fixing merge errors: unary_element_wise_operation.hpp

example/62_convnd_activ/dynamic_unary/CMakeLists.txt

+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+      add_custom_target(example_convnd_activ_dynamic_unary_xdl)
+      # Sigmoid
+      add_example_executable(example_convnd_fwd_xdl_dynamic_sigmoid_fp16 convnd_fwd_xdl_dynamic_sigmoid_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_sigmoid_fp16)
+      # Tanh
+      add_example_executable(example_convnd_fwd_xdl_dynamic_tanh_fp16 convnd_fwd_xdl_dynamic_tanh_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_tanh_fp16)
+      # Relu
+      add_example_executable(example_convnd_fwd_xdl_dynamic_relu_fp16 convnd_fwd_xdl_dynamic_relu_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_relu_fp16)
+      # SoftRelu
+      add_example_executable(example_convnd_fwd_xdl_dynamic_softrelu_fp16 convnd_fwd_xdl_dynamic_softrelu_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_softrelu_fp16)
+      # Abs
+      add_example_executable(example_convnd_fwd_xdl_dynamic_abs_fp16 convnd_fwd_xdl_dynamic_abs_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_abs_fp16)
+      # Pow
+      add_example_executable(example_convnd_fwd_xdl_dynamic_pow_fp16 convnd_fwd_xdl_dynamic_pow_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_pow_fp16)
+      # Clipped Relu
+      add_example_executable(example_convnd_fwd_xdl_dynamic_clippedrelu_fp16 convnd_fwd_xdl_dynamic_clippedrelu_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_clippedrelu_fp16)
+      # Leaky Relu
+      add_example_executable(example_convnd_fwd_xdl_dynamic_leakyrelu_fp16 convnd_fwd_xdl_dynamic_leakyrelu_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_leakyrelu_fp16)
+      # Elu
+      add_example_executable(example_convnd_fwd_xdl_dynamic_elu_fp16 convnd_fwd_xdl_dynamic_elu_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_elu_fp16)
+      # Swish
+      add_example_executable(example_convnd_fwd_xdl_dynamic_swish_fp16 convnd_fwd_xdl_dynamic_swish_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_swish_fp16)
+      # PassThrough
+      add_example_executable(example_convnd_fwd_xdl_dynamic_passthrough_fp16 convnd_fwd_xdl_dynamic_passthrough_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_passthrough_fp16)
+      # Logistic
+      add_example_executable(example_convnd_fwd_xdl_dynamic_logistic_fp16 convnd_fwd_xdl_dynamic_logistic_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_dynamic_unary_xdl example_convnd_fwd_xdl_dynamic_logistic_fp16)
+   
+   set(target 1)
+ endif()
+endforeach()

example/62_convnd_activ/dynamic_unary/convnd_fwd_activ_dynamic_unary_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+constexpr ck::index_t NDimSpatial = 3;
+using InDataType                  = ck::half_t;
+using WeiDataType                 = ck::half_t;
+using AccDataType                 = float;
+using CShuffleDataType            = ck::half_t;
+using OutDataType                 = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InLayout  = ck::tensor_layout::convolution::GNDHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::GNDHWK;
+
+using InElementOp      = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp     = ck::tensor_operation::element_wise::PassThrough;
+using DynamicElementOp = ck::tensor_operation::element_wise::DynamicUnaryOp;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGroupedConvNDActivInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        DynamicElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv(bool do_verification,
+                      int init_method,
+                      bool time_kernel,
+                      const ck::utils::conv::ConvParam& conv_param,
+                      const HostTensorDescriptor& in_g_n_c_wis_desc,
+                      const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                      const HostTensorDescriptor& out_g_n_k_wos_desc,
+                      const InElementOp& in_element_op,
+                      const WeiElementOp& wei_element_op,
+                      const OutElementOp& out_element_op)
+{
+
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.05, 0.05});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("The device op with the specified compilation parameters does "
+                                 "not support this convolution problem.");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device, out_host, "Error: incorrect results!");
+    }
+
+    return true;
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_abs_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::UnaryAbs out_element_op;
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_clippedrelu_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::ClippedRelu out_element_op(0.f, 1.f);
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_elu_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::Elu out_element_op(2.f);
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_leakyrelu_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::LeakyRelu out_element_op(0.f);
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_logistic_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::Logistic out_element_op(1.0f);
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_passthrough_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::PassThrough out_element_op;
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_pow_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::Power out_element_op(4.f, 1.f, 2.f);
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_relu_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::Relu out_element_op;
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_sigmoid_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::Sigmoid out_element_op;
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_softrelu_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::SoftRelu out_element_op;
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_swish_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::Swish out_element_op(1.0f);
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/dynamic_unary/convnd_fwd_xdl_dynamic_tanh_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_activ_dynamic_unary_common.hpp"
+
+#include "../run_convnd_activ_dynamic_example.inc"
+
+int main(int argc, char* argv[])
+{
+
+    ck::tensor_operation::element_wise::TanH out_element_op;
+    return !run_convnd_example(argc, argv, out_element_op);
+}

