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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
//! Cumulative operations for CUDA runtime
use crate::error::{Error, Result};
use crate::ops::{CumulativeOps, reduce_dim_output_shape, reduce_output_shape};
use crate::runtime::cuda::kernels::{
launch_cumprod, launch_cumprod_strided, launch_cumsum, launch_cumsum_strided, launch_logsumexp,
launch_logsumexp_strided,
};
use crate::runtime::cuda::{CudaClient, CudaRuntime};
use crate::runtime::ensure_contiguous;
use crate::tensor::Tensor;
impl CumulativeOps<CudaRuntime> for CudaClient {
fn cumsum(&self, a: &Tensor<CudaRuntime>, dim: isize) -> Result<Tensor<CudaRuntime>> {
let shape = a.shape();
let ndim = shape.len();
// Normalize dimension (handle negative indexing)
let dim = if dim < 0 {
(ndim as isize + dim) as usize
} else {
dim as usize
};
if dim >= ndim {
return Err(Error::InvalidDimension {
dim: dim as isize,
ndim,
});
}
// Handle empty tensor
if a.numel() == 0 {
return Ok(Tensor::<CudaRuntime>::empty(shape, a.dtype(), &self.device));
}
// Ensure contiguous for CUDA kernel
let a_contig = ensure_contiguous(a);
// Calculate dimensions for kernel launch
let scan_size = shape[dim];
let outer_size: usize = shape[..dim].iter().product();
let inner_size: usize = shape[dim + 1..].iter().product();
// Allocate output
let out = Tensor::<CudaRuntime>::empty(shape, a.dtype(), &self.device);
// Choose kernel based on dimension position
if inner_size == 1 {
// Scan along last dimension or effectively contiguous
let outer = outer_size.max(1);
unsafe {
launch_cumsum(
&self.context,
&self.stream,
self.device.index,
a.dtype(),
a_contig.ptr(),
out.ptr(),
scan_size,
outer,
)?;
}
} else {
// Strided scan for non-last dimension
unsafe {
launch_cumsum_strided(
&self.context,
&self.stream,
self.device.index,
a.dtype(),
a_contig.ptr(),
out.ptr(),
scan_size,
outer_size.max(1),
inner_size,
)?;
}
}
Ok(out)
}
fn cumprod(&self, a: &Tensor<CudaRuntime>, dim: isize) -> Result<Tensor<CudaRuntime>> {
let shape = a.shape();
let ndim = shape.len();
// Normalize dimension (handle negative indexing)
let dim = if dim < 0 {
(ndim as isize + dim) as usize
} else {
dim as usize
};
if dim >= ndim {
return Err(Error::InvalidDimension {
dim: dim as isize,
ndim,
});
}
// Handle empty tensor
if a.numel() == 0 {
return Ok(Tensor::<CudaRuntime>::empty(shape, a.dtype(), &self.device));
}
// Ensure contiguous for CUDA kernel
let a_contig = ensure_contiguous(a);
// Calculate dimensions for kernel launch
let scan_size = shape[dim];
let outer_size: usize = shape[..dim].iter().product();
let inner_size: usize = shape[dim + 1..].iter().product();
// Allocate output
let out = Tensor::<CudaRuntime>::empty(shape, a.dtype(), &self.device);
// Choose kernel based on dimension position
if inner_size == 1 {
// Scan along last dimension or effectively contiguous
let outer = outer_size.max(1);
unsafe {
launch_cumprod(
&self.context,
&self.stream,
self.device.index,
a.dtype(),
a_contig.ptr(),
out.ptr(),
scan_size,
outer,
)?;
}
} else {
// Strided scan for non-last dimension
unsafe {
launch_cumprod_strided(
&self.context,
&self.stream,
self.device.index,
a.dtype(),
a_contig.ptr(),
out.ptr(),
scan_size,
outer_size.max(1),
inner_size,
)?;
}
}
Ok(out)
}
fn logsumexp(
&self,
a: &Tensor<CudaRuntime>,
dims: &[usize],
keepdim: bool,
) -> Result<Tensor<CudaRuntime>> {
// Support: F32, F64, F16, BF16
// For F16/BF16: upcast to F32, compute, downcast back
use crate::dtype::DType;
use crate::ops::TypeConversionOps;
let input_dtype = a.dtype();
if !matches!(
input_dtype,
DType::F32 | DType::F64 | DType::F16 | DType::BF16 | DType::FP8E4M3 | DType::FP8E5M2
) {
return Err(Error::UnsupportedDType {
dtype: input_dtype,
op: "logsumexp",
});
}
// F16/BF16/FP8 have native CUDA kernels that accumulate in F32 internally
let (a_compute, needs_cast) = (a.clone(), false);
let shape = a_compute.shape();
let ndim = shape.len();
// Handle empty dims (reduce over all dimensions)
let actual_dims: Vec<usize> = if dims.is_empty() {
(0..ndim).collect()
} else {
dims.to_vec()
};
// Validate dimensions
for &dim in &actual_dims {
if dim >= ndim {
return Err(Error::InvalidDimension {
dim: dim as isize,
ndim,
});
}
}
// Handle empty tensor
if a_compute.numel() == 0 {
let out_shape = reduce_output_shape(shape, &actual_dims, keepdim);
let out = Tensor::<CudaRuntime>::empty(&out_shape, a_compute.dtype(), &self.device);
// Cast back to original dtype if needed
return if needs_cast {
Ok(self.cast(&out, input_dtype)?)
} else {
Ok(out)
};
}
// For multi-dimensional reduction, reduce one dimension at a time
if actual_dims.len() > 1 {
let mut result = a_compute.clone();
// Sort dims in descending order to avoid index invalidation
let mut sorted_dims = actual_dims.clone();
sorted_dims.sort_by(|a, b| b.cmp(a));
for &dim in &sorted_dims {
result = self.logsumexp(&result, &[dim], true)?;
}
// Remove keepdim if not requested
if !keepdim {
let out_shape = reduce_output_shape(shape, &actual_dims, false);
result = result.reshape(&out_shape)?;
}
return Ok(result);
}
// Single dimension reduction
let dim = actual_dims[0];
// Ensure contiguous for CUDA kernel
let a_contig = ensure_contiguous(&a_compute);
// Calculate dimensions for kernel launch
let reduce_size = shape[dim];
let outer_size: usize = shape[..dim].iter().product();
let inner_size: usize = shape[dim + 1..].iter().product();
// Calculate output shape
let out_shape = reduce_dim_output_shape(shape, dim, keepdim);
let out_numel: usize = out_shape.iter().product();
// Allocate output (in F32 if upcast, else in original dtype)
let compute_dtype = a_compute.dtype();
let out = Tensor::<CudaRuntime>::empty(&out_shape, compute_dtype, &self.device);
// Choose kernel based on dimension position
if inner_size == 1 {
// Reduction along last dimension
let outer = outer_size.max(1);
unsafe {
launch_logsumexp(
&self.context,
&self.stream,
self.device.index,
a_compute.dtype(),
a_contig.ptr(),
out.ptr(),
reduce_size,
outer,
)?;
}
} else {
// Strided reduction for non-last dimension
unsafe {
launch_logsumexp_strided(
&self.context,
&self.stream,
self.device.index,
a_compute.dtype(),
a_contig.ptr(),
out.ptr(),
reduce_size,
outer_size.max(1),
inner_size,
)?;
}
}
// Cast back to original dtype if needed
let result = if needs_cast {
self.cast(&out, input_dtype)?
} else {
out
};
// Handle keepdim reshape if needed
if keepdim && result.numel() == out_numel {
Ok(result)
} else {
Ok(result)
}
}
}