9 const device T* src [[buffer(0)]],
10 device T* out [[buffer(1)]],
11 const constant
int* src_shape [[buffer(2)]],
12 const constant
size_t* src_strides [[buffer(3)]],
13 const constant
size_t& src_ndim [[buffer(4)]],
14 const constant
int* slice_sizes [[buffer(5)]],
15 const constant
int* axes [[buffer(6)]],
17 uint2 index [[thread_position_in_grid]],
18 uint2 grid_dim [[threads_per_grid]]) {
19 auto ind_idx = index.x;
20 auto ind_offset = index.y;
23 for (
int i = 0; i < NIDX; ++i) {
27 }
else if (IDX_NDIM == 1) {
28 idx_loc = ind_idx * indices.strides[indices.ndim * i];
32 &indices.shapes[indices.ndim * i],
33 &indices.strides[indices.ndim * i],
37 auto idx_val =
offset_neg_idx(indices.buffers[i][idx_loc], src_shape[ax]);
38 src_idx += idx_val * src_strides[ax];
41 auto src_offset =
elem_to_loc(ind_offset, slice_sizes, src_strides, src_ndim);
43 size_t out_idx = index.y +
static_cast<size_t>(grid_dim.y) * index.x;
44 out[out_idx] = src[src_offset + src_idx];
METAL_FUNC void gather_impl(const device T *src, device T *out, const constant int *src_shape, const constant size_t *src_strides, const constant size_t &src_ndim, const constant int *slice_sizes, const constant int *axes, const thread Indices< IdxT, NIDX > &indices, uint2 index, uint2 grid_dim)
Definition gather.h:8