Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into wmma_op

CMakeLists.txt

@@ -247,6 +247,16 @@ message("CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
 
 add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -C ${CMAKE_CFG_INTDIR})
 
+file(GLOB_RECURSE INSTANCE_FILES "${PROJECT_SOURCE_DIR}/*/device_*_instance.cpp")
+file(GLOB dir_list RELATIVE ${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu ${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/*)
+set(CK_DEVICE_INSTANCES)
+FOREACH(subdir_path ${dir_list})
+    IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}")
+       list(APPEND CK_DEVICE_INSTANCES device_${subdir_path}_instance)
+    ENDIF()
+ENDFOREACH()
+add_custom_target(instances DEPENDS utility;${CK_DEVICE_INSTANCES}  SOURCES ${INSTANCE_FILES})
+
 rocm_package_setup_component(tests
         LIBRARY_NAME composablekernel
         PACKAGE_NAME tests # Prevent -static suffix on package name

README.md

@@ -59,7 +59,7 @@ cmake
 -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
 -D CMAKE_CXX_FLAGS="-O3"                                                                          \
 -D CMAKE_BUILD_TYPE=Release                                                                       \
--D GPU_TARGETS=gfx908;gfx90a                                                                      \
+-D GPU_TARGETS="gfx908;gfx90a"                                                                    \
 ..
 ```
 

client_example/08_fused_attention/CMakeLists.txt

+add_executable(client_fused_attention fused_attention.cpp)
+target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_operations)

client_example/08_fused_attention/fused_attention.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+using ADataType   = ck::half_t;
+using B0DataType  = ck::half_t;
+using B1DataType  = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    int G0 = 48;
+    int G1 = 16;
+    int M  = 1024;
+    int N  = 1024;
+    int K  = 64;
+    int O  = 64;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          ADataType,
+                                                                          B0DataType,
+                                                                          B1DataType,
+                                                                          CDataType,
+                                                                          ck::Tuple<>,
+                                                                          ck::Tuple<>,
+                                                                          AElementOp,
+                                                                          B0ElementOp,
+                                                                          Acc0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CElementOp,
+                                                                          MaskingSpec>;
+
+    // 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;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    G0 * G1;
+
+            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, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best instance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

dev-requirements.txt

 ROCmSoftwarePlatform/rocm-recipes
 RadeonOpenCompute/rocm-cmake@04f694df2a8dc9d7e35fa4dee4ba5fa407ec04f8 --build
-# 1.90+
-danmar/cppcheck@dd05839a7e63ef04afd34711cb3e1e0ef742882f
\ No newline at end of file
+danmar/cppcheck@2.9
\ No newline at end of file

example/12_reduce/reduce_blockwise.cpp

@@ -140,6 +140,10 @@ bool reduce_blockwise_test(bool do_verification,
         if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
             return;
 
+        std::array<int, ShapeType::NumReduceDim_> arrReduceDims;
+
+        std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+
         result = reduce_blockwise_impl<InOutDataType,
                                        AccDataType,
                                        ReduceOpId,
@@ -147,7 +151,7 @@ bool reduce_blockwise_test(bool do_verification,
                                        ShapeType::NumReduceDim_,
                                        PropagateNan,
                                        OutputIndex>(
-            do_verification, init_method, time_kernel, inLengths, reduceDims, alpha, beta);
+            do_verification, init_method, time_kernel, inLengths, arrReduceDims, alpha, beta);
 
         matched = true;
     });

example/12_reduce/reduce_blockwise_impl.hpp

@@ -30,7 +30,7 @@ int reduce_blockwise_impl(bool do_verification,
                           int init_method,
                           bool time_kernel,
                           const std::vector<size_t>& inLengths,
-                          const std::vector<int>& reduceDims,
+                          const std::array<int, NumReduceDim>& reduceDims,
                           float alpha,
                           float beta)
 
@@ -38,6 +38,8 @@ int reduce_blockwise_impl(bool do_verification,
     using namespace ck;
     using namespace ck::tensor_operation::device;
 
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
     constexpr bool op_support_indices =
         (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
          ReduceOpId == ReduceTensorOp::AMAX);
@@ -143,7 +145,7 @@ int reduce_blockwise_impl(bool do_verification,
 
     std::vector<size_t> outLengths;
 
-    std::vector<int> invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
 
     if(invariantDims.empty())
         outLengths.push_back(1);
@@ -256,22 +258,22 @@ int reduce_blockwise_impl(bool do_verification,
                        acc_elementwise_op);
     };
 
-    std::vector<ck::index_t> i_inLengths;
-    std::vector<ck::index_t> i_inStrides;
-    std::vector<ck::index_t> i_outLengths;
-    std::vector<ck::index_t> i_outStrides;
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
 
-    i_inLengths.assign(inLengths.begin(), inLengths.end());
-    i_inStrides.assign(inStrides.begin(), inStrides.end());
-    i_outLengths.assign(outLengths.begin(), outLengths.end());
-    i_outStrides.assign(outStrides.begin(), outStrides.end());
+    std::copy(inLengths.begin(), inLengths.end(), arrInLengths.begin());
+    std::copy(inStrides.begin(), inStrides.end(), arrInStrides.begin());
+    std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
+    std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
 
     auto reduce = DeviceReduceInstance{};
 
-    auto argument_ptr = reduce.MakeArgumentPointer(i_inLengths,
-                                                   i_inStrides,
-                                                   i_outLengths,
-                                                   i_outStrides,
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
                                                    reduceDims,
                                                    alpha,
                                                    beta,

example/12_reduce/reduce_blockwise_two_call.cpp

@@ -90,15 +90,15 @@ static bool time_kernel;
 int main(int argc, char* argv[])
 {
     // used by the device reduction
-    const std::vector<int> reduceDims_1    = {4};
-    const std::vector<int> invariantDims_1 = {0, 1, 2, 3};
+    const std::array<int, 1> reduceDims_1 = {4};
+    // const std::array<int, 4> invariantDims_1 = {0, 1, 2, 3};
 
-    const std::vector<int> reduceDims_2    = {3};
-    const std::vector<int> invariantDims_2 = {0, 1, 2};
+    const std::array<int, 1> reduceDims_2 = {3};
+    // const std::array<int, 3> invariantDims_2 = {0, 1, 2};
 
     // used by the host reduction
-    const std::vector<int> reduceDims    = {3, 4};
-    const std::vector<int> invariantDims = {0, 1, 2};
+    const std::array<int, 2> reduceDims    = {3, 4};
+    const std::array<int, 3> invariantDims = {0, 1, 2};
 
     const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
 
@@ -214,26 +214,26 @@ int main(int argc, char* argv[])
                        acc_elementwise_op);
     };
 
-    std::vector<ck::index_t> i_inLengths_1;
-    std::vector<ck::index_t> i_inStrides_1;
-    std::vector<ck::index_t> i_inLengths_2;
-    std::vector<ck::index_t> i_inStrides_2;
-    std::vector<ck::index_t> i_outLengths;
-    std::vector<ck::index_t> i_outStrides;
+    std::array<index_t, 5> arrInLengths_1;
+    std::array<index_t, 5> arrInStrides_1;
+    std::array<index_t, 4> arrInLengths_2;
+    std::array<index_t, 4> arrInStrides_2;
+    std::array<index_t, 3> arrOutLengths;
+    std::array<index_t, 3> arrOutStrides;
 
-    i_inLengths_1.assign(inLengths_1.begin(), inLengths_1.end());
-    i_inStrides_1.assign(inStrides_1.begin(), inStrides_1.end());
-    i_inLengths_2.assign(inLengths_2.begin(), inLengths_2.end());
-    i_inStrides_2.assign(inStrides_2.begin(), inStrides_2.end());
-    i_outLengths.assign(outLengths.begin(), outLengths.end());
-    i_outStrides.assign(outStrides.begin(), outStrides.end());
+    std::copy(inLengths_1.begin(), inLengths_1.end(), arrInLengths_1.begin());
+    std::copy(inStrides_1.begin(), inStrides_1.end(), arrInStrides_1.begin());
+    std::copy(inLengths_2.begin(), inLengths_2.end(), arrInLengths_2.begin());
+    std::copy(inStrides_2.begin(), inStrides_2.end(), arrInStrides_2.begin());
+    std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
+    std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
 
     auto reduce_1 = DeviceReduceInstance_1{};
 
-    auto argument_ptr_1 = reduce_1.MakeArgumentPointer(i_inLengths_1,
-                                                       i_inStrides_1,
-                                                       i_inLengths_2,
-                                                       i_inStrides_2,
+    auto argument_ptr_1 = reduce_1.MakeArgumentPointer(arrInLengths_1,
+                                                       arrInStrides_1,
+                                                       arrInLengths_2,
+                                                       arrInStrides_2,
                                                        reduceDims_1,
                                                        1.0f,
                                                        0.0f,
@@ -255,10 +255,10 @@ int main(int argc, char* argv[])
 
     auto reduce_2 = DeviceReduceInstance_2{};
 
-    auto argument_ptr_2 = reduce_2.MakeArgumentPointer(i_inLengths_2,
-                                                       i_inStrides_2,
-                                                       i_outLengths,
-                                                       i_outStrides,
