From 261df74754f0feb632f1b604091da927aacb6f7c Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 23 Feb 2021 00:00:42 +0000 Subject: Avoid division with zero on CLGEMM lhs/rhs configuration Resolves COMPMID-4285 Signed-off-by: Georgios Pinitas Change-Id: I414e3ba352ed86345d6355d9ae366ea92e641842 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5153 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp | 15 ++++----- src/runtime/CL/mlgo/Common.h | 38 +++++++++++----------- 2 files changed, 26 insertions(+), 27 deletions(-) (limited to 'src/runtime') diff --git a/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp b/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp index 45a63980ca..489be356d9 100644 --- a/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp +++ b/src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.cpp @@ -105,14 +105,14 @@ GEMMConfigResult select_mlgo_gemm_config_reshaped_only_rhs(const CommonQuery &qu if(valid) { ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("MLGOHeuristics query returns gemm config: %s.", to_string(config).c_str()); + // Setting irrelevant unsigned int parameters to 1 and bool parameters to false as they do no matter + std::tie(lhs_info, rhs_info) = configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, 1, config.h0, false, config.interleave_rhs, !config.transpose_rhs, config.transpose_rhs, + config.export_cl_image); } else { ARM_COMPUTE_LOG_INFO_MSG_CORE("MLGOHeuristics query failed"); } - // Setting irrelevant unsigned int parameters to 1 and bool parameters to false as they do no matter - std::tie(lhs_info, rhs_info) = configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, 1, config.h0, false, config.interleave_rhs, !config.transpose_rhs, config.transpose_rhs, - config.export_cl_image); return GEMMConfigResult{ valid, lhs_info, rhs_info }; } @@ -140,14 +140,13 @@ GEMMConfigResult select_mlgo_gemm_config_reshaped(const CommonQuery &query) if(valid) { ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("MLGOHeuristics query returns gemm config: %s.", to_string(config).c_str()); + std::tie(lhs_info, rhs_info) = configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, config.v0, config.h0, config.interleave_lhs, config.interleave_rhs, !config.transpose_rhs, + config.transpose_rhs, config.export_cl_image); } else { ARM_COMPUTE_LOG_INFO_MSG_CORE("MLGOHeuristics query failed"); } - std::tie(lhs_info, rhs_info) = configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, config.v0, config.h0, config.interleave_lhs, config.interleave_rhs, !config.transpose_rhs, - config.transpose_rhs, - config.export_cl_image); return GEMMConfigResult{ valid, lhs_info, rhs_info }; } @@ -175,13 +174,13 @@ GEMMConfigResult select_mlgo_gemm_config_native(const CommonQuery &query) if(valid) { ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("MLGOHeuristics query returns gemm config: %s.", to_string(config).c_str()); + // Setting irrelevant unsigned int parameters to 1 and bool parameters to false as they do no matter + std::tie(lhs_info, rhs_info) = configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, 1, 1, false, false, false, false, false); } else { ARM_COMPUTE_LOG_INFO_MSG_CORE("MLGOHeuristics query failed"); } - // Setting irrelevant unsigned int parameters to 1 and bool parameters to false as they do no matter - std::tie(lhs_info, rhs_info) = configure_lhs_rhs_info(query.m, query.n, config.m0, config.n0, config.k0, 1, 1, false, false, false, false, false); return GEMMConfigResult{ valid, lhs_info, rhs_info }; } } // namespace auto_heuristics diff --git a/src/runtime/CL/mlgo/Common.h b/src/runtime/CL/mlgo/Common.h index 9e06689cc8..c451bd9062 100644 --- a/src/runtime/CL/mlgo/Common.h +++ b/src/runtime/CL/mlgo/Common.h @@ -45,35 +45,35 @@ using GEMMType = CLGEMMKernelType; /** GEMM Configuration for Native kernel */ struct GEMMConfigNative { - unsigned int m0; /**< Number of rows processed by the matrix multiplication */ - unsigned int n0; /**< Number of columns processed by the matrix multiplication */ - unsigned int k0; /**< Number of partial accumulations performed by the matrix multiplication */ + unsigned int m0{ 1 }; /**< Number of rows processed by the matrix multiplication */ + unsigned int n0{ 1 }; /**< Number of columns processed by the matrix multiplication */ + unsigned int k0{ 1 }; /**< Number of partial accumulations performed by the matrix multiplication */ }; /** GEMM Configuration for Reshaped Only RHS kernel */ struct GEMMConfigReshapedOnlyRHS { - unsigned int m0; /**< Number of rows processed by the matrix multiplication */ - unsigned int n0; /**< Number of columns processed by the matrix multiplication */ - unsigned int k0; /**< Number of partial accumulations performed by the matrix multiplication */ - unsigned int h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool interleave_rhs; /**< True if the h0 (k0xn0) blocks have to be interleaved in the output row */ - bool transpose_rhs; /**< True if the (k0xn0) block has to be transposed before been stored */ - bool export_cl_image; /**< True if the reshaped rhs has to be exported to cl_image. n0 must be equal to 4 */ + unsigned int m0{ 1 }; /**< Number of rows processed by the matrix multiplication */ + unsigned int n0{ 1 }; /**< Number of columns processed by the matrix multiplication */ + unsigned int k0{ 1 }; /**< Number of partial accumulations performed by the matrix multiplication */ + unsigned int h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool interleave_rhs{ false }; /**< True if the h0 (k0xn0) blocks have to be interleaved in the output row */ + bool transpose_rhs{ false }; /**< True if the (k0xn0) block has to be transposed before been stored */ + bool export_cl_image{ false }; /**< True if the reshaped rhs has to be exported to cl_image. n0 must be equal to 4 */ }; /** GEMM Configuration for Reshaped kernel */ struct GEMMConfigReshaped { - unsigned int m0; /**< Number of rows processed by the matrix multiplication */ - unsigned int n0; /**< Number of columns processed by the matrix multiplication */ - unsigned int k0; /**< Number of partial accumulations performed by the matrix multiplication */ - unsigned int v0; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ - unsigned int h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ - bool interleave_lhs; /**< True if the v0 (m0xk0) blocks have to be interleaved in the output row */ - bool interleave_rhs; /**< True if the h0 (k0xn0) blocks have to be interleaved in the output row */ - bool transpose_rhs; /**< True if the (k0xn0) block has to be transposed before been stored */ - bool export_cl_image; /**< True if the reshaped rhs has to be exported to cl_image. n0 must be equal to 4 */ + unsigned int m0{ 1 }; /**< Number of rows processed by the matrix multiplication */ + unsigned int n0{ 1 }; /**< Number of columns processed by the matrix multiplication */ + unsigned int k0{ 1 }; /**< Number of partial accumulations performed by the matrix multiplication */ + unsigned int v0{ 1 }; /**< Number of vertical blocks of size (m0xk0) stored on the same output row */ + unsigned int h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ + bool interleave_lhs{ false }; /**< True if the v0 (m0xk0) blocks have to be interleaved in the output row */ + bool interleave_rhs{ false }; /**< True if the h0 (k0xn0) blocks have to be interleaved in the output row */ + bool transpose_rhs{ false }; /**< True if the (k0xn0) block has to be transposed before been stored */ + bool export_cl_image{ false }; /**< True if the reshaped rhs has to be exported to cl_image. n0 must be equal to 4 */ }; } // namespace mlgo -- cgit v1.2.1