Merge branch 'develop' into gemm_layernorm_welford

client_example/09_quantization/CMakeLists.txt

+add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_operations)
+
 add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
 target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_operations)
 
+add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_operations)
+
 add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
 target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations)

client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout         = ck::tensor_layout::convolution::G_K;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::GNHWK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 4;
+static constexpr ck::index_t K             = 64;
+static constexpr ck::index_t C             = 32;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 71;
+static constexpr ck::index_t Wi            = 71;
+static constexpr ck::index_t Ho            = 36;
+static constexpr ck::index_t Wo            = 36;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        OutElementOp>;
+    // 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(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        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});
+
+            std::size_t flop      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    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(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file

client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp

@@ -6,7 +6,7 @@
 #include <vector>
 
 #include "ck/ck.hpp"
-#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_bias_forward_perlayer_quantization.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"

client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::GNHWK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = PassThrough;
+using OutElementOp       = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 4;
+static constexpr ck::index_t K             = 64;
+static constexpr ck::index_t C             = 32;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 71;
+static constexpr ck::index_t Wi            = 71;
+static constexpr ck::index_t Ho            = 36;
+static constexpr ck::index_t Wo            = 36;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                    InLayout,
+                                                                    WeiLayout,
+                                                                    ck::Tuple<RequantScaleLayout>,
+                                                                    OutLayout,
+                                                                    InDataType,
+                                                                    WeiDataType,
+                                                                    ck::Tuple<RequantScaleDataType>,
+                                                                    OutDataType,
+                                                                    PassThrough,
+                                                                    PassThrough,
+                                                                    OutElementOp>;
+    // 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(),
+                                                        {requant_scale.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {requant_scale_lengths},
+                                                        {requant_scale_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        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});
+
+            std::size_t flop      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            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::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    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(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file

client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp

@@ -6,7 +6,7 @@
 #include <vector>
 
 #include "ck/ck.hpp"
-#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_perlayer_quantization.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"

client_example/13_batchnorm/CMakeLists.txt

 add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
+add_executable(client_batchnorm_bwd_nhwc batchnorm_bwd_nhwc.cpp)
 target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_operations)

client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp"
+
+using XDataType           = ck::half_t;
+using DxDataType          = float;
+using DyDataType          = float;
+using AccDataType         = float;
+using ScaleDataType       = ck::half_t;
+using DscaleDbiasDataType = float;
+using MeanVarDataType     = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon = std::numeric_limits<float>::epsilon();
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem dy(sizeof(DyDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dx(sizeof(DxDataType) * numXYElement);
+    SimpleDeviceMem dscale(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dbias(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                                      DxDataType,
+                                                                      DyDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      DscaleDbiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    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;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 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(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
+                numScaleBiasMeanVarElement *
+                    (sizeof(ScaleDataType) + sizeof(DscaleDbiasDataType) * 2 +
+                     sizeof(MeanVarDataType) * 2);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << 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(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

example/14_gemm_quantization/CMakeLists.txt

+add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
+add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
\ No newline at end of file