+    auto argument_ptr_2 = reduce_2.MakeArgumentPointer(arrInLengths_2,
+                                                       arrInStrides_2,
+                                                       arrOutLengths,
+                                                       arrOutStrides,
                                                        reduceDims_2,
                                                        alpha,
                                                        beta,

example/12_reduce/reduce_example_common.hpp

@@ -5,11 +5,10 @@
 
 #include "ck/ck.hpp"
 
-template <ck::index_t Rank, ck::index_t NumReduceDim>
-std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
 {
-    assert(NumReduceDim == reduceDims.size());
-
     int reduceFlag = 0;
 
     // flag the bits for the reduceDims
@@ -18,13 +17,15 @@ std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
         reduceFlag |= 1 << reduceDims[i];
     };
 
-    std::vector<int> invariantDims;
+    std::array<int, Rank - NumReduceDim> invariantDims;
 
     // collect invariant dimensions
+    int dim = 0;
     for(int i = 0; i < Rank; i++)
         if((reduceFlag & (1 << i)) == 0)
         {
-            invariantDims.push_back(i);
+            invariantDims[dim] = i;
+            dim++;
         };
 
     return invariantDims;

example/12_reduce/reduce_multiblock_atomic_add.cpp

@@ -138,13 +138,17 @@ bool reduce_multiblock_atomic_add_test(bool do_verification,
         if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
             return;
 
+        std::array<int, ShapeType::NumReduceDim_> a_reduceDims;
+
+        std::copy(reduceDims.begin(), reduceDims.end(), a_reduceDims.begin());
+
         result = reduce_multiblock_atomic_add_impl<InOutDataType,
                                                    AccDataType,
                                                    ReduceOpId,
                                                    ShapeType::Rank_,
                                                    ShapeType::NumReduceDim_,
                                                    PropagateNan>(
-            do_verification, init_method, time_kernel, inLengths, reduceDims, alpha, beta);
+            do_verification, init_method, time_kernel, inLengths, a_reduceDims, alpha, beta);
 
         matched = true;
     });

example/12_reduce/reduce_multiblock_atomic_add_impl.hpp

@@ -29,7 +29,7 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
                                       int init_method,
                                       bool time_kernel,
                                       const std::vector<size_t>& inLengths,
-                                      const std::vector<int>& reduceDims,
+                                      const std::array<int, NumReduceDim>& reduceDims,
                                       float alpha,
                                       float beta)
 
@@ -37,6 +37,8 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
     using namespace ck;
     using namespace ck::tensor_operation::device;
 
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
     constexpr bool op_support_atomic_add =
         (ReduceOpId == ReduceTensorOp::ADD || ReduceOpId == ReduceTensorOp::AVG);
 
@@ -84,7 +86,7 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
 
     std::vector<size_t> outLengths;
 
-    std::vector<int> invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
 
     if(invariantDims.empty())
         outLengths.push_back(1);
@@ -169,22 +171,22 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
                        acc_elementwise_op);
     };
 
-    std::vector<ck::index_t> i_inLengths;
-    std::vector<ck::index_t> i_inStrides;
-    std::vector<ck::index_t> i_outLengths;
-    std::vector<ck::index_t> i_outStrides;
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
 
-    i_inLengths.assign(inLengths.begin(), inLengths.end());
-    i_inStrides.assign(inStrides.begin(), inStrides.end());
-    i_outLengths.assign(outLengths.begin(), outLengths.end());
-    i_outStrides.assign(outStrides.begin(), outStrides.end());
+    std::copy(inLengths.begin(), inLengths.end(), arrInLengths.begin());
+    std::copy(inStrides.begin(), inStrides.end(), arrInStrides.begin());
+    std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
+    std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
 
     auto reduce = DeviceReduceInstance{};
 
-    auto argument_ptr = reduce.MakeArgumentPointer(i_inLengths,
-                                                   i_inStrides,
-                                                   i_outLengths,
-                                                   i_outStrides,
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
                                                    reduceDims,
                                                    alpha,
                                                    beta,

example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt

@@ -2,9 +2,11 @@ add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_
 add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
 add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
 add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
 
 add_custom_target(example_gemm_scale_softmax_gemm)
 add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
 add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
 add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
 add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp

@@ -33,9 +33,6 @@ using S = ck::Sequence<Is...>;
 using F16 = ck::half_t;
 using F32 = float;
 
-using Row = ck::tensor_layout::gemm::RowMajor;
-using Col = ck::tensor_layout::gemm::ColumnMajor;
-
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
 using ADataType        = F16;
@@ -44,13 +41,14 @@ using B1DataType       = F16;
 using AccDataType      = F32;
 using CShuffleDataType = F32;
 using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
 
-using ALayout  = Row;
-using B0Layout = Col;
-using B1Layout = Row;
-
-using CPermuteNumDims_G_M_O =
-    S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
 
 using AElementOp    = PassThrough;
 using B0ElementOp   = PassThrough;
@@ -59,17 +57,27 @@ using B1ElementOp   = PassThrough;
 using CElementOp    = PassThrough;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
 
 using DeviceGemmInstance =
     ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
-        ALayout,
-        B0Layout,
-        B1Layout,
-        CPermuteNumDims_G_M_O,
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
         ADataType,
         B0DataType,
         B1DataType,
         CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
         AccDataType,
         CShuffleDataType,
         AElementOp,
@@ -78,6 +86,10 @@ using DeviceGemmInstance =
         B1ElementOp,
         CElementOp,
         GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
         1,
         256,
         128,         // MPerBlock
@@ -118,7 +130,7 @@ using DeviceGemmInstance =
         2,              // CShuffleNXdlPerWavePerShuffle
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
-        true>;          // MaskOutUpperTriangle
+        MaskingSpec>;   // MaskingSpecialization
 
 // Ref Gemm0: fp16 in, fp32 out
 using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
@@ -142,268 +154,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
                                                                                 B1ElementOp,
                                                                                 CElementOp>;
 
-int main(int argc, char* argv[])
-{
-    bool do_verification = true;
-    int init_method      = 1;
-    bool time_kernel     = false;
-
-    // GEMM shape for A/B0/B1/C
-    // C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
-    ck::index_t M             = 512;
-    ck::index_t N             = 512;
-    ck::index_t K             = 64;
-    ck::index_t O             = 128;
-    ck::index_t StrideA       = -1;
-    ck::index_t StrideB0      = -1;
-    ck::index_t StrideB1      = -1;
-    ck::index_t BatchStrideA  = -1;
-    ck::index_t BatchStrideB0 = -1;
-    ck::index_t BatchStrideB1 = -1;
-    float alpha               = 1;
-
-    // Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
-    // C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
-    // C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
-    ck::index_t G0 = 7;
-    ck::index_t G1 = 13;
-
-    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 == 11)
-    {
-        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]);
-        O  = std::stoi(argv[7]);
-        G0 = std::stoi(argv[8]);
-        G1 = std::stoi(argv[9]);
-
-        alpha = std::stof(argv[10]);
-    }
-    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 11: M, N, K, O, G0, G1\n");
-        printf("arg10: scale (alpha)\n");
-        exit(0);
-    }
-
-    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
-    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
-
-    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
-    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
-    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
-
-    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
-    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
-    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
-
-    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
-    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
-    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
-
-    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
-    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
-    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
-
-    const int BatchCount = G0 * G1;
-
-    auto f_host_tensor_descriptor = [](std::size_t batch_count,
-                                       std::size_t row,
-                                       std::size_t col,
-                                       std::size_t stride,
-                                       std::size_t batch_stride,
-                                       auto layout) {
-        if(std::is_same<decltype(layout), Row>::value)
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
-        }
-        else
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
-        }
-    };
-
-    // C_m_o = A_m_k * B0_k_n * B1_n_o
-    Tensor<ADataType> a_g_m_k(
-        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
-    Tensor<B0DataType> b0_g_k_n(
-        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
