diff options
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp | 6 | ||||
-rw-r--r-- | src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp | 29 |
2 files changed, 28 insertions, 7 deletions
diff --git a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp index 59a2a82edf..46eeff3524 100644 --- a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp +++ b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationBifrost.cpp @@ -269,13 +269,13 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedKernelConfiguratio const float workload = (static_cast<float>(m) * static_cast<float>(n) * static_cast<float>(b)) / 20.0f; - if(workload <= 232.8000f) + if(workload <= 323.4000f) { - return configure_lhs_rhs_info(m, n, 2, 4, 4, 4, 4, true, true, true, false, false); + return configure_lhs_rhs_info(m, n, 2, 2, 8, 4, 8, false, false, false, true, false); } else { - return configure_lhs_rhs_info(m, n, 4, 4, 4, 4, 4, true, true, true, false, false); + return configure_lhs_rhs_info(m, n, 4, 8, 4, 2, 2, true, true, true, false, false); } } diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp index a2c1ed2c8e..d5b76d8eaf 100644 --- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp +++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp @@ -322,8 +322,15 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi const float r_mk = static_cast<float>(m) / static_cast<float>(k); const float r_nk = static_cast<float>(n) / static_cast<float>(k); + GEMMLHSMatrixInfo lhs_info_buf; + GEMMRHSMatrixInfo rhs_info_buf; + GEMMLHSMatrixInfo lhs_info_img; + GEMMRHSMatrixInfo rhs_info_img; + if(m == 1) { + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, false, false); + if(r_mk <= 0.0026f) { if(r_nk <= 0.4664f) @@ -332,7 +339,10 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi } else { - return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, 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 @@ -343,12 +353,17 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi } else { - return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, 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 { + std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 5, 8, 4, 1, 2, false, false, false, false, false); + if(workload <= 362.6000f) { return configure_lhs_rhs_info(m, n, 2, 2, 8, 1, 16, false, false, false, true, false); @@ -359,7 +374,10 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi { if(workload <= 708.8000f) { - return configure_lhs_rhs_info(m, n, 5, 4, 4, 1, 2, false, false, false, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 5, 4, 4, 1, 2, false, false, false, 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 { @@ -374,7 +392,10 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfi } else { - return configure_lhs_rhs_info(m, n, 5, 4, 4, 1, 2, false, false, false, false, true); + std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 5, 4, 4, 1, 2, false, false, false, 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); } } } |