example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int32_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using CDEElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using BiasDataType     = I32;
+using DsDataType       = ck::Tuple<BiasDataType>;
+using EDataType        = I8;
+
+using ALayout    = Row;
+using BLayout    = Col;
+using BiasLayout = Row;
+using DsLayout   = ck::Tuple<BiasLayout>;
+using ELayout    = Row;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
+     ALayout,
+     BLayout,
+     DsLayout,
+     ELayout,
+     ADataType,
+     BDataType,
+     AccDataType,
+     CShuffleDataType,
+     DsDataType,
+     EDataType,
+     PassThrough,                // AElementwiseOperation,
+     PassThrough,                // BElementwiseOperation,
+     CDEElementOp,               // CDEElementwiseOperation,
+     GemmDefault,                // GemmSpecialization GemmSpec,
+     1,                          // NumGemmKPrefetchStage,
+     256,                        // BlockSize,
+     256,                        // MPerBlock,
+     128,                        // NPerBlock,
+     64,                         // KPerBlock,
+     16,                         // AK1,
+     16,                         // BK1,
+     32,                         // MPerXDL,
+     32,                         // NPerXDL,
+     4,                          // MXdlPerWave,
+     2,                          // NXdlPerWave,
+     S<4, 64, 1>,                // ABlockTransferThreadClusterLengths_AK0_M_AK1,
+     S<1, 0, 2>,                 // ABlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // ABlockTransferSrcAccessOrder,
+     2,                          // index_t ABlockTransferSrcVectorDim,
+     16,                         // index_t ABlockTransferSrcScalarPerVector,
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1,
+     1,                          // bool ABlockLdsExtraM,
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder,
+     2,                          // index_t BBlockTransferSrcVectorDim,
+     8,                          // index_t BBlockTransferSrcScalarPerVector,
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1,
+     1,                          // bool BBlockLdsExtraN,
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle,
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle,
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        PassThrough>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA    = 1024;
+    ck::index_t StrideB    = 1024;
+    ck::index_t StrideBias = 0;
+    ck::index_t StrideE    = 1024;
+
+    float requant_scale = 0.03;
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1_uz}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1_uz, stride}));
+            }
+        };
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor2d(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 << "bias_n: " << bias_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
+    bias_n.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_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());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+
+    auto a_element_op   = PassThrough{};
+    auto b_element_op   = PassThrough{};
+    auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {bias_device_buf.GetDeviceBuffer()},
+                                      e_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      {StrideBias},
+                                      StrideE,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n(HostTensorDescriptor{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), bias_n(n));
+            }
+        }
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/14_gemm_xdl_quantization/gemm_xdl_relu_quantization_int8.cpp → example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp

@@ -9,7 +9,7 @@
 #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/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/device_memory.hpp"
@@ -22,50 +22,59 @@
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
 
+using I8  = int8_t;
+using I32 = int32_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
 using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
-using ActivationOp = ck::tensor_operation::element_wise::Relu;
-using CElementOp   = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+using ActivationOp = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
 
-using ADataType        = int8_t;
-using BDataType        = int8_t;
-using CDataType        = int8_t;
-using AccDataType      = int32_t;
-using CShuffleDataType = float;
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = I8;
 
-using ALayout = ck::tensor_layout::gemm::RowMajor;
-using BLayout = ck::tensor_layout::gemm::ColumnMajor;
-using CLayout = ck::tensor_layout::gemm::RowMajor;
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 
 // clang-format off
-using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<
-     ALayout,                    // typename ALayout,
-     BLayout,                    // typename BLayout,
-     CLayout,                    // typename CLayout,
-     ADataType,                  // typename ADataType,
-     BDataType,                  // typename BDataType,
-     CDataType,                  // typename CDataType,
-     AccDataType,                // typename GemmAccDataType,
-     CShuffleDataType,           // typename CShuffleDataType,
-     PassThrough,                // typename AElementwiseOperation,
-     PassThrough,                // typename BElementwiseOperation,
-     CElementOp,         // typename CElementwiseOperation,
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
+     ALayout,
+     BLayout,
+     DsLayout,
+     ELayout,
+     ADataType,
+     BDataType,
+     AccDataType,
+     CShuffleDataType,
+     DsDataType,
+     EDataType,
+     PassThrough,                // AElementwiseOperation,
+     PassThrough,                // BElementwiseOperation,
+     CDEElementOp,               // CDEElementwiseOperation,
      GemmDefault,                // GemmSpecialization GemmSpec,
-     1,                          // index_t NumGemmKPrefetchStage,
-     256,                        // index_t BlockSize,
-     256,                        // index_t MPerBlock,
-     128,                        // index_t NPerBlock,
-     64,                         // index_t KPerBlock,
-     16,                         // index_t AK1,
-     16,                         // index_t BK1,
-     32,                         // index_t MPerXDL,
-     32,                         // index_t NPerXDL,
-     4,                          // index_t MXdlPerWave,
-     2,                          // index_t NXdlPerWave,
-     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
-     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder,
-     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder,
+     1,                          // NumGemmKPrefetchStage,
+     256,                        // BlockSize,
+     256,                        // MPerBlock,
+     128,                        // NPerBlock,
+     64,                         // KPerBlock,
+     16,                         // AK1,
+     16,                         // BK1,
+     32,                         // MPerXDL,
+     32,                         // NPerXDL,
+     4,                          // MXdlPerWave,
+     2,                          // NXdlPerWave,
+     S<4, 64, 1>,                // ABlockTransferThreadClusterLengths_AK0_M_AK1,
+     S<1, 0, 2>,                 // ABlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // ABlockTransferSrcAccessOrder,
      2,                          // index_t ABlockTransferSrcVectorDim,
      16,                         // index_t ABlockTransferSrcScalarPerVector,
      16,                         // index_t ABlockTransferDstScalarPerVector_AK1,
@@ -84,53 +93,23 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::
-    ReferenceGemm<ADataType, BDataType, CDataType, float, PassThrough, PassThrough, CElementOp>;
+    ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
 