example/62_convnd_activ/run_convnd_activ_dynamic_example.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <typename OutElementOp>
+bool run_convnd_example(int argc, char* argv[], const OutElementOp& out_element_op)
+{
+    print_helper_msg();
+
+    bool do_verification = true;
+    // Use floats for SoftRelu by default to avoid overflow after e^x.
+    int init_method =
+        std::is_same_v<OutElementOp, ck::tensor_operation::element_wise::SoftRelu> ? 2 : 1;
+    bool time_kernel = false;
+
+    // Following shapes are selected to avoid overflow. Expect inf in case of
+    // size increase for some elementwise ops.
+    ck::utils::conv::ConvParam conv_param{
+        3, 2, 16, 128, 8, {3, 3, 3}, {17, 17, 17}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+
+    const auto run = [&]() {
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_grouped_conv<NDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InElementOp,
+                                WeiElementOp,
+                                OutElementOp,
+                                DeviceGroupedConvNDActivInstance>(do_verification,
+                                                                  init_method,
+                                                                  time_kernel,
+                                                                  conv_param,
+                                                                  in_g_n_c_wis_desc,
+                                                                  wei_g_k_c_xs_desc,
+                                                                  out_g_n_k_wos_desc,
+                                                                  in_element_op,
+                                                                  wei_element_op,
+                                                                  out_element_op);
+    };
+
+    if(conv_param.num_dim_spatial_ == 3)
+    {
+        return run();
+    }
+
+    return false;
+}

example/65_gemm_multiply_multiply/CMakeLists.txt

 add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
 add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp)
 add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)
+add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp)
\ No newline at end of file

example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int;
+using F16 = ck::half_t;
+using FP8 = ck::f8_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = I8;
+using B0DataType       = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = F16;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using D0Layout = Row;
+using D1Layout = Col;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+struct MultiplyMultiply
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, float, float>(
+        ck::half_t& e, const float& c, const float& d0, const float& d1) const
+    {
+        const float x0_f = c * d0 * d1;
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, int, float, float>(
+        ck::half_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
+        ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::bhalf_t>(x0_f);
+    }
+};
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MultiplyMultiply;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
+    // clang-format off
+///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     DsData|     EData|     AccData|         CShuffle|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+///######|         |         |         |        |       Type|       Type|       Type|      Type|        Type|         DataType| Elementwise| Elementwise|  Elementwise| Spacialization|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+///######|         |         |         |        |           |           |           |          |            |                 |   Operation|   Operation|    Operation|               |      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+///######|         |         |         |        |           |           |           |          |            |                 |            |            |             |               |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |    S<C, D0, D1>|
+///###### RRR
+      ///<      Row,      Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   256,   128,    64,  16,   4,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, I8>;
+///###### RCR
+         <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   256,   128,    64,  16,  16,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,     S<4, 64, 1>,    S<1, 0, 2>,     S<1, 0, 2>,             2,              16,             16,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, I8>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    ck::index_t KBatch = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 12)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        KBatch = std::stoi(argv[11]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf(
+            "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, KBatch\n");
+        exit(0);
+    }
+    do_verification = false;
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{0, 2});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{0, 2});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+
+    constexpr auto I0 = ck::Number<0>{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
+                               b0_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
+                                                                   d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, NumDTensor>{I0, I0},
+                               StrideE,
+                               KBatch,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        invoker.Run(argument, StreamConfig{nullptr, false});
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/66_complex_contraction_bilinear/run_complex_contraction_bilinear_example.inc

@@ -127,44 +127,47 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     switch(init_method)
     {
-        case 0: break;
-        case 1:
+    case 0: break;
+    case 1:
 
-            a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-            b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-            d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
 
-            a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-            b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-            d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-            break;
+        a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
 
-        default:
-            a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-            b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-            d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    default:
+        a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
 
-            a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-            b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-            d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
 
-            break;
+        break;
     }
 
     DeviceMem a_device_buf_re(sizeof(ADataType) * a_ms_ks_re.mDesc.GetElementSpaceSize());
     DeviceMem b_device_buf_re(sizeof(BDataType) * b_ns_ks_re.mDesc.GetElementSpaceSize());
     DeviceMem d_device_buf_re(sizeof(DDataType) * d_ms_ns_re.mDesc.GetElementSpaceSize());
-    DeviceMem e_device_buf_re(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_re(sizeof(EDataType) *
+                              e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
 
