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
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
use crate::device::Device;
use crate::dtype::Element;
use crate::gpu_dispatch::GpuBufferHandle;
/// The underlying data buffer for a tensor, tagged with its device.
///
/// Owns the data directly (`Vec<T>` for CPU, `GpuBufferHandle` for GPU).
/// The GPU handle is type-erased -- ferrotorch-gpu provides the concrete
/// implementation via the `GpuBackend` trait.
#[derive(Debug)]
pub struct TensorStorage<T: Element> {
pub(crate) data: StorageBuffer<T>,
pub(crate) device: Device,
}
/// Device-specific data buffer.
pub enum StorageBuffer<T: Element> {
/// CPU heap-allocated data.
Cpu(Vec<T>),
/// GPU device memory, accessed via the registered `GpuBackend`.
Gpu(GpuBufferHandle),
/// Meta storage — no backing memory, only the element count is
/// recorded. Tensors built on this variant carry shape and dtype
/// info but cannot be read or written. Used for shape inference
/// and dry-run model construction. CL-395.
Meta {
numel: usize,
_phantom: std::marker::PhantomData<T>,
},
}
impl<T: Element> TensorStorage<T> {
/// Create a new CPU storage from a `Vec<T>`.
pub fn cpu(data: Vec<T>) -> Self {
Self {
data: StorageBuffer::Cpu(data),
device: Device::Cpu,
}
}
/// Create a meta storage with the given element count. No memory is
/// allocated for the elements; only the size is recorded. Reading
/// the data of a meta tensor returns an error.
pub fn meta(numel: usize) -> Self {
Self {
data: StorageBuffer::Meta {
numel,
_phantom: std::marker::PhantomData,
},
device: Device::Meta,
}
}
/// Create storage on `target_device` from CPU data.
///
/// If `target_device` is CPU, wraps the `Vec` directly (zero-copy).
/// If `target_device` is GPU, uploads the data and returns GPU storage.
///
/// Use this instead of `TensorStorage::cpu(data).to(device)` to avoid
/// injecting a `ToDeviceBackward` node into the autograd graph.
pub fn on_device(data: Vec<T>, target_device: Device) -> crate::error::FerrotorchResult<Self> {
match target_device {
Device::Cpu => Ok(Self::cpu(data)),
Device::Cuda(ordinal) => {
let backend = crate::gpu_dispatch::gpu_backend()
.ok_or(crate::error::FerrotorchError::DeviceUnavailable)?;
let bytes: &[u8] = unsafe {
std::slice::from_raw_parts(
data.as_ptr() as *const u8,
data.len() * std::mem::size_of::<T>(),
)
};
let handle = backend.cpu_to_gpu(bytes, std::mem::size_of::<T>(), ordinal)?;
Ok(Self::gpu(handle))
}
Device::Xpu(ordinal) => Ok(Self::xpu(data, ordinal)),
Device::Meta => {
// Discard the data; only the element count matters.
Ok(Self::meta(data.len()))
}
}
}
/// Create storage on `target_device` from CPU data, using pinned host
/// memory for the CPU→GPU transfer (~2x faster for large tensors).
///
/// Falls back to regular transfer if no GPU backend or if target is CPU.
pub fn on_device_pinned(
data: Vec<T>,
target_device: Device,
) -> crate::error::FerrotorchResult<Self> {
match target_device {
Device::Cpu => Ok(Self::cpu(data)),
Device::Cuda(ordinal) => {
let backend = crate::gpu_dispatch::gpu_backend()
.ok_or(crate::error::FerrotorchError::DeviceUnavailable)?;
let bytes: &[u8] = unsafe {
std::slice::from_raw_parts(
data.as_ptr() as *const u8,
data.len() * std::mem::size_of::<T>(),
)
};
let handle =
backend.cpu_to_gpu_pinned(bytes, std::mem::size_of::<T>(), ordinal)?;
Ok(Self::gpu(handle))
}
Device::Xpu(ordinal) => Ok(Self::xpu(data, ordinal)),
Device::Meta => Ok(Self::meta(data.len())),
}
}
/// Create XPU storage with the given ordinal. The data is held as a
/// CPU-resident `Vec<T>` (no DMA today); the device marker drives
/// dispatch through the `ferrotorch-xpu` crate, which uses a
/// CubeCL wgpu runtime to actually run kernels on Intel GPUs.
