Merge branch 'develop' into gridwise_2d

example/34_batchnorm/batchnorm_forward_nhwc.cpp

@@ -15,13 +15,9 @@
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/utility/host_common_util.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward_nhwc_c.hpp"
-
-#include "batchnorm_forward_impl.hpp"
-
-template <typename InOutDataType, typename AccDataType>
-using ReferenceBatchNormFwdInstance =
-    ck::tensor_operation::host::ReferenceBatchNormFwd_Input_N_H_W_C_Output_C<InOutDataType,
-                                                                             AccDataType>;
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
                                        {"verify", required_argument, nullptr, 'v'},
@@ -41,9 +37,10 @@ class BatchNormFwdArg
     bool updateMovingAverage;
     bool saveMeanAndInvVariance;
 
-    int data_type    = 0;
-    int init_method  = 2;
-    bool time_kernel = false;
+    int data_type               = 0;
+    int init_method             = 2;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
 
     public:
     void show_usage(const char* cmd)
@@ -68,6 +65,7 @@ class BatchNormFwdArg
                      "value, 3=decimal value)"
                   << std::endl;
         std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
     };
 
     int processArgs(int argc, char* argv[])
@@ -110,14 +108,15 @@ class BatchNormFwdArg
             };
         };
 
-        if(optind + 5 > argc)
+        if(optind + 6 > argc)
             throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
 
         data_type              = std::atoi(argv[optind++]);
         updateMovingAverage    = std::atoi(argv[optind++]);
         saveMeanAndInvVariance = std::atoi(argv[optind++]);
         init_method            = std::atoi(argv[optind++]);
-        time_kernel            = static_cast<bool>(std::atoi(argv[optind]));
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
 
         if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
             return (-1);
@@ -128,7 +127,7 @@ class BatchNormFwdArg
 
 using namespace ck;
 
-template <typename InOutDataType, typename AccDataType>
+template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
 bool bnorm_fwd_nhwc_test(bool do_verification,
                          int init_method,
                          bool time_kernel,
@@ -273,73 +272,140 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
               scaleBiasMeanVarStrides.end(),
               i_scaleBiasMeanVarStrides.begin());
 
-    int result = 0;
-
-    // used for saving meansquare
-    DeviceMem workspace(sizeof(AccDataType) * 2 * resultSaveMean_ref.mDesc.GetElementSpaceSize() +
-                        128);
-
-    void* p_tmp_mean = workspace.GetDeviceBuffer();
-    void* p_tmp_meansquare =
-        static_cast<char*>(p_tmp_mean) +
-        (sizeof(AccDataType) * resultSaveMean_ref.mDesc.GetElementSpaceSize() + 63) / 64 * 64;
-
-    result = bnorm_fwd<InOutDataType, AccDataType, Rank, NumReduceDim, false>(
-        time_kernel,
-        updateMovingAverage,
-        saveMeanAndInvVariance,
-        {0, 1, 2},
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormFwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
+                                                             InOutDataType,
+                                                             AccDataType,
+                                                             AccDataType,   // ScaleDataType
+                                                             AccDataType,   // BiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp, // YElementwiseOp
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1>;
+
+    auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
+
+    auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
         i_inOutLengths,
         i_inOutStrides,
         i_inOutStrides,
+        {0, 1, 2},
         i_scaleBiasMeanVarLengths,
         i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
         x_dev.GetDeviceBuffer(),
         bnScale_dev.GetDeviceBuffer(),
         bnBias_dev.GetDeviceBuffer(),
-        y_dev.GetDeviceBuffer(),
-        averageFactor,
-        updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
-        updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr,
         epsilon,
+        PassThroughOp{},
+        y_dev.GetDeviceBuffer(),
         saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
         saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
-        p_tmp_mean,
-        p_tmp_meansquare);
+        averageFactor,
+        updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
+        updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
 
-    if(result < 0)
+    if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
         return (false);
+    };
+
+    size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes +=
+            total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
+
+        // outputing of mean, inv-variance
+        num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        // updating of moving mean, variance
+        num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
     bool pass = true;
 
     if(do_verification)
     {
-        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance<InOutDataType, AccDataType>{};
+
+        using ReferenceBatchNormFwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormFwd_Input_N_H_W_C_Output_C<InOutDataType,
+                                                                                     InOutDataType,
+                                                                                     AccDataType,
+                                                                                     AccDataType,
+                                                                                     AccDataType,
+                                                                                     AccDataType,
+                                                                                     PassThroughOp>;
+
+        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
 
         auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
             i_inOutLengths,
             i_inOutStrides,
             i_inOutStrides,
+            {0, 1, 2},
             i_scaleBiasMeanVarLengths,
             i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
             x.mData.data(),
             bnScale.mData.data(),
             bnBias.mData.data(),
-            y_ref.mData.data(),
-            0.1, // exponentialAverageFactor
-            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr, // resultRunningMean
-            updateMovingAverage ? resultRunningVariance_ref.mData.data()
-                                : nullptr, // resultRunningVariance
             epsilon,
+            PassThroughOp{},
+            y_ref.mData.data(),
             saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
-            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr);
+            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
+            updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
 
         if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
         {
-            std::cout
-                << "The runtime parameters seems not supported by the BatchNorm instance, exiting!"
-                << std::endl;
-            return (-2);
+            std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
+                         "instance, exiting!"
+                      << std::endl;
+            return (false);
         };
 
         auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
@@ -365,6 +431,8 @@ bool bnorm_fwd_nhwc_test(bool do_verification,
 
         if(saveMeanAndInvVariance)
         {
+            using ck::host_common::dumpBufferToFile;
+
             Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
             Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
 
@@ -396,70 +464,129 @@ int main(int argc, char* argv[])
 
         if(arg.data_type == 0)
         {
-            pass = bnorm_fwd_nhwc_test<ck::half_t, float>(arg.do_verification,
-                                                          arg.init_method,
-                                                          arg.time_kernel,
-                                                          arg.inOutLengths,
-                                                          arg.updateMovingAverage,
-                                                          arg.saveMeanAndInvVariance,
-                                                          averageFactor,
-                                                          epsilon);
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.updateMovingAverage,
+                                                                    arg.saveMeanAndInvVariance,
+                                                                    averageFactor,
+                                                                    epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
         }
         else if(arg.data_type == 1)
         {
-            pass = bnorm_fwd_nhwc_test<float, float>(arg.do_verification,
-                                                     arg.init_method,
-                                                     arg.time_kernel,
-                                                     arg.inOutLengths,
-                                                     arg.updateMovingAverage,
-                                                     arg.saveMeanAndInvVariance,
-                                                     averageFactor,
-                                                     epsilon);
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.updateMovingAverage,
+                                                               arg.saveMeanAndInvVariance,
+                                                               averageFactor,
+                                                               epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
         }
         else if(arg.data_type == 3)
         {
-            pass = bnorm_fwd_nhwc_test<int8_t, float>(arg.do_verification,
-                                                      arg.init_method,
-                                                      arg.time_kernel,
-                                                      arg.inOutLengths,
-                                                      arg.updateMovingAverage,
-                                                      arg.saveMeanAndInvVariance,
-                                                      averageFactor,
-                                                      epsilon);
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
         }
         else if(arg.data_type == 5)
         {
-            pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float>(arg.do_verification,
-                                                           arg.init_method,
-                                                           arg.time_kernel,
-                                                           arg.inOutLengths,
-                                                           arg.updateMovingAverage,
-                                                           arg.saveMeanAndInvVariance,
-                                                           averageFactor,
-                                                           epsilon);
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.updateMovingAverage,
+                                                                      arg.saveMeanAndInvVariance,
+                                                                      averageFactor,
+                                                                      epsilon);
         }
         else if(arg.data_type == 6)
         {
-            pass = bnorm_fwd_nhwc_test<double, double>(arg.do_verification,
-                                                       arg.init_method,
-                                                       arg.time_kernel,
-                                                       arg.inOutLengths,
-                                                       arg.updateMovingAverage,
-                                                       arg.saveMeanAndInvVariance,
-                                                       averageFactor,
-                                                       epsilon);
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.updateMovingAverage,
+                                                                  arg.saveMeanAndInvVariance,
+                                                                  averageFactor,
+                                                                  epsilon);
         }
     }
     else
     {
-        pass = bnorm_fwd_nhwc_test<ck::half_t, float>(true,
-                                                      2,
-                                                      false, // don't time kernel
-                                                      {128, 16, 16, 1024},
-                                                      true,
-                                                      false,
-                                                      averageFactor,
-                                                      epsilon);
+        pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            2,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            true,
+                                                            true,
+                                                            averageFactor,
+                                                            epsilon);
+
+        pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     2,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     true,
+                                                                     true,
+                                                                     averageFactor,
+                                                                     epsilon);
     };
 
     return (pass ? 0 : 1);

example/34_batchnorm/batchnorm_infer_impl.hpp

@@ -14,8 +14,12 @@
 
 #include "batchnorm_common.hpp"
 
-template <typename InOutDataType,
+template <typename XDataType,
+          typename YDataType,
           typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
           ck::index_t Rank,
           ck::index_t NumBatchNormReduceDim,
           bool fastest_dim_is_reduced = false>
@@ -26,7 +30,9 @@ int bnorm_infer(
     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 std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
     const void* p_x,
     const void* p_scale,
     const void* p_bias,
@@ -41,11 +47,11 @@ int bnorm_infer(
                   "Invalid number of reduced dimensions for batchnorm!");
 
     using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
-        ck::Tuple<InOutDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
-                                                                                      // variance,
-                                                                                      // scale,
-                                                                                      // bias,
-        ck::Tuple<InOutDataType>,                                                     // y
+        ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
+                                                                                  // variance,
+                                                                                  // scale,
+                                                                                  // bias,
+        ck::Tuple<YDataType>,                                                     // y
         NormalizeInInfer,
         Rank,
         2,                           // MPerthread
@@ -53,14 +59,18 @@ int bnorm_infer(
         ck::Sequence<1>>;            // scalarPerVector: y
 
     auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
-    std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnScaleStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnBiasStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnMeanVarStrides{0};
 
     int i = 0;
     for(auto dim : invariantDims)
     {
         assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
 
-        aligned_scaleBiasMeanVarStrides[dim] = bnScaleBiasMeanVarStrides[i];
+        aligned_bnScaleStrides[dim]   = bnScaleStrides[i];
+        aligned_bnBiasStrides[dim]    = bnBiasStrides[i];
+        aligned_bnMeanVarStrides[dim] = bnMeanVarStrides[i];
         i++;
     };
 
@@ -84,10 +94,10 @@ int bnorm_infer(
     auto argument_ptr1 = dev_normalize.MakeArgumentPointer(
         xyLengths,
         {xStrides,
-         aligned_scaleBiasMeanVarStrides,
-         aligned_scaleBiasMeanVarStrides,
-         aligned_scaleBiasMeanVarStrides,
-         aligned_scaleBiasMeanVarStrides},
+         aligned_bnMeanVarStrides,
+         aligned_bnMeanVarStrides,
+         aligned_bnScaleStrides,
+         aligned_bnBiasStrides},
         {yStrides},
         {p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias},
         {p_y},
@@ -105,8 +115,10 @@ int bnorm_infer(
 
     avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
 
-    num_bytes += (total_length * (1 * sizeof(InOutDataType) + 4 * sizeof(AccDataType)) +
-                  total_length * sizeof(InOutDataType));
+    num_bytes += total_length * sizeof(XDataType) +
+                 invariantLength *
+                     (sizeof(ScaleDataType) + sizeof(BiasDataType) + 2 * sizeof(MeanVarDataType)) +
+                 total_length * sizeof(YDataType);
 
     if(time_kernel)
     {

example/34_batchnorm/batchnorm_infer_nhwc.cpp

@@ -18,11 +18,6 @@
 
 #include "batchnorm_infer_impl.hpp"
 
-template <typename InOutDataType, typename AccDataType>
-using ReferenceBatchNormInferInstance =
-    ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<InOutDataType,
-                                                                               AccDataType>;
-
 static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
                                        {"verify", required_argument, nullptr, 'v'},
                                        {"help", no_argument, nullptr, '?'},
@@ -236,21 +231,30 @@ bool bnorm_infer_nhwc_test(bool do_verification,
 
     int result = 0;
 