     DeviceMem a_device_buf_img(sizeof(ADataType) * a_ms_ks_img.mDesc.GetElementSpaceSize());
     DeviceMem b_device_buf_img(sizeof(BDataType) * b_ns_ks_img.mDesc.GetElementSpaceSize());
     DeviceMem d_device_buf_img(sizeof(DDataType) * d_ms_ns_img.mDesc.GetElementSpaceSize());
-    DeviceMem e_device_buf_img(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_img(sizeof(EDataType) *
+                               e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
 
     // Intermediate Value For E Real and Img
-    DeviceMem e_device_buf_re1(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
-    DeviceMem e_device_buf_img1(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
-
+    DeviceMem e_device_buf_re1(sizeof(EDataType) *
+                               e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_img1(sizeof(EDataType) *
+                                e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
 
     a_device_buf_re.ToDevice(a_ms_ks_re.mData.data());
     b_device_buf_re.ToDevice(b_ns_ks_re.mData.data());
@@ -181,7 +184,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     // set zero for intermediate values
     e_device_buf_re1.SetZero();
     e_device_buf_img1.SetZero();
- 
+
     auto a_element_op   = AElementOp{};
     auto b_element_op   = BElementOp{};
     auto cde_element_op = CDEElementOp{alpha, beta};
@@ -189,23 +192,24 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     // device operation
     // For real Intermediate Value re_1
 
-    auto op       = DeviceOpInstance{};
-    auto invoker  = op.MakeInvoker();
-    auto argument_re1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
-                                    b_device_buf_re.GetDeviceBuffer(),
-                                    std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
-                                    e_device_buf_re1.GetDeviceBuffer(),
-                                    a_ms_ks_lengths,
-                                    a_ms_ks_strides,
-                                    b_ns_ks_lengths,
-                                    b_ns_ks_strides,
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                    e_ms_ns_lengths,
-                                    e_ms_ns_strides,
-                                    a_element_op,
-                                    b_element_op,
-                                    cde_element_op);
+    auto op      = DeviceOpInstance{};
+    auto invoker = op.MakeInvoker();
+    auto argument_re1 =
+        op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
+                        b_device_buf_re.GetDeviceBuffer(),
+                        std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
+                        e_device_buf_re1.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_re1))
     {
@@ -216,7 +220,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     float ave_time_re1 = invoker.Run(argument_re1, StreamConfig{nullptr, time_kernel});
 
-
     alpha = -1.f;
     beta  = 1.f;
 
@@ -228,21 +231,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     // For real Intermediate Value re_2
     // auto op       = DeviceOpInstance{};
     // auto invoker  = op.MakeInvoker();
-    auto argument_re2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
-                                    b_device_buf_img.GetDeviceBuffer(),
-                                    std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
-                                    e_device_buf_re.GetDeviceBuffer(),
-                                    a_ms_ks_lengths,
-                                    a_ms_ks_strides,
-                                    b_ns_ks_lengths,
-                                    b_ns_ks_strides,
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                    e_ms_ns_lengths,
-                                    e_ms_ns_strides,
-                                    a_element_op,
-                                    b_element_op,
-                                    cde_element_op);
+    auto argument_re2 =
+        op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
+                        b_device_buf_img.GetDeviceBuffer(),
+                        std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
+                        e_device_buf_re.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_re2))
     {
@@ -253,7 +257,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     float ave_time_re2 = invoker.Run(argument_re2, StreamConfig{nullptr, time_kernel});
 
-    
     alpha = 1.f;
     beta  = 1.f;
 
@@ -261,22 +264,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     b_element_op   = BElementOp{};
     cde_element_op = CDEElementOp{alpha, beta};
 
-    auto argument_img1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
-                                b_device_buf_img.GetDeviceBuffer(),
-                                std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
-                                e_device_buf_img1.GetDeviceBuffer(),
-                                a_ms_ks_lengths,
-                                a_ms_ks_strides,
-                                b_ns_ks_lengths,
-                                b_ns_ks_strides,
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                e_ms_ns_lengths,
-                                e_ms_ns_strides,
-                                a_element_op,
-                                b_element_op,
-                                cde_element_op);
-
+    auto argument_img1 =
+        op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
+                        b_device_buf_img.GetDeviceBuffer(),
+                        std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
+                        e_device_buf_img1.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_img1))
     {
@@ -290,23 +293,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     alpha = 1.f;
     beta  = 1.f;
 
