diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-06-28 16:29:29 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:54 +0000 |
commit | 215b4ea6c9dee480a22070d5873b0b8cb52531a0 (patch) | |
tree | 398e552c4d01c0b84d03a873098a9183ba8f82e4 /src/core/CL/cl_kernels/convolution_layer.cl | |
parent | ad486e21e5870f41774f30825c270762e08ae71e (diff) | |
download | ComputeLibrary-215b4ea6c9dee480a22070d5873b0b8cb52531a0.tar.gz |
COMPMID-1277 - Optimizing CLIm2ColKernel for NHWC.
This patch includes:
- Im2Col optimizations for NHWC using a new data layout
- Refactoring of CLIm2ColKernel adding validation method and auto-init
- Removed im2col_reduced from CLIm2ColKernel and created a new kernel CLFlattenLayerKernel
Change-Id: I1620640b6796baa268324b33ae92cdd8de53e27c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/141241
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels/convolution_layer.cl')
-rw-r--r-- | src/core/CL/cl_kernels/convolution_layer.cl | 74 |
1 files changed, 2 insertions, 72 deletions
diff --git a/src/core/CL/cl_kernels/convolution_layer.cl b/src/core/CL/cl_kernels/convolution_layer.cl index 9335b047fe..2b75b45fe1 100644 --- a/src/core/CL/cl_kernels/convolution_layer.cl +++ b/src/core/CL/cl_kernels/convolution_layer.cl @@ -53,7 +53,7 @@ * @param[in] total_filters Total number of filters. 4th dimension of the weights matrix * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) */ -__kernel void reshape_to_columns_nchw( +__kernel void reshape_to_columns( TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(dst), #ifdef HAS_BIAS @@ -109,74 +109,4 @@ __kernel void reshape_to_columns_nchw( } } } - -/** This kernel reshapes the tensor's low three dimensions to single column - * - * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short - * - * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 - * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) - * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) - * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) - * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[out] dst_ptr Pointer to the destination tensor. Same as @p src_ptr - * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] bias_ptr Pointer to the bias tensor. Same as @p src_ptr - * @param[in] bias_stride_x Stride of the bias tensor in X dimension (in bytes) - * @param[in] bias_step_x bias_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] bias_offset_first_element_in_bytes The offset of the first element in the source tensor - * @param[in] depth The depth of the input tensor - * @param[in] width The width of the input tensor - * @param[in] height The height of the input tensor - * @param[in] total_filters Total number of filters. 4th dimension of the weights matrix - */ -__kernel void reshape_to_columns_nhwc( - TENSOR3D_DECLARATION(src), - IMAGE_DECLARATION(dst), -#ifdef HAS_BIAS - VECTOR_DECLARATION(bias), -#endif /* HAS_BIAS */ - uint depth, uint width, uint height, uint total_filters, uint dst_stride_z) -{ - Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); - bool is_last_thread = (get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1)); - - __global uchar *tmp_src_ptr = src.ptr; - __global uchar *tmp_dst_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * dst_stride_y + get_global_id(2) * width * dst_stride_y + get_global_id( - 0) * width * height * dst_stride_y; -#ifdef HAS_BIAS - __global uchar *tmp_bias_ptr = bias_ptr + bias_offset_first_element_in_bytes; -#endif /* HAS_BIAS */ - - if(is_last_thread) - { - for(uint i = 0; i < total_filters; ++i) - { - *((__global DATA_TYPE *)tmp_dst_ptr) = *((__global DATA_TYPE *)tmp_src_ptr); - -#ifdef HAS_BIAS - *((__global DATA_TYPE *)(tmp_dst_ptr + dst_stride_y)) = *((__global DATA_TYPE *)(tmp_bias_ptr)); - tmp_bias_ptr += bias_stride_x; -#endif /* HAS_BIAS */ - tmp_src_ptr += height * src_stride_z; - tmp_dst_ptr += dst_stride_x; - } - } - else - { - for(uint i = 0; i < total_filters; ++i) - { - *((__global DATA_TYPE *)tmp_dst_ptr) = *((__global DATA_TYPE *)tmp_src_ptr); - tmp_src_ptr += height * src_stride_z; - tmp_dst_ptr += dst_stride_x; - } - } -} -#endif // defined(DATA_TYPE) && defined(NUM_GROUPS)
\ No newline at end of file +#endif // defined(DATA_TYPE) |