diff --git a/profiler/include/profiler/profile_gemm_universal_streamk_impl.hpp b/profiler/include/profiler/profile_gemm_universal_streamk_impl.hpp old mode 100644 new mode 100755 index e625fae808..38ee978f5d --- a/profiler/include/profiler/profile_gemm_universal_streamk_impl.hpp +++ b/profiler/include/profiler/profile_gemm_universal_streamk_impl.hpp @@ -338,5 +338,309 @@ bool profile_gemm_universal_streamk_impl(int do_verification, return pass; } +bool profile_gemm_universal_streamk_impl(int do_verification, + int init_method, + bool do_log, + bool time_kernel, + int M, + int N, + int K, + int StrideA, + int StrideB, + int StrideC, + int Streamk_sel, + int Grid_size, + int cold_time, + int hot_time, + uint64_t rotating = 0) +{ + bool pass = true; + + auto f_host_tensor_descriptor = + [](std::size_t row, std::size_t col, std::size_t stride, auto layout) { + using namespace ck::literals; + + if(is_same::value) + { + return HostTensorDescriptor({row, col}, {stride, 1_uz}); + } + else + { + return HostTensorDescriptor({row, col}, {1_uz, stride}); + } + }; + + Tensor a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); + Tensor b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); + Tensor c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); + + Tensor c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); + Tensor c_m_n_device_ref_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); + + int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes(); + int rotating_count = std::max( + 1, + std::min(n_iter, + static_cast(std::ceil(static_cast(rotating) / total_gemm_needed)))); + + std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; + std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; + std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl; + std::cout << "rotating count: " << rotating_count << std::endl; + + switch(init_method) + { + case 0: break; + case 1: + a_m_k.GenerateTensorValue(GeneratorTensor_2{-1, 2}); + b_k_n.GenerateTensorValue(GeneratorTensor_2{-1, 2}); + break; + default: + a_m_k.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + b_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + } + + using AElementOp = ck::tensor_operation::element_wise::PassThrough; + using BElementOp = ck::tensor_operation::element_wise::PassThrough; + using CElementOp = ck::tensor_operation::element_wise::PassThrough; + + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto c_element_op = CElementOp{}; + + DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize()); + DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize()); + + DeviceMem c_m_n_device_ref_buf(sizeof(CDataType) * + c_m_n_device_ref_result.mDesc.GetElementSpaceSize()); + + a_device_buf.ToDevice(a_m_k.mData.data()); + b_device_buf.ToDevice(b_k_n.mData.data()); + + using DeviceOp = ck::tensor_operation::device::DeviceGemm_Streamk_V2; + + // get device op instances + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "found " << op_ptrs.size() << " instances" << std::endl; + + // Run reference GEMM + if(do_verification) + { + + // Use CPU validation + // Note: GPU validation is not supported for fp8 !!! + using ReferenceGemmInstanceCPU = ck::tensor_operation::host::ReferenceGemm; + auto ref_gemm_cpu = ReferenceGemmInstanceCPU{}; + auto ref_invoker_cpu = ref_gemm_cpu.MakeInvoker(); + auto ref_argument_cpu = ref_gemm_cpu.MakeArgument( + a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op); + ref_invoker_cpu.Run(ref_argument_cpu); + } + + std::string best_op_name; + float best_ave_time = 0; + float best_tflops = 0; + float best_gb_per_sec = 0; + float best_grid_size = 0; + float best_streamk_sel = 0; + + // profile device GEMM instances + for(auto& op_ptr : op_ptrs) + { + std::vector grid_size_list = {38, 76, 114, 152, 190, 228, 266, 304, 342, 380}; + std::vector streamk_sel_list = { + 0, 1, 2, 3, 4}; // 0: Data Parallel (DP) mode (Stream-K OFF), 1: 1-tile Stream-K+ DP, + // 2:2-tile Stream-K + DP + + if(Grid_size == -1) + { + grid_size_list = {Grid_size}; + } + if(Streamk_sel != -1) + { + streamk_sel_list = {Streamk_sel}; + } + for(std::size_t j = 0; j < streamk_sel_list.size(); j++) + { + for(std::size_t i = 0; i < grid_size_list.size(); i++) + { + auto grid_size_curr = grid_size_list[i]; + index_t streamk_sel_curr = streamk_sel_list[j]; + printf("streamk_sel_curr=%0d\n", streamk_sel_curr); + auto argument_ptr = op_ptr->MakeArgumentPointer( + static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(b_device_buf.GetDeviceBuffer()), + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + StrideA, + StrideB, + StrideC, + streamk_sel_curr, + grid_size_curr, + a_element_op, + b_element_op, + c_element_op); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + + // re-init C to zero before profiling next kernel + c_device_buf.SetZero(); + + invoker_ptr->Run(argument_ptr.get(), + StreamConfig{nullptr, false, 0, n_warmup, n_iter}); + + if(do_verification) + { + c_device_buf.FromDevice(c_m_n_device_result.mData.data()); + + // Always compare against CPU reference results computed earlier + pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result); + + if(do_log) + { + LogRangeAsType(std::cout << "a : ", a_m_k.mData, ",") + << std::endl; + LogRangeAsType(std::cout << "b: ", b_k_n.mData, ",") + << std::endl; + LogRangeAsType( + std::cout << "c_host : ", c_m_n_host_result.mData, ",") + << std::endl; + LogRangeAsType( + std::cout << "c_device: ", c_m_n_device_result.mData, ",") + << std::endl; + } + } + + std::string op_name = op_ptr->GetTypeString(); + + float ave_time = invoker_ptr->Run(argument_ptr.get(), + StreamConfig{nullptr, + time_kernel, + 0, + n_warmup, + n_iter, + rotating_count > 1, + rotating_count}); + + std::size_t flop = std::size_t(2) * M * N * K; + + std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + + sizeof(CDataType) * M * N; + + float tflops = static_cast(flop) / 1.E9 / ave_time; + + float gb_per_sec = num_btype / 1.E6 / ave_time; + + std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops + << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", Grid_size " + << grid_size_curr << ", streamk selection strategy" + << streamk_sel_curr << std::endl; + +#if defined CK_ENABLE_FP8 + // set softer tolerances for fp8 + if constexpr(is_same_v || is_same_v || + is_same_v) + { + std::string msg = "Error: Incorrect results!"; + double rtol = 1e-1; + double atol = 1e-1; + pass = pass & ck::utils::check_err( + c_m_n_device_result, c_m_n_host_result, msg, rtol, atol); + } + else + { +#endif + pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result); +#if defined CK_ENABLE_FP8 + } +#endif + + if(tflops > best_tflops) + { + best_op_name = op_name; + best_tflops = tflops; + best_ave_time = ave_time; + best_gb_per_sec = gb_per_sec; + best_grid_size = grid_size_curr; + best_streamk_sel = streamk_sel_curr; + } + } + else + { + std::cout << op_ptr->GetTypeString() << " does not support this problem" + << std::endl; + } + } + } + } + + if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = f32"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = f16"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = bf16"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = int8"; + } + + if constexpr(is_same::value) + { + std::cout << " ALayout = RowMajor"; + } + else if constexpr(is_same::value) + { + std::cout << " ALayout = ColumnMajor"; + } + + if constexpr(is_same::value) + { + std::cout << " BLayout = RowMajor"; + } + else if constexpr(is_same::value) + { + std::cout << " BLayout = ColumnMajor"; + } + + std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA + << " StrideB = " << StrideB << " StrideC = " << StrideC + << " Grid_size = " << best_grid_size + << " Stream-K selection strategy = " << best_streamk_sel << " : " << best_ave_time + << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, " + << best_op_name << std::endl; + + return pass; +} + } // namespace profiler } // namespace ck