-    auto argument_img2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
-                                b_device_buf_re.GetDeviceBuffer(),
-                                std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
-                                e_device_buf_img.GetDeviceBuffer(),
-                                a_ms_ks_lengths,
-                                a_ms_ks_strides,
-                                b_ns_ks_lengths,
-                                b_ns_ks_strides,
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                e_ms_ns_lengths,
-                                e_ms_ns_strides,
-                                a_element_op,
-                                b_element_op,
-                                cde_element_op);
-
-
+    auto argument_img2 =
+        op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
+                        b_device_buf_re.GetDeviceBuffer(),
+                        std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
+                        e_device_buf_img.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_img2))
     {
@@ -317,7 +319,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     float ave_time_img2 = invoker.Run(argument_img2, StreamConfig{nullptr, time_kernel});
 
-
     ck::index_t M =
         ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
 
@@ -331,9 +332,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
                             sizeof(DDataType) * M * N + sizeof(EDataType) * M * N * 2;
 
-    float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1 ; 
+    float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1;
 
-    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+    float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
     float gb_per_sec = num_btype / 1.E6 / ave_time;
 
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
@@ -343,7 +344,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     e_device_buf_img.FromDevice(e_ms_ns_device_result_img.mData.data());
 
     auto isRealOk = 0;
-    auto isImgOk = 0;
+    auto isImgOk  = 0;
 
     if(do_verification)
     {
@@ -366,17 +367,16 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
         auto ref_op      = ReferenceOpInstance{};
         auto ref_invoker = ref_op.MakeInvoker();
 
-        auto ref_argument_re =
-            ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
+        auto ref_argument_re = ref_op.MakeArgument(
+            a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
 
         ref_invoker.Run(ref_argument_re);
 
         alpha = 1.f;
         beta  = 1.f;
-   
+
         cde_element_op = CDEElementOp{alpha, beta};
 
-       
         for(size_t m0 = 0; m0 < e_ms_ns_host_result_re.mDesc.GetLengths()[0]; ++m0)
         {
             for(size_t m1 = 0; m1 < e_ms_ns_host_result_re.mDesc.GetLengths()[1]; ++m1)
@@ -395,11 +395,11 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
         alpha = 1.f;
         beta  = -1.f;
-   
+
         cde_element_op = CDEElementOp{alpha, beta};
 
-        auto ref_argument_re1 =
-            ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
+        auto ref_argument_re1 = ref_op.MakeArgument(
+            a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
 
         ref_invoker.Run(ref_argument_re1);
 
@@ -419,23 +419,20 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
             }
         }
 
-        isRealOk =  ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
-
-        
-
+        isRealOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
 
         // Img Part Verification
         Tensor<CShuffleDataType> c_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides);
         Tensor<CShuffleDataType> c_ms_ns_host_result_img1(e_ms_ns_lengths, e_ms_ns_strides);
 
-        auto ref_argument_img =
-            ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
-            
+        auto ref_argument_img = ref_op.MakeArgument(
+            a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
+
         ref_invoker.Run(ref_argument_img);
 
         alpha = 1.f;
         beta  = 1.f;
-   
+
         cde_element_op = CDEElementOp{alpha, beta};
 
         for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
@@ -454,9 +451,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
             }
         }
 
-        auto ref_argument_img1 =
-            ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
-            
+        auto ref_argument_img1 = ref_op.MakeArgument(
+            a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
+
         ref_invoker.Run(ref_argument_img1);
 
         for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
@@ -475,7 +472,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
             }
         }
 
-        isImgOk =  ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
+        isImgOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
 
         return (isRealOk && isImgOk);
     }

example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py

@@ -21,6 +21,14 @@ DTYPE_BITS = {
     "bf8" : 8
 }
 
+K0_MAX_SUBMAX_MAP = {
+    32 : 32,
+    64 : 64,
+    96 : 128,
+    128: 128,
+    256: 256
+}
+
 TILE_PARTITIONER_MAP = {
     "shb" : "ck_tile::FmhaFwdTilePartitioner_SHB",
     "hbs" : "ck_tile::FmhaFwdTilePartitioner_HBS",
@@ -35,14 +43,13 @@ FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
 FMHA_FWD_KERNEL_BODY="""
 using fmha_dtype_{F_idx} = {F_dtype};
 
-using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}>;
-using fmha_block_warps_{F_idx} = ck_tile::sequence<{F_rm}, {F_rn}, {F_rk}>;
+using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>;
 using fmha_warp_tile_{F_idx} = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
 
 using fmha_shape_{F_idx} = ck_tile::TileFmhaShape<fmha_block_tile_{F_idx},
-                                      fmha_block_warps_{F_idx},
+                                      ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>,
                                       fmha_warp_tile_{F_idx},
-                                      fmha_block_warps_{F_idx},
+                                      ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>,
                                       fmha_warp_tile_{F_idx},
                                       {F_vlayout}>;
 