-    Tensor<B1DataType> b1_g_n_o(
-        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
-    Tensor<CDataType> c_gs_ms_os_host_result(
-        std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-        std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-    Tensor<CDataType> c_gs_ms_os_device_result(
-        std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-        std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-
-    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
-    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
-    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
-    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
-        break;
-    case 2:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
-        break;
-    case 3:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-        break;
-    default:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-    }
-
-    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
-    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
-    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
-    DeviceMem c_gs_ms_os_device_buf(sizeof(CDataType) *
-                                    c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
-
-    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
-    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
-    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
-
-    auto a_element_op    = AElementOp{};
-    auto b0_element_op   = B0ElementOp{};
-    auto acc0_element_op = Acc0ElementOp{alpha};
-    auto b1_element_op   = B1ElementOp{};
-    auto c_element_op    = CElementOp{};
-
-    // do GEMM
-    auto gemm    = DeviceGemmInstance{};
-    auto invoker = gemm.MakeInvoker();
-    auto argument =
-        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
-                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
-                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
-                          static_cast<CDataType*>(c_gs_ms_os_device_buf.GetDeviceBuffer()),
-                          M,
-                          N,
-                          K,
-                          O,
-                          BatchCount,
-                          c_gs_ms_os_lengths,
-                          c_gs_ms_os_strides,
-                          StrideA,
-                          StrideB0,
-                          StrideB1,
-                          BatchStrideA,
-                          BatchStrideB0,
-                          BatchStrideB1,
-                          a_element_op,
-                          b0_element_op,
-                          acc0_element_op,
-                          b1_element_op,
-                          c_element_op);
-
-    if(!gemm.IsSupportedArgument(argument))
-    {
-        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
-
-        return 0;
-    }
-
-    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
-
-    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
-    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
-                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
-                            BatchCount;
-
-    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, "
-              << gemm.GetTypeString() << std::endl;
-
-    if(do_verification)
-    {
-        c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
-
-        // Output of Gemm0 is input A of Gemm1
-        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-
-        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-
-        Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{BatchCount, M, O},
-                                              std::vector<int>{M * O, O, 1});
-
-        auto ref_gemm0          = ReferenceGemm0Instance{};
-        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
-        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
-            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
-
-        // gemm 0
-        ref_gemm0_invoker.Run(ref_gemm0_argument);
-
-        // mask out upper triangle
-        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
-            if(idx[1] < idx[2])
-                self(idx) = -ck::NumericLimits<float>::Infinity();
-        });
-
-        auto ref_softmax          = ReferenceSoftmaxInstance{};
-        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
-        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
-
-        // softmax
-        ref_softmax_invoker.Run(ref_softmax_argument);
-
-        auto ref_gemm1          = ReferenceGemm1Instance{};
-        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
-        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
-            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
-
-        // gemm1
-        ref_gemm1_invoker.Run(ref_gemm1_argument);
-
-        // permute
-        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
-            const size_t& g0 = idx[0];
-            const size_t& g1 = idx[1];
-
-            const size_t g = g0 * G1 + g1;
-
-            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
-        });
-
-        return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
-                   ? 0
-                   : 1;
-    }
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
 
-    return 0;
-}
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp

@@ -33,9 +33,6 @@ using S = ck::Sequence<Is...>;
 using F16 = ck::half_t;
 using F32 = float;
 
-using Row = ck::tensor_layout::gemm::RowMajor;
-using Col = ck::tensor_layout::gemm::ColumnMajor;
-
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
 using ADataType        = F16;
@@ -44,13 +41,14 @@ using B1DataType       = F16;
 using AccDataType      = F32;
 using CShuffleDataType = F32;
 using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
 
-using ALayout  = Row;
-using B0Layout = Col;
-using B1Layout = Row;
-
-using CPermuteNumDims_G_M_O =
-    S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
 
 using AElementOp    = PassThrough;
 using B0ElementOp   = PassThrough;
@@ -59,17 +57,27 @@ using B1ElementOp   = PassThrough;
 using CElementOp    = PassThrough;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
 
 using DeviceGemmInstance =
     ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
-        ALayout,
-        B0Layout,
-        B1Layout,
-        CPermuteNumDims_G_M_O,
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
         ADataType,
         B0DataType,
         B1DataType,
         CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
         AccDataType,
         CShuffleDataType,
         AElementOp,
@@ -78,6 +86,10 @@ using DeviceGemmInstance =
         B1ElementOp,
         CElementOp,
         GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
         1,
         256,
         128,         // MPerBlock
@@ -118,7 +130,7 @@ using DeviceGemmInstance =
         2,              // CShuffleNXdlPerWavePerShuffle
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
-        false>;         // MaskOutUpperTriangle
+        MaskingSpec>;   // MaskingSpecialization
 
 // Ref Gemm0: fp16 in, fp32 out
 using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
@@ -142,258 +154,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
                                                                                 B1ElementOp,
                                                                                 CElementOp>;
 
-int main(int argc, char* argv[])
-{
-    bool do_verification = true;
-    int init_method      = 1;
-    bool time_kernel     = false;
-
-    // GEMM shape for A/B0/B1/C
-    // C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
-    ck::index_t M             = 120;
-    ck::index_t N             = 1000;
-    ck::index_t K             = 64;
-    ck::index_t O             = 128;
-    ck::index_t StrideA       = -1;
-    ck::index_t StrideB0      = -1;
-    ck::index_t StrideB1      = -1;
-    ck::index_t BatchStrideA  = -1;
-    ck::index_t BatchStrideB0 = -1;
-    ck::index_t BatchStrideB1 = -1;
-    float alpha               = 1;
-
-    // Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
-    // C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
-    // C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
-    ck::index_t G0 = 7;
-    ck::index_t G1 = 13;
-
-    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 == 11)
-    {
-        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]);
-        O  = std::stoi(argv[7]);
-        G0 = std::stoi(argv[8]);
-        G1 = std::stoi(argv[9]);
-
-        alpha = std::stof(argv[10]);
-    }
-    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 11: M, N, K, O, G0, G1\n");
-        printf("arg10: scale (alpha)\n");
-        exit(0);
-    }
-
-    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
-    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
-
-    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
-    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
-    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
-
-    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
-    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
-    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
-
-    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
-    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
-    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
-
-    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
-    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
-    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
-
-    const int BatchCount = G0 * G1;
-
-    auto f_host_tensor_descriptor = [](std::size_t batch_count,
-                                       std::size_t row,
-                                       std::size_t col,
-                                       std::size_t stride,
-                                       std::size_t batch_stride,
-                                       auto layout) {
-        if(std::is_same<decltype(layout), Row>::value)
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
-        }
-        else
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
-        }
-    };
-
-    // C_m_o = A_m_k * B0_k_n * B1_n_o
-    Tensor<ADataType> a_g_m_k(
-        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
-    Tensor<B0DataType> b0_g_k_n(
-        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
