From ec0113dd7749991959ae351934eea0c0d8077dcb Mon Sep 17 00:00:00 2001 From: Gunes Bayir Date: Wed, 9 Nov 2022 09:26:27 +0000 Subject: Optimize Transposed Convolution for CL backend (FP32/16) This patch optimizes transposed convolution for CL backend by rewriting it in a single kernel instead of three (flip_kernel + upsample + conv). The new kernel skips the upsampling step which reduces the input space of convolution by stride_x * stride_y, resulting in significant performance improvement. It also skips the kernel flipping by traversing the weights accordingly, thus reduces the memory footprint. Resolves: COMPMID-5676 Signed-off-by: Gunes Bayir Change-Id: I8a333212dc7c5f7f0597aa58b0d56d44814baa14 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8588 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) (limited to 'arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h') diff --git a/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h index 8ad805492d..0c59e2c86d 100644 --- a/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h +++ b/arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,6 +44,8 @@ public: /** Default constructor */ CLDeconvolutionLayer(std::shared_ptr memory_manager = nullptr); + ~CLDeconvolutionLayer(); + /** Set the input, weights, biases and output tensors. * * Valid data layouts: @@ -105,6 +107,9 @@ public: private: std::shared_ptr _memory_manager; std::unique_ptr _function; + + struct Impl; + std::unique_ptr _impl; }; } // namespace arm_compute #endif /* ARM_COMPUTE_CLDECONVOLUTIONLAYER_H */ -- cgit v1.2.1