-    result = bnorm_infer<InOutDataType, AccDataType, Rank, NumReduceDim, false>(
-        time_kernel,
-        {0, 1, 2},
-        i_inOutLengths,
-        i_inOutStrides,
-        i_inOutStrides,
-        i_scaleBiasMeanVarLengths,
-        i_scaleBiasMeanVarStrides,
-        x_dev.GetDeviceBuffer(),
-        bnScale_dev.GetDeviceBuffer(),
-        bnBias_dev.GetDeviceBuffer(),
-        epsilon,
-        estimatedMean_dev.GetDeviceBuffer(),
-        estimatedVariance_dev.GetDeviceBuffer(),
-        y_dev.GetDeviceBuffer());
+    result = bnorm_infer<InOutDataType,
+                         InOutDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         Rank,
+                         NumReduceDim,
+                         false>(time_kernel,
+                                {0, 1, 2},
+                                i_inOutLengths,
+                                i_inOutStrides,
+                                i_inOutStrides,
+                                i_scaleBiasMeanVarLengths,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                x_dev.GetDeviceBuffer(),
+                                bnScale_dev.GetDeviceBuffer(),
+                                bnBias_dev.GetDeviceBuffer(),
+                                epsilon,
+                                estimatedMean_dev.GetDeviceBuffer(),
+                                estimatedVariance_dev.GetDeviceBuffer(),
+                                y_dev.GetDeviceBuffer());
 
     if(result < 0)
         return (false);
@@ -259,7 +263,15 @@ bool bnorm_infer_nhwc_test(bool do_verification,
 
     if(do_verification)
     {
-        auto batchNormInfer_ref = ReferenceBatchNormInferInstance<InOutDataType, AccDataType>{};
+        using ReferenceBatchNormInferInstance =
+            ck::tensor_operation::host::ReferenceBatchNormInfer_Input_N_H_W_C_Output_C<
+                InOutDataType,
+                InOutDataType,
+                AccDataType,
+                AccDataType,
+                AccDataType,
+                AccDataType>;
+        auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
 
         auto argument_ptr_ref =
             batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
@@ -267,6 +279,8 @@ bool bnorm_infer_nhwc_test(bool do_verification,
                                                    i_inOutStrides,
                                                    i_scaleBiasMeanVarLengths,
                                                    i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
                                                    x.mData.data(),
                                                    bnScale.mData.data(),
                                                    bnBias.mData.data(),

include/ck/ck.hpp

@@ -159,6 +159,11 @@
 // tuning parameter
 #define CK_WORKAROUND_SWDEV_325164 0
 
+// workaround: disable broken fused attention kernel instance that does not pass validation
+// issue found on mi100/#10738 combo when irregular KPerBlock attention kernel has acc0 scaling
+// enabled
+#define CK_WORKAROUND_DISABLE_BROKEN_ATTN_KERNEL_INSTANCE 1
+
 namespace ck {
 
 enum struct InMemoryDataOperationEnum

include/ck/tensor_description/tensor_space_filling_curve.hpp

@@ -14,7 +14,8 @@ namespace ck {
 
 template <typename TensorLengths,
           typename DimAccessOrder,
-          typename ScalarsPerAccess> // # of scalars per access in each dimension
+          typename ScalarsPerAccess,
+          bool SnakeCurved = true> // # of scalars per access in each dimension
 struct SpaceFillingCurve
 {
     static constexpr index_t nDim = TensorLengths::Size();
@@ -136,9 +137,10 @@ struct SpaceFillingCurve
             Index ordered_idx;
 
             static_for<0, nDim, 1>{}([&](auto idim) {
-                ordered_idx(idim) = forward_sweep[idim] ? ordered_access_idx[idim]
-                                                        : ordered_access_lengths[idim] - 1 -
-                                                              ordered_access_idx[idim];
+                ordered_idx(idim) =
+                    !SnakeCurved || forward_sweep[idim]
+                        ? ordered_access_idx[idim]
+                        : ordered_access_lengths[idim] - 1 - ordered_access_idx[idim];
             });
 
             return container_reorder_given_old2new(ordered_idx, dim_access_order) *

include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp

@@ -151,6 +151,27 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         return make_tuple(c_thread_m, c_thread_n);
     }
 
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(Number<m0>{},
+                          Number<n0>{},
+                          waveId_m,
+                          waveId_n,
+                          blk_idx[I0],
+                          blk_idx[I1],
+                          blk_idx[I2],
+                          blk_idx[I3]);
+    }
+
     __host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1()
     {
         static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
@@ -724,6 +745,21 @@ struct BlockwiseGemmXdlops_v2
         return make_tuple(c_thread_m, c_thread_n);
     }
 
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
     using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
 
     __host__ __device__ BlockwiseGemmXdlops_v2(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp

@@ -24,7 +24,8 @@ template <typename ALayout,
           typename B0ElementwiseOperation,
           typename Acc0ElementwiseOperation,
           typename B1ElementwiseOperation,
-          typename CElementwiseOperation>
+          typename CElementwiseOperation,
+          bool MaskOutUpperTriangle> // TODO: enum for mask type
 struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
 {
     virtual std::unique_ptr<BaseArgument>

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp

@@ -7,49 +7,60 @@
 #include <vector>
 
 #include "device_base.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <typename ALayout,
-          typename B0Layout,
-          typename B1Layout,
-          typename CPermuteNumDims_G_M_Gemm1N, // Sequence<>
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
           typename ADataType,
           typename B0DataType,
           typename B1DataType,
           typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
           typename AElementwiseOperation,
           typename B0ElementwiseOperation,
           typename Acc0ElementwiseOperation,
           typename B1ElementwiseOperation,
-          typename CElementwiseOperation>
+          typename CElementwiseOperation,
+          MaskingSpecialization MaskingSpec>
 struct DeviceBatchedGemmSoftmaxGemmPermute : public BaseOperator
 {
-    virtual std::unique_ptr<BaseArgument>
-    MakeArgumentPointer(const void* p_a,
-                        const void* p_b0,
-                        const void* p_b1,
-                        void* p_c,
-                        ck::index_t M,
-                        ck::index_t N,
-                        ck::index_t K,
-                        ck::index_t O,
-                        ck::index_t Batch,
-                        std::vector<index_t> c_gs_ms_os_lengths,
-                        std::vector<index_t> c_gs_ms_os_strides,
-                        ck::index_t StrideA,
-                        ck::index_t StrideB0,
-                        ck::index_t StrideB1,
-                        ck::index_t BatchStrideA,
-                        ck::index_t BatchStrideB0,
-                        ck::index_t BatchStrideB1,
-                        AElementwiseOperation a_element_op,
-                        B0ElementwiseOperation b0_element_op,
-                        Acc0ElementwiseOperation acc0_element_op,
-                        B1ElementwiseOperation b1_element_op,
-                        CElementwiseOperation c_element_op) = 0;
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b0,
+        const void* p_b1,
+        void* p_c,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        B0ElementwiseOperation b0_element_op,
+        Acc0ElementwiseOperation acc0_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };

include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp

@@ -13,31 +13,36 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <index_t Rank, index_t NumBatchNormReduceDim>
+template <index_t Rank, index_t NumBatchNormReduceDim, typename YElementwiseOp>
 struct DeviceBatchNormFwd : public BaseOperator
 {
     virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
         const std::array<index_t, Rank> xyLengths,
         const std::array<index_t, Rank> xStrides,
         const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
         const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
-        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
         const void* p_x,
         const void* bnScale,
         const void* bnBias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
         void* p_y,
+        void* resultSaveMean,
+        void* resultSaveInvVariance,
         double exponentialAverageFactor,
         void* resultRunningMean,
-        void* resultRunningVariance,
-        double epsilon,
-        void* resultSaveMean,
-        void* resultSaveInvVariance) = 0;
+        void* resultRunningVariance) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
-template <index_t Rank, index_t NumBatchNormReduceDim>
-using DeviceBatchNormFwdPtr = std::unique_ptr<DeviceBatchNormFwd<Rank, NumBatchNormReduceDim>>;
+template <index_t Rank, index_t NumBatchNormReduceDim, typename YElementwiseOp>
+using DeviceBatchNormFwdPtr =
+    std::unique_ptr<DeviceBatchNormFwd<Rank, NumBatchNormReduceDim, YElementwiseOp>>;
 
 } // namespace device
 } // namespace tensor_operation

include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp

@@ -21,7 +21,9 @@ struct DeviceBatchNormInfer : public BaseOperator
         const std::array<index_t, Rank> xStrides,
         const std::array<index_t, Rank> yStrides,
         const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
-        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
         const void* p_x,
         const void* bnScale,
         const void* bnBias,

include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedConvFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a, // input image
+                        const void* p_b, // weight
+                        void* p_c,       // output image
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const AElementwiseOperation& a_element_op,
+                        const BElementwiseOperation& b_element_op,
+                        const CElementwiseOperation& c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA, index_t BatchStrideB, index_t BatchStrideC)
+        : BatchStrideA_(BatchStrideA), BatchStrideB_(BatchStrideB), BatchStrideC_(BatchStrideC)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideC_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    index_t BatchStrideC_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename CDataType,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_dl(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            CDataType* __restrict__ p_c_grid,
+            const index_t batch_count,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
+
+    __shared__ ABDataType p_shared[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid + a_batch_offset,
+                      p_b_grid + b_batch_offset,
+                      p_c_grid + c_batch_offset,
+                      p_shared,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = c_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetCPtrOffset(0);
+#endif
+}
+
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <
+    index_t NDimSpatial,
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AccDataType,
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    GemmSpecialization GemmSpec,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t K1,
+    index_t M1PerThread,
+    index_t N1PerThread,
+    index_t KPerThread,
+    typename M1N1ThreadClusterM1Xs,
+    typename M1N1ThreadClusterN1Xs,
+    typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+    typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+    typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+    typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+    typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+    typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+    typename CThreadTransferSrcDstAccessOrder,
+    index_t CThreadTransferSrcDstVectorDim,
+    index_t CThreadTransferDstScalarPerVector,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimSpatial,
+                                                                           ALayout,
+                                                                           BLayout,
+                                                                           CLayout,
+                                                                           ADataType,
+                                                                           BDataType,
+                                                                           CDataType,
+                                                                           AElementwiseOperation,
+                                                                           BElementwiseOperation,
+                                                                           CElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_AK0_M_AK1(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                                  const std::array<index_t, NDimSpatial>& input_left_pads,
+                                  const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        c_g_n_k_wos_lengths,
+                                                                        c_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        const auto M = in_gemmm_gemmk_desc.GetLength(I0);
+        const auto K = in_gemmm_gemmk_desc.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            in_gemmm_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_BK0_N_BK1(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        const auto N = wei_gemmn_gemmk_desc.GetLength(I0);
+        const auto K = wei_gemmn_gemmk_desc.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            wei_gemmn_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename CLay>
+    static auto
+    MakeCGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<CLay>(c_g_n_k_wos_lengths,
+                                                                        c_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    // desc for problem definition
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
+        MakeAGridDescriptor_AK0_M_AK1<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>({}, {}))>;
+    using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N<CLayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
+                                     ADataType,
+                                     AccDataType,
+                                     CDataType,
+                                     InMemoryDataOperationEnum::Set,
+                                     AGridDesc_AK0_M_AK1,
+                                     BGridDesc_BK0_N_BK1,
+                                     CGridDesc_M_N,
+                                     MPerBlock,
+                                     NPerBlock,
+                                     K0PerBlock,
+                                     K1,
+                                     M1PerThread,
+                                     N1PerThread,
+                                     KPerThread,
+                                     M1N1ThreadClusterM1Xs,
+                                     M1N1ThreadClusterN1Xs,
+                                     ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterArrangeOrder,
+                                     ABlockTransferSrcAccessOrder,
+                                     ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                     ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                     ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                     BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterArrangeOrder,
+                                     BBlockTransferSrcAccessOrder,
+                                     BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                     BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                     BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                     CThreadTransferSrcDstAccessOrder,
+                                     CThreadTransferSrcDstVectorDim,
+                                     CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_AK0_M_AK1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_BK0_N_BK1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 void* p_c,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CElementwiseOperation& c_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_c_grid_{static_cast<CDataType*>(p_c)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1<ALayout>(a_g_n_c_wis_lengths,
+                                                                   a_g_n_c_wis_strides,
+                                                                   b_g_k_c_xs_lengths,
+                                                                   b_g_k_c_xs_strides,
+                                                                   c_g_n_k_wos_lengths,
+                                                                   c_g_n_k_wos_strides,
+                                                                   conv_filter_strides,
+                                                                   conv_filter_dilations,
+                                                                   input_left_pads,
+                                                                   input_right_pads)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1<BLayout>(
+                  b_g_k_c_xs_lengths, b_g_k_c_xs_strides)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N<CLayout>(c_g_n_k_wos_lengths,
+                                                                          c_g_n_k_wos_strides)},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              c_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{
+                  a_g_n_c_wis_strides[0], b_g_k_c_xs_strides[0], c_g_n_k_wos_strides[0]},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              c_g_n_k_wos_lengths_{c_g_n_k_wos_lengths},
+              c_g_n_k_wos_strides_{c_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideC_ = c_g_n_k_wos_strides[0];
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, c_grid_desc_m_n_))
+            {
+
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_ak0_m_ak1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_bk0_n_bk1_);
+                c_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[K0, M, K1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B[K0, N, K1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "C[M, N]: " << c_grid_desc_m_n_ << std::endl;
+            std::cout << "num_group: " << num_group_ << std::endl;
+
+            std::cout << "A[k0, m0, m1, k1]: " << a_grid_desc_k0_m0_m1_k1_ << std::endl;
+            std::cout << "B[k0, n0, n1, k1]: " << b_grid_desc_k0_n0_n1_k1_ << std::endl;
+            std::cout << "A[m0, m10, m11, n0, n10, n11]: " << c_grid_desc_m0_m10_m11_n0_n10_n11_
+                      << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // block-to-e-tile map
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            // if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_.GetLength(I0),
+                                                arg.c_grid_desc_m_n_.GetLength(I1)) *
+                arg.num_group_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop;
+
+                const auto kernel =
+                    kernel_grouped_conv_fwd_dl<GridwiseGemm,
+                                               ADataType, // TODO: distiguish A/B datatype
+                                               CDataType,
+                                               DeviceOp::AGridDesc_K0_M0_M1_K1,
+                                               DeviceOp::BGridDesc_K0_N0_N1_K1,
+                                               DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11,
+                                               DefaultBlock2CTileMap,
+                                               ComputePtrOffsetOfStridedBatch,
+                                               has_main_loop,
+                                               has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X
+                              << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad
+                              << " RightPad = " << RightPad << std::endl;
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X
+                              << " LeftPad = " << LeftPad << " RightPad = " << RightPad
+                              << std::endl;
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of C
+        if constexpr(is_same_v<CLayout, ctc::G_NW_K> || is_same_v<CLayout, ctc::G_NHW_K> ||
+                     is_same_v<CLayout, ctc::G_NDHW_K> || is_same_v<CLayout, ctc::GNWK> ||
+                     is_same_v<CLayout, ctc::GNHWK> || is_same_v<CLayout, ctc::GNDHWK> ||
+                     is_same_v<CLayout, ctc::NWGK> || is_same_v<CLayout, ctc::NHWGK> ||
+                     is_same_v<CLayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.c_g_n_k_wos_lengths_[2];
+
+            if(!(K % CThreadTransferDstScalarPerVector == 0 && CThreadTransferSrcDstVectorDim == 5))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             void* p_c,
+                             const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                             const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                             const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                             const std::array<index_t, NDimSpatial>& input_left_pads,
+                             const std::array<index_t, NDimSpatial>& input_right_pads,
+                             const AElementwiseOperation& a_element_op,
+                             const BElementwiseOperation& b_element_op,
+                             const CElementwiseOperation& c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        c_g_n_k_wos_lengths,
+                        c_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const AElementwiseOperation& a_element_op,
+                        const BElementwiseOperation& b_element_op,
+                        const CElementwiseOperation& c_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_c,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          c_g_n_k_wos_lengths,
+                                          c_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp

