Merge branch 'develop' into amd-develop

CMakeLists.txt

@@ -180,7 +180,16 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 500723302)
 endif()
 if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090)
    message("Adding the enable-post-misched=0 compiler flag")
-   add_compile_options(-mllvm -enable-post-misched=0)
+   add_compile_options("SHELL: -mllvm -enable-post-misched=0")
+endif()
+if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000)
+   message("Adding the amdgpu-coerce-illegal-types=1")
+   add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1")
+endif()
+if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
+   message("Adding -amdgpu-early-inline-all=true and -amdgpu-function-calls=false")
+   add_compile_options("SHELL: -mllvm -amdgpu-early-inline-all=true")
+   add_compile_options("SHELL: -mllvm -amdgpu-function-calls=false")
 endif()
 #
 # Seperate linking jobs from compiling
@@ -235,6 +244,16 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
 message("CMAKE_CXX_COMPILER: ${CMAKE_CXX_COMPILER}")
 
+# https://gcc.gnu.org/onlinedocs/libstdc++/manual/using_macros.html
+# _GLIBCXX_ASSERTIONS
+# Undefined by default. When defined, enables extra error checking in the form of
+# precondition assertions, such as bounds checking in strings and null pointer
+# checks when dereferencing smart pointers
+option(USE_GLIBCXX_ASSERTIONS "Turn on additional c++ library checks." OFF)
+if(USE_GLIBCXX_ASSERTIONS)
+  add_compile_options(-Wp,-D_GLIBCXX_ASSERTIONS)
+endif()
+
 ## HIP
 set(CMAKE_HIP_PLATFORM amd)
 set(CMAKE_HIP_COMPILER ${CMAKE_CXX_COMPILER})

Dockerfile

@@ -23,11 +23,11 @@ RUN if [ "$ROCMVERSION" != "6.2" ]; then \
         wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
         sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
         sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
-    elif [ "$ROCMVERSION" = "6.2" ] && [ "$compiler_version" = "rc3" ]; then \
+    elif [ "$ROCMVERSION" = "6.2" ] && [ "$compiler_version" = "rc4" ]; then \
         sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.2-20.04-1_all.deb --no-check-certificate" && \
         apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install dialog libpopt0 rsync && DEBIAN_FRONTEND=noninteractive apt-get install ./amdgpu-install-internal_6.2-20.04-1_all.deb && \
-        sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.2 rel-45 > /etc/apt/sources.list.d/rocm-build.list' && \
-        amdgpu-repo --amdgpu-build=2003709; \
+        sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.2 rel-63 > /etc/apt/sources.list.d/rocm-build.list' && \
+        amdgpu-repo --amdgpu-build=2009461; \
     fi
 
 RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"

Jenkinsfile

@@ -86,17 +86,7 @@ def check_host() {
 
 def build_compiler(){
     def compiler
-    if (params.BUILD_COMPILER == "hipcc"){
-        compiler = '/opt/rocm/bin/hipcc'
-    }
-    else{
-        if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
-            compiler = "/llvm-project/build/bin/clang++"
-        }
-        else{
-            compiler = "/opt/rocm/llvm/bin/clang++"
-        }        
-    }
+    compiler = "${params.BUILD_COMPILER}"
     return compiler
 }
 
@@ -307,7 +297,7 @@ def buildHipClangJob(Map conf=[:]){
         def prefixpath = conf.get("prefixpath", "/opt/rocm")
 
         // Jenkins is complaining about the render group 
-        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
         if (conf.get("enforce_xnack_on", false)) {
             dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
         }
@@ -366,7 +356,7 @@ def runCKProfiler(Map conf=[:]){
         def prefixpath = conf.get("prefixpath", "/opt/rocm")
 
         // Jenkins is complaining about the render group 
-        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
         if (conf.get("enforce_xnack_on", false)) {
             dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
         }
@@ -487,7 +477,7 @@ def Build_CK(Map conf=[:]){
         def prefixpath = conf.get("prefixpath", "/opt/rocm")
 
         // Jenkins is complaining about the render group 
-        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
         if (conf.get("enforce_xnack_on", false)) {
             dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
         }
@@ -600,7 +590,7 @@ def process_results(Map conf=[:]){
     def prefixpath = "/opt/rocm"
 
     // Jenkins is complaining about the render group 
-    def dockerOpts="--cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+    def dockerOpts="--rm --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
     if (conf.get("enforce_xnack_on", false)) {
         dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
     }
@@ -664,8 +654,8 @@ def process_results(Map conf=[:]){
 //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
 CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.1; RUN_CK_TILE_TESTS=true
                                               0 21 * * * % ROCMVERSION=6.1;hipTensor_test=true
-                                              0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;COMPILER_COMMIT=;USE_SCCACHE=false
-                                              0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;COMPILER_COMMIT=;USE_SCCACHE=false
+                                              0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false
+                                              0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false
                                               0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : ""
 
 pipeline {
@@ -699,8 +689,8 @@ pipeline {
             description: 'Specify which commit of compiler branch to use: leave blank to use the latest commit (default), or use some specific commit of llvm-project branch.')
         string(
             name: 'BUILD_COMPILER', 
-            defaultValue: 'clang', 
-            description: 'Specify whether to build CK with hipcc or with clang (default).')
+            defaultValue: '/opt/rocm/llvm/bin/clang++', 
+            description: 'Build CK with /opt/rocm/bin/hipcc, /llvm-project/build/bin/clang++, or with /opt/rocm/llvm/bin/clang++ (default).')
         booleanParam(
             name: "RUN_FULL_QA",
             defaultValue: false,

client_example/24_grouped_conv_activation/CMakeLists.txt

@@ -39,6 +39,10 @@ target_link_libraries(client_grouped_convnd_fwd_bilinear_residual_fp16 PRIVATE c
 add_executable(client_conv3d_fwd_convinvscale_fp8 
                grouped_convnd_fwd_convinvscale/conv3d_fwd_convinvscale_fp8.cpp)
 target_link_libraries(client_conv3d_fwd_convinvscale_fp8 PRIVATE composable_kernel::device_conv_operations)
+# Fwd convscale + Bias
+add_executable(client_conv3d_fwd_convscale_add_fp8
+               grouped_convnd_fwd_convscale_add/conv3d_fwd_convscale_add_fp8.cpp)
+target_link_libraries(client_conv3d_fwd_convscale_add_fp8 PRIVATE composable_kernel::device_conv_operations)
 # Fwd convscale + ReLU
 add_executable(client_conv3d_fwd_convscale_relu_fp8
                grouped_convnd_fwd_convscale_relu/conv3d_fwd_convscale_relu_fp8.cpp)

