From caf7b9578f276a3f8473f9177a15b2de5ef5f500 Mon Sep 17 00:00:00 2001 From: IctHu Date: Thu, 31 Oct 2024 19:11:36 +0800 Subject: [PATCH] [fix] change conv default algo and set with corresponding correct index --- src/ppl/nn/engines/cuda/optimizer/algos/algo_conv_hmma.cc | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/ppl/nn/engines/cuda/optimizer/algos/algo_conv_hmma.cc b/src/ppl/nn/engines/cuda/optimizer/algos/algo_conv_hmma.cc index b49147614..854ba41e6 100644 --- a/src/ppl/nn/engines/cuda/optimizer/algos/algo_conv_hmma.cc +++ b/src/ppl/nn/engines/cuda/optimizer/algos/algo_conv_hmma.cc @@ -80,8 +80,8 @@ double TuringHMMAImpgemm::ExcuteTimer(const ir::Node* node, OptKernelOptions& op return 0.0f; } else { // Give the default kernel #ifdef PPLNN_CUDA_ENABLE_KERNEL_CUT - attr_param_.extra_param.algo_info.algo_name = "nvSwzlSm75Fp16Conv_hmma1688_nhwc_fn_b32x256_w32x64_k8_buf2"; - attr_param_.extra_param.algo_info.kid = 685; + attr_param_.extra_param.algo_info.algo_name = "nvSwzlSm75Fp16Conv_hmma1688_nhwc_fn_b32x256_w32x64_k8_buf1"; + attr_param_.extra_param.algo_info.kid = 622; #else attr_param_.extra_param.algo_info.algo_name = "nvSwzlSm75Fp16Conv_hmma1688_nhwc_fn_b128x64_w64x32_k64_buf2"; attr_param_.extra_param.algo_info.kid = 5197;