@@ -88,7 +95,7 @@ using fmha_kernel_{F_idx} =
                   fmha_pipeline_{F_idx},
                   fmha_epilogue_{F_idx}>;
 
-using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout},
+using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
                         {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
 
 #include <iostream>
@@ -126,7 +133,7 @@ FMHA_FWD_API_PER_HDIM_CASE="""        {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <
 
 FMHA_FWD_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse})  && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) &&
                         ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
-                using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+                using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
                 return fmha_fwd_<trait_>(s, a);
             }}
 """
@@ -143,7 +150,7 @@ class FmhaFwdApiTrait:
     bk0       : int  # tile size along qk gemm unroll
     bn1       : int  # tile size along v head_dim
     bk1       : int  # tile size along kv gemm unroll
-    bk0blen   : int
+    bk0max    : int
     vlayout   : str
     mask      : str
     bias      : str  #
@@ -157,7 +164,7 @@ class FmhaFwdApiTrait:
 
     @property
     def name(self) -> str:
-        return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0blen}-'+\
+        return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0max}-'+\
                     f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.dropout}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}'
 
     @property
@@ -189,8 +196,9 @@ class FmhaFwdApiTrait:
             if self.dpad == 't': return f'a.hdim_q % {vec} == 0'
             else :               assert False
         elif self.pipeline_tag in ['qr']:
-            if self.dpad == 't': return f'true /*a.hdim_q % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
-            else :               return f'a.hdim_q % {self.bk0blen} == 0'
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dpad == 't': return f'true /*a.hdim_q % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :               return f'a.hdim_q % {bk0submax} == 0'
         else:   assert False
 
     @property
@@ -200,8 +208,9 @@ class FmhaFwdApiTrait:
             if self.dvpad == 't': return f'a.hdim_v % {vec} == 0'
             else :                assert False
         elif self.pipeline_tag in ['qr']:
-            if self.dvpad == 't': return f'true /*a.hdim_v % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
-            else :                return f'a.hdim_v % {self.bk0blen} == 0'
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dvpad == 't': return f'true /*a.hdim_v % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.hdim_v % {bk0submax} == 0'
         else:   assert False
 
 @dataclass
@@ -272,7 +281,7 @@ class FmhaFwdApiPool:
                                    F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] ,
                                    F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
                                    F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
-                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen,
+                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
                                    F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
                 if_j = 'if' if j == 0 else 'else if'
                 per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
@@ -290,19 +299,22 @@ class FmhaFwdTileSize:
     F_bk0       : int  # tile size along qk gemm unroll
     F_bn1       : int  # tile size along v head_dim
     F_bk1       : int  # tile size along kv gemm unroll
-    F_bk0blen   : int  # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile)
-    F_rm        : int  # number of warps along q seqlen (block warps)
-    F_rn        : int  # number of warps along k seqlen(not used)
-    F_rk        : int  # number of warps along gemm-k(not used)
+    F_bk0max    : int  # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile)
+    F_rm0       : int  # number of warps for gemm0 along q seqlen
+    F_rn0       : int  # number of warps for gemm0 along k seqlen 
+    F_rk0       : int  # number of warps for gemm0 along head dim q (not used)
+    F_rm1       : int  # number of warps for gemm1 along q seqlen
+    F_rn1       : int  # number of warps for gemm1 along head dim v
+    F_rk1       : int  # number of warps for gemm1 along k seqlen (not used)
     F_wm        : int  # warp size along m (warp size)
     F_wn        : int  # warp size along n
     F_wk        : int  # warp size along k
     F_occupancy : int  # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
     @property
     def name(self) -> str:
-        return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bn1}x{self.F_bk1}x{self.F_bk0blen}" +\
-        f"_r{self.F_rm}x{self.F_rn}x{self.F_rk}_w{self.F_wm}x{self.F_wn}x{self.F_wk}" +\
-            ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
+        return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bn1}x{self.F_bk1}x{self.F_bk0max}" +\
+        f"_r{self.F_rm0}x{self.F_rn0}x{self.F_rk0}_r{self.F_rm1}x{self.F_rn1}x{self.F_rk1}" +\
+        f"_w{self.F_wm}x{self.F_wn}x{self.F_wk}" + ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
 