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_add/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_add.hpp"
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ConvScaleAdd = ck::tensor_operation::element_wise::ConvScaleAdd;
+using F32          = float;
+using BiasDataType = F32;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
+         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths,
+         const std::size_t& ds_size)
+{
+    // G * N * C * <output spatial lengths product> * (2 * K * <filter spatial lengths product> +
+    // <number of scale factors>)
+    ck::index_t G = weights_lengths[0];
+    ck::index_t N = output_lengths[1];
+    ck::index_t K = weights_lengths[1];
+    ck::index_t C = weights_lengths[2];
+
+    return G * N * C *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           (static_cast<std::size_t>(2) * K *
+                std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                                std::end(weights_lengths),
+                                static_cast<std::size_t>(1),
+                                std::multiplies<>()) +
+            ds_size);
+}
+
+template <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * std::accumulate(std::begin(input_lengths),
+                                                std::end(input_lengths),
+                                                static_cast<std::size_t>(1),
+                                                std::multiplies<>());
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * std::accumulate(std::begin(weights_lengths),
+                                                 std::end(weights_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<>());
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * std::accumulate(std::begin(output_lengths),
+                                                 std::end(output_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<std::size_t>());
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumNonSpatialDim = 3,
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
+bool run_grouped_conv_fwd_convscale_add(
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
+{
+
+    namespace ctc = ck::tensor_layout::convolution;
+    static_assert(NumDimSpatial == 3 && ck::is_same_v<InLayout, ctc::NDHWGC> &&
+                      ck::is_same_v<WeiLayout, ctc::GKZYXC> &&
+                      ck::is_same_v<OutLayout, ctc::NDHWGK>,
+                  "Unsupported configuration");
+
+    const ck::index_t G  = in_lengths[4];
+    const ck::index_t N  = in_lengths[0];
+    const ck::index_t K  = wei_lengths[1];
+    const ck::index_t C  = in_lengths[5];
+    const ck::index_t Z  = wei_lengths[2];
+    const ck::index_t Y  = wei_lengths[3];
+    const ck::index_t X  = wei_lengths[4];
+    const ck::index_t Di = in_lengths[1];
+    const ck::index_t Hi = in_lengths[2];
+    const ck::index_t Wi = in_lengths[3];
+    const ck::index_t Do = out_lengths[1];
+    const ck::index_t Ho = out_lengths[2];
+    const ck::index_t Wo = out_lengths[3];
+
+    const std::size_t in_mem_size   = sizeof(InDataType) * N * Di * Hi * Wi * G * C;
+    const std::size_t wei_mem_size  = sizeof(WeiDataType) * G * K * Z * Y * X * C;
+    const std::size_t out_mem_size  = sizeof(OutDataType) * N * Do * Ho * Wo * G * K;
+    const std::size_t bias_mem_size = sizeof(BiasDataType) * N * Do * Ho * Wo * G * K;
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+    SimpleDeviceMem bias(bias_mem_size);
+
+    float scale_in  = float(std::rand()) / float(RAND_MAX);
+    float scale_wei = float(std::rand()) / float(RAND_MAX);
+    float scale_out = float(std::rand()) / float(RAND_MAX);
+
+    // We have NDHWGC/GKZYXC/NDHWGK (x, weight, y) in memory space.
+    // However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
+    // Hence, we need to adjust the order of strides.
+    const std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+    const std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+    const std::array<ck::index_t, NumDimSpatial + 3> weights_lengths{G, K, C, Z, Y, X};
+    const std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+    const std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+    const std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+    const std::array<ck::index_t, NumDimSpatial + 3> bias_lengths{G, N, K, Do, Ho, Wo};
+    const std::array<ck::index_t, NumDimSpatial + 3> bias_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+    const std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    const std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    std::size_t ds_size   = 3 + 1; // 3 element-wise scale multipliers + 1 elementwise Bias
+    std::size_t flop      = GetFlops<NumDimSpatial>(output_lengths, weights_lengths, ds_size);
+    std::size_t num_bytes = in_mem_size + wei_mem_size + sizeof(float) + sizeof(float) +
+                            sizeof(float) + out_mem_size + bias_mem_size;
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                      InLayout,
+                                                                      WeiLayout,
+                                                                      ck::Tuple<OutLayout>,
+                                                                      OutLayout,
+                                                                      InDataType,
+                                                                      WeiDataType,
+                                                                      ck::Tuple<BiasDataType>,
+                                                                      OutDataType,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      ConvScaleAdd,
+                                                                      AComputeType,
+                                                                      BComputeType>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 1>{bias.GetDeviceBuffer()},
+            out.GetDeviceBuffer(),
+            input_lengths,
+            input_strides,
+            weights_lengths,
+            weights_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_lengths}},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_strides}},
+            output_lengths,
+            output_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScaleAdd{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 1>{bias.GetDeviceBuffer()},
+            out.GetDeviceBuffer(),
+            input_lengths,
+            input_strides,
+            weights_lengths,
+            weights_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_lengths}},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 1>{
+                {bias_strides}},
+            output_lengths,
+            output_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            ConvScaleAdd{scale_in, scale_wei, scale_out});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return true;
+}