@@ -7,46 +7,50 @@
 #include <vector>
 
 #include "device_base.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <typename ALayout,
-          typename B0Layout,
-          typename B1Layout,
-          typename CPermuteNumDims_G_M_Gemm1N, // Sequence<>
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
           typename ADataType,
           typename B0DataType,
           typename B1DataType,
           typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
           typename AElementwiseOperation,
           typename B0ElementwiseOperation,
           typename Acc0ElementwiseOperation,
           typename B1ElementwiseOperation,
-          typename CElementwiseOperation>
+          typename CElementwiseOperation,
+          MaskingSpecialization MaskingSpec>
 struct DeviceGroupedGemmSoftmaxGemmPermute : public BaseOperator
 {
     struct ProblemDesc
     {
-        // Overall problem shape
-        index_t M;
-        index_t N;
-        index_t K;
-        index_t O;
-        index_t Batch;
+        std::vector<index_t> a_gs_ms_ks_lengths;
+        std::vector<index_t> a_gs_ms_ks_strides;
 
-        // Stride for A/B0/B1; layout determined by template args
-        index_t StrideA;
-        index_t StrideB0;
-        index_t StrideB1;
-        index_t BatchStrideA;
-        index_t BatchStrideB0;
-        index_t BatchStrideB1;
+        std::vector<index_t> b0_gs_ns_ks_lengths;
+        std::vector<index_t> b0_gs_ns_ks_strides;
+
+        std::vector<index_t> b1_gs_os_ns_lengths;
+        std::vector<index_t> b1_gs_os_ns_strides;
 
-        // Lengths and strides for output C
         std::vector<index_t> c_gs_ms_os_lengths;
         std::vector<index_t> c_gs_ms_os_strides;
+
+        std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_lengths;
+        std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_strides;
+
+        std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_lengths;
+        std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_strides;
     };
 
     virtual std::unique_ptr<BaseArgument>
@@ -54,6 +58,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute : public BaseOperator
                         std::vector<const void*> p_b0_vec,
                         std::vector<const void*> p_b1_vec,
                         std::vector<void*> p_c_vec,
+                        std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                        std::vector<std::vector<const void*>> p_acc1_biases_vec,
                         std::vector<ProblemDesc> problem_desc_vec,
                         AElementwiseOperation a_element_op,
                         B0ElementwiseOperation b0_element_op,

include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp

@@ -14,6 +14,7 @@
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 
@@ -54,9 +55,8 @@ __global__ void
     index_t right    = group_count;
     index_t group_id = index_t((left + right) / 2);
 
-    while((!(block_id >= arg_ptr[group_id].block_start_ &&
-             block_id < arg_ptr[group_id].block_end_)) &&
-          left <= right)
+    while(
+        (!(block_id >= arg_ptr[group_id].block_start_ && block_id < arg_ptr[group_id].block_end_)))
     {
         if(block_id < arg_ptr[group_id].block_start_)
         {
@@ -114,14 +114,17 @@ __global__ void
 // Computes C = A * B0 * B1
 //              ^^^^^^ (Acc0)
 //              ^^^^^^^^^^^ (Acc1)
-template <typename ALayout,
-          typename BLayout, // B0Layout
-          typename B1Layout,
-          typename CPermuteNumDims_G_M_Gemm1N, // Sequence<NumDimG, NumDimM, NumDimGemm1N>
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO, // NumDimGemm1N
           typename ADataType,
           typename BDataType,
           typename B1DataType,
           typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
           typename GemmAccDataType,
           typename CShuffleDataType,
           typename AElementwiseOperation,
@@ -130,6 +133,10 @@ template <typename ALayout,
           typename B1ElementwiseOperation,
           typename CElementwiseOperation,
           GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization B1Spec,
+          TensorSpecialization CSpec,
           index_t NumGemmKPrefetchStage,
           index_t BlockSize,
           index_t MPerBlock,
@@ -170,297 +177,152 @@ template <typename ALayout,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
-          bool MaskOutUpperTriangle,
+          MaskingSpecialization MaskingSpec,
           LoopScheduler LoopSched = LoopScheduler::Default>
 struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
-    : public DeviceGroupedGemmSoftmaxGemmPermute<ALayout,
-                                                 BLayout,
-                                                 B1Layout,
-                                                 CPermuteNumDims_G_M_Gemm1N,
+    : public DeviceGroupedGemmSoftmaxGemmPermute<NumDimG,
+                                                 NumDimM,
+                                                 NumDimN,
+                                                 NumDimK,
+                                                 NumDimO,
                                                  ADataType,
                                                  BDataType,
                                                  B1DataType,
                                                  CDataType,
+                                                 Acc0BiasDataType,
+                                                 Acc1BiasDataType,
                                                  AElementwiseOperation,
                                                  BElementwiseOperation,
                                                  AccElementwiseOperation,
                                                  B1ElementwiseOperation,
-                                                 CElementwiseOperation>
+                                                 CElementwiseOperation,
+                                                 MaskingSpec>
 {
-    using DeviceOp = DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle;
-    using ProblemDesc =
-        typename DeviceGroupedGemmSoftmaxGemmPermute<ALayout,
-                                                     BLayout,
-                                                     B1Layout,
-                                                     CPermuteNumDims_G_M_Gemm1N,
-                                                     ADataType,
-                                                     BDataType,
-                                                     B1DataType,
-                                                     CDataType,
-                                                     AElementwiseOperation,
-                                                     BElementwiseOperation,
-                                                     AccElementwiseOperation,
-                                                     B1ElementwiseOperation,
-                                                     CElementwiseOperation>::ProblemDesc;
+    static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
+                  "Number of dimension must be greater than 0");
+
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    // TODO ANT: implement bias combination
+    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+
+#if 0
+    // TODO ANT: use alias
+    static constexpr index_t NumDimGemm0M = NumDimM;
+    static constexpr index_t NumDimGemm0N = NumDimN;
+    static constexpr index_t NumDimGemm0K = NumDimK;
+    static constexpr index_t NumDimGemm1M = NumDimM;
+    static constexpr index_t NumDimGemm1N = NumDimO;
+    static constexpr index_t NumDimGemm1K = NumDimN;
+#endif
+
+    using DeviceOp    = DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle;
+    using ProblemDesc = typename DeviceGroupedGemmSoftmaxGemmPermute<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     NumDimO,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     B1DataType,
+                                                                     CDataType,
+                                                                     Acc0BiasDataType,
+                                                                     Acc1BiasDataType,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     AccElementwiseOperation,
+                                                                     B1ElementwiseOperation,
+                                                                     CElementwiseOperation,
+                                                                     MaskingSpec>::ProblemDesc;
 
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
 
-    static constexpr auto matrix_padder =
-        GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
-            MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
-
-    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
-    {
-        const auto a_grid_desc_mraw_kraw = [&]() {
-            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
-            {
-                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
-                                                    make_tuple(StrideA, I1));
-            }
-            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
-            {
-                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
-                                                    make_tuple(I1, StrideA));
-            }
-        }();
-
-        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
-
-        const auto M = a_grid_desc_m_k.GetLength(I0);
-        const auto K = a_grid_desc_m_k.GetLength(I1);
-
-        const auto AK0 = K / AK1;
-
-        return transform_tensor_descriptor(a_grid_desc_m_k,
-                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
-                                                      make_pass_through_transform(M)),
-                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-    }
-
-    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
-    {
-        const auto b_grid_desc_nraw_kraw = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(I1, StrideB));
-            }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(StrideB, I1));
-            }
-        }();
-
-        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
-
-        const auto N = b_grid_desc_n_k.GetLength(I0);
-        const auto K = b_grid_desc_n_k.GetLength(I1);
-
-        const auto BK0 = K / BK1;
-
-        return transform_tensor_descriptor(b_grid_desc_n_k,
-                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                                      make_pass_through_transform(N)),
-                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-    }
-
-    // Args: Gemm1KRaw, Gemm1NRaw, StrideB1
-    static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                              const std::vector<index_t>& a_gs_ms_ks_strides_vec)
     {
-        const auto b1_grid_desc_nraw_kraw = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(I1, StrideB));
-            }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(StrideB, I1));
-            }
-        }();
-
-        const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
-
-        const auto N = b1_grid_desc_n_k.GetLength(I0);
-        const auto K = b1_grid_desc_n_k.GetLength(I1);
-
-        const auto B1K0 = K / B1K1;
-
-        return transform_tensor_descriptor(
-            b1_grid_desc_n_k,
-            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
-                       make_pass_through_transform(N)),
-            make_tuple(Sequence<1>{}, Sequence<0>{}),
-            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
+            Number<AK1>{});
     }
 
-    // assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
-    static auto MakeCGridDescriptor_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
-                                        const std::vector<index_t>& c_gs_ms_ns_strides_vec)
+    static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                              const std::vector<index_t>& b_gs_ns_ks_strides_vec)
     {
-        constexpr index_t NumDimG = CPermuteNumDims_G_M_Gemm1N::At(I0);
-        constexpr index_t NumDimM = CPermuteNumDims_G_M_Gemm1N::At(I1);
-        constexpr index_t NumDimN = CPermuteNumDims_G_M_Gemm1N::At(I2); // NumDimGemm1N
-
-        assert(c_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
-               c_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
-
-        const auto to_tuple = [&](auto& vec, auto start, auto end) {
-            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
-        };
-
-        const auto c_ms_ns_lengths = to_tuple(
-            c_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
-        const auto c_ms_ns_strides = to_tuple(
-            c_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
-
-        // dimension Ids for M0, M1, ...
-        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
-
-        // dimension Ids for N0, N1, ...
-        constexpr auto nDimIds =
-            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
-
-        // lengths for M0, M1, ...
-        const auto mLengths = get_container_subset(c_ms_ns_lengths, mDimIds);
-
-        // lengths for K0, K1, ...
-        const auto nLengths = get_container_subset(c_ms_ns_lengths, nDimIds);
-
-        // naive tensor C[M0, M1, M2, ..., N0, N1, N2...]
-        const auto c_grid_desc_ms_ns =
-            make_naive_tensor_descriptor(c_ms_ns_lengths, c_ms_ns_strides);
-
-        // transformed tensor C[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
-        const auto c_grid_desc_mraw_nraw = transform_tensor_descriptor(
-            c_grid_desc_ms_ns,
-            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
-            make_tuple(mDimIds, nDimIds),
-            make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(
+            Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
+            Number<BK1>{});
     }
 