-int main(int argc, char* argv[])
+int main()
 {
     bool do_verification = true;
-    int init_method      = 1;
     bool time_kernel     = false;
 
     // GEMM shape
-    ck::index_t M = 3840;
-    ck::index_t N = 4096;
-    ck::index_t K = 4096;
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
 
-    float quant_multiplier = 0.03;
-
-    if(argc == 4)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        time_kernel     = std::stoi(argv[3]);
-    }
-    else if(argc == 10)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        time_kernel     = std::stoi(argv[3]);
-
-        M = std::stoi(argv[4]);
-        N = std::stoi(argv[5]);
-        K = std::stoi(argv[6]);
-
-        StrideA = std::stoi(argv[7]);
-        StrideB = std::stoi(argv[8]);
-        StrideC = std::stoi(argv[9]);
-    }
-    else
-    {
-        printf("arg1: verification (0=no, 1=yes)\n");
-        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: time kernel (0=n0, 1=yes)\n");
-        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
-        exit(0);
-    }
+    float requant_scale = 0.03;
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
@@ -138,61 +117,56 @@ int main(int argc, char* argv[])
 
             if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1_uz}));
             }
             else
             {
-                return HostTensorDescriptor({row, col}, {1_uz, stride});
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({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<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    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 << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
 
-    switch(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});
-        break;
-    default:
-        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-    }
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
 
-    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
-    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
-    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
 
-    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
-    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
 
     auto a_element_op   = PassThrough{};
     auto b_element_op   = PassThrough{};
-    auto c_element_op = CElementOp{quant_multiplier, ActivationOp{}};
+    auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
 
     // do GEMM
     auto gemm     = DeviceGemmInstance{};
     auto invoker  = gemm.MakeInvoker();
-    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
-                                      static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
-                                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      e_device_buf.GetDeviceBuffer(),
                                       M,
                                       N,
                                       K,
                                       StrideA,
                                       StrideB,
-                                      StrideC,
+                                      {},
+                                      StrideE,
                                       a_element_op,
                                       b_element_op,
-                                      c_element_op);
+                                      cde_element_op);
 
     if(!gemm.IsSupportedArgument(argument))
     {
@@ -205,7 +179,7 @@ int main(int argc, char* argv[])
 
     std::size_t flop = std::size_t(2) * M * N * K;
     std::size_t num_btype =
-        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
 
     float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
@@ -214,7 +188,7 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << gemm.GetTypeString() << std::endl;
 
