add more instances for bfp16 gemm (#1036)

* add more instances for bfp16 * reduce the gemm input values to prevent round-off errors --------- Co-authored-by: Jing Zhang <jizha@amd.com> Co-authored-by: illsilin <Illia.Silin@amd.com>

add more instances for bfp16 gemm (#1036)
* add more instances for bfp16 * reduce the gemm input values to prevent round-off errors --------- Co-authored-by: Jing Zhang <jizha@amd.com> Co-authored-by: illsilin <Illia.Silin@amd.com>
600fc000 · zjing14 · GitHub · 49e52bb3 · 600fc000 · 600fc000
Unverified Commit 600fc000 authored Nov 11, 2023 by zjing14 Committed by GitHub Nov 11, 2023
3 changed files
--- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
--- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
--- a/profiler/include/profiler/profile_gemm_impl.hpp
+++ b/profiler/include/profiler/profile_gemm_impl.hpp
@@ -76,7 +76,7 @@ int profile_gemm_impl(int do_verification,
        break;
    default:
        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 0.1});
-        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.05, 0.05});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.01, 0.01});
    }

    using AElementOp = ck::tensor_operation::element_wise::PassThrough;