client_example/24_grouped_conv_activation/grouped_convnd_fwd_convscale_add/conv3d_fwd_convscale_add_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using CShuffleDataType = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd_convscale_add<NumDimSpatial,
+                                              InDataType,
+                                              WeiDataType,
+                                              OutDataType,
+                                              InLayout,
+                                              WeiLayout,
+                                              OutLayout,
+                                              3,
+                                              AComputeDataType,
+                                              BComputeDataType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

codegen/src/device_grouped_conv_fwd_multiple_abd_operation_xdl_cshuffle.cpp

@@ -233,6 +233,12 @@ extern "C" __global__ void run_${name}(
                                     ${BElementwiseOperation}{},
                                     ${CDEElementwiseOperation}{1.0f, 1.0f});
 
+    if(!DeviceConv::IsSupportedArgument(arg))
+    {
+	printf("Arguement is not supported.\n");
+	return;
+    };    
+
     constexpr ck::LoopScheduler LoopSched = ck::make_default_loop_scheduler();
 
     // GridwiseGemm

docs/sphinx/requirements.in

-rocm-docs-core==1.5.0
+rocm-docs-core==1.6.1
 sphinxcontrib-bibtex==2.6.2

docs/sphinx/requirements.txt

@@ -4,33 +4,33 @@
 #
 #    pip-compile requirements.in
 #
-accessible-pygments==0.0.3
+accessible-pygments==0.0.5
     # via pydata-sphinx-theme
-alabaster==0.7.13
+alabaster==0.7.16
     # via sphinx
-babel==2.12.1
+babel==2.15.0
     # via
     #   pydata-sphinx-theme
     #   sphinx
-beautifulsoup4==4.11.2
+beautifulsoup4==4.12.3
     # via pydata-sphinx-theme
-breathe==4.34.0
+breathe==4.35.0
     # via rocm-docs-core
-certifi==2023.7.22
+certifi==2024.7.4
     # via requests
-cffi==1.15.1
+cffi==1.16.0
     # via
     #   cryptography
     #   pynacl
-charset-normalizer==3.1.0
+charset-normalizer==3.3.2
     # via requests
-click==8.1.3
+click==8.1.7
     # via sphinx-external-toc
-cryptography==41.0.6
+cryptography==43.0.0
     # via pyjwt
-deprecated==1.2.13
+deprecated==1.2.14
     # via pygithub
-docutils==0.16
+docutils==0.21.2
     # via
     #   breathe
     #   myst-parser
@@ -38,35 +38,35 @@ docutils==0.16
     #   pydata-sphinx-theme
     #   sphinx
     #   sphinxcontrib-bibtex
-fastjsonschema==2.18.0
+fastjsonschema==2.20.0
     # via rocm-docs-core
-gitdb==4.0.10
+gitdb==4.0.11
     # via gitpython
-gitpython==3.1.37
+gitpython==3.1.43
     # via rocm-docs-core
-idna==3.4
+idna==3.7
     # via requests
 imagesize==1.4.1
     # via sphinx
-jinja2==3.1.2
+jinja2==3.1.4
     # via
     #   myst-parser
     #   sphinx
-latexcodec==2.0.1
+latexcodec==3.0.0
     # via pybtex
-markdown-it-py==2.2.0
+markdown-it-py==3.0.0
     # via
     #   mdit-py-plugins
     #   myst-parser
-markupsafe==2.1.2
+markupsafe==2.1.5
     # via jinja2
-mdit-py-plugins==0.3.5
+mdit-py-plugins==0.4.1
     # via myst-parser
 mdurl==0.1.2
     # via markdown-it-py
-myst-parser==1.0.0
+myst-parser==3.0.1
     # via rocm-docs-core
-packaging==23.0
+packaging==24.1
     # via
     #   pydata-sphinx-theme
     #   sphinx
@@ -74,48 +74,46 @@ pybtex==0.24.0
     # via
     #   pybtex-docutils
     #   sphinxcontrib-bibtex
-pybtex-docutils==1.0.2
+pybtex-docutils==1.0.3
     # via sphinxcontrib-bibtex
-pycparser==2.21
+pycparser==2.22
     # via cffi
-pydata-sphinx-theme==0.13.3
+pydata-sphinx-theme==0.15.4
     # via
     #   rocm-docs-core
     #   sphinx-book-theme
-pygithub==1.58.1
+pygithub==2.3.0
     # via rocm-docs-core
-pygments==2.15.0
+pygments==2.18.0
     # via
     #   accessible-pygments
     #   pydata-sphinx-theme
     #   sphinx
-pyjwt[crypto]==2.6.0
+pyjwt[crypto]==2.8.0
     # via pygithub
 pynacl==1.5.0
     # via pygithub
-pyyaml==6.0
+pyyaml==6.0.1
     # via
     #   myst-parser
     #   pybtex
     #   rocm-docs-core
     #   sphinx-external-toc
-requests==2.31.0
+requests==2.32.3
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==1.5.0
+rocm-docs-core==1.6.1
     # via -r requirements.in
 six==1.16.0
-    # via
-    #   latexcodec
-    #   pybtex
-smmap==5.0.0
+    # via pybtex
+smmap==5.0.1
     # via gitdb
 snowballstemmer==2.2.0
     # via sphinx
-soupsieve==2.4
+soupsieve==2.5
     # via beautifulsoup4
-sphinx==5.3.0
+sphinx==7.4.7
     # via
     #   breathe
     #   myst-parser
@@ -127,33 +125,39 @@ sphinx==5.3.0
     #   sphinx-external-toc
     #   sphinx-notfound-page
     #   sphinxcontrib-bibtex
-sphinx-book-theme==1.0.1
+sphinx-book-theme==1.1.3
     # via rocm-docs-core
-sphinx-copybutton==0.5.1
+sphinx-copybutton==0.5.2
     # via rocm-docs-core
-sphinx-design==0.4.1
+sphinx-design==0.6.0
     # via rocm-docs-core
-sphinx-external-toc==0.3.1
+sphinx-external-toc==1.0.1
     # via rocm-docs-core
-sphinx-notfound-page==0.8.3
+sphinx-notfound-page==1.0.3
     # via rocm-docs-core
-sphinxcontrib-applehelp==1.0.4
+sphinxcontrib-applehelp==2.0.0
     # via sphinx
 sphinxcontrib-bibtex==2.6.2
     # via -r requirements.in
-sphinxcontrib-devhelp==1.0.2
+sphinxcontrib-devhelp==2.0.0
     # via sphinx
-sphinxcontrib-htmlhelp==2.0.1
+sphinxcontrib-htmlhelp==2.1.0
     # via sphinx
 sphinxcontrib-jsmath==1.0.1
     # via sphinx
-sphinxcontrib-qthelp==1.0.3
+sphinxcontrib-qthelp==2.0.0
     # via sphinx
-sphinxcontrib-serializinghtml==1.1.5
+sphinxcontrib-serializinghtml==2.0.0
     # via sphinx
-typing-extensions==4.5.0
-    # via pydata-sphinx-theme
-urllib3==1.26.18
-    # via requests
-wrapt==1.15.0
+tomli==2.0.1
+    # via sphinx
+typing-extensions==4.12.2
+    # via
+    #   pydata-sphinx-theme
+    #   pygithub
+urllib3==2.2.2
+    # via
+    #   pygithub
+    #   requests
+wrapt==1.16.0
     # via deprecated