-    Tensor<B1DataType> b1_g_n_o(
-        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
-    Tensor<CDataType> c_gs_ms_os_host_result(
-        std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-        std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-    Tensor<CDataType> c_gs_ms_os_device_result(
-        std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-        std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-
-    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
-    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
-    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
-    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
-        break;
-    case 2:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
-        break;
-    case 3:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-        break;
-    default:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-    }
-
-    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
-    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
-    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
-    DeviceMem c_gs_ms_os_device_buf(sizeof(CDataType) *
-                                    c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
-
-    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
-    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
-    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
-
-    auto a_element_op    = AElementOp{};
-    auto b0_element_op   = B0ElementOp{};
-    auto acc0_element_op = Acc0ElementOp{alpha};
-    auto b1_element_op   = B1ElementOp{};
-    auto c_element_op    = CElementOp{};
-
-    // do GEMM
-    auto gemm    = DeviceGemmInstance{};
-    auto invoker = gemm.MakeInvoker();
-    auto argument =
-        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
-                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
-                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
-                          static_cast<CDataType*>(c_gs_ms_os_device_buf.GetDeviceBuffer()),
-                          M,
-                          N,
-                          K,
-                          O,
-                          BatchCount,
-                          c_gs_ms_os_lengths,
-                          c_gs_ms_os_strides,
-                          StrideA,
-                          StrideB0,
-                          StrideB1,
-                          BatchStrideA,
-                          BatchStrideB0,
-                          BatchStrideB1,
-                          a_element_op,
-                          b0_element_op,
-                          acc0_element_op,
-                          b1_element_op,
-                          c_element_op);
-
-    if(!gemm.IsSupportedArgument(argument))
-    {
-        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
-
-        return 0;
-    }
-
-    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
-
-    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
-    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
-                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
-                            BatchCount;
-
-    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, "
-              << gemm.GetTypeString() << std::endl;
-
-    if(do_verification)
-    {
-        c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
-
-        // Output of Gemm0 is input A of Gemm1
-        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-
-        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-
-        Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{BatchCount, M, O},
-                                              std::vector<int>{M * O, O, 1});
-
-        auto ref_gemm0          = ReferenceGemm0Instance{};
-        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
-        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
-            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
-
-        ref_gemm0_invoker.Run(ref_gemm0_argument);
-
-        auto ref_softmax          = ReferenceSoftmaxInstance{};
-        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
-        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
-
-        ref_softmax_invoker.Run(ref_softmax_argument);
-
-        auto ref_gemm1          = ReferenceGemm1Instance{};
-        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
-        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
-            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
-
-        ref_gemm1_invoker.Run(ref_gemm1_argument);
-
-        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
-            const size_t& g0 = idx[0];
-            const size_t& g1 = idx[1];
-
-            const size_t g = g0 * G1 + g1;
-
-            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
-        });
-
-        return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
-                   ? 0
-                   : 1;
-    }
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
 
-    return 0;
-}
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#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/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp

@@ -33,9 +33,6 @@ using S = ck::Sequence<Is...>;
 using F16 = ck::half_t;
 using F32 = float;
 
-using Row = ck::tensor_layout::gemm::RowMajor;
-using Col = ck::tensor_layout::gemm::ColumnMajor;
-
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
 using ADataType        = F16;
@@ -44,13 +41,14 @@ using B1DataType       = F16;
 using AccDataType      = F32;
 using CShuffleDataType = F32;
 using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
 
-using ALayout  = Row;
-using B0Layout = Col;
-using B1Layout = Row;
-
-using CPermuteNumDims_G_M_O =
-    S<1, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_M_O
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
 
 using AElementOp    = PassThrough;
 using B0ElementOp   = PassThrough;
@@ -59,17 +57,27 @@ using B1ElementOp   = PassThrough;
 using CElementOp    = PassThrough;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
 
 using DeviceGemmInstance =
     ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
-        ALayout,
-        B0Layout,
-        B1Layout,
-        CPermuteNumDims_G_M_O,
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
         ADataType,
         B0DataType,
         B1DataType,
         CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
         AccDataType,
         CShuffleDataType,
         AElementOp,
@@ -78,6 +86,10 @@ using DeviceGemmInstance =
         B1ElementOp,
         CElementOp,
         GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
         1,
         256,
         128,         // MPerBlock
@@ -118,7 +130,7 @@ using DeviceGemmInstance =
         2,              // CShuffleNXdlPerWavePerShuffle
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
-        false>;
+        MaskingSpec>;   // MaskingSpecialization
 
 // Ref Gemm0: fp16 in, fp32 out
 using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
@@ -142,303 +154,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
                                                                                 B1ElementOp,
                                                                                 CElementOp>;
 
-int main(int argc, char* argv[])
-{
-    bool do_verification = true;
-    int init_method      = 1;
-    bool time_kernel     = false;
-
-    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
-    {
-        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");
-        exit(0);
-    }
-
-    float alpha = 1; // scaling after 1st gemm
-
-    std::size_t group_count = 13;
-
-    // Problem descs
-    std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
-    std::vector<const void*> p_a;
-    std::vector<const void*> p_b0;
-    std::vector<const void*> p_b1;
-    std::vector<void*> p_c;
-
-    for(std::size_t i = 0; i < group_count; i++)
-    {
-        int M     = 128 * (rand() % 8 + 1);
-        int N     = 128 * (rand() % 8 + 1);
-        int K     = 40;
-        int O     = 40 * (rand() % 2 + 1);
-        int Batch = rand() % 8 + 1;
-
-        const int StrideA  = ck::is_same_v<ALayout, Row> ? K : M;
-        const int StrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
-        const int StrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
-
-        const int BatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
-        const int BatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
-        const int BatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
-
-        std::vector<ck::index_t> c_gs_ms_os_lengths{Batch, M, O};
-        std::vector<ck::index_t> c_gs_ms_os_strides{O, Batch * O, 1};
-
-        problem_descs.push_back({M,
-                                 N,
-                                 K,
-                                 O,
-                                 Batch,
-                                 StrideA,
-                                 StrideB0,
-                                 StrideB1,
-                                 BatchStrideA,
-                                 BatchStrideB0,
-                                 BatchStrideB1,
-                                 c_gs_ms_os_lengths,
-                                 c_gs_ms_os_strides});
-    }
-
-    auto f_host_tensor_descriptor = [](std::size_t batch_count,
-                                       std::size_t row,
-                                       std::size_t col,
-                                       std::size_t stride,
-                                       std::size_t batch_stride,
-                                       auto layout) {
-        if(std::is_same<decltype(layout), Row>::value)
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
-        }
-        else
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
-        }
-    };
-
-    std::vector<Tensor<ADataType>> a_tensors;
-    std::vector<Tensor<B0DataType>> b0_tensors;
-    std::vector<Tensor<B1DataType>> b1_tensors;
-    std::vector<Tensor<CDataType>> c_tensors;
-
-    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
-
-    std::vector<DeviceMemPtr> a_tensors_device;
-    std::vector<DeviceMemPtr> b0_tensors_device;
-    std::vector<DeviceMemPtr> b1_tensors_device;
-    std::vector<DeviceMemPtr> c_tensors_device;
-
-    std::size_t flop = 0, num_byte = 0;
-
-    std::cout << "group count " << group_count << ". printing first 4 groups\n";
