From 78baa48308cba4101b4bcb4680f2f4ca90aeefd7 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Wed, 1 Dec 2021 09:26:14 +0000 Subject: Improve start-up direct convolution on OpenCL - Pass arguments at runtime - Rework ClConv2D heuristic to select direct convolution when OFM < IFM also for small kernel sizes Resolves COMPMID-5000 Change-Id: I9b538e29093829bc366d24d1e904341c247fa22b Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6771 Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena Comments-Addressed: Arm Jenkins --- src/gpu/cl/kernels/ClDirectConv2dKernel.cpp | 12 +++--------- src/gpu/cl/operators/ClConv2d.cpp | 10 ++++++++-- 2 files changed, 11 insertions(+), 11 deletions(-) (limited to 'src/gpu') diff --git a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp index 2d851a6982..7107def8ff 100644 --- a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp +++ b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp @@ -438,14 +438,8 @@ void ClDirectConv2dKernel::configure(const CLCompileContext &compile_context, IT build_options.add_option("-cl-fast-relaxed-math"); build_options.add_option("-DSRC_TENSOR_TYPE=BUFFER"); - build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(width_idx))); - build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(height_idx))); - build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(src->dimension(channel_idx))); build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); build_options.add_option("-DDST_TENSOR_TYPE=BUFFER"); - build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst->dimension(width_idx))); - build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(height_idx))); - build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(channel_idx))); build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type())); build_options.add_option_if_else(export_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER"); build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights->dimension(width_idx))); @@ -613,13 +607,13 @@ void ClDirectConv2dKernel::run_op(ITensorPack &tensors, const Window &window, cl } unsigned int idx = 0; - add_4D_tensor_argument(idx, src, slice); - add_4D_tensor_argument(idx, dst, slice); + add_4d_tensor_nhwc_argument(idx, src); + add_4d_tensor_nhwc_argument(idx, dst); if(export_to_cl_image) { _kernel.setArg(idx++, weights_cl_image); } - add_4D_tensor_argument(idx, weights, slice); + add_4d_tensor_nhwc_argument(idx, weights); if(biases != nullptr) { add_1D_tensor_argument(idx, biases, slice); diff --git a/src/gpu/cl/operators/ClConv2d.cpp b/src/gpu/cl/operators/ClConv2d.cpp index d633c8f738..92b22e758d 100644 --- a/src/gpu/cl/operators/ClConv2d.cpp +++ b/src/gpu/cl/operators/ClConv2d.cpp @@ -257,7 +257,8 @@ ConvolutionMethod ClConv2d::get_convolution_method(const ITensorInfo *src, const { const bool is_large_kernel_sz = (weights->dimension(idx_w) >= kernel_sz_direct_conv_thr) && (weights->dimension(idx_h) >= kernel_sz_direct_conv_thr); const bool is_ifm_ge_16 = src->dimension(idx_c) >= 16; - const bool is_ifm_gt_ofm = src->dimension(idx_c) > weights->dimension(3U); + const bool is_ifm_gt_ofm = weights->dimension(0U) * weights->dimension(1U) * weights->dimension(2U) > weights->dimension(3U); + const bool is_ofm_le_4 = weights->dimension(3U) <= 4; // Run Winograd if valid and IFM >= 16 if(is_wino_valid && is_ifm_ge_16) @@ -265,7 +266,12 @@ ConvolutionMethod ClConv2d::get_convolution_method(const ITensorInfo *src, const return ConvolutionMethod::WINOGRAD; } // Run Direct for Large kernel size - if(is_large_kernel_sz && is_ifm_ge_16 && is_direct_valid && is_ifm_gt_ofm) + if(is_large_kernel_sz && is_ifm_gt_ofm && is_direct_valid) + { + return ConvolutionMethod::DIRECT; + } + + if(is_ofm_le_4 && is_ifm_gt_ofm && is_direct_valid) { return ConvolutionMethod::DIRECT; } -- cgit v1.2.1