example/62_convnd_activ/CMakeLists.txt

@@ -2,6 +2,7 @@ add_subdirectory(binary)
 add_subdirectory(convinvscale)
 add_subdirectory(convscale)
 add_subdirectory(convscale_relu)
+add_subdirectory(convscale_add)
 add_subdirectory(multi_AB)
 add_subdirectory(unary)
 

example/62_convnd_activ/convscale_add/CMakeLists.txt

+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+     add_custom_target(example_convnd_activ_xdl_convscale_add)
+     add_example_executable(example_convnd_fwd_xdl_convscale_add_fp8 convnd_fwd_xdl_convscale_add_fp8.cpp)
+     add_example_dependencies(example_convnd_activ_xdl_convscale_add example_convnd_fwd_xdl_convscale_add_fp8 )
+
+     set(target 1)
+ endif()
+endforeach()

example/62_convnd_activ/convscale_add/convnd_fwd_convscale_add_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ConvScaleAdd = ck::tensor_operation::element_wise::ConvScaleAdd;
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
+         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths,
+         const std::size_t& ds_size)
+{
+    // G * N * C * <output spatial lengths product> * (2 * K * <filter spatial lengths product> +
+    // <number of scale factors>)
+    ck::index_t G = weights_lengths[0];
+    ck::index_t N = output_lengths[1];
+    ck::index_t K = weights_lengths[1];
+    ck::index_t C = weights_lengths[2];
+
+    return G * N * C *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           (static_cast<std::size_t>(2) * K *
+                std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                                std::end(weights_lengths),
+                                static_cast<std::size_t>(1),
+                                std::multiplies<>()) +
+            ds_size);
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<DsDataType> bias(out_g_n_k_wos_desc);
+    Tensor<CShuffleDataType> c(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-1, 1});
+        bias.GenerateTensorValue(GeneratorTensor_2<DsDataType>{-3, 3});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1.0, 1.0});
+        bias.GenerateTensorValue(GeneratorTensor_3<DsDataType>{-3.0, 3.0});
+        break;
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(DsDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), d_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), d_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // random scale values
+    float scale_in  = float(std::rand()) / float(RAND_MAX);
+    float scale_wei = float(std::rand()) / float(RAND_MAX);
+    float scale_out = float(std::rand()) / float(RAND_MAX);
+
+    std::cout << std::endl;
+    std::cout << "scale_in: " << scale_in << std::endl;
+    std::cout << "scale_wei: " << scale_wei << std::endl;
+    std::cout << "scale_out: " << scale_out << std::endl;
+
+    // initialize out_element_op for each iteration
+    const auto out_element_op = OutElementOp{scale_in, scale_wei, scale_out};
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+        out_device_buf.GetDeviceBuffer(),
+        a_g_n_c_wis_lengths,
+        a_g_n_c_wis_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_lengths}},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_strides}},
+        e_g_n_k_wos_lengths,
+        e_g_n_k_wos_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t ds_size = 3 + 1; // 3 element-wise scale multipliers + 1 element-wise add
+    std::size_t flop    = GetFlops<NDimSpatial>(e_g_n_k_wos_lengths, b_g_k_c_xs_lengths, ds_size);
+    std::size_t num_btype =
+        conv_param.GetInputByte<InDataType>() + conv_param.GetWeightByte<WeiDataType>() +
+        sizeof(float) + sizeof(float) + sizeof(float) + conv_param.GetOutputByte<OutDataType>() +
+        conv_param.GetOutputByte<DsDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     CShuffleDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), c(idx), bias(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device,
+                                    out_host,
+                                    "Error: incorrect results!",
+                                    get_rtol<OutDataType>(),
+                                    get_atol<OutDataType>());
+    }
+
+    return true;
+}

example/62_convnd_activ/convscale_add/convnd_fwd_xdl_convscale_add_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/tuple.hpp"
+#include "convnd_fwd_convscale_add_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using DsDataType       = float;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ConvScaleAdd;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<DsLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<DsDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8,
+        AComputeDataType,
+        BComputeDataType>;
+
+#include "run_convnd_fwd_convscale_add_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }

example/62_convnd_activ/convscale_add/run_convnd_fwd_convscale_add_example.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_convnd_fwd_example(int argc, char* argv[])
+{
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    // instantiate in and wei element ops, will
+    // instantiate out_element_op below for every iteration
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+
+    const auto run =
+        [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto ds_layout, auto out_layout) {
+            constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+
+            using InLayout  = decltype(in_layout);
+            using WeiLayout = decltype(wei_layout);
+            using DsLayout  = decltype(ds_layout);
+            using OutLayout = decltype(out_layout);
+
+            const auto in_g_n_c_wis_desc =
+                ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                    conv_param);
+
+            const auto wei_g_k_c_xs_desc =
+                ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                    conv_param);
+
+            const auto out_g_n_k_wos_desc =
+                ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                    conv_param);
+
+            return run_grouped_conv_fwd<ndim_spatial_value,
+                                        InDataType,
+                                        WeiDataType,
+                                        CShuffleDataType,
+                                        DsDataType,
+                                        OutDataType,
+                                        InElementOp,
+                                        WeiElementOp,
+                                        OutElementOp,
+                                        DeviceGroupedConvNDFwdInstance<ndim_spatial_value,
+                                                                       InLayout,
+                                                                       WeiLayout,
+                                                                       DsLayout,
+                                                                       OutLayout>>(
+                do_verification,
+                init_method,
+                time_kernel,
+                conv_param,
+                in_g_n_c_wis_desc,
+                wei_g_k_c_xs_desc,
+                out_g_n_k_wos_desc,
+                in_element_op,
+                wei_element_op);
+        };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{}, ctc::GNWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{}, ctc::GNHWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}

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

@@ -428,11 +428,18 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
                     pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
                     pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
                 else:
-                    pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
-                    pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
-                    pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
-                    pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
-                    if receipt == 1:
+                    if bias == "bias":
+                        # TODO: rocm 6.2 compiler problem if using qr_async for bias case
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    else:
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    if receipt == 1 and bias != "bias":
                         pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
                         pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
         elif dtype in ['fp8', 'bf8']:

include/ck/host_utility/flush_cache.hpp

@@ -14,6 +14,124 @@
 namespace ck {
 namespace utility {
 
+template <typename Argument, typename DsDataType>
+struct RotatingMemWrapperMultiD
+{
+    static constexpr index_t NumDs = DsDataType::Size();
+
+    using ADataType     = decltype(Argument::p_a_grid);
+    using BDataType     = decltype(Argument::p_b_grid);
+    using DsGridPointer = decltype(Argument::p_ds_grid);
+
+    RotatingMemWrapperMultiD() = delete;
+    RotatingMemWrapperMultiD(Argument& arg_,
+                             std::size_t rotating_count_,
+                             std::size_t size_a_,
+                             std::size_t size_b_,
+                             std::array<std::size_t, NumDs> size_ds_)
+        : arg(arg_),
+          rotating_count(rotating_count_),
+          size_a(size_a_),
+          size_b(size_b_),
+          size_ds(size_ds_)
+    {
+        p_a_grids.push_back(arg.p_a_grid);
+        p_b_grids.push_back(arg.p_b_grid);
+        p_ds_grids.push_back(arg.p_ds_grid);
+        for(size_t i = 1; i < rotating_count; i++)
+        {
+            {
+                void* pADeviceBuf;
+                hip_check_error(hipMalloc(static_cast<void**>(&pADeviceBuf), size_a_));
+                hip_check_error(hipMemcpy(static_cast<void*>(pADeviceBuf),
+                                          const_cast<void*>(p_a_grids[0]),
+                                          size_a_,
+                                          hipMemcpyDeviceToDevice));
+                p_a_grids.push_back(pADeviceBuf);
+            }
+
+            {
+                void* pBDeviceBuf;
+                hip_check_error(hipMalloc(static_cast<void**>(&pBDeviceBuf), size_b_));
+                hip_check_error(hipMemcpy(static_cast<void*>(pBDeviceBuf),
+                                          const_cast<void*>(p_b_grids[0]),
+                                          size_b_,
+                                          hipMemcpyDeviceToDevice));
+                p_b_grids.push_back(pBDeviceBuf);
+            }
+
+            {
+
+                DsGridPointer ds_buffer;
+                static_for<0, NumDs, 1>{}([&](auto j) {
+                    void* pDDeviceBuf;
+                    hip_check_error(hipMalloc(static_cast<void**>(&pDDeviceBuf), size_ds_[j]));
+                    hip_check_error(hipMemcpy(static_cast<void*>(pDDeviceBuf),
+                                              static_cast<const void*>(p_ds_grids[0][j]),
+                                              size_ds_[j],
+                                              hipMemcpyDeviceToDevice));
+
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    ds_buffer(j) = static_cast<const DDataType*>(pDDeviceBuf);
+                });
+
+                p_ds_grids.push_back(ds_buffer);
+            }
+        }
+    }
+
+    void Next()
+    {
+        if(rotating_count > 1)
+        {
+            std::size_t idx = iter++ % rotating_count;
+            arg.p_a_grid    = reinterpret_cast<ADataType>(p_a_grids[idx]);
+            arg.p_b_grid    = reinterpret_cast<BDataType>(p_b_grids[idx]);
+            arg.p_ds_grid   = p_ds_grids[idx];
+        }
+    }
+    void Print()
+    {
+        std::cout << "RotatingMemWrapperMultiD: { size_a: " << size_a << ", size_b: " << size_b
+                  << ", rotating_count: " << rotating_count << "}" << std::endl;
+    }
+    ~RotatingMemWrapperMultiD()
+    {
+        if(rotating_count > 1)
+        {
+            // restore ptr
+            arg.p_a_grid  = reinterpret_cast<ADataType>(p_a_grids[0]);
+            arg.p_b_grid  = reinterpret_cast<BDataType>(p_b_grids[0]);
+            arg.p_ds_grid = p_ds_grids[0];
+
+            // free device mem
+            for(size_t i = 1; i < rotating_count; i++)
+            {
+                hip_check_error(hipFree(const_cast<void*>(p_a_grids[i])));
+                hip_check_error(hipFree(const_cast<void*>(p_b_grids[i])));
+
+                static_for<0, NumDs, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+                    hip_check_error(
+                        hipFree(static_cast<void*>(const_cast<DDataType*>(p_ds_grids[i][j]))));
+                });
+            }
+        }
+    }
+
+    private:
+    Argument& arg;
+    std::size_t iter                       = 0;
+    std::size_t rotating_count             = 1;
+    std::size_t size_a                     = 0;
+    std::size_t size_b                     = 0;
+    std::array<std::size_t, NumDs> size_ds = {0};
+    std::vector<const void*> p_a_grids;
+    std::vector<const void*> p_b_grids;
+    std::vector<DsGridPointer> p_ds_grids;
+};
+
 template <typename Argument>
 struct RotatingMemWrapper
 {