-    // assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
-    static auto MakeCGridDescriptor_G_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
-                                          const std::vector<index_t>& c_gs_ms_ns_strides_vec)
+    static auto
+    MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                   const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
     {
-        constexpr index_t NumDimG = CPermuteNumDims_G_M_Gemm1N::At(I0);
-        constexpr index_t NumDimM = CPermuteNumDims_G_M_Gemm1N::At(I1);
-        constexpr index_t NumDimN = CPermuteNumDims_G_M_Gemm1N::At(I2); // NumDimGemm1N
-
-        assert(c_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
-               c_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
-
-        const auto to_tuple = [&](auto& vec, auto start, auto end) {
-            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
-        };
-
-        const auto c_gs_ms_ns_lengths =
-            to_tuple(c_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
-        const auto c_gs_ms_ns_strides =
-            to_tuple(c_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
-
-        // dimension Ids for G0, G1, ...
-        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
-
-        // dimension Ids for M0, M1, ...
-        constexpr auto mDimIds =
-            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
-
-        // dimension Ids for N0, N1, ...
-        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
-                                                                  NumDimG + NumDimM + NumDimN,
-                                                                  1>::type{};
-
-        // lengths for G0, G1, ...
-        const auto gLengths = get_container_subset(c_gs_ms_ns_lengths, gDimIds);
-
-        // lengths for M0, M1, ...
-        const auto mLengths = get_container_subset(c_gs_ms_ns_lengths, mDimIds);
-
-        // lengths for K0, K1, ...
-        const auto nLengths = get_container_subset(c_gs_ms_ns_lengths, nDimIds);
-
-        // naive tensor C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
-        const auto c_grid_desc_gs_ms_ns =
-            make_naive_tensor_descriptor(c_gs_ms_ns_lengths, c_gs_ms_ns_strides);
-
-        // transformed tensor C[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
-        // N2 * ...]
-        const auto c_grid_desc_g_mraw_nraw =
-            transform_tensor_descriptor(c_grid_desc_gs_ms_ns,
-                                        make_tuple(make_merge_transform(gLengths),
-                                                   make_merge_transform(mLengths),
-                                                   make_merge_transform(nLengths)),
-                                        make_tuple(gDimIds, mDimIds, nDimIds),
-                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-
-        // this desc is only for calculating batch offset so no padding needed
-        return c_grid_desc_g_mraw_nraw;
+        return Transform::MakeB1GridDescriptor_BK0_N_BK1(
+            Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                                b1_gs_gemm1ns_gemm1ks_strides_vec),
+            Number<B1K1>{});
     }
 
-    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
-    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
-    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
-    using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N({}, {}));
-    using CGridDesc_G_M_N      = decltype(MakeCGridDescriptor_G_M_N({}, {}));
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
+    using CGridDesc_M_N        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using AGridDesc_G_M_K      = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
+    using BGridDesc_G_N_K      = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
+    using B1GridDesc_G_N_K     = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
+    using CGridDesc_G_M_N      = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
 
-    // to track the points which need to be set to -inf on C0
-    // Note: no need to reset M padding value, because they will not be stored out.
-    struct C0MatrixMask
+    constexpr static auto make_MaskOutPredicate()
     {
-        C0MatrixMask(index_t NRaw) : NRaw_(NRaw) {}
-
-        __host__ __device__ bool IsUpperTriangle(index_t m, index_t n) const { return n > m; }
-
-        __host__ __device__ bool IsNOutOfBound(/*index_t m, */ index_t n) const
+        if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
         {
-            return n >= NRaw_;
+            return MaskDisabledPredicate{};
         }
-
-        __host__ __device__ bool IsMaskedElement(index_t m, index_t n) const
+        else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
         {
-            return IsUpperTriangle(m, n) || IsNOutOfBound(n);
+            return MaskOutUpperTrianglePredicate{};
         }
-
-        private:
-        // index_t MRaw_;
-        index_t NRaw_;
-    };
+    }
+    using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
 
     struct ComputeBasePtrOfStridedBatch
     {
-        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
-                                     index_t BatchStrideB,
-                                     index_t BatchStrideB1,
-                                     CGridDesc_G_M_N c_grid_desc_g_m_n)
-            : BatchStrideA_(BatchStrideA),
-              BatchStrideB_(BatchStrideB),
-              BatchStrideB1_(BatchStrideB1),
+        ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
+                                     const BGridDesc_G_N_K& b_grid_desc_g_n_k,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+            : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
+              b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
               c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
         {
         }
 
         __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
         {
-            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+            return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
         }
 
         __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
         {
-            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+            return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
         }
 
         __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
         {
-            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+            return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
         }
 
         __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
@@ -469,9 +331,9 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
         }
 
         private:
-        index_t BatchStrideA_;
-        index_t BatchStrideB_;
-        index_t BatchStrideB1_;
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
         CGridDesc_G_M_N c_grid_desc_g_m_n_;
     };
 
@@ -535,8 +397,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
         CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CShuffleBlockTransferScalarPerVector_NPerBlock,
         LoopSched,
-        matrix_padder.PadN,
-        MaskOutUpperTriangle>;
+        Transform::matrix_padder.PadN,
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
 
     using Block2CTileMap = OffsettedBlockToCTileMap<typename GridwiseGemm::DefaultBlock2CTileMap>;
 
@@ -570,16 +432,16 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
     struct GroupDeviceArg
     {
-        // problem definiton
-        index_t M;
-        index_t N;
-        index_t K;
-        index_t O;
+        // lengths for the last dimensions of overall problem for sanity check of vector load/store
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
 
-        // Strides for the last dimensions of C for sanity check of vector load/store
-        index_t c_extent_lowest_;
-        index_t c_stride_lowest_;
+        // strides for the last dimensions of each tensor for sanity check of vector load/store
+        std::vector<index_t> a_mz_kz_strides_;
+        std::vector<index_t> b_nz_kz_strides_;
+        std::vector<index_t> b1_nz_kz_strides_;
+        std::vector<index_t> c_mz_gemm1nz_strides_;
 
+        // for gridwise gemm check
         CGridDesc_M_N c_grid_desc_m_n_;
     };
 
@@ -591,6 +453,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
                  std::vector<const void*> p_b_vec,
                  std::vector<const void*> p_b1_vec,
                  std::vector<void*> p_c_vec,
+                 std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                 std::vector<std::vector<const void*>> p_acc1_biases_vec,
                  std::vector<ProblemDesc> problem_desc_vec,
                  AElementwiseOperation a_element_op,
                  BElementwiseOperation b_element_op,
@@ -603,6 +467,7 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
               b1_element_op_{b1_element_op},
               c_element_op_{c_element_op}
         {
+            // TODO ANT: implement bias addition
             group_count_ = problem_desc_vec.size();
 
             if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
@@ -611,6 +476,11 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
                 throw std::runtime_error("wrong! group_count_ != a/b/b1/c_vec.size");
             }
 
+            if(!(p_acc0_biases_vec.size() == p_acc1_biases_vec.size()))
+            {
+                throw std::runtime_error("wrong! acc0_bias_vec.size != acc1_bias_vec.size");
+            }
+
             grid_size_ = 0;
 
             for(std::size_t i = 0; i < group_count_; i++)
@@ -620,14 +490,25 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
                 const auto p_b1_grid = static_cast<const B1DataType*>(p_b1_vec[i]);
                 const auto p_c_grid  = static_cast<CDataType*>(p_c_vec[i]);
 
-                const auto a_grid_desc_ak0_m_ak1 = DeviceOp::MakeAGridDescriptor_AK0_M_AK1(
-                    problem_desc_vec[i].M, problem_desc_vec[i].K, problem_desc_vec[i].StrideA);
-                const auto b_grid_desc_bk0_n_bk1 = DeviceOp::MakeBGridDescriptor_BK0_N_BK1(
-                    problem_desc_vec[i].K, problem_desc_vec[i].N, problem_desc_vec[i].StrideB0);
-                const auto b1_grid_desc_bk0_n_bk1 = DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(
-                    problem_desc_vec[i].N, problem_desc_vec[i].O, problem_desc_vec[i].StrideB1);
-                const auto c_grid_desc_m_n = DeviceOp::MakeCGridDescriptor_M_N(
-                    problem_desc_vec[i].c_gs_ms_os_lengths, problem_desc_vec[i].c_gs_ms_os_strides);
+                const auto& problem_desc = problem_desc_vec[i];
+
+                const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_bk0_n_bk1 = MakeB1GridDescriptor_BK0_N_BK1(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_m_n = Transform::MakeCGridDescriptor_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
+
+                const auto a_grid_desc_g_m_k = Transform::MakeAGridDescriptor_G_M_K(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
 
                 const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
                     GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
@@ -635,25 +516,32 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
                 const index_t BlockStart     = grid_size_;
                 const auto block_2_ctile_map = Block2CTileMap(c_grid_desc_m_n, BlockStart);
-                const index_t grid_size_grp = block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n) *
-                                              problem_desc_vec[i].Batch;
+                const index_t batch_count    = c_grid_desc_g_m_n.GetLength(I0);
+                const index_t grid_size_grp =
+                    block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n) * batch_count;
                 const index_t BlockEnd = grid_size_ + grid_size_grp;
 
                 // batch stride
-                // TODO ANT: only keep batch stride in tensor desc to reduce scalar cache pressure
-                const auto c_grid_desc_g_m_n = DeviceOp::MakeCGridDescriptor_G_M_N(
-                    problem_desc_vec[i].c_gs_ms_os_lengths, problem_desc_vec[i].c_gs_ms_os_strides);
-                const auto compute_base_ptr_of_batch =
-                    ComputeBasePtrOfStridedBatch(problem_desc_vec[i].BatchStrideA,
-                                                 problem_desc_vec[i].BatchStrideB0,
-                                                 problem_desc_vec[i].BatchStrideB1,
-                                                 c_grid_desc_g_m_n);
+                const auto compute_base_ptr_of_batch = ComputeBasePtrOfStridedBatch(
+                    a_grid_desc_g_m_k, b_grid_desc_g_n_k, b1_grid_desc_g_n_k, c_grid_desc_g_m_n);
 
                 // C0 mask
-                const auto c0_matrix_mask = C0MatrixMask(problem_desc_vec[i].N);
+                const auto c0_matrix_mask = C0MatrixMask(b_grid_desc_g_n_k.GetLength(I1));
 
                 grid_size_ += grid_size_grp;
 
+                // for each group, make sure acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias and
+                // so on
+                if(!(problem_desc.acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias &&
+                     problem_desc.acc0_biases_gs_ms_ns_strides.size() == NumAcc0Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_lengths.size() == NumAcc1Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_strides.size() == NumAcc1Bias))
+                {
+                    throw std::runtime_error(
+                        "wrong! number of biases in function argument does not "
+                        "match that in template argument");
+                }
+
                 group_kernel_args_.push_back({p_a_grid,
                                               p_b_grid,
                                               p_b1_grid,
@@ -669,13 +557,20 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
                                               BlockStart,
                                               BlockEnd});
 
-                group_device_args_.push_back({problem_desc_vec[i].M,
-                                              problem_desc_vec[i].N,
-                                              problem_desc_vec[i].K,
-                                              problem_desc_vec[i].O,
-                                              problem_desc_vec[i].c_gs_ms_os_lengths.back(),
-                                              problem_desc_vec[i].c_gs_ms_os_strides.back(),
-                                              c_grid_desc_m_n});
+                group_device_args_.push_back(
+                    {{problem_desc.a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                      problem_desc.b1_gs_os_ns_lengths[NumDimG + NumDimO - 1]},
+                     {problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                      problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+                     {problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+                     {problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO - 1],
+                      problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO + NumDimN - 1]},
+                     {problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM - 1],
+                      problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM + NumDimO - 1]},
+                     c_grid_desc_m_n});
             }
         }
 
@@ -788,6 +683,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
             return false;
         }
 
+        // TODO ANT: Check if tensor specialization & strides mismatch
+
         bool all_has_main_k_block_loop  = true;
         bool some_has_main_k_block_loop = false;
 
@@ -815,19 +712,16 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
             // Note: we need raw lengths since threadwise copy can not handle vector load when
             // part of vector is out of bounds
-            const auto MRaw      = device_arg.M;
-            const auto NRaw      = device_arg.N;
-            const auto KRaw      = device_arg.K;
-            const auto Gemm1NRaw = device_arg.O;
+            const auto MzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+            const auto NzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+            const auto KzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+            const auto Gemm1NzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
 
             // Check scalar per vector requirement
-            const auto a_extent_lowest =
-                is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
-            const auto b_extent_lowest =
-                is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
-            const auto b1_extent_lowest =
-                is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
-            const auto c_extent_lowest = device_arg.c_extent_lowest_;
+            const auto a_extent_lowest  = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
+            const auto b_extent_lowest  = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
+            const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
+            const auto c_extent_lowest  = Gemm1NzRaw;
 
             if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
                  b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
@@ -837,8 +731,22 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
                 return false;
             }
 
