From b30dcc5ab8eb2bd37f0ab742af1ec45113d54296 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Tue, 20 Jun 2017 09:07:21 +0100 Subject: COMPMID-345 - In-place computation for Activation Layer Change-Id: I25ebfccc3d3e758cc8164e0b33805c0bb303891a Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78226 Tested-by: Kaizen Reviewed-by: Pablo Tello Reviewed-by: Georgios Pinitas --- src/core/CL/cl_kernels/activation_layer.cl | 18 ++++++++++++++---- 1 file changed, 14 insertions(+), 4 deletions(-) (limited to 'src/core/CL/cl_kernels') diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl index e3cbb6c801..136191aa22 100644 --- a/src/core/CL/cl_kernels/activation_layer.cl +++ b/src/core/CL/cl_kernels/activation_layer.cl @@ -24,6 +24,8 @@ #include "helpers.h" /** This performs an activation function floating point inputs. + * + * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time * * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short * @note Activation function should be given as a preprocessor argument using -DNAME. e.g. -DTANH @@ -48,12 +50,20 @@ * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image */ __kernel void activation_layer( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output)) + TENSOR3D_DECLARATION(input) +#if !defined IN_PLACE + , + TENSOR3D_DECLARATION(output) +#endif +) { // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); +#if defined IN_PLACE + Tensor3D output = input; +#else Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); +#endif // Load data VEC_DATA_TYPE(DATA_TYPE, 16) @@ -63,7 +73,7 @@ __kernel void activation_layer( #if defined LOGISTIC data = 1 / (1 + exp(-data)); #elif defined TANH - data = (VEC_DATA_TYPE(DATA_TYPE, 16))A * tanh((VEC_DATA_TYPE(DATA_TYPE, 16))B * data); + data = (VEC_DATA_TYPE(DATA_TYPE, 16))A * tanh((VEC_DATA_TYPE(DATA_TYPE, 16))B * data); #elif defined RELU data = max(0, data); #elif defined BRELU -- cgit v1.2.1