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

@@ -727,6 +727,181 @@ struct CodegenDeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
         ck::Array<index_t, NDimSpatial> input_right_pads_;
     };
 
+    static __device__ __host__ bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // 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))
+                {
+                    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))
+                {
+                    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>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 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>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            // FIXME: layout
+            if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
+                         is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
+                         is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
+                         is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+
+                if constexpr(is_same_v<DLayout, ctc::G_K>)
+                {
+                    // G and K must be the same
+                    if(arg.ds_g_n_k_wos_lengths_[i][0] != arg.e_g_n_k_wos_lengths_[0] ||
+                       arg.ds_g_n_k_wos_lengths_[i][2] != arg.e_g_n_k_wos_lengths_[2])
+                    {
+                        valid = false;
+                    }
+                }
+                else
+                {
+                    // E and D must have the same shape
+                    for(index_t d = 0; d < NDimSpatial + 3; d++)
+                    {
+                        if(arg.ds_g_n_k_wos_lengths_[i][d] != arg.e_g_n_k_wos_lengths_[d])
+                        {
+                            valid = false;
+                        }
+                    }
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        if constexpr(isMultiA || isMultiB)
+        {
+            // Genarate tuples with the same descriptors
+            const auto as_grid_desc_ak0_m_ak1 =
+                generate_tuple([&](auto) { return arg.a_grid_desc_m_k_; }, Number<NumATensor>{});
+            const auto bs_grid_desc_bk0_n_bk1 =
+                generate_tuple([&](auto) { return arg.b_grid_desc_n_k_; }, Number<NumBTensor>{});
+            return GridwiseGemm::CheckValidity(as_grid_desc_ak0_m_ak1,
+                                               bs_grid_desc_bk0_n_bk1,
+                                               arg.ds_grid_desc_m_n_,
+                                               arg.e_grid_desc_m_n_,
+                                               arg.block_2_etile_map_);
+        }
+        else
+        {
+            return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                               arg.b_grid_desc_n_k_,
+                                               arg.ds_grid_desc_m_n_,
+                                               arg.e_grid_desc_m_n_,
+                                               arg.block_2_etile_map_);
+        }
+    }
+
     static __device__ __host__ auto MakeArgument(
         APointers p_as,
         BPointers p_bs,

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp

@@ -15,6 +15,7 @@
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/flush_cache.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -163,14 +164,65 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
             const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
 
             const auto Run = [&](const auto& kernel) {
-                if(arg.KBatch > 1)
-                    hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
-                                                     0,
-                                                     arg.M * arg.N * sizeof(CDataType),
-                                                     stream_config.stream_id_));
-
-                ave_time = launch_and_time_kernel(
-                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+                if(stream_config.flush_cache)
+                {
+
+                    std::array<std::size_t, NumDTensor> DsSize;
+
+                    Argument arg_ = arg;
+
+                    const auto ds_grid_desc_m_n = GridwiseGemm::MakeDsGridDescriptor_M_N(
+                        arg_.M, arg_.MPadded, arg_.N, arg_.NPadded, arg_.StrideDs);
+
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+                        DsSize[i] = ds_grid_desc_m_n[i].GetElementSpaceSize() * sizeof(DDataType);
+                    });
+                    ck::utility::RotatingMemWrapperMultiD<Argument, DsDataType> rotating_mem(
+                        arg_,
+                        stream_config.rotating_count,
+                        arg_.M * arg_.K * sizeof(ADataType),
+                        arg_.K * arg_.N * sizeof(BDataType),
+                        DsSize);
+                    rotating_mem.Print();
+
+                    auto run_flush_cache = [&]() {
+                        // flush icache
+                        ck::utility::flush_icache();
+                        // rotating mem
+                        rotating_mem.Next();
+                        // clear c mem
+                        if constexpr(!is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
+                        {
+                            if(arg_.KBatch > 1)
+                                hipGetErrorString(
+                                    hipMemsetAsync(arg_.p_c_grid,
+                                                   0,
+                                                   arg_.M * arg_.N * sizeof(CDataType),
+                                                   stream_config.stream_id_));
+                        }
+                    };
+
+                    ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
+                        stream_config,
+                        run_flush_cache,
+                        kernel,
+                        dim3(gdx, gdy, gdz),
+                        dim3(BlockSize),
+                        0,
+                        arg_);
+                }
+                else
+                {
+                    if(arg.KBatch > 1)
+                        hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
+                                                         0,
+                                                         arg.M * arg.N * sizeof(CDataType),
+                                                         stream_config.stream_id_));
+
+                    ave_time = launch_and_time_kernel(
+                        stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+                }
             };
 
             constexpr index_t minimum_occupancy =

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

@@ -86,7 +86,6 @@ __global__ void
             const AElementwiseOperation a_element_op,
             const BElementwiseOperation b_element_op,
             const CDEElementwiseOperation cde_element_op,
-            const index_t groups_count,
             const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
             const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
             const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
@@ -101,14 +100,11 @@ __global__ void
     defined(__gfx94__))
 
