huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-18 19:47:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support strides in affine.

Browse Source
This commit is contained in:
laurent
2023-06-22 15:38:42 +01:00
parent 836ad5f76c
commit 5b1ab5b687
2 changed files with 26 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
src/cuda_backend.rs
View File

@ -135,19 +135,17 @@ impl CudaStorage {
mul: f64,
add: f64,
) -> Result<Self> {
if !shape.is_contiguous(stride) {
return Err(CudaError::RequiresContiguous { op: "affine" });
}
let elem_count = shape.elem_count();
let cfg = LaunchConfig::for_num_elems(elem_count as u32);
let dims = shape.dims();
let el_count = shape.elem_count();
let cfg = LaunchConfig::for_num_elems(el_count as u32);
let dev = self.device();
let ds = dev.0.htod_copy([dims, stride].concat())?;
match self {
Self::F32(arg) => {
let func = dev.get_or_load_func("affine_f32", kernels::AFFINE)?;
// SAFETY: Set later by running the kernel.
let out = unsafe { dev.0.alloc::<f32>(elem_count) }?;
let params = (elem_count, arg, &out, mul as f32, add as f32);
let out = unsafe { dev.0.alloc::<f32>(el_count) }?;
let params = (el_count, dims.len(), &ds, arg, &out, mul as f32, add as f32);
// SAFETY: ffi.
unsafe { func.launch(cfg, params) }?;
Ok(Self::F32(out))
@ -155,8 +153,8 @@ impl CudaStorage {
Self::F64(arg) => {
let func = dev.get_or_load_func("affine_f64", kernels::AFFINE)?;
// SAFETY: Set later by running the kernel.
let out = unsafe { dev.0.alloc::<f64>(elem_count) }?;
let params = (elem_count, arg, &out, mul, add);
let out = unsafe { dev.0.alloc::<f64>(el_count) }?;
let params = (el_count, dims.len(), &ds, arg, &out, mul, add);
// SAFETY: ffi.
unsafe { func.launch(cfg, params) }?;
Ok(Self::F64(out))
@ -209,13 +207,12 @@ impl CudaStorage {
let dims = shape.dims();
let cfg = LaunchConfig::for_num_elems(elem_count as u32);
let dev = self.device();
let dims_and_strides = [dims, lhs_stride, rhs_stride].concat();
let dims_and_strides = dev.0.htod_copy([dims, lhs_stride, rhs_stride].concat())?;
match (self, rhs) {
(Self::F32(lhs), Self::F32(rhs)) => {
let func = dev.get_or_load_func(B::KERNEL_F32, kernels::BINARY)?;
// SAFETY: Set later by running the kernel.
let out = unsafe { dev.0.alloc::<f32>(elem_count) }?;
let dims_and_strides = dev.0.htod_copy(dims_and_strides)?;
let params = (elem_count, dims.len(), &dims_and_strides, lhs, rhs, &out);
// SAFETY: ffi
unsafe { func.launch(cfg, params) }?;
@ -225,7 +222,6 @@ impl CudaStorage {
// SAFETY: Set later by running the kernel.
let func = dev.get_or_load_func(B::KERNEL_F64, kernels::BINARY)?;
let out = unsafe { dev.0.alloc::<f64>(elem_count) }?;
let dims_and_strides = dev.0.htod_copy(dims_and_strides)?;
let params = (elem_count, dims.len(), &dims_and_strides, lhs, rhs, &out);
// SAFETY: ffi
unsafe { func.launch(cfg, params) }?;