489    }
490
491    fn bitwise_xor_scalar(lhs: IntTensor<Self>, rhs: Scalar) -> IntTensor<Self> {
492        TchOps::bitwise_xor_scalar(lhs, rhs.elem::<i64>())
493    }
494
495    fn bitwise_left_shift(lhs: IntTensor<Self>, rhs: IntTensor<Self>) -> IntTensor<Self> {
496        TchOps::bitwise_left_shift(lhs, rhs)
497    }
498
499    fn bitwise_right_shift(lhs: IntTensor<Self>, rhs: IntTensor<Self>) -> IntTensor<Self> {
500        TchOps::bitwise_right_shift(lhs, rhs)
501    }
502
503    fn bitwise_left_shift_scalar(lhs: IntTensor<Self>, rhs: Scalar) -> IntTensor<Self> {
504        TchOps::bitwise_left_shift_scalar(lhs, rhs.elem::<i64>())
505    }
506
507    fn bitwise_right_shift_scalar(lhs: IntTensor<Self>, rhs: Scalar) -> IntTensor<Self> {
508        TchOps::bitwise_right_shift_scalar(lhs, rhs.elem::<i64>())
509    }
510
511    fn int_cast(tensor: IntTensor<Self>, dtype: IntDType) -> IntTensor<Self> {
512        // NOTE: when dtypes of inputs to an arithmetic operation differ, tch handles type
513        // promotion based on a set of rules: https://pytorch.org/docs/stable/tensor_attributes.html#type-promotion-doc
514
515        // Type promotion is not automatic on all backends so this behavior might differ
516        let kind = dtype.into_kind();
517
518        if tensor.tensor.kind() == kind {
519            tensor
520        } else {
521            TchTensor::new(tensor.tensor.to_kind(kind))
522        }
523    }
524
525    fn int_unfold(
526        tensor: IntTensor<Self>,
527        dim: usize,
528        size: usize,
529        step: usize,
530    ) -> IntTensor<Self> {
531        TchOps::unfold(tensor, dim, size, step)
532    }
533
534    fn int_powi(lhs: IntTensor<Self>, rhs: IntTensor<Self>) -> IntTensor<Self> {
535        TchOps::pow(lhs, rhs)
536    }
537
538    fn int_powi_scalar_impl(lhs: IntTensor<Self>, rhs: Scalar) -> IntTensor<Self> {
539        lhs.unary_ops(
540            |mut tensor| tensor.f_pow_(rhs.elem::<i64>()).unwrap(),
541            |tensor| tensor.pow_tensor_scalar(rhs.elem::<i64>()),
542        )
543    }
544}
burn_tch/ops/int_tensor.rs

burn_tch/ops/
int_tensor.rs
