From 8b84aaa4db093ac08efa96c2cbf766e800465529 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Wed, 18 Nov 2020 14:56:09 +0000 Subject: COMPMID-3981: Fix missing fallback for export to cl_image - Fix missing fallback in the CLGEMMReshaped heuristic on Mali-G77 Change-Id: I0a243c7ed153216966d0809a3b3348f030a845eb Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4463 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas Comments-Addressed: Arm Jenkins --- .../CLGEMMReshapedKernelConfigurationValhall.cpp | 49 ++++++++++++++++++---- 1 file changed, 42 insertions(+), 7 deletions(-) diff --git a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp index 3f82dcab00..4fd446f647 100644 --- a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp +++ b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp @@ -95,6 +95,13 @@ std::pair CLGEMMReshapedKernelConfiguratio const float r_mk = static_cast(m) / static_cast(k); const float r_nk = static_cast(n) / static_cast(k); + GEMMLHSMatrixInfo lhs_info_buf; + GEMMRHSMatrixInfo rhs_info_buf; + GEMMLHSMatrixInfo lhs_info_img; + GEMMRHSMatrixInfo rhs_info_img; + + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 4, 4, false, false, true, false, false); + if(r_mk <= 0.11824845522642136) { if(workload <= 880.0) @@ -111,7 +118,11 @@ std::pair CLGEMMReshapedKernelConfiguratio } else { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + + return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img), + std::make_pair(lhs_info_buf, rhs_info_buf), + n, k, b, DataType::F16); } } else @@ -135,7 +146,11 @@ std::pair CLGEMMReshapedKernelConfiguratio { if(r_mn <= 2.545312523841858) { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + + return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img), + std::make_pair(lhs_info_buf, rhs_info_buf), + n, k, b, DataType::F16); } else { @@ -146,11 +161,19 @@ std::pair CLGEMMReshapedKernelConfiguratio { if(workload <= 2881.199951171875) { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 4, 2, false, false, true, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 4, 2, false, false, true, false, true); + + return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img), + std::make_pair(lhs_info_buf, rhs_info_buf), + n, k, b, DataType::F16); } else { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + + return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img), + std::make_pair(lhs_info_buf, rhs_info_buf), + n, k, b, DataType::F16); } } } @@ -160,16 +183,28 @@ std::pair CLGEMMReshapedKernelConfiguratio { if(r_mn <= 6.010416746139526) { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true); + + return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img), + std::make_pair(lhs_info_buf, rhs_info_buf), + n, k, b, DataType::F16); } else { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, true, false, true, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, true, false, true, false, true); + + return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img), + std::make_pair(lhs_info_buf, rhs_info_buf), + n, k, b, DataType::F16); } } else { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, true, false, true, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, true, false, true, false, true); + + return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img), + std::make_pair(lhs_info_buf, rhs_info_buf), + n, k, b, DataType::F16); } } } -- cgit v1.2.1