Merge branch 'develop' into gemm_layernorm_instance

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp

@@ -10,7 +10,7 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/device_memory.hpp"
@@ -95,7 +95,7 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
 using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
 
 // A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
-using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
+using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
     ck::Tuple<EDataType,
               R0DataType,
               R1DataType,

example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp

@@ -10,7 +10,7 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/device_memory.hpp"
@@ -92,7 +92,7 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
 using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
 
 // A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
-using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
+using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
     ck::Tuple<EDataType,
               R0DataType,
               R1DataType,

example/23_softmax/softmax_blockwise.cpp

@@ -56,8 +56,8 @@ class SimpleAppArgs
     int option_index = 0;
 
     public:
-    std::vector<size_t> inLengths   = {8, 128, 2048};
-    std::vector<AccDataType> scales = {2.0f, 2.0f};
+    std::vector<size_t> inLengths = {8, 128, 2048};
+    std::vector<double> scales    = {2.0, 2.0};
 
     bool do_verification = true;
     int init_method      = 2;
@@ -151,8 +151,8 @@ int main(int argc, char* argv[])
     auto inStrides  = in.mDesc.GetStrides();
     auto outStrides = out.mDesc.GetStrides();
 
-    AccDataType alpha = args.scales[0];
-    AccDataType beta  = args.scales[1];
+    double alpha = args.scales[0];
+    double beta  = args.scales[1];
 
     std::cout << "in: " << in.mDesc << std::endl;
     std::cout << "out: " << out.mDesc << std::endl;
@@ -221,8 +221,8 @@ int main(int argc, char* argv[])
     auto argument_ptr = device_instance.MakeArgumentPointer(i_inLengths,
                                                             i_inStrides,
                                                             reduceDims,
-                                                            &alpha,
-                                                            &beta,
+                                                            alpha,
+                                                            beta,
                                                             in_dev.GetDeviceBuffer(),
                                                             out_dev.GetDeviceBuffer(),
                                                             PassThrough{},

example/33_multiple_reduce/dual_reduce_common.hpp

@@ -217,8 +217,8 @@ int mean_meansquare_dual_reduce_test(size_t n,
     size_t invariant_total_length = n;
     size_t reduce_total_length    = h * w * c;
 
-    const AccDataType alpha = ck::type_convert<AccDataType>(1.0f);
-    const AccDataType beta  = ck::type_convert<AccDataType>(0.0f);
+    const double alpha = 1.0f;
+    const double beta  = 0.0f;
 
     std::size_t num_thread = 1;
 
@@ -267,8 +267,8 @@ int mean_meansquare_dual_reduce_test(size_t n,
         i_outLengths,
         {i_outStrides, i_outStrides},
         reduceDims,
-        {&alpha, &alpha},
-        {&beta, &beta},
+        {alpha, alpha},
+        {beta, beta},
         in_dev.GetDeviceBuffer(),
         {mean_dev.GetDeviceBuffer(), meansquare_dev.GetDeviceBuffer()},
         ck::make_tuple(InElementwiseOperation_Mean{}, InElementwiseOperation_Meansquare{}),

example/34_batchnorm/batchnorm_infer_impl.hpp

@@ -10,7 +10,7 @@
 #include "ck/utility/sequence.hpp"
 #include "ck/utility/tuple.hpp"
 #include "ck/utility/reduction_operator.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 
 #include "batchnorm_common.hpp"
 
@@ -46,7 +46,7 @@ int bnorm_infer(
     static_assert(NumBatchNormReduceDim < Rank,
                   "Invalid number of reduced dimensions for batchnorm!");
 
-    using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
+    using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
         ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
                                                                                   // variance,
                                                                                   // scale,

example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp

@@ -9,7 +9,8 @@
 #include <ctime>
 
 #include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_sparse_embedding3_forward_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -18,53 +19,26 @@
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp"
 
-// using EmbType       = float;
-// using IndexType     = int64_t;
-// using GammaDataType = float;
-// using BetaDataType  = float;
-// using AccDataType   = float;
-// using OutType       = float;
-
+// clang-format off
 using EmbType       = ck::half_t;
 using IndexType     = int64_t;
 using GammaDataType = ck::half_t;
 using BetaDataType  = ck::half_t;
 using AccDataType   = float;
 using OutType       = ck::half_t;
+using EmbElementwiseOperation = ck::tensor_operation::element_wise::AddAdd;
 