-    for(std::size_t i = 0; i < group_count; i++)
-    {
-        const auto& M                  = problem_descs[i].M;
-        const auto& N                  = problem_descs[i].N;
-        const auto& K                  = problem_descs[i].K;
-        const auto& O                  = problem_descs[i].O;
-        const auto& Batch              = problem_descs[i].Batch;
-        const auto& StrideA            = problem_descs[i].StrideA;
-        const auto& StrideB0           = problem_descs[i].StrideB0;
-        const auto& StrideB1           = problem_descs[i].StrideB1;
-        const auto& BatchStrideA       = problem_descs[i].BatchStrideA;
-        const auto& BatchStrideB0      = problem_descs[i].BatchStrideB0;
-        const auto& BatchStrideB1      = problem_descs[i].BatchStrideB1;
-        const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
-        const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
-
-        // C_m_o = A_m_k * B0_k_n * B1_n_o
-        Tensor<ADataType> a_g_m_k(
-            f_host_tensor_descriptor(Batch, M, K, StrideA, BatchStrideA, ALayout{}));
-        Tensor<B0DataType> b0_g_k_n(
-            f_host_tensor_descriptor(Batch, K, N, StrideB0, BatchStrideB0, B0Layout{}));
-        Tensor<B1DataType> b1_g_n_o(
-            f_host_tensor_descriptor(Batch, N, O, StrideB1, BatchStrideB1, B1Layout{}));
-        Tensor<CDataType> c_gs_ms_os_device_result(
-            std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-            std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-
-        flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
-        num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
-                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
-                    Batch;
-
-        if(i < 4)
-        {
-            std::cout << "a_g_m_k[" << i << "]: " << a_g_m_k.mDesc << ", "
-                      << "b0_g_k_n[" << i << "]: " << b0_g_k_n.mDesc << ", "
-                      << "b1_g_n_o[" << i << "]: " << b1_g_n_o.mDesc << ", "
-                      << "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
-        }
-
-        switch(init_method)
-        {
-        case 0: break;
-        case 1:
-            a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-            b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
-            b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
-            break;
-        case 2:
-            a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-            b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
-            b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
-            break;
-        case 3:
-            a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-            b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
-            b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-            break;
-        default:
-            a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-            b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
-            b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-        }
-
-        a_tensors.push_back(a_g_m_k);
-        b0_tensors.push_back(b0_g_k_n);
-        b1_tensors.push_back(b1_g_n_o);
-        c_tensors.push_back(c_gs_ms_os_device_result);
-
-        a_tensors_device.emplace_back(
-            std::make_unique<DeviceMem>(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize()));
-        b0_tensors_device.emplace_back(
-            std::make_unique<DeviceMem>(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize()));
-        b1_tensors_device.emplace_back(
-            std::make_unique<DeviceMem>(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize()));
-        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
-            sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
-
-        a_tensors_device[i]->ToDevice(a_g_m_k.mData.data());
-        b0_tensors_device[i]->ToDevice(b0_g_k_n.mData.data());
-        b1_tensors_device[i]->ToDevice(b1_g_n_o.mData.data());
-
-        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
-        p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
-        p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
-        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
-    }
-
-    auto a_element_op    = AElementOp{};
-    auto b0_element_op   = B0ElementOp{};
-    auto acc0_element_op = Acc0ElementOp{alpha};
-    auto b1_element_op   = B1ElementOp{};
-    auto c_element_op    = CElementOp{};
-
-    // do GEMM
-    auto gemm     = DeviceGemmInstance{};
-    auto invoker  = gemm.MakeInvoker();
-    auto argument = gemm.MakeArgument(p_a,
-                                      p_b0,
-                                      p_b1,
-                                      p_c,
-                                      problem_descs,
-                                      a_element_op,
-                                      b0_element_op,
-                                      acc0_element_op,
-                                      b1_element_op,
-                                      c_element_op);
-
-    // specify workspace for problem_desc
-    DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
-
-    gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
-
-    if(!gemm.IsSupportedArgument(argument))
-    {
-        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
-
-        return 0;
-    }
-
-    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
-
-    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
-
-    float gb_per_sec = num_byte / 1.E6 / ave_time;
-
-    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
-              << gemm.GetTypeString() << std::endl;
-
-    bool pass = true;
-    if(do_verification)
-    {
-        for(std::size_t i = 0; i < group_count; i++)
-        {
-            const auto& M                  = problem_descs[i].M;
-            const auto& N                  = problem_descs[i].N;
-            const auto& O                  = problem_descs[i].O;
-            const auto& Batch              = problem_descs[i].Batch;
-            const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
-            const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
-
-            const auto& a_g_m_k            = a_tensors[i];
-            const auto& b0_g_k_n           = b0_tensors[i];
-            const auto& b1_g_n_o           = b1_tensors[i];
-            auto& c_gs_ms_os_device_result = c_tensors[i];
-            auto& c_gs_ms_os_device_buf    = *c_tensors_device[i];
-
-            Tensor<CDataType> c_gs_ms_os_host_result(
-                std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-                std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-
-            c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
-
-            // Output of Gemm0 is input A of Gemm1
-            Tensor<AccDataType> acc0_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
-
-            Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
-
-            Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{Batch, M, O},
-                                                  std::vector<int>{M * O, O, 1});
-
-            auto ref_gemm0          = ReferenceGemm0Instance{};
-            auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
-            auto ref_gemm0_argument = ref_gemm0.MakeArgument(
-                a_g_m_k, b0_g_k_n, acc0_m_n, a_element_op, b0_element_op, acc0_element_op);
-
-            ref_gemm0_invoker.Run(ref_gemm0_argument);
-
-            auto ref_softmax          = ReferenceSoftmaxInstance{};
-            auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
-            auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_m_n, a1_g_m_n, 1, 0, {2});
-
-            ref_softmax_invoker.Run(ref_softmax_argument);
-
-            auto ref_gemm1          = ReferenceGemm1Instance{};
-            auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
-            auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
-                                                             b1_g_n_o,
-                                                             c_g_m_o_host_result,
-                                                             PassThrough{},
-                                                             b1_element_op,
-                                                             c_element_op);
-
-            ref_gemm1_invoker.Run(ref_gemm1_argument);
-
-            // Note: in this example, we merely permute the dimensions by changing underlying
-            // strides so we simply access data as-is
-            c_gs_ms_os_host_result.ForEach(
-                [&](auto& self, auto idx) { self(idx) = c_g_m_o_host_result(idx); });
-
-            bool pass_ =
-                ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
-            pass &= pass_;
-        }
-    }
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
 
-    return pass ? 0 : 1;
-}
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape for A/B0/B1/C
+    // C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
+    ck::index_t M = 120;
+    ck::index_t N = 1000;
+    ck::index_t K = 64;
+    ck::index_t O = 128;
+
+    // Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
+    // C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
+    // C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
+    ck::index_t G0 = 7;
+    ck::index_t G1 = 13;
+
+    float alpha = 1;
+
+    bool input_permute = false;
+    bool output_permute = true;
+
+    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 == 13)
+    {
+        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]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+
+        input_permute = std::stoi(argv[11]);
+        output_permute = std::stoi(argv[12]);
+    }
+    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 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        printf("arg11 to 12: input / output permute\n");
+        exit(0);
+    }
+
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+            : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+            : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+            : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides =
+        output_permute
+            ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+            : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    // TODO ANT: replace array with vector?