 @dataclass
 class FmhaFwdKernel:
@@ -333,10 +345,13 @@ class FmhaFwdKernel:
                 F_bk0           = self.F_tile.F_bk0,
                 F_bn1           = self.F_tile.F_bn1,
                 F_bk1           = self.F_tile.F_bk1,
-                F_bk0blen       = self.F_tile.F_bk0blen,
-                F_rm            = self.F_tile.F_rm,
-                F_rn            = self.F_tile.F_rn,
-                F_rk            = self.F_tile.F_rk,
+                F_bk0max        = self.F_tile.F_bk0max,
+                F_rm0           = self.F_tile.F_rm0,
+                F_rn0           = self.F_tile.F_rn0,
+                F_rk0           = self.F_tile.F_rk0,
+                F_rm1           = self.F_tile.F_rm1,
+                F_rn1           = self.F_tile.F_rn1,
+                F_rk1           = self.F_tile.F_rk1,
                 F_wm            = self.F_tile.F_wm,
                 F_wn            = self.F_tile.F_wn,
                 F_wk            = self.F_tile.F_wk,
@@ -377,7 +392,7 @@ class FmhaFwdKernel:
                 bk0=self.F_tile.F_bk0,
                 bn1=self.F_tile.F_bn1,
                 bk1=self.F_tile.F_bk1,
-                bk0blen=self.F_tile.F_bk0blen,
+                bk0max=self.F_tile.F_bk0max,
                 vlayout=self.F_pipeline.F_vlayout,
                 mask=self.F_pipeline.F_mask,
                 bias=self.F_pipeline.F_bias,
@@ -394,16 +409,17 @@ class FmhaFwdKernel:
 def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
     if dtype == 'fp16' or dtype == 'bf16':
         return {
-            '32'  : FmhaFwdTileSize(128, 64, 16, 32, 32, 32,     2, 1, 1, 32, 32, 16, -1),
-            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     4, 1, 1, 32, 32, 16, -1),
-            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 16, -1),
-            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 16, -1),
+            '32'  : FmhaFwdTileSize(128, 64, 16, 32, 32, 32,     2, 1, 1,  2, 1, 1,  32, 32, 16, -1),
+            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     4, 1, 1,  4, 1, 1,  32, 32, 16, -1),
+            ## '96'  : FmhaFwdTileSize(128, 128, 32, 128, 32, 96,   4, 1, 1,  4, 1, 1,  32, 32, 16, -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1,  4, 1, 1,  32, 32, 16, -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1,  4, 1, 1,  32, 32, 16, -1),
         }
     elif dtype == 'fp8' or dtype == 'bf8':
         return {
-            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     2, 1, 1, 32, 32, 32, -1),
-            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 32, -1),
-            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 32, -1)
+            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     2, 1, 1,  2, 1, 1,  32, 32, 32, -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1,  4, 1, 1,  32, 32, 32, -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1,  4, 1, 1,  32, 32, 32, -1)
         }
     else:
         return None
@@ -505,4 +521,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im
         _, kernels = get_fwd_blobs(kernel_filter, receipt, mask_impl)
         for kernel in kernels:
             f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
-        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_API_FILENAME) + "\n")
\ No newline at end of file
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_API_FILENAME) + "\n")

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py

@@ -29,6 +29,14 @@ DTYPE_BITS = {
     "bf8" : 8
 }
 
+K0_MAX_SUBMAX_MAP = {
+    32 : 32,
+    64 : 64,
+    96 : 128,
+    128: 128,
+    256: 256
+}
+
 FMHA_FWD_SPLITKV_PIPELINE_MAP = {
     "qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS",
     "qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync",
@@ -41,14 +49,13 @@ using fmha_mask_{F_idx} = {F_mask};
 namespace {{
 template <bool kHasUnevenSplits>
 struct kernel_runner {{
-using fmha_block_tile = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}>;
-using fmha_block_warps = ck_tile::sequence<{F_rm}, {F_rn}, {F_rk}>;
+using fmha_block_tile = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>;
 using fmha_warp_tile = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
 
 using fmha_shape = ck_tile::TileFmhaShape<fmha_block_tile,
-                                          fmha_block_warps,
+                                          ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>,
                                           fmha_warp_tile,
-                                          fmha_block_warps,
+                                          ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>,
                                           fmha_warp_tile,
                                           {F_vlayout}>;
 
@@ -104,7 +111,7 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
 }};
 }}
 
-using trait_{F_idx} = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout},
+using trait_{F_idx} = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
                         {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, 
                         {F_dvpad}>;
 
@@ -162,10 +169,12 @@ using fmha_pipeline_problem = ck_tile::BlockFmhaSplitKVCombinePipelineProblem<
 using fmha_pipeline = ck_tile::BlockFmhaFwdSplitKVCombinePipeline<
     fmha_pipeline_problem>;
 