-    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
 
     if(do_verification)
     {
@@ -222,11 +196,11 @@ int main(int argc, char* argv[])
         auto ref_invoker = ref_gemm.MakeInvoker();
 
         auto ref_argument = ref_gemm.MakeArgument(
-            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+            a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
 
         ref_invoker.Run(ref_argument);
 
-        return ck::utils::check_err(c_m_n_device_result, c_m_n_host_result) ? 0 : 1;
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
     }
 
     return 0;

example/14_gemm_xdl_quantization/CMakeLists.txt

-add_example_executable(example_gemm_xdl_relu_quantization_int8 gemm_xdl_relu_quantization_int8.cpp)
\ No newline at end of file

example/34_batchnorm/CMakeLists.txt

-add_example_executable(example_batchnorm_forward batchnorm_forward_nhwc.cpp)
-add_example_executable(example_batchnorm_infer batchnorm_infer_nhwc.cpp)
+add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
+add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
+add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)

example/34_batchnorm/README.md

@@ -53,4 +53,29 @@ Start running 10 times...
 Perf: 1.28235 ms, 523.329 GB/s
 ```
 
+## Run ```batchnorm backward nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+Arg2 -- 1/0 to indicate whether to use saved mean and invVariance
+Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+Arg4 -- time kernel (0=no, 1=yes)
+Arg5: use multi-block welford (0=n0, 1=yes)
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+```
 
+Result 
+```
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.411026 ms, 91.8702 GB/s
+```

example/34_batchnorm/batchnorm_backward_nhwc.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <getopt.h>
+
+#include "ck/ck.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/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormBwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type               = 0;
+    int init_method             = 3;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg5: use multi-block welford (0=n0, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 5 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar    = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename XDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_bwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool haveSavedMeanInvVar,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr index_t Rank         = 4;
+    constexpr index_t NumReduceDim = 3;
+
+    using ScaleDataType = XDataType;
+
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm backward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<AccDataType> dy(inOutLengths);
+
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> savedMean(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> savedInvVar(scaleBiasMeanVarLengths);
+    // savedVariance is only used for initializing savedInvVar
+    Tensor<AccDataType> savedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm backward algorithm
+    Tensor<AccDataType> dx_ref(inOutLengths);
+    Tensor<AccDataType> dx(inOutLengths);
+
+    Tensor<AccDataType> dscale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> dscale_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = dy.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = dscale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(haveSavedMeanInvVar)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        savedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        savedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+        auto it_src       = savedVariance.mData.begin();
+        auto it_dst       = savedInvVar.mData.begin();
+        float tmp_epsilon = std::numeric_limits<float>::epsilon();
+
+        while(it_src != savedVariance.mData.end())
+        {
+            *it_dst = type_convert<AccDataType>(
+                1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
+
+            it_src++;
+            it_dst++;
+        };
+    }
+    else
+    {
+        const float x_mean   = 0.0f;
+        const float x_stddev = 1.0f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            dy.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            break;
+        case 1:
+            dy.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            break;
+        case 2:
+            dy.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-2, 2}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            dy.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-0.2f, 0.2f}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
+        }
+    };
+
+    // input data of the batchnorm backward algorithm
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem dy_dev(sizeof(AccDataType) * dy.mDesc.GetElementSpaceSize());
+
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+
+    DeviceMem savedMean_dev(sizeof(AccDataType) * savedMean.mDesc.GetElementSpaceSize());
+    DeviceMem savedInvVar_dev(sizeof(AccDataType) * savedInvVar.mDesc.GetElementSpaceSize());
+
+    // output data of the batchnorm backward algorithm
+    DeviceMem dx_dev(sizeof(AccDataType) * dx.mDesc.GetElementSpaceSize());
+
+    DeviceMem dscale_dev(sizeof(AccDataType) * dscale.mDesc.GetElementSpaceSize());
+    DeviceMem dbias_dev(sizeof(AccDataType) * dbias.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    dy_dev.ToDevice(dy.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+
+    if(haveSavedMeanInvVar)
+    {
+        savedMean_dev.ToDevice(savedMean.mData.data());
+        savedInvVar_dev.ToDevice(savedInvVar.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              i_scaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormBwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormBwdImpl<XDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             ScaleDataType, // ScaleDataType
+                                                             AccDataType,   // DscaleDbiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,  // XSrcVectorSize