-// clang-format off
-//                                                                                                         BlockSize, DimClusterSize, RowClusterSize, DimPerBlock, RowPerBlock, DimThreadSize, RowVectorSize
-using DeviceInstance_fp32_e256   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  256,   1, 1>;
-using DeviceInstance_fp32_e512   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  512,   1, 1>;
-using DeviceInstance_fp32_e768   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  768,   1, 1>;
-using DeviceInstance_fp32_e1024  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1024,  1, 1>;
-using DeviceInstance_fp32_e1536  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1536,  1, 1>;
-using DeviceInstance_fp32_e2048  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  2048,  1, 4>;
-using DeviceInstance_fp32_e4096  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  4096,  1, 4>;
-using DeviceInstance_fp32_e8192  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  8192,  1, 4>;
-using DeviceInstance_fp32_e16384 = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  16384, 1, 4>;
-
-using DeviceInstance_fp16_e256   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  256,   1, 1>;
-using DeviceInstance_fp16_e512   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  512,   1, 2>;
-using DeviceInstance_fp16_e768   = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  768,   1, 1>;
-using DeviceInstance_fp16_e1024  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1024,  1, 2>;
-using DeviceInstance_fp16_e1536  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  1536,  1, 2>;
-using DeviceInstance_fp16_e2048  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  2048,  1, 2>;
-using DeviceInstance_fp16_e4096  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  4096,  1, 8>;
-using DeviceInstance_fp16_e8192  = ck::tensor_operation::device::DeviceSparseEmbedding3ForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, 256,  1,  256, 1,  8192,  1, 8>;
+using DeviceInstance_fp16_e256   = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  256,   1, 1, 3>;
+using DeviceInstance_fp16_e512   = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  512,   1, 2, 3>;
+using DeviceInstance_fp16_e768   = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  768,   1, 1, 3>;
+using DeviceInstance_fp16_e1024  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  1024,  1, 2, 3>;
+using DeviceInstance_fp16_e1536  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  1536,  1, 2, 3>;
+using DeviceInstance_fp16_e2048  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  2048,  1, 2, 3>;
+using DeviceInstance_fp16_e4096  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  4096,  1, 8, 3>;
+using DeviceInstance_fp16_e8192  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  8192,  1, 8, 3>;
 
 template<typename emb_type, ck::index_t dim> struct emb_kernel{};
 
-template<> struct emb_kernel<float, 256>  { using kernel_type = DeviceInstance_fp32_e256; };
-template<> struct emb_kernel<float, 512>  { using kernel_type = DeviceInstance_fp32_e512; };
-template<> struct emb_kernel<float, 768>  { using kernel_type = DeviceInstance_fp32_e768; };
-template<> struct emb_kernel<float, 1024> { using kernel_type = DeviceInstance_fp32_e1024;};
-template<> struct emb_kernel<float, 1536> { using kernel_type = DeviceInstance_fp32_e1536;};
-template<> struct emb_kernel<float, 2048> { using kernel_type = DeviceInstance_fp32_e2048;};
-template<> struct emb_kernel<float, 4096> { using kernel_type = DeviceInstance_fp32_e4096;};
-template<> struct emb_kernel<float, 8192> { using kernel_type = DeviceInstance_fp32_e8192;};
-template<> struct emb_kernel<float, 16384>{ using kernel_type = DeviceInstance_fp32_e16384;};
-
 template<> struct emb_kernel<ck::half_t, 256>  { using kernel_type = DeviceInstance_fp16_e256; };
 template<> struct emb_kernel<ck::half_t, 512>  { using kernel_type = DeviceInstance_fp16_e512; };
 template<> struct emb_kernel<ck::half_t, 768>  { using kernel_type = DeviceInstance_fp16_e768; };
@@ -152,19 +126,20 @@ int main()
         beta_dev.ToDevice(beta.mData.data());
 