+/// FIXME: use {F_spad}/{F_dvpad} as kPadM/kPadN parameters after solving
+///        store_tile_raw() data corruption issue
 using fmha_epilogue =
     ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
                                            typename FmhaFwdTypeConfig<{F_dtype}>::ODataType,
-                                           {F_spad}, {F_dvpad}>>;
+                                           false, false>>;
 
 using fmha_kernel =
     ck_tile::FmhaFwdSplitKVCombineKernel<ck_tile::FmhaFwdSplitKVCombineTilePartitioner<{F_bm0}, {F_bn1}>,
@@ -191,7 +200,9 @@ using trait_{F_idx} = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_m
 template<>
 void fmha_fwd_splitkv_combine_oneshot_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
 {{
-    if (a.num_splits <= 16) {{
+    if (a.num_splits <= 8) {{
+        kernel_runner<3>::run(s, a);
+    }} else if (a.num_splits <= 16) {{
         kernel_runner<4>::run(s, a);
     }} else if (a.num_splits <= 32) {{
         kernel_runner<5>::run(s, a);
@@ -238,8 +249,8 @@ float fmha_fwd_splitkv(fmha_fwd_splitkv_traits t, fmha_fwd_splitkv_args a, const
 
 FMHA_FWD_SPLITKV_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse}) && (t.do_fp8_static_quant == {F_squant}) &&
                         ((a.block_table_ptr != nullptr) == {F_pagedkv}) && ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
-                using traits_ = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
-                using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}, {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
+                using traits_ = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+                using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}/2, {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
 
                 return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
             }}
@@ -257,7 +268,7 @@ class FmhaFwdSplitKVApiTrait:
     bk0       : int  # tile size along qk gemm unroll
     bn1       : int  # tile size along v head_dim
     bk1       : int  # tile size along kv gemm unroll
-    bk0blen   : int
+    bk0max    : int
     vlayout   : str
     mask      : str
     bias      : str  #
@@ -267,11 +278,11 @@ class FmhaFwdSplitKVApiTrait:
     skpad     : str
     dpad      : str
     dvpad     : str
-    pagedkv : str
+    pagedkv   : str
 
     @property
     def name(self) -> str:
-        return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0blen}-'+\
+        return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0max}-'+\
                     f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-'+\
                     f'{self.dvpad}-{self.pagedkv}'
 
@@ -304,8 +315,9 @@ class FmhaFwdSplitKVApiTrait:
             if self.dpad == 't': return f'a.hdim_q % {vec} == 0'
             else :               assert False
         elif self.pipeline_tag in ['qr']:
-            if self.dpad == 't': return f'true /*a.hdim_q % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
-            else :               return f'a.hdim_q % {self.bk0blen} == 0'
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dpad == 't': return f'true /*a.hdim_q % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :               return f'a.hdim_q % {bk0submax} == 0'
         else:   assert False
 
     @property
@@ -315,8 +327,9 @@ class FmhaFwdSplitKVApiTrait:
             if self.dvpad == 't': return f'a.hdim_v % {vec} == 0'
             else :                assert False
         elif self.pipeline_tag in ['qr']:
-            if self.dvpad == 't': return f'true /*a.hdim_v % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
-            else :                return f'a.hdim_v % {self.bk0blen} == 0'
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dvpad == 't': return f'true /*a.hdim_v % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.hdim_v % {bk0submax} == 0'
         else:   assert False
 