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+        static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        {}, // std::array<void*, 1> p_acc0_biases;
+        {}, // std::array<void*, 1> p_acc1_biases;
+        a_gs_ms_ks_lengths,
+        a_gs_ms_ks_strides,
+        b0_gs_ns_ks_lengths,
+        b0_gs_ns_ks_strides,
+        b1_gs_os_ns_lengths,
+        b1_gs_os_ns_strides,
+        c_gs_ms_os_lengths,
+        c_gs_ms_os_strides,
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+        a_element_op,
+        b0_element_op,
+        acc0_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    ck::index_t BatchCount = G0 * G1;
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    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, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+
+        // gemm 0
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // masking
+        const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(mask.IsMaskedElement(idx[1], idx[2]))
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        // softmax
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        // gemm1
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    bool input_permute = false;
+    bool output_permute = true;
+
+    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 == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        input_permute = std::stoi(argv[4]);
+        output_permute = std::stoi(argv[5]);
+    }
+    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 5: input / output permute\n");
+        exit(0);
+    }
+
+    float alpha = 1; // scaling after 1st gemm
+
+    std::size_t group_count = 7;
+
+    // Problem descs
+    std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
+    std::vector<const void*> p_a;
+    std::vector<const void*> p_b0;
+    std::vector<const void*> p_b1;
+    std::vector<void*> p_c;
+    std::vector<std::vector<int>> g0_g1_m_n_k_o;
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<B0DataType>> b0_tensors;
+    std::vector<Tensor<B1DataType>> b1_tensors;
+    std::vector<Tensor<CDataType>> c_tensors;
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_tensors_device;
+    std::vector<DeviceMemPtr> b0_tensors_device;
+    std::vector<DeviceMemPtr> b1_tensors_device;
+    std::vector<DeviceMemPtr> c_tensors_device;
+
+    std::size_t flop = 0, num_byte = 0;
+
+    std::cout << "group count " << group_count << ". printing first 4 groups\n";
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        int M     = 128 * (rand() % 8 + 1);
+        int N     = 128 * (rand() % 8 + 1);
+        int K     = 40;
+        int O     = 40 * (rand() % 2 + 1);
+        int G0 = rand() % 3 + 1;
+        int G1 = rand() % 5 + 1;
+
+        g0_g1_m_n_k_o.push_back({G0, G1, M, N, K, O});
+
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+                : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+                : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+                : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides =
+            output_permute
+                ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+                : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+        problem_descs.push_back({a_gs_ms_ks_lengths,
+                                 a_gs_ms_ks_strides,
+                                 b0_gs_ns_ks_lengths,
+                                 b0_gs_ns_ks_strides,
+                                 b1_gs_os_ns_lengths,
+                                 b1_gs_os_ns_strides,
+                                 c_gs_ms_os_lengths,
+                                 c_gs_ms_os_strides,
+                                 {},   // acc0_biases_gs_ms_ns_lengths
+                                 {},   // acc0_biases_gs_ms_ns_strides
+                                 {},   // acc1_biases_gs_ms_os_lengths
+                                 {}}); // acc1_biases_gs_ms_os_strides
+
+        // C_m_o = A_m_k * B0_k_n * B1_n_o
+        Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+        Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+        Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+        Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+        int Batch = G0 * G1;
+        flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
+        num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                    Batch;
+
+        if(i < 4)
+        {
+            std::cout << "a_gs_ms_ks[" << i << "]: " << a_gs_ms_ks.mDesc << ", "
+                      << "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
+                      << "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
+                      << "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
+        }
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+            break;
+        case 2:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+            break;
+        case 3:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+            break;
+        default:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        }
+
+        a_tensors.push_back(a_gs_ms_ks);
+        b0_tensors.push_back(b0_gs_ns_ks);
+        b1_tensors.push_back(b1_gs_os_ns);
+        c_tensors.push_back(c_gs_ms_os_device_result);
+
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize()));
+        b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize()));
+        b1_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
+
+        a_tensors_device[i]->ToDevice(a_gs_ms_ks.mData.data());
+        b0_tensors_device[i]->ToDevice(b0_gs_ns_ks.mData.data());
+        b1_tensors_device[i]->ToDevice(b1_gs_os_ns.mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
+        p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
+        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(p_a,
+                                      p_b0,
+                                      p_b1,
+                                      p_c,
+                                      {}, // p_acc0_biases
+                                      {}, // p_acc1_biases
+                                      problem_descs,
+                                      a_element_op,
+                                      b0_element_op,
+                                      acc0_element_op,
+                                      b1_element_op,
+                                      c_element_op);
+
+    // specify workspace for problem_desc
+    DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < group_count; i++)
+        {
+            const int& G0                  = g0_g1_m_n_k_o[i][0];
+            const int& G1                  = g0_g1_m_n_k_o[i][1];
+            const int& M                   = g0_g1_m_n_k_o[i][2];
+            const int& N                   = g0_g1_m_n_k_o[i][3];
+            const int& K                   = g0_g1_m_n_k_o[i][4];
+            const int& O                   = g0_g1_m_n_k_o[i][5];
+
+            const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
+            const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
+
+            const auto& a_gs_ms_ks         = a_tensors[i];
+            const auto& b0_gs_ns_ks        = b0_tensors[i];
+            const auto& b1_gs_os_ns        = b1_tensors[i];
+            auto& c_gs_ms_os_device_result = c_tensors[i];
+            auto& c_gs_ms_os_device_buf    = *c_tensors_device[i];
+
+            c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+            Tensor<ADataType> a_g_m_k({G0 * G1, M, K});
+            Tensor<B0DataType> b0_g_k_n({G0 * G1, K, N});
+            Tensor<B1DataType> b1_g_n_o({G0 * G1, N, O});
+            Tensor<AccDataType> acc0_g_m_n({G0 * G1, M, N});        // scratch object after gemm0
+            Tensor<ADataType> a1_g_m_n({G0 * G1, M, N});            // scratch object after softmax
+            Tensor<CDataType> c_g_m_o_host_result({G0 * G1, M, O}); // scratch object after gemm1
+            Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+            // permute
+            a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+                a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+            });
+            b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+                b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+            b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+                b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+
+            // gemm 0
+            auto ref_gemm0          = ReferenceGemm0Instance{};
+            auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+            auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+                a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+            ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+            // masking
+            const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+            acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+                if(mask.IsMaskedElement(idx[1], idx[2]))
+                    self(idx) = -ck::NumericLimits<float>::Infinity();
+            });
+
+            // softmax
+            auto ref_softmax          = ReferenceSoftmaxInstance{};
+            auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+            auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+            ref_softmax_invoker.Run(ref_softmax_argument);
+
+            // gemm 1
+            auto ref_gemm1          = ReferenceGemm1Instance{};
