From 34a92ab41a407679d454f437f1f3118b81dd1b34 Mon Sep 17 00:00:00 2001 From: hedaoyuan Date: Mon, 28 Aug 2017 14:58:00 +0800 Subject: [PATCH] ExpandConvLayer adds support of arm-neon acceleration. --- paddle/gserver/layers/ExpandConvLayer.cpp | 23 ++++++++++++++++++----- 1 file changed, 18 insertions(+), 5 deletions(-) diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp index 0ece2799318..0e845817694 100644 --- a/paddle/gserver/layers/ExpandConvLayer.cpp +++ b/paddle/gserver/layers/ExpandConvLayer.cpp @@ -29,6 +29,10 @@ namespace paddle { REGISTER_LAYER(exconv, ExpandConvLayer); REGISTER_LAYER(exconvt, ExpandConvLayer); +inline bool isDepthwiseConv(int channels, int groups) { + return channels == groups; +} + bool ExpandConvLayer::init(const LayerMap &layerMap, const ParameterMap ¶meterMap) { /* Initialize the basic convolutional parent class */ @@ -47,14 +51,23 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, std::vector paddings = {(size_t)paddingY_[i], (size_t)padding_[i]}; std::vector strides = {(size_t)strideY_[i], (size_t)stride_[i]}; - if (useGpu_ && (size_t)groups_[i] == (size_t)channels_[i] && !isDeconv_) { + // Convolution Layer uses the GemmConv function by default. + convType = "GemmConv"; + convGradInputType = "GemmConvGradInput"; + convGradFilterType = "GemmConvGradFilter"; + + // If depth wise convolution and useGpu == true + if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { convType = "DepthwiseConv"; convGradInputType = "DepthwiseConvGradInput"; convGradFilterType = "DepthwiseConvGradFilter"; - } else { - convType = "GemmConv"; - convGradInputType = "GemmConvGradInput"; - convGradFilterType = "GemmConvGradFilter"; + } + + // If depth wise convolution and useGpu == false and ARM-NEON + if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + convType = "NeonDepthwiseConv"; +#endif } if (FLAGS_use_nnpack && !isDeconv_) { -- GitLab