-            // Check vector store requirement; assumes last dimension in N to be contiguous
-            if(device_arg.c_stride_lowest_ != 1)
+            // Check vector load/store requirement
+            const auto a_stride_lowest = ABlockTransferSrcVectorDim == 2
+                                             ? device_arg.a_mz_kz_strides_[1]
+                                             : device_arg.a_mz_kz_strides_[0];
+            const auto b_stride_lowest = BBlockTransferSrcVectorDim == 2
+                                             ? device_arg.b_nz_kz_strides_[1]
+                                             : device_arg.b_nz_kz_strides_[0];
+            const auto b1_stride_lowest = B1BlockTransferSrcVectorDim == 2
+                                              ? device_arg.b1_nz_kz_strides_[1]
+                                              : device_arg.b1_nz_kz_strides_[0];
+            const auto c_stride_lowest =
+                device_arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be
+                                                     // contiguous
+
+            if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
+                 c_stride_lowest == 1))
             {
                 return false;
             }
@@ -873,6 +781,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
                              std::vector<const void*> p_b_vec,
                              std::vector<const void*> p_b1_vec,
                              std::vector<void*> p_c_vec,
+                             std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                             std::vector<std::vector<const void*>> p_acc1_biases_vec,
                              std::vector<ProblemDesc> problem_desc_vec,
                              AElementwiseOperation a_element_op,
                              BElementwiseOperation b_element_op,
@@ -884,6 +794,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
                         p_b_vec,
                         p_b1_vec,
                         p_c_vec,
+                        p_acc0_biases_vec,
+                        p_acc1_biases_vec,
                         problem_desc_vec,
                         a_element_op,
                         b_element_op,
@@ -895,21 +807,26 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
     static auto MakeInvoker() { return Invoker{}; }
 
     // polymorphic
-    std::unique_ptr<BaseArgument> MakeArgumentPointer(std::vector<const void*> p_a_vec,
-                                                      std::vector<const void*> p_b_vec,
-                                                      std::vector<const void*> p_b1_vec,
-                                                      std::vector<void*> p_c_vec,
-                                                      std::vector<ProblemDesc> problem_desc_vec,
-                                                      AElementwiseOperation a_element_op,
-                                                      BElementwiseOperation b_element_op,
-                                                      AccElementwiseOperation acc_element_op,
-                                                      B1ElementwiseOperation b1_element_op,
-                                                      CElementwiseOperation c_element_op) override
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<const void*> p_b1_vec,
+                        std::vector<void*> p_c_vec,
+                        std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                        std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                        std::vector<ProblemDesc> problem_desc_vec,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        AccElementwiseOperation acc_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) override
     {
         return std::make_unique<Argument>(p_a_vec,
                                           p_b_vec,
                                           p_b1_vec,
                                           p_c_vec,
+                                          p_acc0_biases_vec,
+                                          p_acc1_biases_vec,
                                           problem_desc_vec,
                                           a_element_op,
                                           b_element_op,
@@ -942,7 +859,12 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
             << Gemm1NPerBlock << ", "
             << Gemm1KPerBlock << ", "
             << B1K1 << ", "
-            << getGemmSpecializationString(GemmSpec) << ">";
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << "ASpec" << getTensorSpecializationString(ASpec) << ", "
+            << "B0Spec" << getTensorSpecializationString(BSpec) << ", "
+            << "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
+            << "CSpec" << getTensorSpecializationString(CSpec) << ", "
+            << getMaskingSpecializationString(MaskingSpec) << ">";
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp

@@ -130,8 +130,11 @@ namespace device {
 //   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
 //   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
 
-// FIXME: TensorSpecialization::Packed specialization does not cover all packed tensor cases, it
-// merely degenerates into TensorSpecialization::Default with NumDimG/M/N/K = 1
+// NOTE: TensorSpecialization::Packed specialized tensor is "packed" in a sense that each inner
+// dimension in a dimension group (eg [G0, G1] in Gs, [M0, M1, M2] in Ms, etc.) are contiguous and
+// ordered. Not in a sense that the tensor [G0, G1, ..., M0, M1, ..., N0, N1...] can be permuted
+// while still being a contiguous, unpadded tensor. In other words, it merely degenerates into
+// TensorSpecialization::Default with NumDimG/M/N/K = 1
 //
 // Detail- Packed tensor satisfies
 //   stride_0 = 1
@@ -147,7 +150,7 @@ namespace device {
 // essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
 //
 // Might need to expose dimension order to the interface to fully support
-// TensorSpecialization::Packed.
+// TensorSpecialization::Packed in a traditional sense of "packed" tensor
 template <index_t NumDimG,
           index_t NumDimM,
           index_t NumDimN,

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp

@@ -14,6 +14,7 @@
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 
@@ -116,14 +117,17 @@ __global__ void
 // Computes C = A * B0 * B1
 //              ^^^^^^ (Acc0)
 //              ^^^^^^^^^^^ (Acc1)
-template <typename ALayout,
-          typename BLayout, // B0Layout
-          typename B1Layout,
-          typename CPermuteNumDims_G_M_Gemm1N, // Sequence<NumDimG, NumDimM, NumDimGemm1N>
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO, // NumDimGemm1N
           typename ADataType,
           typename BDataType,
           typename B1DataType,
           typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
           typename GemmAccDataType,
           typename CShuffleDataType,
           typename AElementwiseOperation,
@@ -132,6 +136,10 @@ template <typename ALayout,
           typename B1ElementwiseOperation,
           typename CElementwiseOperation,
           GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization B1Spec,
+          TensorSpecialization CSpec,
           index_t NumGemmKPrefetchStage,
           index_t BlockSize,
           index_t MPerBlock,
@@ -172,283 +180,135 @@ template <typename ALayout,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
-          bool MaskOutUpperTriangle,
+          MaskingSpecialization MaskingSpec,
           LoopScheduler LoopSched = LoopScheduler::Default>
 struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
-    : public DeviceBatchedGemmSoftmaxGemmPermute<ALayout,
-                                                 BLayout,
-                                                 B1Layout,
-                                                 CPermuteNumDims_G_M_Gemm1N,
+    : public DeviceBatchedGemmSoftmaxGemmPermute<NumDimG,
+                                                 NumDimM,
+                                                 NumDimN,
+                                                 NumDimK,
+                                                 NumDimO,
                                                  ADataType,
                                                  BDataType,
                                                  B1DataType,
                                                  CDataType,
+                                                 Acc0BiasDataType,
+                                                 Acc1BiasDataType,
                                                  AElementwiseOperation,
                                                  BElementwiseOperation,
                                                  AccElementwiseOperation,
                                                  B1ElementwiseOperation,
-                                                 CElementwiseOperation>
+                                                 CElementwiseOperation,
+                                                 MaskingSpec>
 {
+    static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
+                  "Number of dimension must be greater than 0");
+
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    // TODO ANT: implement bias combination
+    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+
+#if 0
+    // TODO ANT: use alias
+    static constexpr index_t NumDimGemm0M = NumDimM;
+    static constexpr index_t NumDimGemm0N = NumDimN;
+    static constexpr index_t NumDimGemm0K = NumDimK;
+    static constexpr index_t NumDimGemm1M = NumDimM;
+    static constexpr index_t NumDimGemm1N = NumDimO;
+    static constexpr index_t NumDimGemm1K = NumDimN;
+#endif
+
     using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle;
 
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
 
-    static constexpr auto matrix_padder =
-        GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
-            MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
-
-    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                              const std::vector<index_t>& a_gs_ms_ks_strides_vec)
     {
-        const auto a_grid_desc_mraw_kraw = [&]() {
-            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
-            {
-                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
-                                                    make_tuple(StrideA, I1));
-            }
-            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
-            {
-                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
-                                                    make_tuple(I1, StrideA));
-            }
-        }();
-
-        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
-
-        const auto M = a_grid_desc_m_k.GetLength(I0);
-        const auto K = a_grid_desc_m_k.GetLength(I1);
-
-        const auto AK0 = K / AK1;
-
-        return transform_tensor_descriptor(a_grid_desc_m_k,
-                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
-                                                      make_pass_through_transform(M)),
-                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-    }
-
-    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
-    {
-        const auto b_grid_desc_nraw_kraw = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(I1, StrideB));
-            }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(StrideB, I1));
-            }
-        }();
-
-        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
-
-        const auto N = b_grid_desc_n_k.GetLength(I0);
-        const auto K = b_grid_desc_n_k.GetLength(I1);
-
-        const auto BK0 = K / BK1;
-
-        return transform_tensor_descriptor(b_grid_desc_n_k,
-                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                                      make_pass_through_transform(N)),
-                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-    }
-
-    // Args: Gemm1KRaw, Gemm1NRaw, StrideB1
-    static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
-    {
-        const auto b1_grid_desc_nraw_kraw = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(I1, StrideB));
-            }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
-            {
-                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
-                                                    make_tuple(StrideB, I1));
-            }
-        }();
-
-        const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
-
-        const auto N = b1_grid_desc_n_k.GetLength(I0);
-        const auto K = b1_grid_desc_n_k.GetLength(I1);
-
-        const auto B1K0 = K / B1K1;
-
-        return transform_tensor_descriptor(
-            b1_grid_desc_n_k,
-            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
-                       make_pass_through_transform(N)),
-            make_tuple(Sequence<1>{}, Sequence<0>{}),
-            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
+            Number<AK1>{});
     }
 
-    // assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
-    static auto MakeCGridDescriptor_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
-                                        const std::vector<index_t>& c_gs_ms_ns_strides_vec)
+    static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                              const std::vector<index_t>& b_gs_ns_ks_strides_vec)
     {
-        constexpr index_t NumDimG = CPermuteNumDims_G_M_Gemm1N::At(I0);
-        constexpr index_t NumDimM = CPermuteNumDims_G_M_Gemm1N::At(I1);
-        constexpr index_t NumDimN = CPermuteNumDims_G_M_Gemm1N::At(I2); // NumDimGemm1N
-
-        assert(c_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
-               c_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
-
-        const auto to_tuple = [&](auto& vec, auto start, auto end) {
-            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
-        };
-
-        const auto c_ms_ns_lengths = to_tuple(
-            c_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
-        const auto c_ms_ns_strides = to_tuple(
-            c_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
-
-        // dimension Ids for M0, M1, ...
-        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
-
-        // dimension Ids for N0, N1, ...
-        constexpr auto nDimIds =
-            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
-
-        // lengths for M0, M1, ...
-        const auto mLengths = get_container_subset(c_ms_ns_lengths, mDimIds);
-
-        // lengths for K0, K1, ...
-        const auto nLengths = get_container_subset(c_ms_ns_lengths, nDimIds);
-
-        // naive tensor C[M0, M1, M2, ..., N0, N1, N2...]
-        const auto c_grid_desc_ms_ns =
-            make_naive_tensor_descriptor(c_ms_ns_lengths, c_ms_ns_strides);
-
-        // transformed tensor C[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
-        const auto c_grid_desc_mraw_nraw = transform_tensor_descriptor(
-            c_grid_desc_ms_ns,
-            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
-            make_tuple(mDimIds, nDimIds),
-            make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(
+            Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
+            Number<BK1>{});
     }
 
-    // assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
-    static auto MakeCGridDescriptor_G_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
-                                          const std::vector<index_t>& c_gs_ms_ns_strides_vec)
+    static auto
+    MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                   const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
     {
-        constexpr index_t NumDimG = CPermuteNumDims_G_M_Gemm1N::At(I0);
-        constexpr index_t NumDimM = CPermuteNumDims_G_M_Gemm1N::At(I1);
-        constexpr index_t NumDimN = CPermuteNumDims_G_M_Gemm1N::At(I2); // NumDimGemm1N
-
-        assert(c_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
-               c_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
-
-        const auto to_tuple = [&](auto& vec, auto start, auto end) {
-            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
-        };
-
-        const auto c_gs_ms_ns_lengths =
-            to_tuple(c_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
-        const auto c_gs_ms_ns_strides =
-            to_tuple(c_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
-
-        // dimension Ids for G0, G1, ...
-        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
-
-        // dimension Ids for M0, M1, ...
-        constexpr auto mDimIds =
-            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
-
-        // dimension Ids for N0, N1, ...
-        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
-                                                                  NumDimG + NumDimM + NumDimN,
-                                                                  1>::type{};
-
-        // lengths for G0, G1, ...
-        const auto gLengths = get_container_subset(c_gs_ms_ns_lengths, gDimIds);
-
-        // lengths for M0, M1, ...
-        const auto mLengths = get_container_subset(c_gs_ms_ns_lengths, mDimIds);
-
-        // lengths for K0, K1, ...
-        const auto nLengths = get_container_subset(c_gs_ms_ns_lengths, nDimIds);
-
-        // naive tensor C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
-        const auto c_grid_desc_gs_ms_ns =
-            make_naive_tensor_descriptor(c_gs_ms_ns_lengths, c_gs_ms_ns_strides);
-
-        // transformed tensor C[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
-        // N2 * ...]
-        const auto c_grid_desc_g_mraw_nraw =
-            transform_tensor_descriptor(c_grid_desc_gs_ms_ns,
-                                        make_tuple(make_merge_transform(gLengths),
-                                                   make_merge_transform(mLengths),
-                                                   make_merge_transform(nLengths)),
-                                        make_tuple(gDimIds, mDimIds, nDimIds),
-                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-
-        // this desc is only for calculating batch offset so no padding needed
-        return c_grid_desc_g_mraw_nraw;
+        return Transform::MakeB1GridDescriptor_BK0_N_BK1(
+            Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                                b1_gs_gemm1ns_gemm1ks_strides_vec),
+            Number<B1K1>{});
     }
 
