diff options
Diffstat (limited to 'libs/ARMComputeEx/src/core/CL/cl_kernels/batch_to_space_nd.cl')
-rw-r--r-- | libs/ARMComputeEx/src/core/CL/cl_kernels/batch_to_space_nd.cl | 70 |
1 files changed, 70 insertions, 0 deletions
diff --git a/libs/ARMComputeEx/src/core/CL/cl_kernels/batch_to_space_nd.cl b/libs/ARMComputeEx/src/core/CL/cl_kernels/batch_to_space_nd.cl new file mode 100644 index 000000000..ad6a48a02 --- /dev/null +++ b/libs/ARMComputeEx/src/core/CL/cl_kernels/batch_to_space_nd.cl @@ -0,0 +1,70 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * Copyright (c) 2016, 2017 ARM Limited. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include "helpers.h" + +#if defined(DATA_TYPE) && defined(DEPTH_OUT) && defined(BLOCK_SIZE0) && defined(BLOCK_SIZE1) && defined(BATCH_OUT) +/** Perform batch to space rearrangement of tensor + * + * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float + * @attention Output tensor batch should be given as a preprocessor argument using -DBATCH_OUT=size. e.g. -DBATCH_OUT=16 + * @attention block size should be given as a preprocessor argument using -DBLOCK_SIZE0=size. e.g. -DBLOCK_SIZE0=1 + * + * @param[in] input_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 + * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor + * + * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p inpu +t_ptr + * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in +bytes) + * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes) + * @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void batch_to_space_nd( + TENSOR4D_DECLARATION(input), + TENSOR4D_DECLARATION(output)) + { + Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0); + Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(output, DEPTH_OUT); + + int out_index[4]={0}; + int in_index[4]={0}; + + out_index[0] = get_global_id(0);//W + out_index[1] = get_global_id(1);//H + out_index[2] = get_global_id(2) % DEPTH_OUT;//C + out_index[3] = get_global_id(2) / DEPTH_OUT;//N + + in_index[0] = out_index[0]/BLOCK_SIZE1; + in_index[1] = out_index[1]/BLOCK_SIZE0; + in_index[2] = out_index[2]; + in_index[3] = out_index[3] + ((out_index[1] % BLOCK_SIZE0) * BLOCK_SIZE0 + out_index[0] % BLOCK_SIZE1) * BATCH_OUT; + + *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_index[0], in_index[1], in_index[2], in_index[3])); + } +#endif // defined(DATA_TYPE) && defined(DEPTH_IN) && defined(BLOCK_SIZE0) && defined(BLOCK_SIZE1) && defined(BATCH_OUT) |