+            auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+            auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
+                                                             b1_g_n_o,
+                                                             c_g_m_o_host_result,
+                                                             PassThrough{},
+                                                             b1_element_op,
+                                                             c_element_op);
+
+            ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+            // permute
+            c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+                const size_t& g0 = idx[0];
+                const size_t& g1 = idx[1];
+
+                const size_t g = g0 * G1 + g1;
+
+                self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+            });
+
+            bool pass_ =
+                ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
+            pass &= pass_;
+        }
+    }
+
+    return pass ? 0 : 1;
+}

example/34_batchnorm/batchnorm_common.hpp

@@ -10,131 +10,17 @@
 
 #include "ck/utility/data_type.hpp"
 
-// binary operation used to calculate invVariance from mean and meansquare
-struct InvVariance
-{
-    InvVariance(double epsilon) : epsilon_(epsilon){};
-
-    template <typename T>
-    __host__ __device__ constexpr void operator()(T& y, const T& mean, const T& meansquare) const
-    {
-        static_assert(std::is_same<T, float>::value || std::is_same<T, double>::value,
-                      "Data type is not supported by this operation!");
-
-        using ck::type_convert;
-        using ck::math::sqrt;
-
-        T tmp_epsilon = type_convert<T>(epsilon_);
-
-        y = meansquare - mean * mean;
-        y = 1.0f / sqrt(tmp_epsilon + y);
-    };
-
-    double epsilon_;
-};
-
-// (4-in, 2-out) element-wise operation used to update the moving average of mean and variance
-struct MovingAverage
-{
-    MovingAverage(double factor) : factor_(factor){};
-
-    template <typename T>
-    __host__ __device__ constexpr void operator()(T& y0,
-                                                  T& y1,
-                                                  const T& mean,
-                                                  const T& runningMean,
-                                                  const T& meansquare,
-                                                  const T& runningVariance) const
-    {
-        static_assert(std::is_same<T, float>::value || std::is_same<T, double>::value,
-                      "Data type is not supported by this operation!");
-
-        using ck::type_convert;
-
-        T tmp_factor = type_convert<T>(factor_);
-        T variance   = meansquare - mean * mean;
-
-        y0 = runningMean * (type_convert<T>(1.0f) - tmp_factor) + mean * tmp_factor;
-        y1 = runningVariance * (type_convert<T>(1.0f) - tmp_factor) + variance * tmp_factor;
-    };
-
-    double factor_;
-};
-
-struct MovingAverageAndInvVariance
-{
-    MovingAverageAndInvVariance(double epsilon, double factor)
-        : epsilon_(epsilon), factor_(factor){};
-
-    template <typename T>
-    __host__ __device__ constexpr void operator()(T& y0, // resultRunningMean
-                                                  T& y1, // resultRunningVariance
-                                                  T& y2, // saveInvVariance
-                                                  const T& mean,
-                                                  const T& runningMean,
-                                                  const T& meansquare,
-                                                  const T& runningVariance) const
-    {
-        static_assert(std::is_same<T, float>::value || std::is_same<T, double>::value,
-                      "Data type is not supported by this operation!");
-
-        using ck::type_convert;
-        using ck::math::sqrt;
-
-        T tmp_epsilon = type_convert<T>(epsilon_);
-        T tmp_factor  = type_convert<T>(factor_);
-        T variance    = meansquare - mean * mean;
-
-        y0 = runningMean * (type_convert<T>(1.0f) - tmp_factor) + mean * tmp_factor;
-        y1 = runningVariance * (type_convert<T>(1.0f) - tmp_factor) + variance * tmp_factor;
-
-        y2 = 1.0f / sqrt(tmp_epsilon + variance);
-    };
-
-    double epsilon_;
-    double factor_;
-};
-
 struct NormalizeInInfer
 {
     NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
 
-    template <typename T1, typename T2>
+    template <typename T1, typename T2, typename T3, typename T4>
     __host__ __device__ constexpr void operator()(T1& y,
                                                   const T1& x,
                                                   const T2& mean,
                                                   const T2& variance,
-                                                  const T2& gamma,
-                                                  const T2& beta) const
-    {
-        static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
-                      "Data type is not supported by this operation!");
-
-        using ck::type_convert;
-        using ck::math::sqrt;
-
-        T2 tmp_x, tmp_y;
-
-        tmp_x = type_convert<T2>(x);
-
-        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) * gamma + beta;
-        y     = type_convert<T1>(tmp_y);
-    };
-
-    double epsilon_;
-};
-
-struct NormalizeInForward
-{
-    NormalizeInForward(double epsilon = 1e-4) : epsilon_(epsilon) {}
-
-    template <typename T1, typename T2>
-    __host__ __device__ constexpr void operator()(T1& y,
-                                                  const T1& x,
-                                                  const T2& mean,
-                                                  const T2& meansquare,
-                                                  const T2& gamma,
-                                                  const T2& beta) const
+                                                  const T3& gamma,
+                                                  const T4& beta) const
     {
         static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
                       "Data type is not supported by this operation!");
@@ -143,11 +29,12 @@ struct NormalizeInForward
         using ck::math::sqrt;
 
         T2 tmp_x, tmp_y;
-        T2 variance = meansquare - mean * mean;
 
         tmp_x = type_convert<T2>(x);
 
-        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) * gamma + beta;
+        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
+                    type_convert<T2>(gamma) +
+                type_convert<T2>(beta);
         y = type_convert<T1>(tmp_y);
     };
 

example/34_batchnorm/batchnorm_forward_impl.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <cassert>
-#include <vector>
-
-#include "ck/ck.hpp"
-#include "ck/utility/reduction_operator.hpp"
-#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
-
-#include "batchnorm_common.hpp"
-
-template <typename InOutDataType,
-          typename AccDataType,
-          ck::index_t Rank,
-          ck::index_t NumBatchNormReduceDim,
-          bool fastest_dim_is_reduced = false>
-int bnorm_fwd(bool time_kernel,
-              bool updateMovingAverage,
-              bool saveMeanAndInvVariance,
-              const std::array<int, NumBatchNormReduceDim> reduceDims,
-              const std::array<ck::index_t, Rank> xyLengths,
-              const std::array<ck::index_t, Rank> xStrides,
-              const std::array<ck::index_t, Rank> yStrides,
-              const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
-              const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarStrides,
-              const void* p_x,
-              const void* p_scale,
-              const void* p_bias,
-              void* p_y,
-              double exponentialAverageFactor,
-              void* p_runningMean,
-              void* p_runningVariance,
-              double epsilon,
-              void* p_saveMean,
-              void* p_saveInvVariance,
-              void* p_tmp_mean,
-              void* p_tmp_meansquare)
-{
-    static_assert(NumBatchNormReduceDim < Rank,
-                  "Invalid number of reduced dimensions for batchnorm!");
-
-    constexpr ck::index_t NumScaleBiasMeanVarDim = Rank - NumBatchNormReduceDim;
-
-    using InElementwiseOperation_Mean  = ck::tensor_operation::element_wise::PassThrough;
-    using AccElementwiseOperation_Mean = ck::tensor_operation::element_wise::UnaryDivide;
-
-    using InElementwiseOperation_Meansquare  = ck::tensor_operation::element_wise::UnarySquare;
-    using AccElementwiseOperation_Meansquare = ck::tensor_operation::element_wise::UnaryDivide;
-
-    using DeviceMeanAndMeansquareInstance =
-        ck::tensor_operation::device::DeviceMultipleReduceMultiBlock<
-            2,
-            InOutDataType,
-            AccDataType,
-            ck::Tuple<AccDataType, AccDataType>,
-            Rank,
-            NumBatchNormReduceDim,
-            ck::reduce::Add,
-            ck::Tuple<InElementwiseOperation_Mean, InElementwiseOperation_Meansquare>,
-            ck::Tuple<AccElementwiseOperation_Mean, AccElementwiseOperation_Meansquare>,
-            ck::InMemoryDataOperationEnum::Set,
-            false, // PropagateNan
-            256,
-            16,
-            16,
-            1,
-            1,
-            fastest_dim_is_reduced ? 1 : 0,
-            1,
-            ck::Sequence<1, 1>>;
-
-    using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