-    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
-    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
-    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
-    using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N({}, {}));
-    using CGridDesc_G_M_N      = decltype(MakeCGridDescriptor_G_M_N({}, {}));
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
+    using CGridDesc_M_N        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using AGridDesc_G_M_K      = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
+    using BGridDesc_G_N_K      = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
+    using B1GridDesc_G_N_K     = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
+    using CGridDesc_G_M_N      = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
 
-    // to track the points which need to be set to -inf on C0
-    // Note: no need to reset M padding value, because they will not be stored out.
-    struct C0MatrixMask
+    constexpr static auto make_MaskOutPredicate()
     {
-        C0MatrixMask(index_t NRaw) : NRaw_(NRaw) {}
-
-        __host__ __device__ bool IsUpperTriangle(index_t m, index_t n) const { return n > m; }
-
-        __host__ __device__ bool IsNOutOfBound(/*index_t m, */ index_t n) const
+        if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
         {
-            return n >= NRaw_;
+            return MaskDisabledPredicate{};
         }
-
-        __host__ __device__ bool IsMaskedElement(index_t m, index_t n) const
+        else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
         {
-            return IsUpperTriangle(m, n) || IsNOutOfBound(n);
+            return MaskOutUpperTrianglePredicate{};
         }
-
-        private:
-        // index_t MRaw_;
-        index_t NRaw_;
-    };
+    }
+    using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
 
     struct ComputeBasePtrOfStridedBatch
     {
-        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
-                                     index_t BatchStrideB,
-                                     index_t BatchStrideB1,
-                                     CGridDesc_G_M_N c_grid_desc_g_m_n)
-            : BatchStrideA_(BatchStrideA),
-              BatchStrideB_(BatchStrideB),
-              BatchStrideB1_(BatchStrideB1),
+        ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
+                                     const BGridDesc_G_N_K& b_grid_desc_g_n_k,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+            : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
+              b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
               c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
         {
         }
 
         __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
         {
-            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+            return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
         }
 
         __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
         {
-            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+            return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
         }
 
         __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
         {
-            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+            return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
         }
 
         __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
@@ -457,9 +317,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         }
 
         private:
-        index_t BatchStrideA_;
-        index_t BatchStrideB_;
-        index_t BatchStrideB1_;
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
         CGridDesc_G_M_N c_grid_desc_g_m_n_;
     };
 
@@ -523,47 +383,59 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CShuffleBlockTransferScalarPerVector_NPerBlock,
         LoopSched,
-        matrix_padder.PadN,
-        MaskOutUpperTriangle>;
+        Transform::matrix_padder.PadN,
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
 
     // Argument
     // FIXME: constness
     struct Argument : public BaseArgument
     {
-        Argument(const ADataType* p_a_grid,
-                 const BDataType* p_b_grid,
-                 const B1DataType* p_b1_grid,
-                 CDataType* p_c_grid,
-                 index_t MRaw,
-                 index_t NRaw,
-                 index_t KRaw,
-                 index_t Gemm1NRaw, // = ORaw
-                 index_t Batch,
-                 std::vector<index_t> c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
-                 std::vector<index_t> c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
-                 index_t StrideA,
-                 index_t StrideB,
-                 index_t StrideB1,
-                 index_t BatchStrideA,
-                 index_t BatchStrideB,
-                 index_t BatchStrideB1,
-                 AElementwiseOperation a_element_op,
-                 BElementwiseOperation b_element_op,
-                 AccElementwiseOperation acc_element_op,
-                 B1ElementwiseOperation b1_element_op,
-                 CElementwiseOperation c_element_op)
+        Argument(
+            const ADataType* p_a_grid,
+            const BDataType* p_b_grid,
+            const B1DataType* p_b1_grid,
+            CDataType* p_c_grid,
+            const std::array<void*, NumAcc0Bias> p_acc0_biases,
+            const std::array<void*, NumAcc1Bias> p_acc1_biases,
+            const std::vector<index_t>& a_gs_ms_ks_lengths,
+            const std::vector<index_t>& a_gs_ms_ks_strides,
+            const std::vector<index_t>& b_gs_ns_ks_lengths,
+            const std::vector<index_t>& b_gs_ns_ks_strides,
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+            const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+            const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+                acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+                acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+            AElementwiseOperation a_element_op,
+            BElementwiseOperation b_element_op,
+            AccElementwiseOperation acc_element_op,
+            B1ElementwiseOperation b1_element_op,
+            CElementwiseOperation c_element_op)
             : p_a_grid_{p_a_grid},
               p_b_grid_{p_b_grid},
               p_b1_grid_{p_b1_grid},
               p_c_grid_{p_c_grid},
-              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
-              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
-              b1_grid_desc_bk0_n_bk1_{
-                  DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(NRaw, Gemm1NRaw, StrideB1)},
-              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
-                                                                 c_gs_ms_gemm1ns_strides)},
-              c_grid_desc_g_m_n_{DeviceOp::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
-                                                                     c_gs_ms_gemm1ns_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              b1_grid_desc_bk0_n_bk1_{DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              c_grid_desc_m_n_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                  c_gs_ms_gemm1ns_strides)},
+              a_grid_desc_g_m_k_{
+                  Transform::MakeAGridDescriptor_G_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_g_n_k_{
+                  Transform::MakeB0GridDescriptor_G_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              b1_grid_desc_g_n_k_{Transform::MakeB1GridDescriptor_G_N_K(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              c_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                      c_gs_ms_gemm1ns_strides)},
               c_grid_desc_mblock_mperblock_nblock_nperblock_{},
               block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
               a_element_op_{a_element_op},
@@ -571,14 +443,31 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
               acc_element_op_{acc_element_op},
               b1_element_op_{b1_element_op},
               c_element_op_{c_element_op},
-              batch_count_(Batch),
+              c0_matrix_mask_{b_grid_desc_g_n_k_.GetLength(I1)},
+              raw_lengths_mz_nz_kz_gemm1nz_{a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                                            b1_gs_gemm1ns_gemm1ks_lengths[NumDimG + NumDimO - 1]},
+              a_mz_kz_strides_{a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                               a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+              b_nz_kz_strides_{b_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                               b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+              b1_nz_kz_strides_{b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO - 1],
+                                b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
+              c_mz_gemm1nz_strides_{c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
+                                    c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
+              batch_count_{c_grid_desc_g_m_n_.GetLength(I0)},
               compute_base_ptr_of_batch_{
-                  BatchStrideA, BatchStrideB, BatchStrideB1, c_grid_desc_g_m_n_},
-              c0_matrix_mask_{NRaw},
-              raw_lengths_m_n_k_o_{MRaw, NRaw, KRaw, Gemm1NRaw},
-              c_extent_lowest_{c_gs_ms_gemm1ns_lengths.back()},
-              c_stride_lowest_{c_gs_ms_gemm1ns_strides.back()}
+                  a_grid_desc_g_m_k_, b_grid_desc_g_n_k_, b1_grid_desc_g_n_k_, c_grid_desc_g_m_n_}
         {
+            // TODO ANT: implement bias addition
+            ignore = p_acc0_biases;
+            ignore = p_acc1_biases;
+            ignore = acc0_biases_gs_ms_ns_lengths;
+            ignore = acc0_biases_gs_ms_ns_strides;
+            ignore = acc1_biases_gs_ms_gemm1ns_lengths;
+            ignore = acc1_biases_gs_ms_gemm1ns_strides;
+
             if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
                                            b_grid_desc_bk0_n_bk1_,
                                            b1_grid_desc_bk0_n_bk1_,
@@ -591,34 +480,66 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             }
         }
 
-        //  private:
+        void Print() const
+        {
+            std::cout << "a_grid_desc_g_m_k_: " << a_grid_desc_g_m_k_.GetLength(I0) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I1) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I2) << '\n';
+            // a_grid_desc_g_m_k_.Print();
+            std::cout << "b_grid_desc_g_n_k_: " << b_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            // b_grid_desc_g_n_k_.Print();
+            std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            // b1_grid_desc_g_n_k_.Print();
+            std::cout << "c_grid_desc_g_m_n_: " << c_grid_desc_g_m_n_.GetLength(I0) << ", "
+                      << c_grid_desc_g_m_n_.GetLength(I1) << ", "
+                      << c_grid_desc_g_m_n_.GetLength(I2) << '\n';
+            // c_grid_desc_g_m_n_.Print();
+        }
+
+        // pointers
         const ADataType* p_a_grid_;
         const BDataType* p_b_grid_;
         const B1DataType* p_b1_grid_;
         CDataType* p_c_grid_;
+
+        // tensor descriptor
         AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
         BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
         B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
         CGridDesc_M_N c_grid_desc_m_n_;
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
         CGridDesc_G_M_N c_grid_desc_g_m_n_;
         typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
             c_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-c-tile map
         typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // element-wise op
         AElementwiseOperation a_element_op_;
         BElementwiseOperation b_element_op_;
         AccElementwiseOperation acc_element_op_;
         B1ElementwiseOperation b1_element_op_;
         CElementwiseOperation c_element_op_;
-        index_t batch_count_;
-        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
 
         // check C0 masking and padding
         C0MatrixMask c0_matrix_mask_;
 
         // For robust IsSupportedArgument() check
-        std::vector<index_t> raw_lengths_m_n_k_o_;
-        index_t c_extent_lowest_;
-        index_t c_stride_lowest_;
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
+        std::vector<index_t> a_mz_kz_strides_;
+        std::vector<index_t> b_nz_kz_strides_;
+        std::vector<index_t> b1_nz_kz_strides_;
+        std::vector<index_t> c_mz_gemm1nz_strides_;
+
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
     };
 
     // Invoker
@@ -628,13 +549,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
-                                            arg.b_grid_desc_bk0_n_bk1_,
-                                            arg.b1_grid_desc_bk0_n_bk1_,
-                                            arg.c_grid_desc_m_n_,
-                                            arg.block_2_ctile_map_))
+            if(!DeviceOp::IsSupportedArgument(arg))
             {
-                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+                throw std::runtime_error("wrong! unsupported argument");
             }
 
             const index_t grid_size =
@@ -719,17 +636,24 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+#if 0
+        arg.Print();
+#endif
+
         if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
         {
             return false;
         }
 
+        // TODO ANT: Check if tensor specialization & strides mismatch
+
         // Check if C permute dimension matches GEMM + GEMM shape
         const index_t c_g       = arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
         const index_t c_m       = arg.c_grid_desc_m_n_.GetLength(I0);
         const index_t c_gemm1n  = arg.c_grid_desc_m_n_.GetLength(I1);
         const index_t a_m       = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
         const index_t b1_gemm1n = arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
+
         if(!(c_g == arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n))
         {
             return false;
@@ -737,19 +661,17 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
         // Note: we need raw lengths since threadwise copy can not handle vector load when part of
         // vector is out of bounds
-        const auto MRaw      = arg.raw_lengths_m_n_k_o_[0];
-        const auto NRaw      = arg.raw_lengths_m_n_k_o_[1];
-        const auto KRaw      = arg.raw_lengths_m_n_k_o_[2];
-        const auto Gemm1NRaw = arg.raw_lengths_m_n_k_o_[3];
+        // Note: need lowest dim in Ms/Ns/Ks/Os, not merged M/N/K/O
+        const auto MzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+        const auto NzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+        const auto KzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+        const auto Gemm1NzRaw = arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
 
         // Check scalar per vector requirement
-        const auto a_extent_lowest =
-            is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
-        const auto b_extent_lowest =
-            is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
-        const auto b1_extent_lowest =
-            is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
-        const auto c_extent_lowest = arg.c_extent_lowest_;
+        const auto a_extent_lowest  = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
+        const auto b_extent_lowest  = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
+        const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
+        const auto c_extent_lowest  = Gemm1NzRaw;
 
         if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
              b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
@@ -759,8 +681,18 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             return false;
         }
 
