diff options
Diffstat (limited to 'src/core/CL/cl_kernels/concatenate.cl')
| -rw-r--r-- | src/core/CL/cl_kernels/concatenate.cl | 17 |
1 files changed, 11 insertions, 6 deletions
diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl index 4281e675d7..0b211a6d1f 100644 --- a/src/core/CL/cl_kernels/concatenate.cl +++ b/src/core/CL/cl_kernels/concatenate.cl @@ -383,10 +383,13 @@ __kernel void concatenate_height( #endif /* defined(HEIGHT_OFFSET) && defined(DEPTH) */ +#if defined(VEC_SIZE_LEFTOVER) + /** This kernel concatenates the input tensor into the output tensor along the third dimension * * @note The data type has to be passed at compile time using -DDATA_TYPE. i.e. -DDATA_TYPE=float * @note Vector size has to be passed at compile time using -DVEC_SIZE. i.e. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE=3. It is defined as the remainder between the input's first dimension and VEC_SIZE * * @param[in] src_ptr Pointer to the source tensor. Supported data types: All * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) @@ -411,17 +414,19 @@ __kernel void concatenate( TENSOR3D_DECLARATION(dst), int offset) { - Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); - Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); + uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0); + + __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x_offs + get_global_id(1) * src_stride_y + get_global_id(2) * src_stride_z; + __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_offs + get_global_id(1) * dst_stride_y + get_global_id(2) * dst_stride_z; VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) - source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr); + source_values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src_addr); #if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) - source_values = requantize(source_values, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT); + source_values0 = requantize(source_values0, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT); #endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */ - VSTORE(VEC_SIZE) - (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offset)); + STORE_VECTOR_SELECT(source_values, DATA_TYPE, dst_addr + offset, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) } +#endif /* defined(VEC_SIZE_LEFTOVER) */ #endif /* defined(DATA_TYPE) && defined(VEC_SIZE) */ |