diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2020-08-07 15:36:30 +0100 |
---|---|---|
committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2020-08-07 18:19:52 +0000 |
commit | 4aed4aafa2ddb0b6f4b76aef5008c8bb45599ea4 (patch) | |
tree | 6b4cc56b305ec0966aac40998494075d8569e429 /src/core | |
parent | 9c7fed85d339df64937e8edac3b591b8571ccce8 (diff) | |
download | ComputeLibrary-4aed4aafa2ddb0b6f4b76aef5008c8bb45599ea4.tar.gz |
COMPMID-3683: Fix performance regression on Mali-G76 (Fully connected)
COMPMID-3682: Fix performance regression on Mali-G76 (Convolution)
Updated the heuristic for GEMMReshapedOnlYRHS for Mali-G76 in order to
take into account small workload cases
Change-Id: I99fccbd0e94e4e21c0d1b88e23f02af06ef16ee9
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3689
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp | 35 |
1 files changed, 26 insertions, 9 deletions
diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp index 581c2d2199..f9b65dc931 100644 --- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp +++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp @@ -149,34 +149,51 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi GEMMLHSMatrixInfo lhs_info_img; GEMMRHSMatrixInfo rhs_info_img; + const bool is_workload_big = ((m * n * b) / 16) >= 2048; // Get lhs_info/rhs_info in case of OpenCL buffer if(m == 1) { - const unsigned int h0 = std::max(n / 2, 1U); - std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true); + if((n / 4) >= 2048) + { + const unsigned int h0 = std::max(n / 4, 1U); + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 4, 8, 1, h0, false, true, false, true); + } + else + { + const unsigned int h0 = std::max(n / 2, 1U); + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true); + } } else { - std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, false, true, false, true); + const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(16)), static_cast<int>(1)); + if(is_workload_big) + { + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, h0, false, true, false, true); + } + else + { + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, h0, false, true, false, true); + } } // Get lhs_info/rhs_info in case of OpenCL image - if(m == 1) + const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(16)), static_cast<int>(1)); + if(is_workload_big) { - std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, 4, false, true, false, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, h0, false, true, false, false, true); } else { - const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(16)), static_cast<int>(1)); - std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, h0, false, true, false, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, h0, false, true, false, true, true); } const TensorInfo tensor_rhs_info(TensorShape(n, k, b), 1, DataType::F32); const TensorShape shape = compute_rhs_reshaped_shape(tensor_rhs_info, rhs_info_img); const TensorInfo tensor_reshaped_info(shape, 1, DataType::F32); - // In case of vector by matrix with few work-items, we use the OpenCL buffer rather than the OpenCL image2d - const bool use_cl_image2d = (m == 1 && n <= 4096) ? false : true; + // In case of vector by matrix or small workloads, we use the OpenCL buffer rather than the OpenCL image2d + const bool use_cl_image2d = ((m == 1) || ((((m * n * b) / 16) < 2048) && n < 128)) ? false : true; if(bool(validate_image2d_support_on_rhs(tensor_reshaped_info, rhs_info_img)) && use_cl_image2d) { |