-        // Check vector store requirement; assumes last dimension in N to be contiguous
-        if(arg.c_stride_lowest_ != 1)
+        // Check vector load/store requirement
+        const auto a_stride_lowest =
+            ABlockTransferSrcVectorDim == 2 ? arg.a_mz_kz_strides_[1] : arg.a_mz_kz_strides_[0];
+        const auto b_stride_lowest =
+            BBlockTransferSrcVectorDim == 2 ? arg.b_nz_kz_strides_[1] : arg.b_nz_kz_strides_[0];
+        const auto b1_stride_lowest =
+            B1BlockTransferSrcVectorDim == 2 ? arg.b1_nz_kz_strides_[1] : arg.b1_nz_kz_strides_[0];
+        const auto c_stride_lowest =
+            arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be contiguous
+
+        if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
+             c_stride_lowest == 1))
         {
             return false;
         }
@@ -778,46 +710,51 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
     }
 
-    static auto MakeArgument(const ADataType* p_a,
-                             const BDataType* p_b,
-                             const B1DataType* p_b1,
-                             CDataType* p_c,
-                             index_t MRaw,
-                             index_t NRaw,
-                             index_t KRaw,
-                             index_t Gemm1NRaw,
-                             index_t Batch,
-                             std::vector<index_t> c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
-                             std::vector<index_t> c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
-                             index_t StrideA,
-                             index_t StrideB,
-                             index_t StrideB1,
-                             index_t BatchStrideA,
-                             index_t BatchStrideB,
-                             index_t BatchStrideB1,
-                             AElementwiseOperation a_element_op,
-                             BElementwiseOperation b_element_op,
-                             AccElementwiseOperation acc_element_op,
-                             B1ElementwiseOperation b1_element_op,
-                             CElementwiseOperation c_element_op)
+    static auto MakeArgument(
+        const ADataType* p_a,
+        const BDataType* p_b,
+        const B1DataType* p_b1,
+        CDataType* p_c,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        AccElementwiseOperation acc_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op)
     {
         return Argument{p_a,
                         p_b,
                         p_b1,
                         p_c,
-                        MRaw,
-                        NRaw,
-                        KRaw,
-                        Gemm1NRaw,
-                        Batch,
-                        c_gs_ms_gemm1ns_lengths,
-                        c_gs_ms_gemm1ns_strides,
-                        StrideA,
-                        StrideB,
-                        StrideB1,
-                        BatchStrideA,
-                        BatchStrideB,
-                        BatchStrideB1,
+                        p_acc0_biases,
+                        p_acc1_biases,
+                        a_gs_ms_ks_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                        b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                        c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                        c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                        acc0_biases_gs_ms_ns_lengths,
+                        acc0_biases_gs_ms_ns_strides,
+                        acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+                        acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
                         a_element_op,
                         b_element_op,
                         acc_element_op,
@@ -829,47 +766,51 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
     // polymorphic
     // FIXME: constness
-    std::unique_ptr<BaseArgument>
-    MakeArgumentPointer(const void* p_a,
-                        const void* p_b,
-                        const void* p_b1,
-                        void* p_c,
-                        index_t MRaw,
-                        index_t NRaw,
-                        index_t KRaw,
-                        index_t Gemm1NRaw,
-                        index_t Batch,
-                        std::vector<index_t> c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
-                        std::vector<index_t> c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
-                        index_t StrideA,
-                        index_t StrideB,
-                        index_t StrideB1,
-                        index_t BatchStrideA,
-                        index_t BatchStrideB,
-                        index_t BatchStrideB1,
-                        AElementwiseOperation a_element_op,
-                        BElementwiseOperation b_element_op,
-                        AccElementwiseOperation acc_element_op,
-                        B1ElementwiseOperation b1_element_op,
-                        CElementwiseOperation c_element_op) override
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const void* p_b1,
+        void* p_c,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        AccElementwiseOperation acc_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op) override
     {
         return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                           static_cast<const BDataType*>(p_b),
                                           static_cast<const B1DataType*>(p_b1),
                                           static_cast<CDataType*>(p_c),
-                                          MRaw,
-                                          NRaw,
-                                          KRaw,
-                                          Gemm1NRaw,
-                                          Batch,
-                                          c_gs_ms_gemm1ns_lengths,
-                                          c_gs_ms_gemm1ns_strides,
-                                          StrideA,
-                                          StrideB,
-                                          StrideB1,
-                                          BatchStrideA,
-                                          BatchStrideB,
-                                          BatchStrideB1,
+                                          p_acc0_biases, // cast in struct Argument
+                                          p_acc1_biases, // cast in struct Argument
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                                          b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                                          c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                                          c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                                          acc0_biases_gs_ms_ns_lengths,
+                                          acc0_biases_gs_ms_ns_strides,
+                                          acc1_biases_gs_ms_gemm1ns_lengths,
+                                          acc1_biases_gs_ms_gemm1ns_strides,
                                           a_element_op,
                                           b_element_op,
                                           acc_element_op,
@@ -901,7 +842,12 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             << Gemm1NPerBlock << ", "
             << Gemm1KPerBlock << ", "
             << B1K1 << ", "
-            << getGemmSpecializationString(GemmSpec) << ">";
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << "ASpec" << getTensorSpecializationString(ASpec) << ", "
+            << "B0Spec" << getTensorSpecializationString(BSpec) << ", "
+            << "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
+            << "CSpec" << getTensorSpecializationString(CSpec) << ", "
+            << getMaskingSpecializationString(MaskingSpec) << ">";
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp

@@ -12,6 +12,7 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
 #include "ck/host_utility/device_prop.hpp"
@@ -196,7 +197,8 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
                                           BElementwiseOperation,
                                           AccElementwiseOperation,
                                           B1ElementwiseOperation,
-                                          CElementwiseOperation>
+                                          CElementwiseOperation,
+                                          MaskOutUpperTriangle>
 {
     using DeviceOp = DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle;
 
@@ -315,29 +317,6 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
         return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
     }
 
-    // to track the points which need to be set to -inf on C0
-    // Note: no need to reset M padding value, because they will not be stored out.
-    struct C0MatrixMask
-    {
-        C0MatrixMask(index_t NRaw) : NRaw_(NRaw) {}
-
-        __host__ __device__ bool IsUpperTriangle(index_t m, index_t n) const { return n > m; }
-
-        __host__ __device__ bool IsNOutOfBound(/*index_t m, */ index_t n) const
-        {
-            return n >= NRaw_;
-        }
-
-        __host__ __device__ bool IsMaskedElement(index_t m, index_t n) const
-        {
-            return IsUpperTriangle(m, n) || IsNOutOfBound(n);
-        }
-
-        private:
-        // index_t MRaw_;
-        index_t NRaw_;
-    };
-
     struct ComputeBasePtrOfStridedBatch
     {
         ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
@@ -383,6 +362,10 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
     using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
     using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
 
+    using C0MatrixMask = conditional_t<MaskOutUpperTriangle,
+                                       C0MatrixMask_impl<MaskOutUpperTrianglePredicate>,
+                                       C0MatrixMask_impl<MaskDisabledPredicate>>;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
         ADataType, // TODO: distinguish A/B datatype

