huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-18 11:37:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add a cuda kernel for embeddings.

Browse Source
This commit is contained in:
laurent
2023-06-26 11:47:57 +01:00
parent 5952c3fa91
commit 16f0f5b9d2
5 changed files with 101 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
kernels/src/embeddings.cu Normal file
View File

@ -0,0 +1,34 @@
// WARNING: THIS IS ONLY VALID ASSUMING THAT inp IS CONTIGUOUS!
// TODO: proper error reporting when ids are larger than v_size.
#include "cuda_utils.cuh"
#include<stdint.h>
#define EMB_OP(TYPENAME, FN_NAME) \
extern "C" __global__ void FN_NAME( \
const size_t numel, \
const size_t num_dims, \
const size_t *info, \
const uint32_t *ids, \
const TYPENAME *inp, \
TYPENAME *out, \
const size_t h_size, \
const size_t v_size \
) { \
const size_t *dims = info; \
const size_t *strides = info + num_dims; \
if (is_contiguous(num_dims, dims, strides)) { \
for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
memcpy(out + i * h_size, inp + ids[i], h_size); \
} \
} \
else { \
for (unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; i < numel; i += blockDim.x * gridDim.x) { \
unsigned strided_i = get_strided_index(i, num_dims, dims, strides); \
memcpy(out + i * h_size, inp + ids[i], h_size); \
} \
} \
} \
EMB_OP(float, emb_f32)
EMB_OP(double, emb_f64)
EMB_OP(uint32_t, emb_u32)

1
kernels/src/lib.rs
View File

@ -1,4 +1,5 @@
pub const AFFINE: &str = include_str!(concat!(env!("OUT_DIR"), "/affine.ptx"));
pub const BINARY: &str = include_str!(concat!(env!("OUT_DIR"), "/binary.ptx"));
pub const EMBEDDINGS: &str = include_str!(concat!(env!("OUT_DIR"), "/embeddings.ptx"));
pub const FILL: &str = include_str!(concat!(env!("OUT_DIR"), "/fill.ptx"));
pub const UNARY: &str = include_str!(concat!(env!("OUT_DIR"), "/unary.ptx"));