         auto device_instance = typename emb_kernel<EmbType, current_dim>::kernel_type{};
-        auto argument_ptr    = device_instance.MakeArgumentPointer(out_dev.GetDeviceBuffer(),
-                                                                emb_a_dev.GetDeviceBuffer(),
-                                                                emb_b_dev.GetDeviceBuffer(),
-                                                                emb_c_dev.GetDeviceBuffer(),
-                                                                index_a_dev.GetDeviceBuffer(),
-                                                                index_b_dev.GetDeviceBuffer(),
-                                                                index_c_dev.GetDeviceBuffer(),
-                                                                gamma_dev.GetDeviceBuffer(),
-                                                                beta_dev.GetDeviceBuffer(),
-                                                                num_rows,
-                                                                current_dim,
-                                                                index_length,
-                                                                epsilon);
+        auto argument_ptr    = device_instance.MakeArgumentPointer(
+            out_dev.GetDeviceBuffer(),
+            {ck::type_convert<EmbType*>(emb_a_dev.GetDeviceBuffer()),
+             ck::type_convert<EmbType*>(emb_b_dev.GetDeviceBuffer()),
+             ck::type_convert<EmbType*>(emb_c_dev.GetDeviceBuffer())},
+            {ck::type_convert<IndexType*>(index_a_dev.GetDeviceBuffer()),
+             ck::type_convert<IndexType*>(index_b_dev.GetDeviceBuffer()),
+             ck::type_convert<IndexType*>(index_c_dev.GetDeviceBuffer())},
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            current_dim,
+            index_length,
+            epsilon,
+            EmbElementwiseOperation{});
         std::cout << "Dim:" << current_dim << ", kernel:" << device_instance.GetTypeString()
                   << std::endl
                   << std::flush;

example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp

@@ -3,7 +3,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -19,13 +19,13 @@ using BDataType = F16;
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using DeviceElementwisePermuteInstance =
-    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
-                                                    ck::Tuple<BDataType>,
-                                                    PassThrough,
-                                                    4,
-                                                    8,
-                                                    ck::Sequence<8>,
-                                                    ck::Sequence<1>>;
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>,
+                                                        ck::Tuple<BDataType>,
+                                                        PassThrough,
+                                                        4,
+                                                        8,
+                                                        ck::Sequence<8>,
+                                                        ck::Sequence<1>>;
 
 template <typename HostTensorA, typename HostTensorB, typename Functor>
 void host_elementwise4D(HostTensorB& B_nhwc, const HostTensorA& A_nchw, Functor functor)

example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp

@@ -3,7 +3,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/device_elementwise_2d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -17,15 +17,15 @@ using BDataType = F16;
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using DeviceElementwisePermuteInstance =
-    ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
-                                                    ck::Tuple<BDataType>,
-                                                    PassThrough,
-                                                    3, // NumDim_M
-                                                    1, // NumDim_N
-                                                    8,
-                                                    8,
-                                                    ck::Sequence<8>,
-                                                    ck::Sequence<8>>;
+    ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>,
+                                                          ck::Tuple<BDataType>,
+                                                          PassThrough,
+                                                          3, // NumDim_M
+                                                          1, // NumDim_N
+                                                          8,
+                                                          8,
+                                                          ck::Sequence<8>,
+                                                          ck::Sequence<8>>;
 
 template <typename HostTensorA, typename HostTensorB, typename Functor>
 void host_elementwise4D(HostTensorB& B_nhwc,

example/46_gemm_add_multiply/CMakeLists.txt

+add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
+add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)

example/46_gemm_add_multiply/README.md

+# Instructions for ```example_gemm_add_multiply_dl_fp16```
+
+## Run ```example_gemm_add_multiply_dl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
+./bin/example_gemm_add_multiply_dl_fp16 1 1 1
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
+d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
+d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m0_m1_k1_{2048, 3840, 2}
+arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
+arg.e_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {960, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 3.99904 ms, 32.22 TFlops, 31.9913 GB/s, DeviceGemmMultipleD_Dl<256, 128, 128, 16, 2, 4, 4, 1>
+```

example/46_gemm_add_multiply/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cstddef>
+#include <iostream>
+#include <stdexcept>
+#include <string>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+using I8   = int8_t;
+using I32  = int32_t;
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline bool
+parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 12)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA  = std::stoi(argv[7]);
+        problem_size.StrideB  = std::stoi(argv[8]);
+        problem_size.StrideD0 = std::stoi(argv[9]);
+        problem_size.StrideD1 = std::stoi(argv[10]);
+        problem_size.StrideE  = std::stoi(argv[11]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
+                     "StrideE"
+                  << std::endl;
+        return false;
+    }
+
+    return true;
+}

