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

Stride support for unary ops.

Browse Source
This commit is contained in:
laurent
2023-06-22 15:46:34 +01:00
parent 5b1ab5b687
commit 065b7a19c7
2 changed files with 14 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
kernels/src/unary.cu
View File

@ -3,11 +3,16 @@
#define UNARY_OP(TYPENAME, FN_NAME, FUNC) \ #define UNARY_OP(TYPENAME, FN_NAME, FUNC) \
extern "C" __global__ void FN_NAME( \ extern "C" __global__ void FN_NAME( \
const size_t numel, \ const size_t numel, \
const size_t num_dims, \
const size_t *info, \
const TYPENAME *inp, \ const TYPENAME *inp, \
TYPENAME *out \ TYPENAME *out \
) { \ ) { \
const size_t *dims = info; \
const size_t *strides = info + num_dims; \
for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \ for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
TYPENAME x = inp ? inp[i] : out[i]; \ unsigned strided_i = get_strided_index(i, num_dims, dims, strides); \
TYPENAME x = inp ? inp[strided_i] : out[i]; \
out[i] = FUNC; \ out[i] = FUNC; \
} \ } \
} \ } \

18
src/cuda_backend.rs
View File

@ -167,19 +167,17 @@ impl CudaStorage {
shape: &Shape, shape: &Shape,
stride: &[usize], stride: &[usize],
) -> Result<Self> { ) -> Result<Self> {
if !shape.is_contiguous(stride) { let dims = shape.dims();
return Err(CudaError::RequiresContiguous { op: "affine" }); let el_count = shape.elem_count();
} let cfg = LaunchConfig::for_num_elems(el_count as u32);
let elem_count = shape.elem_count();
let cfg = LaunchConfig::for_num_elems(elem_count as u32);
let dev = self.device(); let dev = self.device();
let ds = dev.0.htod_copy([dims, stride].concat())?;
match self { match self {
Self::F32(arg) => { Self::F32(arg) => {
let func = dev.get_or_load_func(U::KERNEL_F32, kernels::UNARY)?; let func = dev.get_or_load_func(U::KERNEL_F32, kernels::UNARY)?;
// SAFETY: Set later by running the kernel. // SAFETY: Set later by running the kernel.
let out = unsafe { dev.0.alloc::<f32>(elem_count) }?; let out = unsafe { dev.0.alloc::<f32>(el_count) }?;
let params = (elem_count, arg, &out); let params = (el_count, dims.len(), &ds, arg, &out);
// SAFETY: ffi. // SAFETY: ffi.
unsafe { func.launch(cfg, params) }?; unsafe { func.launch(cfg, params) }?;
Ok(Self::F32(out)) Ok(Self::F32(out))
@ -187,8 +185,8 @@ impl CudaStorage {
Self::F64(arg) => { Self::F64(arg) => {
let func = dev.get_or_load_func(U::KERNEL_F64, kernels::UNARY)?; let func = dev.get_or_load_func(U::KERNEL_F64, kernels::UNARY)?;
// SAFETY: Set later by running the kernel. // SAFETY: Set later by running the kernel.
let out = unsafe { dev.0.alloc::<f64>(elem_count) }?; let out = unsafe { dev.0.alloc::<f64>(el_count) }?;
let params = (elem_count, arg, &out); let params = (el_count, dims.len(), &ds, arg, &out);
// SAFETY: ffi. // SAFETY: ffi.
unsafe { func.launch(cfg, params) }?; unsafe { func.launch(cfg, params) }?;
Ok(Self::F64(out)) Ok(Self::F64(out))