/// CL-452.
pub fn xpu(data: Vec<T>, ordinal: usize) -> Self {
Self {
data: StorageBuffer::Cpu(data),
device: Device::Xpu(ordinal),
}
}
/// Create a new GPU storage from a handle.
pub fn gpu(handle: GpuBufferHandle) -> Self {
let device = Device::Cuda(handle.device_ordinal());
Self {
data: StorageBuffer::Gpu(handle),
device,
}
}
/// The device this storage resides on.
#[inline]
pub fn device(&self) -> Device {
self.device
}
/// Total number of elements in the buffer.
pub fn len(&self) -> usize {
match &self.data {
StorageBuffer::Cpu(v) => v.len(),
StorageBuffer::Gpu(h) => h.len(),
StorageBuffer::Meta { numel, .. } => *numel,
}
}
/// Whether the buffer is empty.
pub fn is_empty(&self) -> bool {
self.len() == 0
}
/// Borrow the data as a slice. Only available for CPU storage.
///
/// # Panics
/// Panics if the tensor is on a GPU device. Call `.cpu()` first.
/// Panics if the tensor is a meta tensor.
pub fn as_slice(&self) -> &[T] {
match &self.data {
StorageBuffer::Cpu(v) => v.as_slice(),
StorageBuffer::Gpu(_) => {
panic!("cannot access GPU tensor as CPU slice -- call .cpu() first")
}
StorageBuffer::Meta { .. } => {
panic!("cannot access meta tensor as a slice -- meta tensors carry no data")
}
}
}
/// Borrow the data as a mutable slice. Only available for CPU storage.
pub fn as_mut_slice(&mut self) -> &mut [T] {
match &mut self.data {
StorageBuffer::Cpu(v) => v.as_mut_slice(),
StorageBuffer::Gpu(_) => {
panic!("cannot mutate GPU tensor as CPU slice -- call .cpu() first")
}
StorageBuffer::Meta { .. } => {
panic!("cannot mutate meta tensor as a slice -- meta tensors carry no data")
}
}
}
/// Returns `true` if this storage is on CPU.
#[inline]
pub fn is_cpu(&self) -> bool {
matches!(&self.data, StorageBuffer::Cpu(_))
}
/// Returns `true` if this storage is on a GPU.
#[inline]
pub fn is_gpu(&self) -> bool {
matches!(&self.data, StorageBuffer::Gpu(_))
}
/// Returns `true` if this storage is a meta (no-data) tensor.
#[inline]
pub fn is_meta(&self) -> bool {
matches!(&self.data, StorageBuffer::Meta { .. })
}
/// Get the GPU buffer handle. Returns `None` for CPU and Meta storage.
pub fn gpu_handle(&self) -> Option<&GpuBufferHandle> {
match &self.data {
StorageBuffer::Gpu(h) => Some(h),
StorageBuffer::Cpu(_) | StorageBuffer::Meta { .. } => None,
}
}
/// Get a mutable GPU buffer handle. Returns `None` for CPU and Meta storage.