-        ck::Tuple<InOutDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
-                                                                                      // meansquare,
-                                                                                      // scale, bias
-        ck::Tuple<InOutDataType>,                                                     // y
-        NormalizeInForward,
-        Rank,
-        2,                           // MPerthread
-        ck::Sequence<1, 1, 1, 1, 1>, // scalarPerVector: x, mean, meansquare, scale, bias
-        ck::Sequence<1>>;            // scalarPerVector: y
-
-    using DeviceInvVarianceInstance = ck::tensor_operation::device::DeviceElementwise<
-        ck::Tuple<AccDataType, AccDataType>, // mean, meansquare
-        ck::Tuple<AccDataType>,              // invVariance
-        InvVariance,
-        NumScaleBiasMeanVarDim,
-        2,                  // MPerthread
-        ck::Sequence<1, 1>, // scalarPerVector: mean, meansquare
-        ck::Sequence<1>>;   // scalarPerVector: invVariance
-
-    using DeviceMovingAverageInstance = ck::tensor_operation::device::DeviceElementwise<
-        ck::Tuple<AccDataType, AccDataType, AccDataType, AccDataType>, // old moving mean, new mean,
-                                                                       // old moving variance, new
-                                                                       // meansquare
-        ck::Tuple<AccDataType, AccDataType>, // updated moving mean, updated moving variance
-        MovingAverage,
-        NumScaleBiasMeanVarDim,
-        4,                        // MPerthread
-        ck::Sequence<1, 1, 1, 1>, // scalarPerVector: old moving mean, new mean, old moving
-                                  // variance, new meansquare
-        ck::Sequence<1, 1>>;      // scalarPerVector: updated moving mean, updated moving variance
-
-    using DeviceMovingAverageAndInvVarianceInstance =
-        ck::tensor_operation::device::DeviceElementwise<
-            ck::Tuple<AccDataType, AccDataType, AccDataType, AccDataType>, // old moving mean, new
-                                                                           // mean, old moving
-                                                                           // variance, new
-                                                                           // meansquare
-            ck::Tuple<AccDataType, AccDataType, AccDataType>, // updated moving mean, updated moving
-                                                              // variancem, invVariance
-            MovingAverageAndInvVariance,
-            NumScaleBiasMeanVarDim,
-            4,                        // MPerthread
-            ck::Sequence<1, 1, 1, 1>, // scalarPerVector: old moving mean, new mean, old moving
-                                      // variance, new meansquare
-            ck::Sequence<1, 1, 1>>; // scalarPerVector: updated moving mean, updated moving variance
-
-    auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
-    std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
-
-    int i = 0;
-    for(auto dim : invariantDims)
-    {
-        assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
-
-        aligned_scaleBiasMeanVarStrides[dim] = bnScaleBiasMeanVarStrides[i];
-        i++;
-    };
-
-    int32_t reduceLength = 1;
-
-    for(auto dim : reduceDims)
-        reduceLength *= xyLengths[dim];
-
-    int32_t invariantLength = 1;
-
-    for(auto dim : invariantDims)
-        invariantLength *= xyLengths[dim];
-
-    size_t total_length = static_cast<size_t>(invariantLength) * reduceLength;
-
-    float avg_time        = 0.0f;
-    std::size_t num_bytes = 0;
-
-    auto dev_mean_and_meansquare = DeviceMeanAndMeansquareInstance{};
-
-    void* p_mean = saveMeanAndInvVariance ? p_saveMean : p_tmp_mean;
-
-    const AccDataType alpha = ck::type_convert<AccDataType>(1.0f);
-    const AccDataType beta  = ck::type_convert<AccDataType>(0.0f);
-
-    auto argument_ptr1 = dev_mean_and_meansquare.MakeArgumentPointer(
-        xyLengths,
-        xStrides,
-        bnScaleBiasMeanVarLengths,
-        {bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
-        reduceDims,
-        {&alpha, &alpha},
-        {&beta, &beta},
-        p_x,
-        {p_mean, p_tmp_meansquare},
-        ck::make_tuple(InElementwiseOperation_Mean{}, InElementwiseOperation_Meansquare{}),
-        ck::make_tuple(AccElementwiseOperation_Mean{reduceLength},
-                       AccElementwiseOperation_Meansquare{reduceLength}));
-
-    auto dev_normalize = DeviceNormalizeInstance{};
-
-    auto argument_ptr2 =
-        dev_normalize.MakeArgumentPointer(xyLengths,
-                                          {xStrides,
-                                           aligned_scaleBiasMeanVarStrides,
-                                           aligned_scaleBiasMeanVarStrides,
-                                           aligned_scaleBiasMeanVarStrides,
-                                           aligned_scaleBiasMeanVarStrides},
-                                          {yStrides},
-                                          {p_x, p_mean, p_tmp_meansquare, p_scale, p_bias},
-                                          {p_y},
-                                          NormalizeInForward{epsilon});
-
-    if(!dev_mean_and_meansquare.IsSupportedArgument(argument_ptr1.get()) ||
-       !dev_normalize.IsSupportedArgument(argument_ptr2.get()))
-    {
-        std::cout << "The runtime parameters seems not supported by the Devic, exiting!"
-                  << std::endl;
-
-        return (-1);
-    };
-
-    auto invoker_ptr1 = dev_mean_and_meansquare.MakeInvokerPointer();
-    auto invoker_ptr2 = dev_normalize.MakeInvokerPointer();
-
-    avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
-    avg_time += invoker_ptr2->Run(argument_ptr2.get(), StreamConfig{nullptr, time_kernel});
-
-    num_bytes +=
-        (total_length * sizeof(InOutDataType) + invariantLength * 2 * sizeof(AccDataType)) + // No.1
-        (total_length * (1 * sizeof(InOutDataType) + 4 * sizeof(AccDataType)) +
-         total_length * sizeof(InOutDataType)); // No.2
-
-    if(saveMeanAndInvVariance && updateMovingAverage)
-    {
-        auto dev_moving_average_inv_variance = DeviceMovingAverageAndInvVarianceInstance{};
-
-        auto argument_ptr3 = dev_moving_average_inv_variance.MakeArgumentPointer(
-            bnScaleBiasMeanVarLengths,
-            {bnScaleBiasMeanVarStrides,
-             bnScaleBiasMeanVarStrides,
-             bnScaleBiasMeanVarStrides,
-             bnScaleBiasMeanVarStrides},
-            {bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
-            {p_mean, p_runningMean, p_tmp_meansquare, p_runningVariance},
-            {p_runningMean, p_runningVariance, p_saveInvVariance},
-            MovingAverageAndInvVariance{epsilon, exponentialAverageFactor});
-
-        if(!dev_moving_average_inv_variance.IsSupportedArgument(argument_ptr3.get()))
-        {
-            std::cout << "Runtime parameters not supported by the Device, exiting!" << std::endl;
-
-            return (-1);
-        };
-
-        auto invoker_ptr3 = dev_moving_average_inv_variance.MakeInvokerPointer();
-
-        avg_time += invoker_ptr3->Run(argument_ptr3.get(), StreamConfig{nullptr, time_kernel});
-
-        num_bytes += invariantLength * (4 + 3) * sizeof(AccDataType) * 2; // No.5
-    }
-    else if(saveMeanAndInvVariance)
-    {
-        auto dev_inv_variance = DeviceInvVarianceInstance{};
-        auto argument_ptr3    = dev_inv_variance.MakeArgumentPointer(
-            bnScaleBiasMeanVarLengths,
-            {bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
-            {bnScaleBiasMeanVarStrides},
-            {p_mean, p_tmp_meansquare},
-            {p_saveInvVariance},
-            InvVariance{epsilon});
-
-        if(!dev_inv_variance.IsSupportedArgument(argument_ptr3.get()))
-        {
-            std::cout << "Runtime parameters not supported by the Device, exiting!" << std::endl;
-
-            return (-1);
-        };
-
-        auto invoker_ptr3 = dev_inv_variance.MakeInvokerPointer();
-
-        avg_time += invoker_ptr3->Run(argument_ptr3.get(), StreamConfig{nullptr, time_kernel});
-
-        num_bytes += invariantLength * (2 + 1) * sizeof(AccDataType);
-    }
-    else if(updateMovingAverage)
-    {
-        auto dev_moving_average = DeviceMovingAverageInstance{};
-
-        auto argument_ptr3 = dev_moving_average.MakeArgumentPointer(
-            bnScaleBiasMeanVarLengths,
-            {bnScaleBiasMeanVarStrides,
-             bnScaleBiasMeanVarStrides,
-             bnScaleBiasMeanVarStrides,
-             bnScaleBiasMeanVarStrides},
-            {bnScaleBiasMeanVarStrides, bnScaleBiasMeanVarStrides},
-            {p_mean, p_runningMean, p_tmp_meansquare, p_runningVariance},
-            {p_runningMean, p_runningVariance},
-            MovingAverage{exponentialAverageFactor});
-
-        if(!dev_moving_average.IsSupportedArgument(argument_ptr3.get()))
-        {
-            std::cout << "Runtime parameters not supported by the Device, exiting!" << std::endl;
-
-            return (-1);
-        };
-
-        auto invoker_ptr3 = dev_moving_average.MakeInvokerPointer();
-
-        avg_time += invoker_ptr3->Run(argument_ptr3.get(), StreamConfig{nullptr, time_kernel});
-
-        num_bytes += invariantLength * (4 + 2) * sizeof(AccDataType) * 2; // No.5
-    };
-
-    if(time_kernel)
-    {
-        float gb_per_sec = num_bytes / 1.E6 / avg_time;
-
-        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
-    };
-
-    return (0);
-};