     // offset base pointer for each work-group
-    const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(gridDim.y / groups_count);
-    const index_t& num_blocks_per_n    = groups_count;
-    const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_batch);
-    const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_n);
-
-    const long_index_t e_batch_offset =
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y);
+    const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.z);
+    const long_index_t e_group_offset =
         amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
-    const auto& ds_batch_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
+    const auto& ds_group_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
 
     const long_index_t e_n_offset =
         amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
@@ -121,14 +117,14 @@ __global__ void
         DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
 
     static_for<0, NumDTensor, 1>{}(
-        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_group_offset[i]; });
 
     if constexpr(isMultiA || isMultiB)
     {
         AsPointer p_as_grid_grp;
         BsPointer p_bs_grid_grp;
 
-        const auto& as_batch_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx);
+        const auto& as_group_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx);
 
         // compute_ptr_offset_of_n_ not need BatchStrideB so
         // in case of MultiA is false but isMultiB is true
@@ -139,27 +135,27 @@ __global__ void
 
             static constexpr index_t NumATensor = AGridDesc_AK0_M_AK1::Size();
             static_for<0, NumATensor, 1>{}([&](auto i) {
-                p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + as_n_offset[i];
+                p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + as_n_offset[i];
             });
         }
         else
         {
             const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
             static_for<0, 1, 1>{}(
-                [&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + a_n_offset; });
+                [&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + a_n_offset; });
         }
 
-        const auto& bs_batch_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx);
+        const auto& bs_group_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx);
 
         static constexpr index_t NumBTensor = BGridDesc_BK0_N_BK1::Size();
         static_for<0, NumBTensor, 1>{}(
-            [&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_batch_offset[i]; });
+            [&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_group_offset[i]; });
 
         GridwiseGemm::template Run<HasMainKBlockLoop>(
             p_as_grid_grp,
             p_bs_grid_grp,
             p_ds_grid_grp,
-            p_e_grid + e_batch_offset + e_n_offset,
+            p_e_grid + e_group_offset + e_n_offset,
             p_shared,
             a_element_op,
             b_element_op,
@@ -172,19 +168,19 @@ __global__ void
     }
     else
     {
-        const long_index_t a_batch_offset =
+        const long_index_t a_group_offset =
             amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
-        const long_index_t b_batch_offset =
+        const long_index_t b_group_offset =
             amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
 
         const long_index_t a_n_offset =
             amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
 
         GridwiseGemm::template Run<HasMainKBlockLoop>(
-            p_as_grid + a_batch_offset + a_n_offset,
-            p_bs_grid + b_batch_offset,
+            p_as_grid + a_group_offset + a_n_offset,
+            p_bs_grid + b_group_offset,
             p_ds_grid_grp,
-            p_e_grid + e_batch_offset + e_n_offset,
+            p_e_grid + e_group_offset + e_n_offset,
             p_shared,
             a_element_op,
             b_element_op,
@@ -200,7 +196,6 @@ __global__ void
     ignore = p_bs_grid;
     ignore = p_ds_grid;
     ignore = p_e_grid;
-    ignore = groups_count;
     ignore = a_grid_desc_k0_m_k1;
     ignore = b_grid_desc_k0_n_k1;
     ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
@@ -287,7 +282,8 @@ template <index_t NDimSpatial,
                                                      // in tuple for MultiAB), unpack if tuple was
                                                      // passed
           typename BComputeDataType = AComputeDataType,
-          LoopScheduler LoopSched   = make_default_loop_scheduler()>
+          LoopScheduler LoopSched   = make_default_loop_scheduler(),
+          index_t NumGroupsToMerge  = 1>
 struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
     : public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
                                              ALayout,
@@ -306,6 +302,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
 {
     using DeviceOp = DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle;
 
+    static_assert(NumGroupsToMerge >= 1);
+
     static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
     static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value;
 
@@ -322,7 +320,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                                                             ConvForwardSpecialization,
                                                             true /*SplitN*/,
                                                             ALayout,
-                                                            ELayout>;
+                                                            ELayout,
+                                                            NumGroupsToMerge>;
 
     static constexpr auto matrix_padder =
         MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
@@ -521,7 +520,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
             {
                 static_for<0, NumATensor, 1>{}([&](auto i) {
                     // Init compute_ptr_offset_of_groups_ for multiple AB
-                    compute_ptr_offset_of_groups_.BatchStrideA_(i) = a_g_n_c_wis_strides[0];
+                    compute_ptr_offset_of_groups_.BatchStrideA_(i) =
+                        a_g_n_c_wis_strides[0] * NumGroupsToMerge;
 
                     // Use GemmADataType/GemmBDataType to iterate over tuple (even if passed data
                     // type is not tuple)
@@ -549,7 +549,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                 });
                 static_for<0, NumBTensor, 1>{}([&](auto i) {
                     // Init compute_ptr_offset_of_groups_ for multiple AB
-                    compute_ptr_offset_of_groups_.BatchStrideB_(i) = b_g_k_c_xs_strides[0];
+                    compute_ptr_offset_of_groups_.BatchStrideB_(i) =
+                        b_g_k_c_xs_strides[0] * NumGroupsToMerge;
 
                     using DataType = remove_cvref_t<tuple_element_t<i.value, GemmBDataType>>;
                     // It is possible that one of the AB is a pointer and one is a tuple.
@@ -569,8 +570,10 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
             }
             else
             {
-                compute_ptr_offset_of_groups_.BatchStrideA_ = a_g_n_c_wis_strides[0];
-                compute_ptr_offset_of_groups_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+                compute_ptr_offset_of_groups_.BatchStrideA_ =
+                    a_g_n_c_wis_strides[0] * NumGroupsToMerge;
+                compute_ptr_offset_of_groups_.BatchStrideB_ =
+                    b_g_k_c_xs_strides[0] * NumGroupsToMerge;
                 compute_ptr_offset_of_n_.BatchStrideA_ = a_g_n_c_wis_strides[1] * conv_N_per_block_;
 