///
/// # Safety note
///
/// Callers must ensure exclusive access to the storage (e.g. via the
/// same unsafe contract as `update_data`).
pub fn gpu_handle_mut(&mut self) -> Option<&mut GpuBufferHandle> {
match &mut self.data {
StorageBuffer::Gpu(h) => Some(h),
StorageBuffer::Cpu(_) | StorageBuffer::Meta { .. } => None,
}
}
/// Fallible clone — same as `Clone::clone` but returns `Result` instead
/// of panicking when the GPU backend is missing or a CUDA call fails.
pub fn try_clone(&self) -> crate::error::FerrotorchResult<Self> {
match &self.data {
StorageBuffer::Cpu(v) => Ok(Self {
data: StorageBuffer::Cpu(v.clone()),
device: self.device,
}),
StorageBuffer::Gpu(h) => {
let backend = crate::gpu_dispatch::gpu_backend()
.ok_or(crate::error::FerrotorchError::DeviceUnavailable)?;
let cloned = backend.clone_buffer(h)?;
Ok(Self {
data: StorageBuffer::Gpu(cloned),
device: self.device,
})
}
StorageBuffer::Meta { numel, .. } => Ok(Self::meta(*numel)),
}
}
/// Clone a contiguous sub-region `[offset..offset+numel]` of this storage.
///
/// For CPU, slices the `Vec` directly. For GPU, round-trips through the
/// host to extract the sub-region (correct, not yet optimized with D2D
/// memcpy). Returns an error instead of panicking on GPU failures.
pub fn try_clone_subregion(
&self,
offset: usize,
numel: usize,
) -> crate::error::FerrotorchResult<Self> {
if offset == 0 && numel == self.len() {
return self.try_clone();
}
match &self.data {
StorageBuffer::Cpu(v) => {
let slice = &v[offset..offset + numel];
Ok(Self {
data: StorageBuffer::Cpu(slice.to_vec()),
device: self.device,
})
}
StorageBuffer::Gpu(h) => {
let backend = crate::gpu_dispatch::gpu_backend()
.ok_or(crate::error::FerrotorchError::DeviceUnavailable)?;
let bytes = backend.gpu_to_cpu(h)?;
let elem_size = std::mem::size_of::<T>();
let start = offset * elem_size;
let end = (offset + numel) * elem_size;
let handle =
backend.cpu_to_gpu(&bytes[start..end], elem_size, h.device_ordinal())?;
Ok(Self {
data: StorageBuffer::Gpu(handle),
device: self.device,
})
}
StorageBuffer::Meta { .. } => Ok(Self::meta(numel)),
}
}
}
impl<T: Element> Clone for TensorStorage<T> {
fn clone(&self) -> Self {
match &self.data {
StorageBuffer::Cpu(v) => Self {
data: StorageBuffer::Cpu(v.clone()),
device: self.device,
},
StorageBuffer::Gpu(h) => {
// Clone GPU buffer via the registered backend
if let Some(backend) = crate::gpu_dispatch::gpu_backend() {
match backend.clone_buffer(h) {
Ok(cloned) => Self {
data: StorageBuffer::Gpu(cloned),
device: self.device,
},
Err(_) => panic!("failed to clone GPU buffer"),
}
} else {
panic!("no GPU backend registered -- cannot clone GPU tensor")
}
}
StorageBuffer::Meta { numel, .. } => Self::meta(*numel),
}
}
}
impl<T: Element> Drop for TensorStorage<T> {
fn drop(&mut self) {
// Return CPU buffers to the pool for reuse.
if let StorageBuffer::Cpu(ref mut v) = self.data {
if !v.is_empty() {
// Take the Vec out, replacing with an empty one (no alloc).
let buf = std::mem::take(v);
crate::cpu_pool::pool_return_cpu(buf);
}
}
// GPU buffers are dropped normally (returned to GPU pool by CudaBuffer's Drop).
}
}
impl<T: Element> std::fmt::Debug for StorageBuffer<T> {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match self {
StorageBuffer::Cpu(v) => write!(f, "Cpu({} elements)", v.len()),
StorageBuffer::Gpu(h) => write!(f, "Gpu({h:?})"),
StorageBuffer::Meta { numel, .. } => write!(f, "Meta({numel} elements)"),
}
}
}