example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using D0DataType  = F16;
+using D1DataType  = F16;
+using DsDataType  = ck::Tuple<D0DataType, D1DataType>;
+using EDataType   = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::
+        //  ##################| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|     DsData|     EData|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ##################|        |        |         |        |      Type|      Type|        Type|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ##################|        |        |         |        |          |          |            |           |          |   Operation|   Operation|    Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ##################|        |        |         |        |          |          |            |           |          |            |            |             |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmMultipleD_Dl< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_multiply_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_multiply_example(argc, argv); }

example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using D0DataType  = F16;
+using D1DataType  = F16;
+using DsDataType  = ck::Tuple<D0DataType, D1DataType>;
+using EDataType   = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::
+        //##############################|      A|      B|       Ds|      E| AData| BData| AccData| CShuffle|     DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|   Layout| Layout|  Type|  Type|    Type| DataType|       Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |         |       |      |      |        |         |           |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |         |       |      |      |        |         |           |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, DsLayout,    Row,   F16,   F16,     F32,      F16, DsDataType,   F16, PassThrough, PassThrough, CDEElementOp,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_multiply_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_multiply_example(argc, argv); }

example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc

+#pragma once
+
+bool run_gemm_add_multiply(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+    auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+	    std::cout << "wrong! this device_op instance does not support this problem" << std::endl;
+	    return true;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = 2_uz * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(D0DataType) * N + sizeof(D1DataType) * M * N +
+                            sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+    }
+
+    return true;
+}
+
+bool run_gemm_add_multiply_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) ||
+           run_gemm_add_multiply(problem_size, config);
+}

include/ck/ck.hpp

@@ -18,8 +18,13 @@
 #define CK_USE_LAUNCH_BOUNDS 1
 
 #ifdef CK_USE_LAUNCH_BOUNDS
+// for most kernels
 #define CK_MAX_THREAD_PER_BLOCK 256
 #define CK_MIN_BLOCK_PER_CU 2
+
+// for wavelet GEMM kernel
+#define CK_WAVELET_MAX_THREAD_PER_BLOCK 512
+#define CK_WAVELET_MIN_BLOCK_PER_CU 2
 #endif
 
 // check GPU target

include/ck/tensor_operation/gpu/device/device_elementwise_base.hpp → include/ck/tensor_operation/gpu/device/device_elementwise.hpp

@@ -17,7 +17,7 @@ template <typename InDataTypeTuple,
           typename OutDataTypeTuple,
           typename ElementwiseOperation,
           index_t NumDim>
-struct DeviceElementwiseBase : public BaseOperator
+struct DeviceElementwise : public BaseOperator
 {
     static constexpr int NumInput  = InDataTypeTuple::Size();
     static constexpr int NumOutput = OutDataTypeTuple::Size();
@@ -37,8 +37,8 @@ template <typename InDataTypeTuple,
           typename OutDataTypeTuple,
           typename ElementwiseOperation,
           index_t NumDim>
-using DeviceElementwiseBasePtr = std::unique_ptr<
-    DeviceElementwiseBase<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>>;
+using DeviceElementwisePtr = std::unique_ptr<
+    DeviceElementwise<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>>;
 
 } // namespace device
 } // namespace tensor_operation

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

@@ -32,7 +32,7 @@ struct DeviceElementwiseNormalization : public BaseOperator
                         const std::vector<index_t> betaStrides,
                         const std::vector<index_t> yStrides,
                         const std::vector<index_t> reduceDims,
-                        AccDataType epsilon,
+                        double epsilon,
                         const std::array<const void*, NumInput> in_dev_buffers,
                         const void* p_gamma,
                         const void* p_beta,

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

@@ -32,8 +32,8 @@ struct DeviceMultipleReduce : public BaseOperator
         const std::array<index_t, NumOutputDim> outLengths,
         const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStrides,
         const std::array<int, NumReduceDim> reduceDims,
-        const std::array<const void*, NumReduction> alphas,
-        const std::array<const void*, NumReduction> betas,
+        const std::array<double, NumReduction> alphas,
+        const std::array<double, NumReduction> betas,
         const void* in_dev,
         const std::array<void*, NumReduction> out_dev_buffers,
         const InElementwiseOperationTuple in_elementwise_op_tuple,

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

@@ -28,7 +28,7 @@ struct DeviceNormalization : public BaseOperator
                         const std::vector<index_t> betaStrides,
                         const std::vector<index_t> yStrides,
                         const std::vector<index_t> reduceDims,
-                        AccDataType epsilon,
+                        double epsilon,
                         const void* p_x,
                         const void* p_gamma,
                         const void* p_beta,