 @dataclass
@@ -411,7 +424,7 @@ class FmhaFwdSplitKVApiPool:
                                    F_lse=BOOL_MAP[trait.lse], F_squant=BOOL_MAP[trait.squant], F_pagedkv=BOOL_MAP[trait.pagedkv], 
                                    F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
                                    F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
-                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen,
+                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
                                    F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
                 if_j = 'if' if j == 0 else 'else if'
                 per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
@@ -455,10 +468,13 @@ class FmhaFwdSplitKVKernel:
                 F_bk0           = self.F_tile.F_bk0,
                 F_bn1           = self.F_tile.F_bn1,
                 F_bk1           = self.F_tile.F_bk1,
-                F_bk0blen       = self.F_tile.F_bk0blen,
-                F_rm            = self.F_tile.F_rm,
-                F_rn            = self.F_tile.F_rn,
-                F_rk            = self.F_tile.F_rk,
+                F_bk0max        = self.F_tile.F_bk0max,
+                F_rm0           = self.F_tile.F_rm0,
+                F_rn0           = self.F_tile.F_rn0,
+                F_rk0           = self.F_tile.F_rk0,
+                F_rm1           = self.F_tile.F_rm1,
+                F_rn1           = self.F_tile.F_rn1,
+                F_rk1           = self.F_tile.F_rk1,
                 F_wm            = self.F_tile.F_wm,
                 F_wn            = self.F_tile.F_wn,
                 F_wk            = self.F_tile.F_wk,
@@ -498,7 +514,7 @@ class FmhaFwdSplitKVKernel:
                 bk0=self.F_tile.F_bk0,
                 bn1=self.F_tile.F_bn1,
                 bk1=self.F_tile.F_bk1,
-                bk0blen=self.F_tile.F_bk0blen,
+                bk0max=self.F_tile.F_bk0max,
                 vlayout=self.F_pipeline.F_vlayout,
                 mask=self.F_pipeline.F_mask,
                 bias=self.F_pipeline.F_bias,
@@ -551,16 +567,17 @@ class FmhaFwdSplitKVCombineKernel:
 def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
     if dtype == 'fp16' or dtype == 'bf16':
         return {
-                '32'  : FmhaFwdTileSize(128, 64, 16, 32, 32, 32,     2, 1, 1, 32, 32, 16, -1),
-                '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     4, 1, 1, 32, 32, 16, -1),
-                '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 16, -1),
-                '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 16, -1),
+            '32'  : FmhaFwdTileSize(32, 64,  16, 32,  32,  32,   2, 1, 1,  2, 1, 1,  16, 16, 16, -1),
+            '64'  : FmhaFwdTileSize(64, 64,  32, 64,  32,  64,   4, 1, 1,  4, 1, 1,  16, 16, 16, -1),
+            ## '96'  : FmhaFwdTileSize(64, 128, 32, 128, 32,  96,   4, 1, 1,  4, 1, 1,  16, 16, 16, -1),
+            '128' : FmhaFwdTileSize(64, 128, 32, 128, 32,  128,  4, 1, 1,  4, 1, 1,  16, 16, 16, -1),
+            '256' : FmhaFwdTileSize(64, 128, 32, 256, 32,  256,  4, 1, 1,  4, 1, 1,  16, 16, 16, -1),
         }
     elif dtype == 'fp8' or dtype == 'bf8':
         return {
-            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     2, 1, 1, 32, 32, 32, -1),
-            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 32, -1),
-            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 32, -1)
+            '64'  : FmhaFwdTileSize(128, 64,  32, 64,  32,  64,   2, 1, 1,  2, 1, 1,  32, 32, 32, -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32,  128,  4, 1, 1,  4, 1, 1,  32, 32, 32, -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32,  256,  4, 1, 1,  4, 1, 1,  32, 32, 32, -1)
         }
     else:
         return None
@@ -568,16 +585,17 @@ def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
 def get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
     if dtype == 'fp16' or dtype == 'bf16':
         return {
-                '32'  : FmhaFwdSplitKVCombineTileSize(64, 32, -1),
-                '64'  : FmhaFwdSplitKVCombineTileSize(64, 64, -1),
-                '128' : FmhaFwdSplitKVCombineTileSize(64, 128, -1),
-                '256' : FmhaFwdSplitKVCombineTileSize(64, 256, -1),
+            '32'  : FmhaFwdSplitKVCombineTileSize(16, 16,  -1),
+            '64'  : FmhaFwdSplitKVCombineTileSize(32, 32,  -1),
+            ## '96' : FmhaFwdSplitKVCombineTileSize(32, 64,  -1),
+            '128' : FmhaFwdSplitKVCombineTileSize(32, 64,  -1),
+            '256' : FmhaFwdSplitKVCombineTileSize(32, 128, -1),
     }
     elif dtype == 'fp8' or dtype == 'bf8':
         return {
-                '64'  : FmhaFwdSplitKVCombineTileSize(64, 64, -1),
-                '128' : FmhaFwdSplitKVCombineTileSize(64, 128, -1),
-                '256' : FmhaFwdSplitKVCombineTileSize(64, 256, -1),
+            '64'  : FmhaFwdSplitKVCombineTileSize(64, 32,  -1),
+            '128' : FmhaFwdSplitKVCombineTileSize(64, 64,  -1),
+            '256' : FmhaFwdSplitKVCombineTileSize(64, 128, -1),
         }
     else:
         return None
@@ -598,7 +616,7 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
         if dtype in ['fp16', 'bf16']:
             for mask, bias, lse, pagedkv in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"]):
                 # TODO: use async pipeline when compiler is more stable 
-                if hdim == 256 or hdim in [32, 64, 128]:
+                if hdim == 256 or hdim in [32, 64, 128]:         ### [32, 64, 96, 128]:
                 # if True:
                     pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
                     pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
@@ -737,4 +755,4 @@ def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_im
         _, kernels = get_fwd_splitkv_blobs(kernel_filter, receipt, mask_impl)
         for kernel in kernels:
             f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
-        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_SPLITKV_API_FILENAME) + "\n")
\ No newline at end of file
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_SPLITKV_API_FILENAME) + "\n")