+                                                             1,  // DySrcVectorSize
+                                                             1,  // DxDstVectorSize
+                                                             1,  // ScaleSrcVectorSize
+                                                             1,  // DscaleDbiasDstVectorSize
+                                                             1>; // MeanVarSrcVectorSize
+
+    auto batchnorm_bwd = DeviceBatchNormBwdInstance{};
+
+    auto argument_ptr = batchnorm_bwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2},
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        dy_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
+        haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
+        epsilon,
+        PassThroughOp{},
+        dx_dev.GetDeviceBuffer(),
+        dscale_dev.GetDeviceBuffer(),
+        dbias_dev.GetDeviceBuffer());
+
+    if(!batchnorm_bwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_bwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_bwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_bwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, dy, scale, outputing of dx, dscale, dbias
+        num_bytes +=
+            total_length * sizeof(XDataType) * 3 + invariant_length * sizeof(AccDataType) * 3;
+
+        // outputing of mean, inv-variance
+        num_bytes += haveSavedMeanInvVar ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormBwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              ScaleDataType, // ScaleDataType
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
+
+        auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2},
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            dy.mData.data(),
+            bnScale.mData.data(),
+            haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_ref.mData.data(),
+            dscale_ref.mData.data(),
+            dbias_ref.mData.data());
+
+        if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout
+                << "The runtime parameters seems not supported by the device instance, exiting!"
+                << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        dx_dev.FromDevice(dx.mData.data());
+        dscale_dev.FromDevice(dscale.data());
+        dbias_dev.FromDevice(dbias.data());
+
+        // clang-format off
+        pass = pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:");
+        // clang-format on
+    };
+
+    return (pass);
+};
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormBwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.haveSavedMeanInvVar,
+                                                                    epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.haveSavedMeanInvVar,
+                                                               epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.haveSavedMeanInvVar,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.haveSavedMeanInvVar,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.haveSavedMeanInvVar,
+                                                                 epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.haveSavedMeanInvVar,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            3,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            false,
+                                                            epsilon);
+
+        pass = pass && bnorm_bwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     3,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     false,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}

example/44_conv2d_fwd_quant/CMakeLists.txt → example/44_conv2d_fwd_quantization/CMakeLists.txt

+add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
 add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
 add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)

example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using CShuffleDataType     = int32_t;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+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 BiasLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // 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
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        8>;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          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& bias_g_k_desc,
+                          const HostTensorDescriptor& requant_scale_g_k_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_k_desc);
+    Tensor<RequantScaleDataType> requant_scale(requant_scale_g_k_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 << "requant_scale: " << requant_scale.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-128, 127});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-128, 127});
+    bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
+    requant_scale.GenerateTensorValue(GeneratorTensor_2<RequantScaleDataType>{0, 1});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem requant_scale_device_buf(sizeof(RequantScaleDataType) *
+                                       requant_scale.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());
+    requant_scale_device_buf.ToDevice(requant_scale.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> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_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(bias_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(requant_scale_g_k_desc.GetLengths(), d1_g_n_k_wos_lengths);
+    copy(requant_scale_g_k_desc.GetStrides(), d1_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);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        {bias_device_buf.GetDeviceBuffer(), requant_scale_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,
+        {d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths},
+        {d0_g_n_k_wos_strides, d1_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 flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        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_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            out_element_op(out_host(idx), c_host(idx), bias(idx), requant_scale(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int main()