include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/device/welford_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim,
+          bool UseMultiblockInK,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct DeviceBatchNormFwdImpl
+    : public DeviceBatchNormFwd<Rank, NumBatchNormReduceDim, YElementwiseOp>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeXY2dDescriptor(const std::array<index_t, Rank>& xyLengths,
+                                   const std::array<index_t, Rank>& xyStrides,
+                                   int blkGroupSize,
+                                   int numBlockTileIteration)
+    {
+        const auto tupleXYLengths =
+            generate_tuple([&](auto I) { return xyLengths[I]; }, Number<Rank>{});
+        const auto tupleXYStrides =
+            generate_tuple([&](auto I) { return xyStrides[I]; }, Number<Rank>{});
+
+        const auto raw_grid_desc = make_naive_tensor_descriptor(tupleXYLengths, tupleXYStrides);
+
+        const auto grid_desc_m_k = [&]() {
+            using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+            using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+            const auto reduceDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[NumInvariantDim + I]; },
+                               Number<NumBatchNormReduceDim>{});
+            const auto invariantDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[I]; }, Number<NumInvariantDim>{});
+
+            return transform_tensor_descriptor(raw_grid_desc,
+                                               make_tuple(make_merge_transform(invariantDimLengths),
+                                                          make_merge_transform(reduceDimLengths)),
+                                               make_tuple(InvariantDims{}, ReduceDims{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        const auto invariantLength = grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = grid_desc_m_k.GetLength(Number<1>{});
+
+        const int workSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad = workSizePerBlock * blkGroupSize - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto MakeMeanVarCountOutputMG2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto grid_desc_m_g =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, blkGroupSize));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_g_padded =
+            transform_tensor_descriptor(grid_desc_m_g,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_pass_through_transform(blkGroupSize)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_g_padded);
+    };
+
+    static auto MakeMeanVarCountInputMK2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto reduceLength = blkGroupSize;
+        const auto grid_desc_m_k =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, reduceLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad =
+            math::integer_least_multiple(reduceLength, KThreadClusterSize) - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto
+    MakeScaleBiasMeanVar1dDescriptor(const std::array<index_t, NumInvariantDim>& lengths,
+                                     const std::array<index_t, NumInvariantDim>& strides)
+    {
+        const auto tupleLengths =
+            generate_tuple([&](auto I) { return lengths[I]; }, Number<NumInvariantDim>{});
+        const auto tupleStrides =
+            generate_tuple([&](auto I) { return strides[I]; }, Number<NumInvariantDim>{});
+
+        auto raw_grid_desc = make_naive_tensor_descriptor(tupleLengths, tupleStrides);
+
+        auto grid_desc_m = transform_tensor_descriptor(
+            raw_grid_desc,
+            make_tuple(make_merge_transform(tupleLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded =
+            transform_tensor_descriptor(grid_desc_m,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (grid_desc_m_padded);
+    };
+
+    using XYGridDesc_M_K             = decltype(MakeXY2dDescriptor({1}, {1}, 1, 1));
+    using ScaleBiasMeanVarGridDesc_M = decltype(MakeScaleBiasMeanVar1dDescriptor({1}, {1}));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* p_scale,
+                 const BiasDataType* p_bias,
+                 const YElementwiseOp y_elementwise_op,
+                 double epsilon,
+                 YDataType* p_y,
+                 MeanVarDataType* resultSaveMean,
+                 MeanVarDataType* resultSaveInvVariance,
+                 double averageFactor,
+                 MeanVarDataType* resultRunningMean,
+                 MeanVarDataType* resultRunningVariance)
+            : bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_scale_(p_scale),
+              p_bias_(p_bias),
+              y_elementwise_op_(y_elementwise_op),
+              p_y_(p_y),
+              resultSaveMean_(resultSaveMean),
+              resultSaveInvVariance_(resultSaveInvVariance),
+              resultRunningMean_(resultRunningMean),
+              resultRunningVariance_(resultRunningVariance)
+        {
+            xyLengths_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xyLengths, reduceDims);
+            xStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xStrides, reduceDims);
+            yStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(yStrides, reduceDims);
+
+            std::tie(invariant_length_, reduce_length_) =
+                get_2d_lengths<Rank, NumBatchNormReduceDim>(xyLengths_);
+
+            epsilon_       = type_convert<AccDataType>(epsilon);
+            averageFactor_ = type_convert<AccDataType>(averageFactor);
+
+            updateMovingAverage_ =
+                (resultRunningMean != nullptr && resultRunningVariance != nullptr);
+            saveMeanInvVariance_ = (resultSaveMean != nullptr && resultSaveInvVariance_ != nullptr);
+
+            if(UseMultiblockInK)
+            {
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                           (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize_ = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                (K_BlockTileSize * iterations);
+
+                numBlockTileIteration_ = iterations;
+            }
+            else
+            {
+                blkGroupSize_          = 1;
+                numBlockTileIteration_ = (reduce_length_ + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize_ = (invariant_length_ + M_BlockTileSize - 1) / M_BlockTileSize * blkGroupSize_;
+
+            x_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+            y_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+            scale_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnScaleStrides_);
+            bias_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnBiasStrides_);
+            mean_var_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnMeanVarStrides_);
+        }
+
+        AccDataType epsilon_;
+        AccDataType averageFactor_;
+
+        bool updateMovingAverage_;
+        bool saveMeanInvVariance_;
+
+        std::array<index_t, Rank> xyLengths_;
+        std::array<index_t, Rank> xStrides_;
+        std::array<index_t, Rank> yStrides_;
+
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* p_scale_;
+        const BiasDataType* p_bias_;
+        const YElementwiseOp y_elementwise_op_;
+        YDataType* p_y_;
+
+        MeanVarDataType* resultSaveMean_;
+        MeanVarDataType* resultSaveInvVariance_;
+
+        MeanVarDataType* resultRunningMean_;
+        MeanVarDataType* resultRunningVariance_;
+
+        long_index_t invariant_length_;
+        long_index_t reduce_length_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        XYGridDesc_M_K x_grid_desc_m_k_;
+        XYGridDesc_M_K y_grid_desc_m_k_;
+        ScaleBiasMeanVarGridDesc_M scale_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M bias_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M mean_var_grid_desc_m_;
+
+        void* workspace_mean_;
+        void* workspace_variance_;
+        void* workspace_count_;
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+            // workspace for welford intermediate mean
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate variance
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate count
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(int32_t) + 64;
+        }
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+
+            // setup buffer used for intermediate welford mean
+            pArg_->workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
+
+            index_t mean_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            mean_space_sz = math::integer_least_multiple(mean_space_sz, 64);
+
+            // setup buffer used for intermediate welford varirance
+            pArg_->workspace_variance_ =
+                reinterpret_cast<char*>(pArg_->workspace_mean_) + mean_space_sz;
+
+            index_t variance_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            variance_space_sz = math::integer_least_multiple(variance_space_sz, 64);
+
+            // setup buffer used for intermediate welfor count
+            pArg_->workspace_count_ =
+                reinterpret_cast<char*>(pArg_->workspace_variance_) + variance_space_sz;
+        };
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float avg_time = 0;
+
+            if(UseMultiblockInK && arg.blkGroupSize_ > 1)
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForMultiblockWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.blkGroupSize_, arg.numBlockTileIteration_, arg.reduce_length_);
+
+                const auto mean_var_count_grid_desc_m_g =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountOutputMG2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                const auto mean_var_count_grid_desc_m_k =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountInputMK2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                using MeanVarCountGridDesc_M_G = decltype(mean_var_count_grid_desc_m_g);
+                using MeanVarCountGridDesc_M_K = decltype(mean_var_count_grid_desc_m_k);
+
+                using GridwiseMultiblockWelfordFirstHalf_ =
+                    GridwiseMultiblockWelfordFirstHalf<XDataType,
+                                                       AccDataType,
+                                                       MeanVarDataType,
+                                                       XYGridDesc_M_K,
+                                                       MeanVarCountGridDesc_M_G,
+                                                       GetReduceCountPerThreadFunctor,
+                                                       BlockSize,
+                                                       MThreadClusterSize,
+                                                       KThreadClusterSize,
+                                                       MThreadSliceSize,
+                                                       KThreadSliceSize,
+                                                       XSrcYDstVectorDim,
+                                                       XSrcVectorSize>;
+
+                using GridwiseWelfordSecondHalfBatchNormForwardFinal_ =
+                    GridwiseWelfordSecondHalfBatchNormForwardFinal<XDataType,
+                                                                   YDataType,
+                                                                   AccDataType,
+                                                                   ScaleDataType,
+                                                                   BiasDataType,
+                                                                   MeanVarDataType,
+                                                                   YElementwiseOp,
+                                                                   XYGridDesc_M_K,
+                                                                   MeanVarCountGridDesc_M_K,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   BlockSize,
+                                                                   MThreadClusterSize,
+                                                                   KThreadClusterSize,
+                                                                   MThreadSliceSize,
+                                                                   KThreadSliceSize,
+                                                                   XSrcYDstVectorDim,
+                                                                   XSrcVectorSize,
+                                                                   YDstVectorSize,
+                                                                   ScaleSrcVectorSize,
+                                                                   BiasSrcVectorSize,
+                                                                   MeanVarSrcDstVectorSize>;
+
+                index_t numMeanVarCountBlockTileIteration =
+                    (arg.blkGroupSize_ + KThreadClusterSize - 1) / KThreadClusterSize;
+
+                const auto kern_multiblock_welford_first_half =
+                    kernel_multiblock_welford_first_half<GridwiseMultiblockWelfordFirstHalf_,
+                                                         XDataType,
+                                                         MeanVarDataType,
+                                                         XYGridDesc_M_K,
+                                                         MeanVarCountGridDesc_M_G,
+                                                         GetReduceCountPerThreadFunctor>;
+
+                const auto kern_welford_second_half_batchnorm_forward_final =
+                    kernel_welford_second_half_batchnorm_forward_final<
+                        GridwiseWelfordSecondHalfBatchNormForwardFinal_,
+                        XDataType,
+                        YDataType,
+                        AccDataType,
+                        ScaleDataType,
+                        BiasDataType,
+                        MeanVarDataType,
+                        YElementwiseOp,
+                        XYGridDesc_M_K,
+                        MeanVarCountGridDesc_M_K,
+                        ScaleBiasMeanVarGridDesc_M,
+                        ScaleBiasMeanVarGridDesc_M>;
+
+                avg_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kern_multiblock_welford_first_half,
+                                           dim3(arg.gridSize_),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.x_grid_desc_m_k_,
+                                           mean_var_count_grid_desc_m_g,
+                                           get_reduce_count_per_thread,
+                                           arg.numBlockTileIteration_,
+                                           arg.p_x_,
+                                           static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                                           static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                                           static_cast<int32_t*>(arg.workspace_count_));
+
+                avg_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kern_welford_second_half_batchnorm_forward_final,
+                                           dim3(arg.gridSize_),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.x_grid_desc_m_k_,
+                                           arg.y_grid_desc_m_k_,
+                                           mean_var_count_grid_desc_m_k,
+                                           arg.scale_grid_desc_m_,
+                                           arg.bias_grid_desc_m_,
+                                           arg.mean_var_grid_desc_m_,
+                                           arg.blkGroupSize_,
+                                           arg.numBlockTileIteration_,
+                                           numMeanVarCountBlockTileIteration,
+                                           arg.epsilon_,
+                                           static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                                           static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                                           static_cast<int32_t*>(arg.workspace_count_),
+                                           arg.p_x_,
+                                           arg.p_scale_,
+                                           arg.p_bias_,
+                                           arg.y_elementwise_op_,
+                                           arg.p_y_,
+                                           arg.updateMovingAverage_,
+                                           arg.averageFactor_,
+                                           arg.resultRunningMean_,
+                                           arg.resultRunningVariance_,
+                                           arg.saveMeanInvVariance_,
+                                           arg.resultSaveMean_,
+                                           arg.resultSaveInvVariance_);
+            }
+            else
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForBlockwiseWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.numBlockTileIteration_, arg.reduce_length_);
+
+                using GridwiseBatchNormForwardWithBlockwiseWelford_ =
+                    GridwiseBatchNormForwardWithBlockwiseWelford<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 XYGridDesc_M_K,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 GetReduceCountPerThreadFunctor,
+                                                                 BlockSize,
+                                                                 MThreadClusterSize,
+                                                                 KThreadClusterSize,
+                                                                 MThreadSliceSize,
+                                                                 KThreadSliceSize,
+                                                                 XSrcYDstVectorDim,
+                                                                 XSrcVectorSize,
+                                                                 YDstVectorSize,
+                                                                 ScaleSrcVectorSize,
+                                                                 BiasSrcVectorSize,
+                                                                 MeanVarSrcDstVectorSize>;
+
+                const auto kern_batchnorm_fwd = kernel_batchnorm_forward_with_blockwise_welford<
+                    GridwiseBatchNormForwardWithBlockwiseWelford_,
+                    XDataType,
+                    YDataType,
+                    AccDataType,
+                    ScaleDataType,
+                    BiasDataType,
+                    MeanVarDataType,
+                    YElementwiseOp,
+                    XYGridDesc_M_K,
+                    ScaleBiasMeanVarGridDesc_M,
+                    ScaleBiasMeanVarGridDesc_M,
+                    GetReduceCountPerThreadFunctor>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kern_batchnorm_fwd,
+                                                   dim3(arg.gridSize_),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.x_grid_desc_m_k_,
+                                                   arg.y_grid_desc_m_k_,
+                                                   arg.scale_grid_desc_m_,
+                                                   arg.bias_grid_desc_m_,
+                                                   arg.mean_var_grid_desc_m_,
+                                                   get_reduce_count_per_thread,
+                                                   arg.numBlockTileIteration_,
+                                                   arg.epsilon_,
+                                                   arg.p_x_,
+                                                   arg.p_scale_,
+                                                   arg.p_bias_,
+                                                   arg.y_elementwise_op_,
+                                                   arg.p_y_,
+                                                   arg.updateMovingAverage_, // true or false
+                                                   arg.averageFactor_,
+                                                   arg.resultRunningMean_,
+                                                   arg.resultRunningVariance_,
+                                                   arg.saveMeanInvVariance_, // true or false
+                                                   arg.resultSaveMean_,
+                                                   arg.resultSaveInvVariance_);
+            };
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* pArg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(pArg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* pArg) override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        if constexpr(XSrcYDstVectorDim == 0)
+        {
+            if(pArg_->xStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->yStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[NumInvariantDim - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % YDstVectorSize != 0)
+                return false;
+        }
+        else
+        {
+            if(pArg_->xStrides_[Rank - 1] != 1 || pArg_->yStrides_[Rank - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[Rank - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % YDstVectorSize != 0)
+                return false;
+        };
+
+        if(pArg_->bnScaleStrides_[NumInvariantDim - 1] != 1 && ScaleSrcVectorSize != 1)
+            return false;
+        if(pArg_->bnBiasStrides_[NumInvariantDim - 1] != 1 && BiasSrcVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % ScaleSrcVectorSize != 0)
+            return false;
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % BiasSrcVectorSize != 0)
+            return false;
+
+        if(pArg_->bnMeanVarStrides_[NumInvariantDim - 1] != 1 && MeanVarSrcDstVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % MeanVarSrcDstVectorSize != 0)
+            return false;
+
+        bool is_valid = true;
+
+        static_for<0, NumInvariantDim, 1>{}([&](auto I) {
+            if(pArg_->xyLengths_[I] != pArg_->bnScaleBiasMeanVarLengths_[I])
+                is_valid = false;
+        });
+
+        if(!is_valid)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* p_scale,
+        const void* p_bias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        void* p_y,
+        void* resultSaveMean,
+        void* resultSaveInvVariance,
+        double averageFactor,
+        void* resultRunningMean,
+        void* resultRunningVariance) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const BiasDataType*>(p_bias),
+                                          y_elementwise_op,
+                                          epsilon,
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<MeanVarDataType*>(resultSaveMean),
+                                          static_cast<MeanVarDataType*>(resultSaveInvVariance),
+                                          averageFactor,
+                                          static_cast<MeanVarDataType*>(resultRunningMean),
+                                          static_cast<MeanVarDataType*>(resultRunningVariance));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchNormFwdImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XSrcYDstVectorDim_" << XSrcYDstVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_scale_" << ScaleSrcVectorSize << "_bias_" << BiasSrcVectorSize << "_mean_var_" << MeanVarSrcDstVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp

@@ -214,6 +214,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
                                      MPerBlock,
                                      NPerBlock,
                                      K0PerBlock,
+                                     K1,
                                      M1PerThread,
                                      N1PerThread,
                                      KPerThread,

include/ck/tensor_operation/gpu/device/masking_specialization.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct MaskingSpecialization
+{
+    MaskDisabled,
+    MaskOutUpperTriangle
+};
+
+inline std::string getMaskingSpecializationString(const MaskingSpecialization& s)
+{
+    switch(s)
+    {
+    case MaskingSpecialization::MaskDisabled: return "MaskDisabled";
+    case MaskingSpecialization::MaskOutUpperTriangle: return "MaskOutUpperTriangle";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+struct MaskDisabledPredicate
+{
+    __host__ __device__ constexpr bool operator()(index_t /*m*/, index_t /*n*/) const
+    {
+        return false;
+    };
+
+    __host__ __device__ constexpr bool
+        IsTileSkippable(index_t /*m*/, index_t /*n*/, index_t /*m_tile*/, index_t /*n_tile*/) const
+    {
+        return false;
+    }
+};
+
+struct MaskOutUpperTrianglePredicate
+{
+    __host__ __device__ constexpr bool operator()(index_t m, index_t n) const { return n > m; }
+
+    __host__ __device__ constexpr bool
+    IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t /*n_tile*/) const
+    {
+        return operator()(m + m_tile - 1, n);
+    }
+};
+
+// to track the points which need to be set to -inf on C0
+// Note: no need to reset M padding value, because they will not be stored out.
+template <typename MaskOutPredicate>
+struct C0MatrixMask_impl
+{
+    C0MatrixMask_impl(index_t NRaw) : NRaw_(NRaw), predicate_(MaskOutPredicate{}) {}
+
+    __host__ __device__ constexpr bool IsNOutOfBound(/*index_t m, */ index_t n) const
+    {
+        return n >= NRaw_;
+    }
+
+    __host__ __device__ constexpr bool IsMaskedElement(index_t m, index_t n) const
+    {
+        return predicate_(m, n) || IsNOutOfBound(n);
+    }
+
+    __host__ __device__ constexpr bool
+    IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t n_tile) const
+    {
+        return predicate_.IsTileSkippable(m, n, m_tile, n_tile);
+    }
+
+    private:
+    // index_t MRaw_;
+    index_t NRaw_;
+    MaskOutPredicate predicate_;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck