diff --git a/src/solver/conv_asm_implicit_gemm_wrw_gtc_dynamic_xdlops.cpp b/src/solver/conv_asm_implicit_gemm_wrw_gtc_dynamic_xdlops.cpp index aa72abc424..855568145f 100644 --- a/src/solver/conv_asm_implicit_gemm_wrw_gtc_dynamic_xdlops.cpp +++ b/src/solver/conv_asm_implicit_gemm_wrw_gtc_dynamic_xdlops.cpp @@ -732,17 +732,15 @@ static inline std::tuple 1) { - if(c % gemm_n_per_block != 0 || gemm_m % gemm_m_per_block != 0) + if(c % gemm_n_per_block != 0) { continue; } } - else - { - if(cfg.tensor_a_thread_lengths[2] * cfg.tensor_a_thread_lengths[3] > 1) - if(gemm_m % gemm_m_per_block != 0) - continue; - } + + if(cfg.tensor_a_thread_lengths[2] * cfg.tensor_a_thread_lengths[3] > 1) + if(gemm_m % gemm_m_per_block != 0) + continue; if(wo % cfg.tensor_b_thread_lengths[1] != 0) {