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        32,           // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+
+    using InLayout           = ck::tensor_layout::convolution::GNHWC;
+    using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+    using BiasLayout         = ck::tensor_layout::convolution::G_K;
+    using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+    using OutLayout          = ck::tensor_layout::convolution::GNHWK;
+
+    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);
+
+    // TODO - make_bias_host_tensor_descriptor_g_n_k_wos_packed()
+    const auto bias_g_k_desc = HostTensorDescriptor({conv_param.G_,
+                                                     conv_param.N_,
+                                                     conv_param.K_,
+                                                     conv_param.output_spatial_lengths_[0],
+                                                     conv_param.output_spatial_lengths_[1]},
+                                                    {
+                                                        conv_param.K_, // g
+                                                        0,             // n
+                                                        1,             // k
+                                                        0,             // ho
+                                                        0              // wo
+                                                    });
+
+    const auto requant_scale_g_k_desc = bias_g_k_desc;
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    std::cout << out_g_n_k_wos_desc << std::endl;
+
+    using deviceOp = DeviceGroupedConvNDFwdInstance<ndim_spatial,
+                                                    InLayout,
+                                                    WeiLayout,
+                                                    BiasLayout,
+                                                    RequantScaleLayout,
+                                                    OutLayout>;
+
+    return run_grouped_conv_fwd<ndim_spatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InElementOp,
+                                WeiElementOp,
+                                OutElementOp,
+                                deviceOp>(do_verification,
+                                          time_kernel,
+                                          conv_param,
+                                          in_g_n_c_wis_desc,
+                                          wei_g_k_c_xs_desc,
+                                          bias_g_k_desc,
+                                          requant_scale_g_k_desc,
+                                          out_g_n_k_wos_desc,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+}

example/44_conv2d_fwd_quant/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp → example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp

@@ -11,6 +11,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/utility/convolution_parameter.hpp"
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
@@ -163,17 +164,16 @@ bool run_grouped_conv_fwd(bool do_verification,
     // do Conv
     auto conv     = DeviceConvNDFwdInstance{};
     auto invoker  = conv.MakeInvoker();
-    auto argument = conv.MakeArgument(
-        in_device_buf.GetDeviceBuffer(),
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
                                       wei_device_buf.GetDeviceBuffer(),
-        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+                                      {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>{{d0_g_n_k_wos_lengths}},
-        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d0_g_n_k_wos_strides}},
+                                      {d0_g_n_k_wos_lengths},
+                                      {d0_g_n_k_wos_strides},
                                       e_g_n_k_wos_lengths,
                                       e_g_n_k_wos_strides,
                                       conv_filter_strides,
@@ -235,8 +235,8 @@ bool run_grouped_conv_fwd(bool do_verification,
 
         out_device_buf.FromDevice(out_device.mData.data());
 
-        pass &= ck::utils::check_err(
-            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
     }
 
     return (pass ? 0 : 1);

example/44_conv2d_fwd_quant/conv2d_fwd_xdl_perlayer_quantization_int8.cpp → example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp

@@ -11,6 +11,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/utility/convolution_parameter.hpp"
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
@@ -150,14 +151,14 @@ bool run_grouped_conv_fwd(bool do_verification,
     auto invoker  = conv.MakeInvoker();
     auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
                                       wei_device_buf.GetDeviceBuffer(),
-                                      std::array<const void*, 0>{},
+                                      {},
                                       out_device_buf.GetDeviceBuffer(),
                                       a_g_n_c_wis_lengths,
                                       a_g_n_c_wis_strides,
                                       b_g_k_c_xs_lengths,
                                       b_g_k_c_xs_strides,
-                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
-                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      {},
+                                      {},
                                       e_g_n_k_wos_lengths,
                                       e_g_n_k_wos_strides,
                                       conv_filter_strides,
@@ -213,8 +214,8 @@ bool run_grouped_conv_fwd(bool do_verification,
 
         out_device_buf.FromDevice(out_device.mData.data());
 
-        pass &= ck::utils::check_err(
-            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
     }
 
     return (pass ? 0 : 1);