                 // p_as and p_bs are pointers
@@ -587,7 +590,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                 p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
 
                 // D batch stride
-                compute_ptr_offset_of_groups_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+                compute_ptr_offset_of_groups_.BatchStrideDs_(i) =
+                    ds_g_n_k_wos_strides[i][0] * NumGroupsToMerge;
                 compute_ptr_offset_of_n_.BatchStrideDs_(i) =
                     ds_g_n_k_wos_strides[i][1] * conv_N_per_block_;
 
@@ -606,7 +610,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                 ds_grid_desc_m_n_(i) =
                     DeviceOp::MakeEGridDescriptor_M_N<DLayout>(conv_to_gemm_transformer_d);
             });
-            compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+            compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0] * NumGroupsToMerge;
             compute_ptr_offset_of_n_.BatchStrideE_ = e_g_n_k_wos_strides[1] * conv_N_per_block_;
 
             // populate desc for Ds/E
@@ -730,8 +734,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                 arg.a_g_n_c_wis_lengths_[I1] / arg.conv_N_per_block_;
 
             const index_t gdx = arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
-            const index_t gdy = arg.num_group_ * num_workgroups_per_Conv_N;
-            const index_t gdz = 1;
+            const index_t gdy = arg.num_group_ / NumGroupsToMerge;
+            const index_t gdz = num_workgroups_per_Conv_N;
 
             const auto K =
                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
@@ -780,7 +784,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                         arg.a_element_op_,
                         arg.b_element_op_,
                         arg.cde_element_op_,
-                        arg.a_g_n_c_wis_lengths_[0], // Group count
                         as_grid_desc_ak0_m_ak1,
                         bs_grid_desc_bk0_n_bk1,
                         arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
@@ -824,7 +827,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                         arg.a_element_op_,
                         arg.b_element_op_,
                         arg.cde_element_op_,
-                        arg.a_g_n_c_wis_lengths_[0], // Group count
                         arg.a_grid_desc_ak0_m_ak1_,
                         arg.b_grid_desc_bk0_n_bk1_,
                         arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
@@ -856,6 +858,10 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
     {
         namespace ctc = tensor_layout::convolution;
 
+        const index_t G = arg.b_g_k_c_xs_lengths_[I0];
+        const index_t K = arg.b_g_k_c_xs_lengths_[I1];
+        const index_t C = arg.b_g_k_c_xs_lengths_[I2];
+
         // check device
         if(get_device_name() == "gfx908")
         {
@@ -904,6 +910,42 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                 }
             }
         }
+        else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::Filter3x3)
+        {
+            if(C != 1)
+            {
+                return false;
+            }
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t filter_spatial_dim = arg.b_g_k_c_xs_lengths_[i + I3];
+
+                if(filter_spatial_dim != I3)
+                {
+                    return false;
+                }
+            }
+            if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
+            {
+                return false;
+            }
+        }
+
+        if constexpr(NumGroupsToMerge > 1)
+        {
+            if(!(C == 1))
+            {
+                return false;
+            }
+            if(G % NumGroupsToMerge != 0)
+            {
+                return false;
+            }
+            if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
+            {
+                return false;
+            }
+        }
 
         // check vector access of A
         // FIXME: layout
@@ -913,11 +955,16 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                      is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
                      is_same_v<ALayout, ctc::NDHWGC>)
         {
-            const index_t C = arg.a_g_n_c_wis_lengths_[2];
-
+            // Check access per C
             if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
             {
-                return false;
+                // If not possible, check access per G
+                if(!(ABlockTransferSrcVectorDim == 1 && C == 1 &&
+                     is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>() &&
+                     G % ABlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
             }
         }
         else
@@ -934,8 +981,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                      is_same_v<BLayout, ctc::KZYXGC>)
 
         {
-            const index_t C = arg.b_g_k_c_xs_lengths_[2];
-
             if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
             {
                 return false;
@@ -959,8 +1004,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                          is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
                          is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>)
             {
-                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
-
                 if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
                 {
                     valid = false;
@@ -1005,8 +1048,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                      is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
                      is_same_v<ELayout, ctc::NDHWGK>)
         {
-            const index_t K = arg.e_g_n_k_wos_lengths_[2];
-
             if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
             {
                 return false;
@@ -1157,7 +1198,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
             << BBlockTransferSrcScalarPerVector << ", "
             << CDEBlockTransferScalarPerVector_NPerBlock << ", "
             << CShuffleMXdlPerWavePerShuffle << ", "
-            << CShuffleNXdlPerWavePerShuffle
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << NumGroupsToMerge
             << ">";
         // clang-format on
 

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

@@ -638,6 +638,32 @@ struct AddSilu
     }
 };
 
+struct ConvScaleAdd
+{
+    __host__ __device__ ConvScaleAdd(float scale_in  = 1.f,
+                                     float scale_wei = 1.f,
+                                     float scale_out = 1.f)
+        : scale_in_(scale_in), scale_wei_(scale_wei), scale_out_(scale_out)
+    {
+    }
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ void
+    operator()<f8_t, float, float>(f8_t& e, const float& c, const float& d) const
+    {
+        float x;
+        Add{}.template operator()<float>(x, c * scale_in_ * scale_wei_, d);
+        e = type_convert<f8_t>(x * scale_out_);
+    };
+
+    float scale_in_;
+    float scale_wei_;
+    float scale_out_;
+};
+
 } // namespace element_wise
 } // namespace tensor_operation
 } // namespace ck